A spatially explicit approach to measuring carbon dynamics for reducing emissions from deforestation and forest degradation: a case study of Chinese forests

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Request accessible transcript

Transcript (if available)

Content A SPATIALLY EXPLICIT APPROACH TO MEASURING CARBON DYNAMICS FOR REDUCING EMISSIONS FROM DEFORESTATION AND FOREST DEGRADATION: A CASE STUDY OF CHINESE FORESTS by Oh Seok Kim A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (GEOGRAPHY) August 2013 Copyright 2013 Oh Seok Kim Dedication The dissertation is dedicated to my multicultural family and friends, espe- cially to my son whose good fortune includes three names: one Korean, one Chinese, and one English. If he hadn’t been born, I bet I would not be finish- ing my Ph.D. so soon, and so I am also thankful to my lovely wife. Without her, there would be no son or the extraordinary rewards of our family life. ii Acknowledgments The true value of a dissertation is something more than to prove if the author is capable of conducting quality research meaningful to a particular discipline. It is in fact a won- derful invitation of life sent by a greater community to the author. Not all invitations are guest-friendly, and that is why I have needed endless supports from my disserta- tion committee members and have had countless dialogues with them in order for me to figure it out whether I truly wanted to be part of that greater community—academia. Without their warm-hearted patience and sincere mentoring, I would not have been able to accept the invitation of life sent to me, let alone recognize its great value. I am gratefully indebted to my knowledgeable dissertation committee: Professors Andrew Curtis, Jeff Nugent, Eric Heikkila, Rod McKenzie, and my advisor, Josh Newell. I feel honored and fortunate to become a USC alumnus as Newell’s first student to earn a doctorate. Associate Provost Mark Todd deserves special mention for his uncom- mon support. Professor Zhi L ¨ u and her team members Hao Wang, Ai Chen, Shan Sun, Yi He, and Fangyi Yang have shared their knowledge and resources during my stay in China. Professor Jianchu Xu and Zhuangfang Yi have shared their knowledge and data. Professors Dietrich Schmidt-Vogt, Rhett Harrison, Gil Pontius, Ron Eastman, and Tom Gillespie; Clarkies Zhe Li, Victor Hugo Guti´ errez-V´ elez, and the Idrisians; and the editor S deserve my gratitude for their support. Financial assistance was provided by the Association of American Geographers, USC Center for Sustainable Cities, and USC Dana and David Dornsife College of Letters, Arts and Sciences. iii Table of Contents Dedication ii Acknowledgments iii List of Tables vii List of Figures viii Abstract x Chapter 1: Introduction 1 1.1 Background and Problem Statement . . . . . . . . . . . . . . . . . . . . . 1 1.1.1 Definitional issues . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.1.2 Modeling issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.2 Research Objectives and Questions . . . . . . . . . . . . . . . . . . . . . . 10 1.3 Organization of Subsequent Chapters . . . . . . . . . . . . . . . . . . . . 15 Chapter 2: Literature Review 17 2.1 Forest Carbon Flux and Deforestation . . . . . . . . . . . . . . . . . . . . 17 2.1.1 Concepts and definition of terms . . . . . . . . . . . . . . . . . . . 18 2.1.2 Forestland change: observation and projection . . . . . . . . . . . 24 2.2 Methodology of Reference Emission Level . . . . . . . . . . . . . . . . . . 33 2.2.1 Verified carbon standard . . . . . . . . . . . . . . . . . . . . . . . . 34 2.2.2 Forestland carbon change . . . . . . . . . . . . . . . . . . . . . . . 40 Chapter 3: Model Development 46 3.1 Model Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.1.1 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.1.2 Streamlined definition . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.1.3 Rationale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.1.4 Geographic Emission Benchmark and GEOMOD . . . . . . . . . 52 3.1.5 Calibration and validation . . . . . . . . . . . . . . . . . . . . . . . 54 3.2 Methods and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2.1 Forest-cover data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.2.2 Receiver operating characteristic . . . . . . . . . . . . . . . . . . . 57 iv 3.2.3 Change detection analysis . . . . . . . . . . . . . . . . . . . . . . . 60 3.2.4 Local indicator of spatial association . . . . . . . . . . . . . . . . . 61 3.2.5 Business-as-usual forest-cover change . . . . . . . . . . . . . . . . 62 3.2.5.1 Linear extrapolation . . . . . . . . . . . . . . . . . . . . . 63 3.2.5.2 Spatial allocation . . . . . . . . . . . . . . . . . . . . . . . 64 3.2.5.3 Deforestation driver data . . . . . . . . . . . . . . . . . . 66 3.2.6 Biomass estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.2.7 Biomass data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 3.2.8 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 3.2.8.1 Figure of merit . . . . . . . . . . . . . . . . . . . . . . . . 73 3.2.8.2 Forest-use data . . . . . . . . . . . . . . . . . . . . . . . . 74 3.2.9 Reference emission level . . . . . . . . . . . . . . . . . . . . . . . . 74 Chapter 4: Case Study: Xishuangbanna, China 77 4.1 Study Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 4.2.1 Receiver operating characteristic . . . . . . . . . . . . . . . . . . . 83 4.2.2 Change detection analysis . . . . . . . . . . . . . . . . . . . . . . . 83 4.2.3 Local indicator of spatial association . . . . . . . . . . . . . . . . . 86 4.2.4 Business-as-usual forest-cover change . . . . . . . . . . . . . . . . 88 4.2.5 Biomass estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 4.2.6 Figure of merit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 4.2.7 Forest-use versus forest-cover . . . . . . . . . . . . . . . . . . . . . 90 4.2.8 Reference emission level . . . . . . . . . . . . . . . . . . . . . . . . 92 4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 4.4.1 Forest-cover and forest carbon stock changes . . . . . . . . . . . . 95 4.4.2 Purpose of land change model . . . . . . . . . . . . . . . . . . . . 96 4.4.3 What do we mean by “business-as-usual?” . . . . . . . . . . . . . 97 Chapter 5: A Step Forward 101 5.1 Research Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 5.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Chapter 6: Concluding Remarks 126 6.1 Intellectual Merit of Empirical Results . . . . . . . . . . . . . . . . . . . . 126 6.2 Broader Implications of Findings . . . . . . . . . . . . . . . . . . . . . . . 126 6.3 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6.3.1 Comparative RELs . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6.3.2 Forest transition theory . . . . . . . . . . . . . . . . . . . . . . . . 128 6.3.3 The carbon footprint of forest transition . . . . . . . . . . . . . . . 128 v Appendix A: Glossary of Acronyms 131 Appendix B: Geographic Coordinate System 133 Appendix C: Sensitivity of ROC and AUC 135 Appendix D: Specifications of GEOMOD 137 Appendix E: Sensitivity of OLS 138 Bibliography 140 vi List of Tables 2.1 Carbon stock in forest carbon pools (in million tonnes) . . . . . . . . . . 23 2.2 Different types of “forest” biomass . . . . . . . . . . . . . . . . . . . . . . 28 3.1 Different characterizations of “forest” by Food and Agriculture Organi- zation of the United Nations (FAO) and the Geographic Emission Bench- mark (GEB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.2 Different characterizations of “deforestation” and “forest degradation” . 50 3.3 Remotely sensed forest-cover data . . . . . . . . . . . . . . . . . . . . . . 55 3.4 Receiver Operating Characteristic’s (ROC’s) contingency table . . . . . . 59 3.5 Vector data for (spatial) allocation . . . . . . . . . . . . . . . . . . . . . . . 69 3.6 Remotely sensed data for (spatial) allocation . . . . . . . . . . . . . . . . 69 4.1 Amounts of observed forest-cover loss between 2000 and 2005 at the pre- fecture level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 4.2 Amounts and rates of observed forest-cover loss between 2000 and 2005 at different spatial levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 4.3 Observed above-ground living biomass (AGB) loss due to deforestation and forest degradation between 2000 and 2005 at the prefecture and province levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 4.4 Projected and observed above-ground living biomass (AGB) loss due to deforestation between 2005 and 2010 at the prefecture level . . . . . . . . 91 4.5 Figure of Merits (FoMs) of the Geographic Emission Benchmark (GEB) and GEOMOD modeling at the prefecture level . . . . . . . . . . . . . . . 92 4.6 Amounts of observed forest-cover and forest-use changes at the prefec- ture level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 4.7 Observed above-ground living biomass (AGB) loss due to distinct defor- estations between 2000 and 2005 at the prefecture level (1 Tg = 1 terra- gram = 1 million tonnes) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 4.8 Projected and observed carbon emissions due to deforestation between 2000 and 2010 at the prefecture level . . . . . . . . . . . . . . . . . . . . . 93 4.9 Projected and observed carbon emissions due to deforestation between 2010 and 2040 at the prefecture level . . . . . . . . . . . . . . . . . . . . . 95 5.1 Coefficients of Ordinary Least Squares regression (OLS) . . . . . . . . . . 105 5.2 Coefficients of Geographically Weighted Regression (GWR) . . . . . . . 106 2 Areal information by different sources . . . . . . . . . . . . . . . . . . . . 133 3 Comparison of Area Under the Curves (AUCs) . . . . . . . . . . . . . . . 135 vii List of Figures 1.1 Concept of reference emission level (REL) . . . . . . . . . . . . . . . . . . 6 1.2 Time window of deforestation rate calibration . . . . . . . . . . . . . . . 12 1.3 Spatial extent of deforestation rate calibration . . . . . . . . . . . . . . . . 13 1.4 Map of the study areas: Banna prefecture (black) and Yunnan province (gray) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.1 Ideal characterization of forest . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2 System of carbon flux, reprinted from Kim (2013) . . . . . . . . . . . . . . 23 2.3 Forest carbon pools in Banna prefecture, photographed by the author in December 2011 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.4 Structure of the literature review on deforestation research . . . . . . . . 27 2.5 Structure of reference emission level (REL) . . . . . . . . . . . . . . . . . 36 3.1 Structural difference of the Geographic Emission Benchmark (GEB) and GEOMOD modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.2 (a) Globcover-based binary map of forest and non-forest and (b) Vegeta- tion Continuous Field (VCF) . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.3 Logic of linear extrapolation . . . . . . . . . . . . . . . . . . . . . . . . . . 64 3.4 Spatial variables for the GEOMOD run . . . . . . . . . . . . . . . . . . . . 67 3.5 Night-light imagery for the Geographic Emission Benchmark (GEB) . . . 68 3.6 Spatially explicit above-ground living biomass (AGB) estimates . . . . . 72 3.7 Progressive expansion of rubber plantations in Banna prefecture . . . . . 75 4.1 Rubber yield in Banna prefecture and Yunnan province . . . . . . . . . . 80 4.2 Rubber plantations in Yunnan province . . . . . . . . . . . . . . . . . . . 81 4.3 Life-cycle of rubber plantations in Banna prefecture, photographed by the author in December 2011 . . . . . . . . . . . . . . . . . . . . . . . . . . 82 4.4 Area Under the Curve (AUC) of the Globcover and Vegetation Continu- ous Field (VCF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 4.5 Observed deforestation, forest degradation, and forest re-growth charac- terized by forest-cover between 2000 and 2005 in Banna prefecture . . . 85 4.6 Hotspots of observed deforestation and forest degradation between 2000 and 2005 in Yunnan province . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.7 Transition potential maps of the (a) Geographic Emission Benchmark (GEB) and (b) GEOMOD modeling . . . . . . . . . . . . . . . . . . . . . . 89 4.8 Business-as-usual loss of forestlands projected by the (a) Geographic Emis- sion Benchmark (GEB) and (b) GEOMOD modeling . . . . . . . . . . . . 90 viii 4.9 Observed forest-cover change between 2000 and 2005 overlaid with the rubber plantations in 2006 . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.10 Observed forest-cover change between 2005 and 2010 overlaid with the rubber plantations in 2006 . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 4.11 Visual comparison of rubber plantations and a tropical raintree in Man- dan village of Banna, photographed by the author in December 2011 . . 100 5.1 Visual comparison of rubber and banana plantations in Banna prefecture, photographed by the author in December 2011 . . . . . . . . . . . . . . . 109 5.2 Spatial drift of intercept . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.3 Spatial drift of age of rubber tree . . . . . . . . . . . . . . . . . . . . . . . 111 5.4 Spatial drift of percent of clay in soil . . . . . . . . . . . . . . . . . . . . . 112 5.5 Spatial drift of soil pH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 5.6 Spatial drift of elevation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 5.7 Spatial drift of aspect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 5.8 Spatial drift of slope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 5.9 Spatial drift of distance from stream networks . . . . . . . . . . . . . . . 117 5.10 Spatial drift of annual precipitation . . . . . . . . . . . . . . . . . . . . . . 118 5.11 Spatial drift of annual mean temperature . . . . . . . . . . . . . . . . . . 119 5.12 Spatial drift of interaction (percent of clay in soil and annual precipitation) 120 5.13 Spatial drift of squared elevation . . . . . . . . . . . . . . . . . . . . . . . 121 5.14 Spatial drift of squared aspect . . . . . . . . . . . . . . . . . . . . . . . . . 122 5.15 Spatial drift of squared slope . . . . . . . . . . . . . . . . . . . . . . . . . 123 5.16 Spatial drift of squared annual precipitation . . . . . . . . . . . . . . . . . 124 5.17 Spatial drift of squared annual mean temperature . . . . . . . . . . . . . 125 ix Abstract This dissertation proposes a Geographic Emission Benchmark (GEB) for reducing emis- sions from deforestation and forest degradation (REDD+). The GEB is designed to func- tion as a “baseline” to indicate reference (carbon) emission levels as business-as-usual if REDD+ does not take place. The GEB is also a new land change model that produces prospective and spatially explicit outcomes. Unlike other land change models forecast- ing business-as-usual deforestation, the GEB internally (1) characterizes “forest” in a general sense based on remotely sensed data and (2) identifies “deforestation hotspots” using a spatial clustering technique in order to transparently estimate regional rates of deforestation. As a byproduct, it also portrays a historic trend of forest degradation. In terms of predictive accuracy, the GEB produces a more accurate projection than GEO- MOD, measured by Figure of Merit, at least for this case study of Chinese tropics. Thus, the research concludes the GEB is a reasonably accurate baseline for current and future reference emission levels. If other land change models are to be used to set a prospective and spatially explicit reference emission level for a particular region, their predictive accuracies must outperform that of the regional GEB. Finally, this dissertation demon- strates an application of spatial regression attempting to identify drivers of deforesta- tion in the study area. The result shows that biophysical factors have more influence on the regional deforestation than other factors such as population or infrastructure. Thus, when forecasting business-as-usual forest loss for a future REDD+ project, the biophys- ical drivers must be considered more seriously than other drivers of deforestation in the Chinese tropics. x Chapter 1 Introduction 1.1 Background and Problem Statement According to the Intergovernmental Panel on Climate Change (IPCC), climate change is a scientific fact (IPCC Working Group I, 2007), and humans are largely responsible for it because we have emitted greenhouse gases (GHGs) to the atmosphere for centuries, in particular since the Industrial Revolution. The GHG emissions alter the biophysical properties of Earth, e.g., its albedo and temperature, and such alteration is expected to substantially affect the human habitat in a degree that is more radical than the usual situation with no climate change. This radical shift is the main reason why many per- ceive climate change as a global threat if we do not have enough time and resources to understand and cope with the repercussions. What then can humans do about climate change? There are two strategies to deal with it, namely, adaptation and mitigation. The former refers to the coping strategy that stresses humanity’s adjustment against the climate change; basically it is about enhancing human capacity to become resilient to the unexpected future driven by the climate change. The latter, mitigation, indicates an anthropogenic attempt to minimize the magnitude of the climate change itself. Controlling GHG emissions exemplifies a mitigation strategy for ensuring fewer GHGs in the atmosphere, thus likely decelerat- ing the pace of climatic change. While these two strategies are equally important and often used in tandem, this dissertation solely focuses on the mitigation side of climate change, through saving forests in particular. The role of forests has been much emphasized in recent intergovernmental meet- ings. Particularly, protecting tropical rainforests, especially the old growth ones, has 1 been a global concern for many decades. Its importance cannot be overemphasized since the rainforests are home to diverse fauna and flora (Gibson et al., 2011), which include rare medical ingredients that might cure cancer (Shanley and Luz, 2003). The rainforests also host numerous ecosystem services that are indispensable to humans (Robbins, 2004). Watershed protection, for example, preserves and secures freshwater on which humans rely (Zhang et al., 2010; Johnson and Lewis, 2007). If tropical rain- forests are removed from an ecosystem, it will experience severe topsoil loss driven by the consistent rainfall that is quite common in any tropical regions. Consequently, the ecosystem’s freshwater will become muddy, which is a waste from the human per- spective. Another ecosystem service of rainforests is absorbing water from the ground and then transfering it to the atmosphere as evapotranspiration. Such hydrological flux makes the regional climate cooler and more humid, which in turn causes more rain so that the fauna and flora in the region can survive or thrive (Moutinho and Schwartz- man, 2005). Further, the rainforests not only play a key role in sustaining nutrient and energy balances of ecosystems (Malhi et al., 2008; Pongratz et al., 2006), but also function as “biological carbon engines” that control terrestrial carbon flux, while simultaneously storing a huge portion of carbon stock in them (Ramankutty et al., 2007; Baker et al., 2004; Cramer et al., 2004). Without the carbon engines functioning, society very well might not be able to mitigate anthropogenic climate change efficiently. To effectively mitigate climate change, it is critical to sustain forests. One way to do so is by planting additional trees either on bare grounds (i.e., afforestation) or defor- ested areas (i.e., reforestation), as these will sequester more carbon. Another way would be to protect existing forests that are vulnerable to future deforestation and/or forest degradation. Once the forests are removed from the ground, their carbon stocks will eventually be released to the atmosphere through, for instance, burning (which emits carbon dioxide, CO 2 ) or decaying (which emits methane, CH 4 ). Afforestation and reforestation activities have been supported by the Kyoto Pro- tocol since 1998 (UNFCCC, 1998). What is unique about the Kyoto Protocol is that it 2 officially mandates developed countries to reduce their GHG emissions and simulta- neously allows trading “carbon credits” between developed countries and developing and least developed countries if a developed country is not capable of reducing GHG emissions of its own (i.e., “cap-and-trade”). The carbon credit is a new monetary sys- tem that allows countries to sell and buy the rights to emit GHGs, and it can be gener- ated when afforestation and reforestation are done based on internationally approved methodologies, where those are evaluated by numerous experts. Encouraged by such an emission trading mechanism, numerous developing and least developed countries have planted a large amount of forests to sequester carbon that is additional to the business-as-usual situation and, more importantly, to generate carbon credits so as to sell those to the developed countries that are not successful in reducing their GHG emissions. The Kyoto Protocol, however, did not allow generating and trading carbon credits through reducing deforestation and/or forest degradation (Brown et al., 2007), where such exclusion turned out to be problematic because it is known that old growth forests are better at sequestrating carbon than artificially planted trees such as planta- tions (Baker et al., 2004). When the Kyoto Protocol was proposed, methodologies were scientifically limited in their ability to prove that activities were avoiding deforestation. That is, they were not capable of dictating an accurate quantity of potential carbon emis- sions in the future if society was to do nothing to halt or reduce deforestation and forest degradation (Brown et al., 2007; Dushku and Brown, 2003). In 2008, the United Nations Collaborative Programme on Reducing Emissions from Deforestation and Forest Degradation in Developing Countries (UN-REDD) was launched (FAO et al., 2008) to make up such limitation of the Kyoto Protocol. UN- REDD, or more generally REDD, attempts to achieve the aim listed in its title by encour- aging developed countries to provide financial supports through a carbon market sys- tem to those developing and least developed countries that are trying to reduce defor- estation and forest degradation in areas with the most substantial deforestation and forest degradation (The Economist, 2010). 3 Before “REDD” was coined, deforestation and forest degradation were usually pre- vented through conservation activities financially subsidized by governments, interna- tional organizations, or affluent individuals. In order to get the subsidies, a solid pric- ing mechanism must be applied to a particular landscape to come up with its monetary value, and Payment for Ecosystem Services (PES) and The Economics of Ecosystems and Biodiversity (TEEB) have played an important role in this respect (The Economist, 2010). By nature, however, the subsidy mechanism can only conserve limited regions because its source of finance is not always given nor stable, so it usually targets the regions that are already well known to the public because in this way it is easier to secure funding. That is to say by nature, such a subsidy mechanism is likely to exclude other marginalized regions that are not popular to the public (M. Park, pers. comm.). While such conservation activities have merit, those regions that appear to have less appeal to the public may still very well be environmentally and culturally valuable and thus important to conserve. On the contrary, REDD is not aiming to become another form of nature pricing mechanism that relies on one-way subsidies. The mechanism of REDD is rather geared towards generating and supporting trades of carbon credits through a market system in order to maximize the resulting environmental and finan- cial benefits and share these globally and locally. Thus, it is expected to secure forests in any remote and small areas unlike the PES and TEEB. As the mechanism is designed in a sense to make great use of existing sciences, multiple components constitute REDD including, but not limited to, reference emis- sion level (VCS, 2012c, 2010c; LEAF, 2011); measurement, reporting and verification (MRV); safeguard; and multiple benefits (UN-REDD, 2013). The reference emission level (REL) basically plays a crucial role in determining the amount of carbon credits, while the MRV refers to a systematic auditing to make sure forest carbon stocks are well protected, after any carbon credits are generated based on that protection. The safeguard indicates a suite of institutional and legal protections for local people who rely on forests to sustain their livelihoods. Lastly, the multiple benefits relate to the 4 other potential accomplishments that REDD may be able to facilitate and support (UN- REDD, 2013). For instance, reducing deforestation and forest degradation will not only conserve forest carbon stocks to mitigate climate change, but also support the forest’s biodiversity and other ecosystem services that were previously mentioned to benefit humans. As such, it sounds truly fascinating what REDD aims to achieve, but it still has a long way to go. Especially, the REL component appears to be problematic, so it has to be fixed in order for society to successfully implement REDD. That said, address- ing existing problems in REL setting and suggesting alternatives are the purpose of this dissertation. Many studies have reviewed different components/modules of REDD, ranging from its general framework (Venter and Koh, 2011; Phelps et al., 2010; Mollicone et al., 2007), equity and financing issues (Sandker et al., 2010; Zaks et al., 2009), forest gover- nance and resource management (Chhatre and Agrawal, 2009; Agrawal et al., 2008), and biodiversity conservation (Busch et al., 2011; Harvey et al., 2010; Gaveau et al., 2009), but very few have worked on issues of how an REL may vary when (1) different def- initions of “deforestation” and “forest degradation” and (2) different sizes of “spatial extent” are involved, where an REL refers to a business-as-usual carbon emission when deforestation and forest degradation would continue indefinitely (Figure 1.1). Having a sound REL is important for any REDD projects because their carbon credits will be generated based on that figure. Without having a solid REL, REDD may not benefit all developing and least devel- oped countries fairly because the amounts of annual deforestation and forest degrada- tion are quantified differently, and the mechanism of REDD does not adequately man- age such problems. Particularly, forest degradation (e.g., conversion of intact forests to secondary forests) is more difficult to define and measure than deforestation (e.g., con- version of forests to croplands), according to IPCC (Penman et al., 2003). So it is quite common that ongoing REDD projects are more about reducing deforestation rather than reducing forest degradation. As a result, the current version of REDD may be 5 CO2 Time Present Future Business-as-usual REDD implementation Figure 1.1: Concept of reference emission level (REL) in fact providing distortionary incentives, unfairly favoring countries that have more deforestation relative to those with forest degradation even when the carbon emis- sions of both deforestation and forest degradation have been thoroughly compared and analyzed with existing methodologies. In other words, the ongoing REDD projects— approved by the Climate, Community, and Biodiversity Alliance (CCBA) so their car- bon credits can be traded in global carbon markets—cannot be compared accurately because they all follow REDD protocols in different ways; that is, the estimates are in fact based on different methodologies (SCS, 2011; FARGC, 2009; Greenoxx, 2009; FCO, 2009; INBio et al., 2009; REDD Forests, 2009; IE, 2008; Aceh, 2008; FAS, 2008). Such dis- tortions and inequities will eventually limit the carbon emission reduction benefits of 6 REDD and thereby allow forest carbon losses to endure at the global level. In addi- tion, if REDD should become partially successful in Southeast Asian countries such as Laos or Thailand (LEAF, 2011), it may actually make matters worse by inducing local forest degradation (e.g., conversion of tropical rainforests to rubber plantations) to be exported to adjacent areas with similar geographic and climatic conditions (Lambin and Meyfroidt, 2011, 2010), such as tropical areas in China or Myanmar, which do not have REDD projects; such displacement of forest degradation and, of course, of deforestation were not the aim of REDD. 1.1.1 Definitional issues Ideally, deforestation and forest degradation in different regions should be monitored simultaneously because if deforestation in a region is exported to another region, then such displacement of forest carbon loss (i.e., “leakage”) must be taken into account when generating carbon credits, so the emission trading mechanism can be truly ben- eficial to climate change mitigation. Using an unified methodology may be suitable in this case because REDD is actually about collectively aggregating localized mitigation efforts to fight against climate change, so there is a fundamental need of inventing the unified methodology to facilitate any trades of carbon credits among different locales, where the methodology must have a comprehensive definition of “forest degradation” and, of course, of “deforestation.” The definitions should be applicable to different geo- graphic regions at multiple spatial levels that are likely to emit carbon in the future regardless of national borders. There are about 90 different definitions of forest around the world (ICRAF, 2012; Ramankutty et al., 2007; Lepers et al., 2005), and this may be one of the greatest chal- lenges for REDD stakeholders in terms of inventing the unified methodology to account for changes of forest carbon stock and to systematically compare the changes at the global level (ICRAF, 2012). Quite naturally, the different definitions of forest lead to different forest carbon accounting methodologies. More customized methodologies are 7 needed as well to assure better accuracy than using the unified methodology for diverse tree species in different ecosystems, for instance. According to Food and Agriculture Organization of the United Nations (FAO), both tropical rainforests in southwest China and boreal forests in Mongolia are considered forests indiscriminately even though the way the tropical rainforests store carbon in their biomass is quite different from the boreal forests (FAO, 2007). The tropical rainforests allocate more carbon in their above- ground living biomass, such as canopy, than their below-ground living biomass, such as root, while the situation of boreal forests is just the opposite (Gower, 2003). Therefore, it seems somewhat unavoidable to employ different methodologies in reality, and if that is the case, then both unified and customized methodologies must assure the greatest transparency; otherwise, it is virtually impossible to encourage trading carbon credits generated through REDD projects in global carbon markets because buyers would not be able to tell how credible the credits are. To assure that level of methodological trans- parency, the first step would be to clearly specify and characterize what a “forest” is in the context of REDD. 1.1.2 Modeling issues The major difference between the other forestry carbon offset projects, such as afforesta- tion and reforestation, is that REDD projects include predictive land change modeling as part of their project plans while the other forestry projects do not. A carbon offset project that is mainly about planting additional trees does not need to prove its car- bon sequestration’s outcomes by simulating carbon stocks of the future (IPCC Working Group III, 2007). The project would be mainly about measuring carbon stocks of newly built forests with existing methods and data as well as monitoring them in a timely and systematic manner (Brown, 1997). Consequently, its carbon sequestration outcomes are measured by calculating the difference between those carbon stocks at different time periods. Such logic, however, is not applicable for REDD projects because their out- comes have to be estimated based on business-as-usual forestland loss in the future, 8 where such loss is unknown and, therefore, often simulated via predictive (quantita- tive) modeling. In short, REDD projects must include accurate measures of future loss of forestlands as well as valid approaches to prevent such loss, while their carbon stocks and the associated emissions are monitored regularly (Calmel et al., 2010; VCS, 2010c,d; World Bank BioCarbon Fund, 2008). The modeling of business-as-usual forestland loss for REDD implementation has two main problems: (1) spatial extents are not specified objectively when estimating rates of deforestation and/or forest degradation (Brown et al., 2007; Paladino and Pon- tius Jr., 2004); and (2) the models’ predictive accuracies in terms of areal change are usually only compared to determine the “best” model, where there is no sound way at the moment of writing to test a statistical difference of such predictive accuracies (Pontius Jr. and Millones, 2011; Kim, 2010). In the literature of land change modeling, the “business-as-usual” refers to a (future) trend that is similar, if not identical, to an observed current and/or historic trend. Without addressing the simple problems, it is virtually impossible to make good use of any land change models for REDD pur- poses. Paladino and Pontius Jr. (2004) mentioned that the spatial extent of a study area will affect any modeling outcomes, but very few studies have addressed this issue systematically. In the case of Brown et al. (2007), the authors demonstrate that differ- ent REL methodologies produce outcomes substantially different from each other, and such fluctuation could get worse when different sizes of spatial extent are employed. For instance, their comparative studies find that even for one specific study area and for one time period there could be an almost 40% difference in terms of forest-cover area because spatial extents and levels of data aggregation vary among existing REL methodologies. Thus, there is a need to explore the multiscalar attribute of REL through a systematic experiment. Assessing and comparing predictive accuracy of land change models may be useful as well for setting RELs in an accurate manner, and this can be coupled with the multiscalar experiment. In this way, research about identifying a more accurate land change model could become more convincing; however, finding a more 9 or the most accurate land change model may not contribute to the most accurate REL immediately. The ultimate purpose of employing spatially explicit land change mod- els is to estimate any future forest carbon loss as accurate as possible in terms of mass, where the land change models produce outcomes only in terms of area. So there is a gap between finding the most accurate land change model and estimating the most accu- rate measure of future carbon loss, in other words, REL. This research gap as well as the varying spatial extent issue must be addressed systematically in order for REDD stakeholders to cost-effectively produce a sound and transparent REL for their REDD projects. 1.2 Research Objectives and Questions Before attempting sophisticated modeling approaches, scientists first must be clear about exactly what they are measuring and must understand their data in a thorough manner; research that monitors past and present deforestation often disregards this principle. Without this crucial first step, conclusions are likely to rest on shaky founda- tions indeed. A striking number of forest carbon accounting and deforestation moni- toring studies, however, do not follow this crucial first step, while the situation in sim- ulating business-as-usual future deforestation has been similar or even worse due to additional modeling components and the associated uncertainties. The research outcomes are to bridge gaps among distinct fields of forest change monitoring, future deforestation simulation, and forest carbon stock measurement through clarifying the terms—“forest (i.e., forestland),” “deforestation,” and “forest degradation”—and amending the fluid spatial extent issue in an objective manner. Fill- ing those gaps is crucial because the terms and spatial extents are often employed arbi- trarily, while researchers consistently rely on each other’s findings for further advanced modeling approaches. This dissertation attempts to demonstrate what kind of aca- demic miscommunication exists in literature and provides a scientific foundation to 10 better specify, measure, and understand forest carbon dynamics in pan-tropical regions for REDD. The overarching research questions are as follows: How does a reference emission level fluctuate when different definitions of defor- estation are used with varying sizes of spatial extent? Does a more accurate measure of business-as-usual deforestation projection (e.g., in hectares) result in a more accurate estimation of forest carbon stock change (e.g., in tonnes)? More specifically, this dissertation addresses a few limitations of existing predictive land change models that are often used to project business-as-usual future deforesta- tion and estimate the resulting carbon emissions. Further, a suite of alternatives to overcome such limitations is suggested in order for scientists and/or policymakers to better implement an emerging global forestry policy, REDD. In addition, the research provides novel insights into the procedures relevant to future REDD projects in pan- tropical regions, in particular with respect to an REL setting that is prospective (i.e., forward-looking), spatially explicit, and nested at different spatial extents (i.e., nested REDD). Through a case study in southwest China, the limitations are articulated in a visual and spatially explicit manner, and the alternatives are demonstrated. The first limitation is about not having a clear definition of deforestation. In short, scientists have monitored different types of forest and also have simulated different types of future deforestation, while it was not always clear what they meant by “for- est” and “deforestation.” This is problematic because when the terms are not defined precisely, then any modeling results become hardly interpretable. Therefore, a system- atic way is proposed to specify the terms, namely “forest,” “deforestation,” and “forest degradation,” based on two important factors that determine a forestland: forest-cover and forest-use. Interchangeably using the two factors is highly likely to yield mislead- ing estimates of forest carbon stock change. Thus, it is crucial to distinguish the two and 11 Forest Time Higher rate Lower rate Figure 1.2: Time window of deforestation rate calibration treat them independently when specifying the terms and reporting the corresponding carbon emission outcomes for REDD purposes. The second limitation is about arbitrarily specifying a spatial extent when project- ing business-as-usual future deforestation at a sub-national level. This is troublesome because different time windows produce varying rates of deforestation (Figure 1.2), where the rate is arguably the most critical factor in determining the amount of carbon credits to be produced through REDD projects. This logic is also true for different sizes of spatial extent (e.g., province or county level), which is why a spatial extent should not be specified arbitrarily (Figure 1.3). Therefore, a quantitative approach is proposed to delineate hotspots of deforestation and forest degradation, where the approach spec- ifies a spatial extent in a data-driven way. It is possible to delineate different types of 12 Deforested Higher rate Lower rate Forested Figure 1.3: Spatial extent of deforestation rate calibration hotspots that are statistically significant, where such a hotspot dictates the spatial extent when calibrating rates of deforestation. It is verified that different sizes of spatial extent results in different rates of deforestation even when the identical time period is used for the calculation. The third limitation is about employing a popular yet unverified premise—a more accurate simulation of business-as-usual deforestation (e.g., in hectares) will result in a more accurate estimation of forest carbon stock change (e.g., in tonnes)—when land change modelers attempt to identify what would be considered the “best” land change model in the context of REDD. The premise, however, is still unverified in this disserta- tion mainly because errors embedded in biomass data/estimates turn out to be larger 13 than expected. Therefore, more accurate biomass data should be used to compare dif- ferent RELs. This finding implies that research merely comparing predictive accuracies of different land change models for REDD purposes may have in fact limited contribu- tion to REDD. Finally, a new land change model, entitled “Geographic Emission Benchmark (GEB),” is proposed to forecast trends of business-as-usual future deforestation and estimate the associated forest carbon stock change in a more transparent manner than other existing ones. The new approach relies on peer-reviewed global datasets to objec- tively characterize “deforestation” and “forest degradation” and open-sourced spatial methods to control varying spatial extents, but it omits selected sensitivity analyses. Through the case study in southwest China, the approach (1) clarifies what it means by “forest,” “deforestation,” and “forest degradation;” (2) controls spatial extent setting when estimating rates of deforestation and those of forest degradation; and (3) demon- strates predictive accuracies of land change models cannot be conclusively compared. Specifically, business-as-usual loss of forestlands is simulated by the GEB and GEO- MOD modeling, where the results are compared and validated with the observed change of forestlands. The outcome shows the regional forest biomass loss due to deforestation and forest degradation between 2000 and 2010. Their predictive accu- racies, measured by Figure of Merit, might differ insignificantly even with the clarified definitions of “forest” and “deforestation.” Thus, it may be more economical to employ the GEB, first and foremost, before applying other more time-consuming land change models because simulating business-as-usual forestland loss may not be the most criti- cal factor that determines an accurate measure of forest carbon loss in future. Or at least, the GEB’s modeling outcome should be used as a baseline for other REL outcomes pro- duced by more sophisticated land change models. A case study is conducted in Xishuangbanna Dai Autonomous Prefecture (here- after, Banna) located in southwest Yunnan, China (Figure 1.4), where rubber planta- tions are regarded as “natural” forests by the central government (Xu, 2011). In other 14 words, from the government’s perspective, intact tropical rainforests are identical to the rubber plantations. As a result, the latter has expanded rapidly along with the grow- ing national economy since 1970s (Qiu, 2009; Li et al., 2008, 2007) because there have been no penalties or regulations for producing rubber, while its domestic demands to manufacture tires for automobiles and trucks have drastically increased. From a con- servationist’s perspective, however, such expansion does not only threaten habitats of endangered animal species (e.g., white elephants) but also may risk the local carbon sequestration activities because the intact tropical rainforests are larger carbon pools compared to the rubber plantations (Xi, 2009). This site appears to be a suitable place to test the newly proposed REL setting approach because the landscape includes forests that are under distinct uses (e.g., nature reserves versus rubber plantations). Thus, the site offers an apt basis of comparison for demonstrating how differently character- ized “deforestations” will produce different carbon outcomes. Moreover, because the study area exemplifies massive rubber production in the Greater Mekong Subregion (Li and Fox, 2012), this dissertations approach could therefore be tested in a larger region, including Laos, Thailand, Vietnam, Cambodia, and Myanmar. 1.3 Organization of Subsequent Chapters The remainder of this dissertation is concisely introduced as follows: In Chapter 2, the literature review is conducted, showing challenges that REDD faces with respect to modeling forest carbon dynamics. This chapter also provides a thorough discus- sion of the quantitative methods used to model those dynamics along with internation- ally accepted protocols of Verified Carbon Standard (VCS). In Chapter 3, the new land change model is proposed to overcome the challenges pointed out in Chapter 2. In Chapter 4, performance of the newly proposed model is compared with that of a GEO- MOD modeling through a case study of Banna in southwest Yunnan, China. Chapter 5 demonstrates an alternative to further improve land change models by reinforcing their 15 Xishuangbanna Counties of Yunnan Other provinces in China 0 200 400 100 Miles Figure 1.4: Map of the study areas: Banna prefecture (black) and Yunnan province (gray) spatial structures. A geographically weighted regression model is used to explain rub- ber productivity in Banna to infer how rubber plantations expand. Finally, in Chapter 6, intellectual merit and broader impacts of the dissertation are summarized, while the chapter also hints of future research through comparative REL studies with respect to forest transitions in China and Southeast Asia. 16 Chapter 2 Literature Review 2.1 Forest Carbon Flux and Deforestation The current section is twofold: Section 2.1.1 introduces and explains basic concepts and terminologies that are frequently used in general forestry-based carbon offset projects or GHG inventory analyses. It first explains how a “forest (i.e., forestland)” is charac- terized by two factors, namely forest-cover and forest-use, and points out the difference between the two. It then articulates why systematically specifying the “forest” is crucial in the context of REDD. Provided also is a description of how a system of carbon flux operates and what types of forest carbon pools exist and could be taken into account when measuring forest carbon stocks or potential carbon emissions when forests are disturbed. Section 2.1.2 presents a comprehensive review of literature regarding observed changes of forestland and their simulations using quantitative methods. The former is referred to as the “observation” literature as it is mostly about monitoring changes of forestland, mainly through remote sensing, so it does not include further land change modeling approaches. The latter is referred to as the “projection” literature because it mainly deals with land change modeling approaches that project business-as-usual changes of forest-use and/or forest-cover over time and space. The term “projection” seems more appropriate to denote the literature as opposed to terms such as simula- tion, prediction, and modeling because the literature is in fact about summarizing the land change forecasting methodologies merely casting the past and current trends of forestland change to the future. This is actually what land change modelers mean by 17 “business-as-usual.” That is to say, the projection literature is not about simulating dif- ferent scenarios of deforestation nor modeling complicated socioeconomic/biophysical systems or processes of forestland change. Although the research to understand such business-as-usual trends of deforestation might seem too simple to grasp the reality of tropical deforestation, these do play an important role in REDD because by far these are the methods supported by VCS. Especially, GEOMOD modeling is arguably the most popular approach for setting an REL for REDD projects (Sloan and Pelletier, 2012; Kim, 2010; Harris et al., 2008; Brown et al., 2007; Brown, 2005, 2002b; Dushku and Brown, 2003). The literature review points out how such a simple approach can be even mis- used to produce invalid outcomes. 2.1.1 Concepts and definition of terms Characterizing a “forest.” First, it is imperative to define what a “forest” is (the term is used interchangeably with a “forestland” throughout the dissertation). Theoretically speaking, to define “forest,” three parameters, namely (1) forest-cover, (2) forest-use, and (3) ecological aspects of forest ecosystem, should be taken into account (Herold et al., 2006). First, forest-cover refers to the biophysical aspect of a forestland, such as percent of crown-cover or tree height, and this aspect is often measured remotely via satellite imagery or directly in the field by a hypsometer. Second, forest-use reflects human intention to generate any economic profits or other benefits from a specific forestland, such as ecotourism in nature reserves or timber harvesting in plantations; this aspect can be measured by conducting surveys or interviews or by analyzing forest- cover data as proxy to understand the corresponding forest-use. Finally, ecological aspects of forest ecosystem indicate a suite of biophysical services that a forestland is able to provide to the landscape; these services cannot be provided by any arbitrary group of trees that are not forests (B. L. Turner Jr., pers. comm.). This third aspect is most challenging to measure and often difficult even to identify as it requires multi- ple sets of theories, methods, and measurements. As shown in Figure 2.1, although it 18 Forest-cover (Biophysical) Forest-use (Human) Ecological functions of forest Forest(land) area .> .5 ha height > 5 m crown.>.10 % Figure 2.1: Ideal characterization of forest is known that those three parameters are required to define a “forest,” they have not always been thoroughly considered by researchers in the literature. Forest-cover versus forest-use. When FAO (2007) defines a “forest,” at least, both forest-cover and forest-use dimensions are taken into account, and it is crucial to dis- tinguish the two and treat them as distinct parameters. FAO and United Nations Envi- ronmental Programme (UNEP) define “land-use,” where the forest-use (i.e., forestland- use) is regarded as a type of land-use, as “the arrangements, activities and inputs peo- ple undertake in a certain land[-]cover type to produce, change or maintain it,” while “land-cover” is considered “the observed (bio)physical cover on the earth’s surface” (FAO and UNEP, 1999), so they should not be used interchangeably. This separation is also reasonable from a data-collecting perspective because measuring the forest-cover 19 requires different data-collecting methodologies compared to measuring the forest-use. Remote sensing is arguably the most popular methodology to collect forest-cover data, especially for a large region (Achard et al., 2010, 2007; GOFC-GOLD, 2010; Olander et al., 2008; DeFries et al., 2007), and it is sometimes possible to differentiate different forest-uses with this method. Differentiating forest-uses in this respect, however, is technically about distinguishing different tree species, and they actually are a member of forest-cover data as those are distinguished based on their biophysical properties. The data indirectly inform certain forest-uses. Therefore, it does not seem perfectly log- ical to consider this remote sensing methodology for collecting forest-use data. Survey and interview, on the other hand, can actually collect data about forest-uses in a direct way because any forest-use has to be operated by certain “agents”—people or organi- zations who use the forest—so it is reasonable to directly ask the agents what they do about their forests, and this information cannot be collected by remote sensing (Rob- bins, 2001; Robbins and Maddock, 2000). In the study area, some forests are protected as nature reserves so they are used for local tourism and/or academic research, while many others are rubber plantations. The nature reserves and rubber plantations may have similar forest-covers, but their uses are completely different, so it appears to be reasonable to employ both forest-cover and forest-use to specify distinct forests. When tropical rainforests are replaced by rubber plantations, the regional carbon flux will change simultaneously although both of them are regarded as “forests” by the central government indiscriminately. That is, measur- ing forest carbon stocks cannot avoid potential measurement errors because the govern- ment’s definition of “forest” is not specific enough to depict the associated forest carbon flux dynamics as it only takes forest-cover into consideration when determining a “for- est.” If forest-use is considered in addition to the forest-cover, the carbon flux of both tropical rainforests and rubber plantations, then, can be better distinguished. Conse- quently, it becomes systematic to specify “deforestation” and “forest degradation.” For example, one can answer the following question systematically and transparently: Is 20 the conversion from the tropical rainforests to the rubber plantations considered defor- estation or forest degradation? On the one hand, it might be considered deforestation as the land-use has changed rather permanently; on the other hand, it might be for- est degradation because the land-cover stays identical but with a smaller amount of biomass. Forest and REDD. Why then is it so important to clarify the term “forest” in the context of REDD? That is simply because REDD is largely about measuring business- as-usual “forest” carbon stock loss due to de“forest”ation, “forest” degradation, and other environmental degradation, which is related to any post-deforestation, while such terms cannot be specified if a “forest” is not adequately defined (ICRAF, 2012). Cop- ing with these definitional issues is crucial because, by nature, measuring forest carbon stocks involves various measurement errors since it combines various datasets from different sources (e.g., satellite imagery or field), so it seems rational at least to mini- mize the errors that are originated from any unclear terms. As pointed out in the pre- vious paragraph, it is challenging to clearly specify a “forest” by utilizing all of those three parameters, especially at the landscape (or global) scale, but the urgency of REDD overwhelms such pursuit of academic perfection. That being said, it seems reasonable to first crudely characterize a “forest” that is measurable and quantifiable both in terms of area and carbon stock and then to discuss how to improve the clarification of the term later on. In fact, this is how many scientists have accounted change of forest stocks in the literature (Hansen et al., 2010, 2009, 2008, 1998; DeFries et al., 2002, 2000, 1998); usually, they make good use of canopy information to characterize forests and to mea- sure those, although “deforestation” and “forest degradation” are not always clarified in the literature (Penman et al., 2003). Such limitations are addressed in the following paragraphs to provide a better foundation to support REDD. System of carbon flux. In order to precisely understand and measure any carbon emissions, it is fundamental to grasp the full mechanism of carbon flux system. Carbon literally exists everywhere. Some exists in the atmosphere, while some is embedded 21 in wood and paper products that humans frequently use and dispose of. Carbon is also mobile. That is, it does not stay in a particular (carbon) pool; rather, it moves around from one pool to another and to another endlessly. When carbon accumulates in a biomass, for instance, basically the carbon is transferred from the atmosphere to the biomass. This process could be rephrased as a carbon transfer from one carbon pool (i.e., the atmosphere) to another (i.e., the biomass). Hence from the biomass per- spective, it sequesters the carbon, while momentarily the carbon was removed from the atmosphere; “carbon removal” and “carbon sequestration” in fact indicate the identi- cal direction of carbon flux. Besides, the biomass in this case is regarded as a “carbon sink.” Conversely, if the biomass emits or releases carbon to the atmosphere, the biomass is then considered a “carbon source.” Figure 2.2 visualizes the system of such carbon flux (Kim, 2013). Forest carbon pools. A forest (i.e., forestland) has multiple carbon pools. Accord- ing to IPCC (2006, Vol. 4, Ch. 1), one is constituted by five carbon pools: above-ground living biomass (AGB), below-ground living biomass (BGB), deadwood, litter, and soil organic matter (Table 2.1 and Figure 2.3). First, the AGB refers to alive trunks, branches, and foliage; these are some of the things that lay people notice immediately when think- ing of a forest. The AGB has been measured and accounted relatively more vigorously than the other forest carbon pools especially through remote sensing (Asner et al., 2012, 2010; Lefsky, 2010; Saatchi et al., 2011, 2007). The BGB indicates alive roots, while these are difficult to measure directly because they are usually covered by soil. Hence, the BGB is actually often estimated using the AGB as proxy because it was found there is a distinct relationship between the two, e.g. a larger AGB usually has a larger BGB (Skole et al., 2011; Damuth, 2001). The deadwood and litter are not part of the AGB or BGB, but instead are part of the dead organic matter; the deadwood includes stumps and dead branches, while the litter indicates “forest floor” including all non-living biomass that is larger than the soil organic matter. While most of the forest carbon pools mainly 22 CARBON FLUX CARBON POOL Carbon Carbon Carbon Sink (removal, sequestration) Carbon Source (emission, release) Figure 2.2: System of carbon flux, reprinted from Kim (2013) store carbon in their biomass via photosynthesis, the soil organic matter removes car- bon from the atmosphere through humification (Lal, 2008). Table 2.1 shows the amount of carbon stocks embedded in each carbon pool, where the litter category also encom- passes deadwood (FAO, 2010). Table 2.1: Carbon stock in forest carbon pools (in million tonnes) Region Biomass Litter Soil Total East Asia 8,754 1,836 17,270 27,860 South and Southeast Asia 25,204 1,051 16,466 42,722 World 288,821 71,888 291,662 652,371 Hardwood product (HWP) refers to the merchandized goods constituted by woody biomass acquired by humans at harvest sites, while the remainder is considered dead 23 organic matter (IPCC, 2006, Vol. 4, Ch. 12); thus, technically the HWP should be cou- pled with the five carbon pools introduced above to more comprehensively understand the forest carbon flux (Gower et al., 2006; Gower, 2003). Figure 2.3 shows them all together. Fuelwood is also worth considering as it emits carbon relatively immediately, and the fuelwood combustion is common in the study area. For instance, coated paper (Newell and Vos, 2012, 2011) will store carbon for a shorter time period than round- woods used for building construction (Perez-Garcia et al., 2006, 2005). The HWP and fuelwood pools are not analyzed in this dissertation (e.g., which stores carbon more than the other), but still it is important to keep the complete list of forest carbon pools in mind because it is likely that a future REDD might mandate to include them as they are already required by other non-REDD carbon protocols (Newell and Vos, 2012, 2011; BSI British Standards et al., 2011; WRI and WBCSD, 2011). In sum, the forest carbon pools are all intertwined to each other, and maximizing the amount of carbon stock residing in the forest carbon pools and the duration is a key strategy to mitigate climate change in forestry. 2.1.2 Forestland change: observation and projection One unique thing about the literature review in this dissertation is that it not only sum- marizes how forestlands are characterized but also provides the overview about how the forestlands are measured. This is different from the literature simply listing various for- est accounting methodologies (Achard et al., 2007; DeFries et al., 2007) or showing dif- ferent definitions of forest (FAO, 2007; IPCC, 2006). It specifically targets the literature that talks about both definitions and measurement and how those two are employed in tandem to account either areal change of deforestation or voluminous/mass change of forest carbon stocks. As previously described, forest-cover and forest-use are the distinct parameters that characterize a “forest”; therefore, they should not be used interchangeably because they do not always have a one-to-one relationship. According to Brown and Duh (2004), 24 Above-ground living biomass Below-ground living biomass Deadwood Litter & Soil organic matter Hardwood product Fuelwood CARBON POOLS OF “FOREST” area .> .5 ha height > 5 m crown.>.10 % Figure 2.3: Forest carbon pools in Banna prefecture, photographed by the author in December 2011 the semantic relationship between the two is quite complicated in urban areas, so the authors argue that a conversion from one land-use to another may involve a series of land-cover changes that can be expressed as a conversion matrix with stochastic ele- ments. For example, if a patch of land becomes a residential area, the land-use con- version will trigger the associated land-cover change to include paved impervious sur- face, lawn, trees, and sometimes swimming pools. Thus if a land-cover and land-use were assumed to be identical, then the outcome cannot avoid the associated semantic uncertainties. The situation is quite similar in rural areas of tropical regions. The World Agroforestry Centre (ICRAF) points out that there are many trees observed in croplands due to the practice of agroforestry, and it is unclear whether or not such trees should 25 be treated as forests if they happen to meet the criteria of FAO’s definition of forest- cover (Zomer et al., 2009). A tea terrace in Assam, India, exemplifies this complicated relationship because the place utilizes trees/forests to benefit the local tea production (J. Xu, pers. comm.). In this situation, the tea terrace might include some forests—one type of land-use (i.e., tea production) having two types of land-cover (i.e., tea terraces and forests). Thus, it is truly difficult to identify forests solely using either forest-cover or forest-use. Both of them are required. The terms, however, are frequently used interchangeably in the literature, and this is where problems start. By separating the literature of deforestation research into “observation” and “projection,” the problems occurred due to such interchangeable use becomes apparent (Figure 2.4). Most studies about observing forest stock change tend to overlook the forest-use dimension, hence in fact accounting only for forest-cover when monitoring “deforestation,” which is basically identical to monitoring “forest- cover loss.” The projection literature largely relies on the data produced by the obser- vation literature, mainly forest-cover maps, to simulate business-as-usual deforestation of the future using the variables that explain forest-use change where the associated semantic uncertainties are usually overlooked or omitted. This is a serious problem because the situation interferes with academic research to comprehend the logical rela- tionship among forest-cover, forest-use, their changes, and what drives such changes. It is problematic for REDD stakeholders as well because it is likely that they may have incorrect information about drivers of deforestation or might project business-as-usual deforestation in an erroneous way. Observation. Spatially explicit benchmark estimates of “forest” biomass at the pan- tropical level are published in Science, Nature Climate Change, and Proceedings of the National Academy of Sciences (Baccini et al., 2012; Harris et al., 2012; Saatchi et al., 2011); it is crucial to thoroughly comprehend the published estimates because many carbon emission studies are likely to rely on their findings. The estimates are created by fus- ing plot-level information, which is usually measured in the field, with landscape-level 26 Observation Projection ... often measures forest-cover and rarely forest-use. ... employs drivers of forest-use to model forest-cover change. units: hectare or percent “Forest” and “deforestation” are loosely specified. FOREST CHANGE Figure 2.4: Structure of the literature review on deforestation research information, which is usually acquired through remote sensing; advanced statistical methods, such as maximum entropy approach (Elith et al., 2011), are used to merge the information. In other words, the benchmark estimates are produced by combining hundreds of plot data scattered around in pan-tropical regions with respect to remotely sensed vegetation indices to extrapolate (and also to interpolate) the plot level knowl- edge towards the pan-tropical level. This data fusion is useful because it is geared towards producing carbon estimates that are most accurate at the broadest scale in an optimal fashion. The estimates vary partly because the authors employ different defi- nitions of forest (Table 2.2), and as a result it is likely that this may cause further error propagation. 27 Table 2.2: Different types of “forest” biomass Baccini et al. (2012) Harris et al. (2012) Saatchi et al. (2011) Area Not clear >.09 ha Not clear Canopy >50 % >25 % 10%, 20%, and 30% Height >15 m >5 m >0 m Diameter >5 cm >10 cm >10 cm Harris et al. (2012) clearly specifies what they mean by “forest” although their def- inition is somewhat different from FAO’s definition (FAO, 2007). It is neither more conservative nor generous than FAO’s because, on the one hand, the definition is more generous in terms of area (i.e.,>.09 ha rather than>.5 ha), and this may allow inclusion of other wooded lands that are not forests; on the other hand, the definition is more con- servative in terms of canopy (i.e.,>25% rather than>10%), which excludes the biomass that is considered forests according to FAO’s criteria. Baccini et al. (2012) use the term “forest” less carefully compared to Harris et al. (2012) as they do not mention any areal component when specifying their forest, but at least the authors acknowledge the dif- ference between forest and other wooded land, while the term “woody vegetation” (182) is used to indicate the two as a whole. In contrast, Saatchi et al. (2011) define their forest solely using the canopy-cover information acquired from data of Vegetation Continuous Field (Hansen et al., 2003), and it is not certain whether or not the authors include information of forestland’s area or height. Thus, when Saatchi et al. (2011) pro- duce a carbon stock estimate for AGB, it is not very clear what he and his colleagues mean by AGB without clearly specifying what a “forest” is, given the AGB is a type of carbon pool of a forestland (Figure 2.3). Hence, the estimate may include carbon stocks of other biomasses such as other wooded lands or fragmented trees. Moreover, Baccini et al. (2012), Harris et al. (2012), and Saatchi et al. (2011) use different minimum sizes of diameter at breath height ([DBH], i.e., 1.3 m), which also interferes with direct com- parison of the biomass estimates (Table 2.2). The DBH is the most important metric in 28 measuring carbon stocks of biomass (Laar and Akc ¸a, 2007), and it basically measures the thickness of biomass trunks. Such varying characterizations of “forest” may confuse scientists who study trop- ical deforestation and the resulting carbon flux dynamics, and this situation is prob- lematic for two main reasons: first and foremost, it is highly likely that the benchmark carbon estimates may include carbon stocks of other woody biomasses that are not for- est or may exclude some forest carbon stocks; this results in initial errors, and such errors are likely to propagate when those are used for other complicated carbon flux modeling research. As a result, the research outcomes would be unreliable regardless of the advanced and improved methods used. Second, it becomes difficult to specify what “deforestation” is. Harris et al. (2012), for example, defines “deforestation” as “the area of forest cover removed because of any human or naturally induced distur- bance” (1574), so the authors do not include the forest-use perspective when defining their deforestation. As such, it is quite likely the research outcomes may not distin- guish “temporarily unstocked forests” (ICRAF, 2012, 3) from a real deforestation. They refer to harvested sites in plantations since in some cases, the plantations are considered “forests” although the biomasses are harvested regularly. It is even more unclear what Baccini et al. (2012) and Saatchi et al. (2011) mean by “deforestation,” and the main rea- son why the term “deforestation” is unclear is due to the absence of a clear definition of “forest.” Except for Broich et al. (2011), tropical deforestation research that is based on change detection analysis tends to loosely define “deforestation” as a forest-cover conversion from forest to non-forest (Sangermano et al., 2012; Sloan and Pelletier, 2012; Kim, 2010; Hansen et al., 2009). Broich et al. (2011) differentiate “deforestation” as opposed to “forest-cover loss,” and the authors acknowledge the difference can affect any result- ing forest carbon profiles because from their perspective, the deforestation implies a permanent land-use change while the forest-cover loss can be a temporary incidence. While forest-cover and forest-use characterize a “forest,” the forest-cover is constituted 29 by three biophysical factors: (1) an area of forestland, (2) its canopy- or crown-cover, and (3) the forestland’s height (FAO, 2007). Hence, in order to properly characterize a forest- cover—of course to ultimately characterize a “forest (i.e, forestland)”—the three factors should be taken into account simultaneously. Although in other literature, forest-cover is considered similar to “tree-,” “canopy-,” or “crown-cover,” in this dissertation they are clearly distinguished. Limited studies, however, consider those three factors thor- oughly, and a majority usually employs one or two factors to characterize a forest-cover (Asner et al., 2009; van der Werf et al., 2009; Hansen et al., 2008; Defries et al., 2006; Miles et al., 2006; Lepers et al., 2005; Morton et al., 2005), and it is more common even not to distinguish forest-cover and forest-use (Macedo et al., 2012; Baker et al., 2010; DeFries and Rosenzweig, 2010; DeFries et al., 2010; Rudel et al., 2009; Morton et al., 2008; Foley et al., 2007; Ramankutty et al., 2007; Anderson et al., 2005; Jasinski et al., 2005; Bonan et al., 2004; Lawrence et al., 1995). Instead, they usually characterize “forests” by analyz- ing spectral information of remotely-sensed data, while the outcomes might not always be ground-truthed with the three factors in mind. For example, the researchers may have validated crown-cover information but are likely to ignore area or height informa- tion. Lastly, the terms “crown-cover” and “canopy-cover” must be distinguished as well. Saatchi et al. (2011) provide carbon estimates of forestlands in pan-tropical areas. Using different thresholds of percent canopy-cover, their research shows how forest carbon stocks vary over space. However, their estimates would make sense only if “canopy- cover” is identical to either “tree-” or “crown-cover.” The terms are often used inter- changeably because IPCC (2006) treats the canopy-cover as identical to the crown-cover, whereas Hansen et al. (2003), who are the producers of the Vegetation Continuous Field (VCF) layers, distinguished the terms based on different measuring methodolo- gies. According to the authors, 80% of canopy-cover is equivalent to 100% of crown- cover, where the “canopy-cover” refers to “the amount of skylight obstructed by tree canopies equal to or greater than 5m in height and is different than per cent crown 30 cover” (Hansen et al., 2003, 6), while the “crown-cover” is defined as “the percentage of the ground covered by a vertical projection of the outermost perimeter of the natural spread of the foliage” (IPCC, 2006, 4.73). As Saatchi et al. (2011) employ the VCF layers to define distinct forests with respect to different levels of canopy-cover, their carbon estimates for each forest may in fact contain errors because they did not use the terms clearly. Projection. The projection portion is actually a subset of land change modeling literature, where the land change modeling refers to a suite of quantitative method- ologies to test and verify knowledge of land change processes by conducting scientific experiments (Aspinall, 2006; Veldkamp and Lambin, 2001). Due to the interdisciplinary nature of land-use and land-cover change, the topic has been studied by various disci- plines such as Agricultural Economics and Geography, and it is one of the most impor- tant subfields in Land Change Science (Turner Jr. and Robbins, 2008; Turner Jr. et al., 2007; Heistermann et al., 2006; Rindfuss et al., 2004; Agarwal et al., 2002). Heistermann et al. (2006) categorize numerous land change models to three major categories: (1) geo- graphic, (2) economic, and (3) integrated geographic and economic models, where a “land change model” is defined as “a tool to compute the change of area allocated to at least one specific land-use type” (142). According to the authors, the geographic model is geared towards analyzing land suitability and its spatial association with various spa- tial factors, such as proximity to roads, while the economic model is about identifying drivers of land change at a more aggregated level in an aspatial fashion; this approach allows engaging more diverse macro data (e.g., market and population structures) than the geographic model. The integrated model seeks to combine the advantages of both approaches, and it often aims at incorporating the interactions of terrestrial environ- ment and global economy. Geographic Information Science/Systems (GIS) often plays an important role in dealing with various data from different sources (Goodchild, 2003), and satellite imagery in particular is frequently employed to characterize different land 31 categories and their conversions. In a way, the land change modeling is a type of time- series analysis that has a unique data structure. Unlike traditional time-series analyses in other social sciences, it has a relatively limited number of temporal observations, often a few time periods. Therefore, it seems natural to utilize additional spatial infor- mation acquired from the satellite images or population census data that contain spatial contexts to understand the associated temporal dynamics to more accurately predict the future. That is to say, the land change modeling is, by nature, spatial, and this litera- ture review is focused on the geographic model as this is more preferred than the other models in the context of REDD (VCS, 2012c). The geographic model is constituted by two components: quantity and (spatial) allocation (Guti´ errez-V´ elez and Pontius Jr., 2012; Pontius Jr. and Millones, 2011; Pontius Jr., 2000). The former refers to an amount of land change (i.e., how many hectares of forestland will be lost?), while the latter indicates the amount’s spatial allocation (i.e., where does the deforestation occur?). Basically, how the geographic land change model works is that when a quantity of land change is estimated, the estimate is converted to a number of pixels and those pixels are sprinkled on a rank map that shows different degrees of transition potential. The “transition potential” is defined as “a degree to which locations might potentially change in a future period of time” (Eastman et al., 2005, 345). The quantity module is estimated sometimes through simple linear models (Kim, 2010; Eastman, 2007; Pontius Jr. and Chen, 2006; Brown, 2005, 2002b) and mainly through econometric models due to their strength in estimating how much land would be converted with respect to numerous factors that drive and determine land change, such as rent, local climate conditions, global market systems, or demographic structures (CIFOR, 2008; Heistermann et al., 2006; Lubowski et al., 2006; Polsky, 2004; Geoghe- gan et al., 2001; Wood and Skole, 1998). Estimating the amount of land change is also strongly related to the observation literature as it includes many findings about the loss of tropical forestlands (Achard et al., 2007, 2002, 2004; Ramankutty et al., 2007; 32 Houghton, 2005, 1999, 1996; Houghton et al., 1985, 1983; Lepers et al., 2005; Lambin, 1997; DeFries et al., 2002). It has been found that such rates can vary substantially when different geographical boundaries, that is, spatial extents, (e.g., the Amazon vs. pan- tropics), are involved and, hence, impede a direct comparison of the different regional deforestation rates (Ramankutty et al., 2007). The (spatial) allocation module is closely related to species distribution and ecologi- cal niche modeling (Sehgal et al., 2011; Gillespie et al., 2009, 2008; Buermann et al., 2008; Peterson et al., 2002) because identifying future deforestation is basically about delineat- ing niches of human species, where such niches embody our interests and willingness to use lands differently. The allocation module is estimated via numerous (quantitative) methods ranging from, but are not limited to, logistic regression (Kim, 2010), artificial neural network (Sangermano et al., 2012; Kim, 2010), and machine-learning (Elith et al., 2011; Sehgal et al., 2011); these are all geared towards fusing various spatial variables that can explain shaping niches of human species. That said, the spatial variables are geared towards explaining forest-use change (e.g., Sangermano et al., 2012; Sloan and Pelletier, 2012; or Kim, 2010) but not necessarily forest-cover, whereas numerous forest- cover maps have been subsidized by the observation literature to simulate future defor- estation in a spatially explicit manner. Thus, it is highly likely the associated modeling outcome would embed semantic errors. That is why a land change model must clearly specify the forest-cover and forest-use components and not interchangeably use those. 2.2 Methodology of Reference Emission Level This section includes the specifics of REDD methodologies in terms of setting a ref- erence emission level (REL), so by nature, this literature review includes both policy and science aspects of the REL construction. The policy aspect is about comprehending 33 existing requirements so that REDD stakeholders can officially set an REL to gener- ate valid carbon credits that can be listed and traded via global/domestic carbon mar- kets, whereas the science aspect is geared towards constructing the REL most accurately since the existing methodologies do not provide the greatest detail. That is to say, Sec- tion 2.2.1 shows the specifics of the VCS methodology, including its requirements to set an REL and concepts of basic terms; the VCS methodology is the most frequently used protocol for REDD projects (Diaz et al., 2011). Section 2.2.2 points out scientific limi- tations of the VCS methodology. Therefore, the section is based on the peer-reviewed journal articles that function as the scientific foundation of the VCS methodology. The literature of GEOMOD modeling is also summarized. 2.2.1 Verified carbon standard Baseline and reference emission level. Technically speaking, there is a clear distinction between “baseline” and “REL” (Sloan and Pelletier, 2012). A baseline is always set based on an REL, as the latter is more about scientifically measured carbon emissions driven by deforestation and/or forest degradation (e.g., predicted through land change mod- els), whereas the former is a subjective adjustment of the REL by policymakers based on their qualitative knowledge and/or political decisions. Hence, a baseline could be higher or lower than an REL, and the difference is caused by the policymakers’ sub- jective intervention. The term, reference level (RL), is also often used nowadays, and this is different from the REL. As REDD has evolved to REDD+, by incorporating the role of conservation, sustainable management of forests, and enhancement of forest carbon stocks, in addition to deforestation and forest degradation reduction (UNFCCC COP, 2011), it appears reasonable to have a new reference methodology (i.e., RL) com- pared to the REL. Both REL and RL include the business-as-usual outcomes of carbon emissions if there is no REDD or REDD+ projects to reduce deforestation and forest degradation, respectively. Unlike the REL, the RL takes a step forward to include the increased carbon stocks due to conservation and enhancement of forest carbon stocks 34 with sustainable management (LEAF, 2011), therefore, likely claiming a larger addition- ality, where “additionality” is defined as “the extent to which project interventions lead to GHG benefits that are additional to ‘business-as-usual’ ” (IPCC, 2000, 304). In this dissertation, only the REL is discussed. Structure of REL. As stated above, the dissertation aims to address the limitations of REL setting and proposing alternatives to overcome the limitations. In order to do so, it is fundamental to understand the structure of REL and how its components func- tion, first and foremost. An REL is basically threefold: first, areal change information of forestlands under the business-as-usual (BAU) scenario is required along with the associated forest carbon density information (Brown et al., 2007). When the areal and density data are multiplied, the outcome will yield mass information. Then, the mass information should be converted to tonnes of carbon dioxide equivalent (tCO 2 e). The final conversion is crucial because it is known that different GHGs have different cli- mate change impacts, so there is a need for using a unified measurement that allows comparing varying climate change impacts of different GHGs. For example, CH 4 can heat up Earth 21 times more efficiently than CO 2 in a given time period (usually 100 years). That is, if there is 1 tonne of methane in the atmosphere, its greenhouse effect is actually equivalent to that of 25 tonnes of carbon dioxide, i.e., 25 tCO 2 e. The last com- ponent could be also useful for accounting indirect GHG emissions (Newell and Vos, 2012) although that is not yet officially mandated (VCS, 2012c). Figure 2.5 shows how an REL is calculated. The components of REL are actually based on three distinct sciences. Identifying areal change of forestlands is a major research topic in Land Change Science (Aspinall, 2006; Rindfuss et al., 2004; Veldkamp and Lambin, 2001), while measuring forest car- bon stocks is intensively studied in Forest Science (Asner et al., 2010; Piao et al., 2009; Brown, 1997). Carbon emission sources and patterns are analyzed in Industrial Ecology (Newell and Vos, 2012, 2011; Graedel and Allenby, 2003). In sum, to produce a reliable REL, all of these components should be considered in tandem systematically. However, 35 Reference Emission Level (in tonnes of CO2e) BAU Forestland Loss and Degradation (in hectares) Carbon Stocks of Forestland (in tonnes per hectare) Carbon Emission Sources and Patterns (in CO2e) X X = Figure 2.5: Structure of reference emission level (REL) that may not be enough when one is to generate REL-based carbon credits and trade those in global carbon markets because there are additional methodological require- ments mandated by international policies/protocols. When RELs are not produced based on such requirements, then no carbon credits would be able to get registered for carbon markets, thus there would be no trading. The most popular REDD protocol is managed by VCS, and more than half of the entire REDD projects at the global level rely on VCS’s methodology (Diaz et al., 2011). VCS provides its methodologies in mod- ular forms; in other words, for each REDD component, such as REL or MRV , distinct methodological modules are prepared so that REDD stakeholders can employ those modules immediately. In this review, only the REL methodology (VCS, 2012c) is dis- cussed, where it is a streamlined outcome of many other carbon accounting protocols combined (VCS, 2012a,b, 2010a,b). Planned and unplanned deforestation. Not all deforestation is unpredictable as there is also much planned deforestation, particularly for commercial purposes (e.g., timber harvesting for HWPs or fiber acquisition for paper products); hence, according to VCS (2011, 2010c), deforestation is twofold: planned and unplanned. First, in order to prove an additionality of the planned loss of forestlands, it is necessary to acquire legal documents that permit harvesting of timber and fiber from those forestlands. Once it is proven there will be some future biomass loss, then a REDD project in those forestlands 36 could target reducing the amount of timber or fiber to be harvested by providing other incentives to the agent(s) with the legal right to remove biomasses from the forestlands; the amount of reduction becomes the carbon credits generated by the REDD project. Apparently, an additionality of the unplanned loss of forestlands is more challenging to prove than that of the planned one simply because it is unknown and arbitrary. This is why predictive methods to simulate business-as-usual future land-use and land-cover change are necessary, where the “business-as-usual” refers to the situation having no organized project interventions to reduce or halt deforestation. Definitions of terms. VCS’s definition of “forest” and of “deforestation” are actu- ally from the Sourcebook drafted by the Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD, 2010). The Sourcebook appears to include the most thorough discussion about definitions of the terms for REDD purposes, and it suggests guidelines in characterizing “forest,” “deforestation,” and “forest degradation.” While the defini- tion of “forest” is largely based on IPCC’s (2006) and FAO’s (2007, 2006) definitions, the Sourcebook recommends that two parameters, namely forest-cover and forest-use, be considered when determining a “forest (i.e., forestland).” According to the Source- book, the forest-cover, which basically indicates the biophysical structure of forestland, needs to meet the following criteria: “Minimum forest area: 0.05 to 1 ha” “Potential to reach a minimum height at maturity in situ of 2-5 m” “Minimum tree crown cover (or equivalent stocking level): 10 to 30 %” (GOFC- GOLD, 2010, 1-4) One value has to be chosen within the range for each criterion. Such flexibil- ity is granted because GOFC-GOLD acknowledges that terms and definitions could be employed quite differently over countries. While the requirements to characterize forest-cover seem relatively systematic and transparent because at least the Sourcebook 37 presents the list of criteria, the situation of forest-use does not seem very promising. The Sourcebook briefly notes “presence of a predominant forest-use is crucial land[-]use clas- sification” (GOFC-GOLD, 2010, 1-3), but the actual meaning of “forest-use” is unclear and there is no tangible list of various forest-uses (such as plantations or national parks). Although it advises that an urban park with trees meeting the criteria of forest-cover must not be regarded as a forestland, it is unclear why it should not be treated as such. To clarify, it seems reasonable to have an official list of forest-uses to better characterize a forestland for REDD purposes. As a forestland could not be specified very clearly, definitions of “deforestation” and “forest degradation” are unclear as well, although the Sourcebook somewhat defines those terms. Its definition of “deforestation” is in fact identical to United Nations Framework Convention on Climate Change’s (UNFCCC’s) definition: “the direct, human-induced conversion of forested land to non-forested land” even though it is unclear how forest-cover and forest-use are considered here (GOFC-GOLD, 2010, 1-4). According to this definition, any temporarily unstocked forests have potential to be regarded as “deforestation.” On the other hand, if a temporary decrease in percent of crown-cover is due to logging or some similar activity, the decrease is not considered “deforestation” according to the Sourcebook, so this naturally implies that any biomass removal at plantation sites may not be considered “deforestation.” Thus, the definition lacks clarity and consistency. The concept of “temporarily unstocked forest” (ICRAF, 2012, 3) hardly makes sense because if a forestland is “unstocked,” then it virtually indi- cates the forest biomass is relocated somewhere else, and such relocation will for sure affect the climate change mitigation result. However, the Sourcebook does not grasp such carbon dynamics. It appears reasonable for the Sourcebook to focus more on changes of crown-cover to see whether or not deforestation occurs because it is likely that forest height will be the same if the forestland is not deforested, so what matters more is definitely the per- cent of crown-cover rather than the height, although there might be some exceptions. 38 Unfortunately, the Sourcebook seems to overlook the areal component, which could be problematic given that now the newest VCS methodology requires high spatial resolu- tion data (at least 30 by 30 meters) be employed to delineate forestlands (VCS, 2012c). Because a pixel of Landsat imagery is equivalent to .09 hectare, a forestland must be constituted with 6 pixels when a country’s minimum forest area is set to .5 hectare. If there are 5 or less pixels agglomerated, those pixels should not be considered a forest- land; thus, removing those hectares would not count as “deforestation.” In sum, these are the potential uncertainties that are very likely to be embedded when one is to set an REL based on the Sourcebook and VCS methodology. The situation of characterizing “forest degradation” is even more vague. The Source- book largely relies on a report published by IPCC that exclusively talks about forest degradation issues (Penman et al., 2003), where it basically says there is no clear to way to define “forest degradation” or to measure it even though a “forest” is somehow defined (GOFC-GOLD, 2010). Carbon emissions due to forest degradation are usually more minor compared to those due to deforestation per unit area, whereas it is more costly to identify and monitor the forest degradation. However, this does not neces- sarily mean the forest degradation emits less carbon than deforestation at the global level. This is in fact why RED has evolved to REDD where the second “D” indicates forest degradation. However, the Sourcebook and VCS methodology present no guide- lines for accounting carbon loss by forest degradation when setting an REL (VCS, 2012c; GOFC-GOLD, 2010). Reference region, leakage belt, and project area. In addition to the definitions of the terms, the VCS methodology provides guidelines to prepare a suite of spatial extents for a REDD project. The spatial extents are threefold: (1) reference region, (2) leakage belt, and (3) project area (VCS, 2012c). The reference region denotes the spatial extent of any REL when estimating a rate of historical deforestation for a given time period. The rate is a crucial parameter in estimating/projecting the corresponding future business- as-usual deforestation. The leakage belt refers to areas that are likely to become more 39 vulnerable than its business-as-usual situation when a REDD project is to be launched nearby. The project area is where the REDD project actually targets to protect. For all of those three boundaries, their percent crown-cover must be 100% in the beginning. It is also required that those must be grouped into a homogeneous region in terms of forest- cover (VCS, 2012c). The methodological requirements are seemingly systematic, but it does not give specifics or objective methods to delineate those boundaries, especially the reference region. Therefore, a reference region could be determined in an arbitrary manner to maximize the associated REL to claim as many carbon credits as possible. At the same time, however, the actual climate change mitigation outcome could stay identical. Specifying temporal extents is relatively clearer than specifying the spatial ones, so if there are objective ways to control such spatial extent of REL, the associated additionality would become clearer to interpret than it is now. 2.2.2 Forestland carbon change REL methodology. According to Huettner et al. (2009), the methods for setting RELs for REDD are mainly twofold: (1) retrospective and (2) prospective. The former may be the simplest as it linearly connects the existing measurements of carbon emissions in the past, where those are plotted in a Cartesian coordinate system, so as to extrapolate the linear trend of the observed carbon emissions to dictate the future carbon emissions. This simplest method could be improved more realistically by controlling factors indi- vidually that determine the carbon emission estimates. For instance, if it is known that a region is vulnerable to frequent wildfire, which disturbs the regional forest biomass, then this could be factored in to adjust the future carbon emission estimate; hence the regional REL will be somewhat lowered compared to the REL produced by the simplest method as it now is employed. The prospective method is more sophisticated than the former, and this category includes predictive land change modeling approaches (Brown et al., 2007). Using land-cover and land-use maps acquired from satellite imagery is the 40 main strength of this method, while it is often accompanied by advanced spatial ana- lyses. According to IPCC (2006), such spatially explicit approaches are considered the most advanced way to measure GHG emissions (i.e., Tier 3), and this is why the spa- tially explicit prospective method is strongly suggested by VCS (2011). GEOMOD modeling. Led by Hall et al. (1995), GEOMOD was originally devel- oped to understand and mimic spatial and temporal patterns of land-use and land- cover change, so as to improve spatial and temporal resolutions of carbon emission estimates in tropical areas. On the one hand, the authors’ ultimate goal was to “esti- mat[e] carbon exchange as a function of space and time” (Hall et al., 1995, 755) by over- coming limitations of aggregated and aspatial carbon models, e.g., book-keeping model (Houghton et al., 1983); on the other hand, its outcomes were used as input of the Gen- eral Circulation Models (GCMs), where the GCMs are a suite of process-based dynamic models that simulate future climate change. The simulation results are frequently used by IPCC. GEOMOD is comprised of two parts, GEOMOD1 and GEOMOD2, where the separation helps researchers to understand land-use and land-cover change more sys- tematically. GEOMOD1 is crafted based on the authors’ hypotheses and theories on how people might use land, whereas GEOMOD2 is an empirical analysis that depicts the actual realizations of the people’s land-use. The modeling results of the two GEO- MODs do not always correspond to each other over different study areas. (Pontius Jr., 2001; Hall et al., 1995). GEOMOD modeling (Pontius Jr. and Chen, 2006) in fact refers to an extended ver- sion of GEOMOD2 (Eastman et al., 2005; Pontius Jr., 2001) by taking into considera- tion a quantity estimation. GEOMOD2 is implemented in different GIS computer pro- grams, namely IDRISI (Eastman, 2012) and ArcGIS (Hong et al., 2012), so the GEO- MOD module in IDRISI is actually GEOMOD2 (Eastman, 2012; Pontius Jr. and Chen, 2006; Eastman et al., 2005). GEOMOD has established its reputation due to the frequent uses by Winrock International and the United States Environmental Protection Agency (EPA) for testing their forestry-based carbon offset projects after the Kyoto Protocol 41 was initiated. The Noel Kempff Mercado Climate Action Project in Bolivia, which was the largest avoided deforestation project at the time, demonstrates the performance of GEOMOD (Brown, 2005, 2002b; Dushku and Brown, 2003). Since then, GEOMOD has been used and tested for many REDD applications (Sloan and Pelletier, 2012; Kim, 2010; Harris et al., 2008; Brown et al., 2007; Eastman et al., 2007; Sathaye and Andrasko, 2007; Paladino and Pontius Jr., 2004). Predictive accuracy. As GEOMODs have been employed for REDD projects to accu- rately dictate business-as-usual loss of forestlands, it is fundamental to review how their accuracies were evaluated. To compare GEOMOD1 and GEOMOD2 applications, Hall et al. (1995) employ Kappa indices that measure the statistical differences between dif- ferent categorical maps. However, the Kappa indices do not differentiate the disagree- ment due to pixel quantity and disagreement due to pixel (spatial) allocation; hence, it is difficult to obtain precise predictive accuracy information from the Kappa indices since the information derived from those is the mixture of two different disagreements (Pontius Jr., 2000). This indicates that the Kappa indices employed in Hall et al. (1995) and Pontius Jr. (2001) could be misleading (i.e., they are not sufficient to measure pre- dictive accuracy of land change modeling and hence, to compare different modeling approaches). Brown (2005, 2002b) and Dushku and Brown (2003) employ Klocation statistics (i.e., a variant of Kappa indices for pixel allocation) to assess predictive accu- racy of GEOMOD2 applications. Although the Klocation statistics are an improved version of the traditional Kappa indices, by removing the disagreement due to pixel quantity portion, this measurement is still insufficient to determine precisely the pre- dictive accuracy of land change modeling. Pontius Jr. and Millones (2011) argue that the Kappa indices and the variants, such as Khisto, Kquantity and Klocation, are all misleading. The authors mathematically prove the problematic attributes of the Kappa family, through mapping the attributes on what they refer to as “disagreement space” (4410). In this way, it becomes clear what type of predictive accuracy can be measured (or differentiated from different types of error) and cannot be measured by the Kappa 42 family. Overall, the Kappa and its variants do not seem to provide any useful infor- mation and simply provide a vague sense of accuracy; thus, they must not be used in assessing performance of land change models (Pontius Jr. and Millones, 2011). Figure of Merit (FoM) is a newer metric to measure predictive accuracy of different land change models, which is, on the one hand, quite similar to the Kappa family but better (Pontius Jr. et al., 2008); on the other hand, the Producer’s and User’s Accuracies (Congalton, 1991) in remote sensing could be considered variants of the FoM in land change modeling. This is also mandated by VCS (2011) when drafting a project design document in order to launch a REDD project. According to VCS (2011), a FoM of a spatially explicit REL must fall within the range of 40% to 80% depending on types of forestland loss. Otherwise, the REL is considered not accurate enough to assure there will be a valid additionality; hence, the REDD project will not be able to produce carbon credits that can be traded in voluntary carbon markets. The FoM’s range is generated based on the datasets of Brown et al. (2007) and Harris et al. (2008); while the main land change model employed is GEOMOD; the geographical areas include Belize, Bolivia, Brazil, Mexico, and Indonesia. Even though the VCS protocol was based on scientific experiments conducted on the different tropical regions, there is little to no point eval- uating a model’s predictive accuracy based on the fixed range of FoM, because the FoM itself does not provide the full picture of a land change model’s accuracy. That fig- ure should always be accompanied with the net change of land-cover and/or land-use; otherwise, the measurement is incomplete, hence cannot be used for any comparison (Sloan and Pelletier, 2012; Pontius Jr. et al., 2008). That is, if the VCS is to specify and offer the range of FoM, then the protocol must specify the range of net change of forest- cover and forest-use accordingly to accurately measure a model’s predictive accuracy. Pontius Jr. et al. (2008) employ a Cartesian coordinate system to compare different land change models’ predictive accuracy where the y-axis denotes FoMs of numerous land change models, e.g., GEOMOD or CLUE-S, and the x-axis denotes observed net changes of land-cover and land-use for each case study. Then it becomes vivid that a larger net 43 change tends to guarantee a higher FoM, so the net areal change must be controlled. The improvement has been made recently in a newer version (VCS, 2012c). There are many statistical measurements to validate a land change model’s predic- tive accuracy, other than the Kappa variants or FoMs (Pontius Jr. and Millones, 2011; Robin et al., 2011; Peterson et al., 2008; Pontius Jr. et al., 2008; Pontius Jr. and Spencer, 2005). Some are geared towards validating accuracies of (spatial) allocation (Robin et al., 2011; Peterson et al., 2008), while others are to evaluate decreasing or increasing accu- racies with respect to different spatial and temporal resolutions (Pontius Jr. et al., 2008; Pontius Jr. and Spencer, 2005). However, there are mainly two reasons why they are not just good enough to validate accuracies of spatially explicit RELs of REDD projects. First and foremost, the metrics in general do not have mechanisms to test a statistical difference of two (or more) distinct values generated by one accuracy assessing mea- surement. In other words, for instance, there is no way of comparing two FoMs to test whether or not the numerical difference of the two values is statistically signifi- cant. The situation is identical even for the state-of-art map comparison measurement as the probability distribution of the measurement has not yet been found; hence it is unknown (Pontius Jr. and Millones, 2011). The other reason why map comparing methodologies are not enough is that errors of land-use and land-cover change predic- tions are not the only error sources that REDD stakeholders need to cope with when launching a REDD project. An amount of carbon stocks (e.g., in tonnes or petagrams) is what they need and want to know, not the corresponding amount of areal change of forestlands (e.g., in hectares). The land change models are only for estimating the amount of areal change of forestlands and where the change occurs; it does not take into consideration the associated carbon stocks directly. Usually, data of carbon stocks are collected independently either in field or via remote sensing, and they are subject to their own errors (Saatchi et al., 2011; Asner et al., 2010; Asner, 2009; Brown, 2002a, 1997; Brown and Lugo, 1992; Fang et al., 2001; Heath et al., 1996). That is to say, the REDD stakeholders are dealing with at least two independent errors: (1) errors in projections 44 of business-as-usual future deforestation and (2) errors in measurements of forest car- bon stocks (Guti´ errez-V´ elez and Pontius Jr., 2012); thus, apparently it is not enough to determine the best land change model for REDD purposes by solely considering the former (Kim, 2010). To conclude, the projection of business-as-usual deforestation and the measurements of forest carbon stocks, and the associated errors for each, must be evaluated in tandem when validating an accuracy of a spatially explicit REL. 45 Chapter 3 Model Development 3.1 Model Design This chapter describes the proposal of a new land change model to address and fix some of the existing problems in land change modeling for REDD implementation. Section 3.1.1 includes a series of research questions in a structured manner. In Section 3.1.2, the streamlined definitions of forest, deforestation, and forest degradation are introduced, which then are employed by the new land change model. The overall analytical logic of the new land change modeling approach is articulated in Section 3.1.3, and its structure is compared to that of GEOMOD modeling in Section 3.1.4. Lastly, the conventions of calibration and of validation in predictive land change modeling are explained in Section 3.1.5. 3.1.1 Research questions To justify inventing a new land change model that is specifically geared towards setting RELs, it is logical to answer the following research questions, first and foremost: 1. How do disturbed areas of forest differ when those are estimated by distinct char- acterizations of “forest”? 2. How do rates of deforestation compare with different sizes of spatial extent? Given that estimation of RELs is subject to historic trends of deforestation that have been monitored over time, those questions could be answered without including any predictive attributes or modeling components. If the estimates of deforested area and deforestation rate vary substantially with respect to the distinct characterizations of 46 “forest” and varying sizes of spatial extent, such volatility must be dealt with properly in order for REDD to generate carbon credits in a solid fashion. The new land change model is entitled Geographic Emission Benchmark (GEB), and it is designed to function as a “baseline” of RELs. Unlike other land change models forecasting business-as-usual future deforestation, the GEB internally (1) characterizes “forest” in a general sense based on remotely sensed data and (2) identifies “defor- estation hotspots” using a spatial clustering technique in order to more transparently estimate regional rates of deforestation than do the existing models. The new model is accompanied by a series of research questions as follows: 1. Is it possible to characterize “deforestation” and/or “forest degradation” based on remotely sensed global data, namely Globcover and Vegetation Continuous Field? 2. Is it possible to objectively distinguish hotspots of deforestation and/or forest degradation using Local Indicator of Spatial Association? 3. How accurate is the GEB compared to a typical GEOMOD modeling in terms of predictive accuracy measured by Figure of Merit? Furthermore, by comparing the performance of GEB and GEOMOD, it is possible to answer more fundamental research questions: Does a more accurate business-as- usual forest-cover loss projection (e.g., in hectares) result in a more accurate measure of business-as-usual forest carbon stock loss (e.g., in tonnes)? In other words, is there a clear relationship between the forest-cover change and the associated forest carbon stock change? Answering the questions will help REDD stakeholders to better under- stand how a predictive land change model works in an REL setting. Based on the more accurate business-as-usual forest-cover loss projection, the following question can be finally answered: What would be a reference emission estimate for Banna’s future REL(s) between 2010 and 2040? 47 3.1.2 Streamlined definition In order to couple the definitional and modeling issues, it is necessary to streamline the three parameters—forest-cover, forest-use, and ecological aspects of forest ecosystem— to specify a “forest” because one cannot collect complete data for each parameter so as to perfectly define one, and without any solid definition specified, no modeling out- come can be meaningful. This reduced version of characterization is used throughout this dissertation. In this way, interpreting results could become more systematic and transparent than what occurs in other approaches. The interpretation also can explic- itly show any uncertainty embedded in the forest carbon loss estimation produced. It is useful to start with the definition that FAO suggested in this regard because that is reduced already in order to make it applicable for different countries, so it can be globally operative in a simultaneous manner. While “other wooded lands” refer to the trees that do not meet FAO’s criteria of forest (FAO, 2007; IPCC, 2006), FAO (2007) defines a “forest (i.e., forestland)” as follows: Land spanning more than 0.5 hectares with trees higher than 5 metres and a canopy cover of more than 10 percent, or trees able to reach these thresholds in situ. It does not include land that is predominantly under agriculture or urban use. In China, the central government follows the definition determined by FAO (2007) although the definition’s limitation has been acknowledged by conservation commu- nities since it does not differentiate rubber plantations as opposed to tropical rain- forests. Such indifference is particularly problematic in the study area because the rub- ber plantations consume much fresh water and facilitate soil erosion of the region (Hu et al., 2008), and these plantations have been expanded rapidly due to their profitable nature. Apparently, the Chinese definition of “forest” does not seem to benefit the study area’s conservation of tropical rainforests. ICRAF (2012) points out “vegetation-based definition” (3) of land-cover (i.e., forest-cover) is likely to treat unfairly “temporarily 48 unstocked forestlands” (3) (i.e., there may be no biomass, but the land could be still considered “forested”). That is exactly what is happening in the study area because rubber trees are harvested on a regular basis (e.g., every 25 or 35 years). Forests with- out forest carbon stock must be treated more reasonably. The definitions in Table 3.1 solely include the forest-cover dimension, and there are two practical reasons for this: first, it was mentioned that forest-cover and forest-use are often used interchangeably, while the consequences of such interchangeable use are not thoroughly discussed in the literature. Therefore, it seems novel to demonstrate the consequences when the terms are employed incorrectly. Perhaps one way is to characterize a “forest” in a rather incomplete way and measure the forest carbon loss accordingly. The assumption here is that forest-cover is equivalent to forest-use, and that is actually the main reason why the real forest-use dimension is excluded from Table 3.1. Second, it is not very clear how “agricultural or urban use” is defined by FAO (2007), so it does not seem reasonable to employ such vague concepts when one is to transparently characterize a term systematically. This is also the reason why it does not seem clear at the moment of writing whether the rubber plantations are regarded as agricultural use, such as croplands, or mere forests. Table 3.1: Different characterizations of “forest” by Food and Agriculture Organization of the United Nations (FAO) and the Geographic Emission Benchmark (GEB) FAO GEB Area >.5 ha >9 ha Canopy >10 % >15 % Height >5 m >5 m Given the characterization of forest is fixed, solely dependent on the forest-cover dimension, two types of “deforestation” are characterized and compared (Table 3.2). On the one hand, the first deforestation is regarded as a land-cover conversion from forest to non-forest categories, which is how research on the topic typically regards 49 deforestation. At the same time, this deforestation refers to a high level of forest degra- dation that exceeds a certain threshold, where the threshold is empirically determined in this chapter based on remotely sensed data and quantitative methods. In this con- text, “forest degradation” indicates other types of biomass loss that are more minor than deforestation; thus, the forest will maintain the same land-cover. Since the for- est degradation only occurs in forests, degradation of other wooded lands is excluded from this research. On the other hand, the second type of deforestation is determined based on land-use change—from tropical rainforests to rubber plantations—which is the most dominant conversion in the study area. The associated forest degradation portion is excluded here because it is difficult to define one. In short, matured rubber trees are treated differently by the two distinct characterizations of “deforestation” because with the former characterization of deforestation, the matured rubber trees are forests whereas with the latter characterization they are not considered forests. Table 3.2: Different characterizations of “deforestation” and “forest degradation” Forest-cover Forest-use Deforestation Forest to non-forest Tropical rainforests to rubber plantations Forest degradation Minor biomass loss Not applicable other than deforestation Only six of the Globcover forest-subclasses—i.e., Categories 40–100 (Bicheron et al., 2008; Bontemps et al., 2011)—are used in this dissertation as these are the only cate- gories that include information of tree height (i.e., trees taller than 5 meters in height). The greater detail about the Globcover is given in Section 3.2.1. In terms of area and canopy-cover, the definition used in this dissertation is in fact more conservative than FAO’s definition because Globcover’s spatial resolution is 300 by 300 meters; thus, a pixel (the smallest areal unit of Globcover data) indicates 9 hectares of land, while the minimum percent of canopy-cover of the six forest-subclasses is 15%. The definitions are summarized and compared in Table 3.1. Such specification seems more reasonable 50 than that of Harris et al. (2012), for instance, which was somewhat similar to FAO’s definition but not necessarily either more generous or conservative, and therefore not particularly useful. The definition used in this dissertation is clearly more conservative than FAO’s; therefore, one should be able to tell that the actual forest carbon stock esti- mated made in this dissertation would be somewhat smaller than the estimation made based on FAO’s definition of forest. 3.1.3 Rationale As explained in Section 2.2, an REL is constructed by multiple scientific modules, where each module is subject to its own error. Such error is never avoidable because every single module includes the knowledge that is produced by combining scientific data with distinct statistical methods. While the data sources have their own measurement errors as always, the different methods also tend to induce methodological errors. In other words, in order to produce an REL as accurately as possible, it is fundamental to understand different types of errors and find a way to minimize those or at least prevent the propagation of such errors. Then employing the simplest land change model may be optimal for that purpose because if a model is simple, then it also means the model may include minimal methodological errors, and then it becomes easier to comprehend the nature of errors embedded in scientific data used for constructing an REL. Thus, the methodological transparency is the most important attribute that the GEB pursues. Unlike the existing land change models, the GEB factors in two additional modules: (1) the characterization of basic terms and (2) the systematic control over varying spatial extents, as part of its land change modeling. While these two modules are, by design, data-driven, they rely on open-sourced quantitative spatial methods and the spatial data used are accessible free of charge, which will be explained further in Section 3.2. In sum, the GEB may contain the minimal mathematical complexity; however, it extends the methodological spectrum of land change models to encompass the two additional 51 modules, which have not been considered so far in the literature, to ultimately support constructing an accurate REL. 3.1.4 Geographic Emission Benchmark and GEOMOD GEOMOD is arguably the most frequently used land change model in REDD (Benito and Pe˜ nas, 2011; Brown, 2005, 2002b; Dushku and Brown, 2003; Harris et al., 2008; Kim, 2010; Sathaye and Andrasko, 2007, 2006; Sloan and Pelletier, 2012), and it may be per- haps the simplest among existing models in terms of mathematical complexity (Kim, 2010). GEOMOD’s logic is already embedded in popular GIS computer programs such as IDRISI (Eastman, 2012) and ArcGIS (Hong et al., 2012). The detailed explanation of the approach is not given in this dissertation since GEOMOD modeling is explained comprehensively in the primer published by Clark Labs at Clark University, MA, in the United States (Pontius Jr. and Chen, 2006). The major structural similarity and difference between the new approach and GEOMOD are laid out in Figure 3.1. They both allow researchers to portray the regional forest carbon dynamics in a spatially and temporally explicit manner with the similar structure, and their maintenance of such spatial and temporal details is also why the kind of approaches are considered one of the most accurate ways to account carbon emissions (i.e., Tier 3), according to IPCC (2006). The GEB should be compared to the GEOMOD both in terms of structure and modeling outcome because, in this way, one may be able to tell which specific part of the new approach actually contributes to the associated modeling outcome so that it can be verified whether or not it actually improves the result. Unlike GEOMOD, the GEB is modeled via open-source computer programs, namely R (www.r-project.org), OpenGeoDa (geodacenter.asu.edu/ogeoda), and Quantum GIS (www.qgis.org). 52 Business-as-usual PROJECTION Remotely sensed OBSERVATION (SPATIAL) ALLOCATION QUANTITY + = Geographic Emission Benchmark GEOMOD Modeling Input data Methods Input data Methods Globcover VCF Night-light •ROC •LISA •Linear extrapolation •Ranked allocation •Random allocation •Linear extrapolation Forest/ Non-forest •GEOMOD module in IDRISI Spatial variables Calibration Validation Methods •Figure of merit Figure 3.1: Structural difference of the Geographic Emission Benchmark (GEB) and GEOMOD modeling 53 3.1.5 Calibration and validation The validation process is essential in any predictive modeling and is mandated by the Verified Carbon Standard (VCS) if one is to list REDD carbon credits in global carbon markets (VCS, 2011). In the literature of land change modeling, the process is frequently referred to as the “validation” process (Pontius Jr. and Spencer, 2005), but in protocols of VCS, the term “confirmation” is used instead (VCS, 2012c). The key element of the validation process is separating data for calibration and validation. The lower part of Figure 3.1 shows how the validation procedure is employed in this dissertation. The data used in this dissertation are therefore twofold: the first set is for calibration, while the second set is for validation. The calibration set includes forest-cover data and other socio-economic spatial variables that function as input data for both the GEB and GEOMOD, whereas the validation set is geared towards verifying whether or not the modeling outcomes are accurate enough to use for a real REDD project. The biomass data are also included in the validation set because such information is indispensable when validating any business-as-usual projections of forestland change for REDD pur- poses (Kim, 2010). More details are provided in later sections along with the methods used for each dataset. 3.2 Methods and Data The section provides descriptions about the methods and data used in this dissertation and also rationalizes why those are determined to be used. The structure of the section may be similar to a primer that generates a GEB. The organization is a mixture of the methods and data, where these are laid out in a manner that shows the order of the dissertation’s entire analytical procedure. 54 3.2.1 Forest-cover data Forest-cover maps, namely Globcover and VCF, are obtained from the two freely accessible data servers: (1) Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD, http://www.gofc-gold.uni-jena.de/sites/globcover.php)—the Euro- pean database supported by European Space Agency (ESA), FAO, UNEP , and Inter- national Geosphere–Biosphere Programme (IGBP)—and (2) Global Land Cover Facil- ity (GLCF, http://glcf.umiacs.umd.edu)—the United States’ database supported by National Aeronautics and Space Administration (NASA). Both of these servers pro- vide the historical forest-cover maps in raster format generated from different sources of satellite imagery (Figures 3.2a–b). Their major difference is that the Globcover shows its forest-cover information in a categorical form (e.g., the value 50 indicates “closed broadleaved deciduous forest”), while the VCF shows the information in a continu- ous form (i.e., percent of canopy-cover). GLCF provides VCF from 2000 to 2005 on an annual basis; GOCF-GOLD offers Globcover only for two time periods (i.e., 2005–2006 and 2009). Their spatial resolutions and spatial/temporal coverages are summarized and compared in Table 3.3. More details about the maps, such as data acquisition, processing, and their accuracy reports, can be found from the following literature: Bon- temps et al. (2011); Bicheron et al. (2008); Hansen et al. (2003, 2002). The reason why those data are used in tandem is explained in the following paragraphs. Table 3.3: Remotely sensed forest-cover data Data type Pixel (in meters) Temporal/spatial coverage Globcover 300 by 300 2005–2006 and 2009, Global VCF 500 by 500 2000–2005 (annually), Global 55 0 50 100 25 Miles 0 50 100 25 Miles 81 0 Vegetation Continuouse Field Forest Non-forest Globcover (Unit: percent of canopy-cover) (a) (b) Figure 3.2: (a) Globcover-based binary map of forest and non-forest and (b) Vegetation Continuous Field (VCF) 56 3.2.2 Receiver operating characteristic The present section is exclusively designed to specify “deforestation” and “forest degra- dation” based the forest-cover data, so it is worth noting this has little to do with forest- use. Using the Globcover to portray a trend of deforestation and/or forest degradation might be quite useful for REDD purposes because the data explicitly characterize a “forest” (Table 3.1); hence, REDD stakeholders may find the Globcover-based results easy to interpret. The downside, however, is that the data cannot smoothly visualize forest-cover heterogeneity, which is fundamental to deal with any forest degradation issues where the “forest degradation” refers to more minor biomass loss than defor- estation per area unit (Table 3.2). The situation of VCF is the opposite. The data do not contain any specifics of forest characterization because the data merely measure the physical amount of sunlight, penetrating layers of foliage, from the ground (Hansen et al., 2003). Given the percent of canopy-cover is an important parameter that deter- mines a “forest,” according to FAO (2007) and IPCC (2006), VCF’s pixel value appears to be a reasonable proxy, albeit perhaps not a complete one, that can smoothly portray the varying heterogeneity of forest-cover. However, that itself is not sufficient because its forest is not clearly characterized, so there is a risk to include other wooded lands or exclude forests when one solely uses VCF to define forest but without the Globcover data. Thus, the Globcover and VCF should be used in tandem to supplement their lim- itations; then finally, it becomes a matter of how to combine these two distinct datasets to characterize a “forest.” The simplest way to use the two datasets may be to merely overlay those so as to identify how much percent of canopy-cover is actually equivalent to the threshold of being “forest.‘’ Once the threshold in terms canopy-cover is identified, vegetation or biomass below the threshold should be regarded as other wooded lands; as such, these cannot be deforested as they are not forests in the first place. In order to apply this simplest algorithm, one has to make sure the Globcover and VCF are detecting the same 57 objects, i.e., forests. Then it becomes necessary to compare the two datasets to measure their similarity. In other words, are they really detecting the same type of “forest”? Receiver Operating Characteristic (ROC) is suitable in this regard because it can assess the agreement of the two forest-cover maps, where one map must be binary and the other one has to be continuous. The binary map refers to the forest and non- forest map that is reclassified from the Globcover of 2005 (Figure 3.2a). More specif- ically, the “forest” category is constituted by closed/open, broad-/needle-leaved, and evergreen/deciduous forestlands that are taller than 5 meters in height, larger than 9 hectares in area, and larger than 15 percent in crown-cover (i.e., Globcover classes 40– 100). On the other hand, the “non-forest” category includes the rest (Bontemps et al., 2011). In sum, the definition of “forest” becomes identical to the more conservative definition in Table 3.1. Lastly, the continuous map indicates the VCF of 2005 (Figure 3.2b). It might be helpful to imagine a 3D model of the VCF layer (Figure 3.2b)—like a 3D digital elevation model—to understand how the ROC works. Then start slicing the 3D model from the top to the bottom so each slice generates a threshold in which each threshold actually determines the corresponding binary map. In other words, for each threshold, one can assign “1” when the pixel values are larger than the threshold (indicating “forest”); otherwise “0” means that the pixel values are smaller than the threshold (indicating “non-forest”). Basically, one binary map is produced for each threshold; thus in total, multiple binary maps are generated. The binary maps are now compared to the forest and non-forest binary map (Figure 3.2a) individually through a simple change detection analysis. So there will be numerous change detection analyses because there are numerous binary maps produced from the VCF layer that are ready to be compared to the forest and non-forest binary map. Each analysis will result in some agreements and disagreements at the pixel level. More specifically, the pixel agreements can be either “forest to forest” or “non-forest to non-forest”; likewise, the disagreements 58 can be either “forest to non-forest” or “non-forest to forest.” The coordinates of an ROC curve are actually derived from the proportion of such agreements and disagreements. Table 3.4: Receiver Operating Characteristic’s (ROC’s) contingency table Forest-cover map Forest (“1”) Non-Forest (“0”) Total VCF Forest (within threshold) A B A+B Non-Forest (otherwise) C D C+D Total A+C B+D A+B+C+D For each threshold, one data point(x;y) is generated wherex is called “specificity,” that is, “the proportion of correctly classified negative observations” (Robin et al., 2011, 1), and y is referred to as “sensitivity,” that is, “the proportion of correctly classified positive observations” (Robin et al., 2011, 1). These data points are plotted and con- nected to create an ROC curve. The sensitivity is derived from A=(A+C) while the specificity is derived fromD=(B+D), whereA;B;C;D are pixel counts in Table 3.4 for each threshold (Robin et al., 2011). When the ROC refers to “relative” operating charac- teristic, the specificity is replaced byB=(B+D), that is, percent of false positive, hence resulting in the opposite direction of thex-axis (Pontius Jr. and Schneider, 2001). The Area Under the Curve (AUC), which summarizes the information of an ROC curve, is calculated according to the following equation: AUC = n X i=1 (x i x i+1 ) ( y i + (y i+1 y i ) 2 ) ; (3.1) where x i is the specificity for the threshold i, y i is the sensitivity for threshold i, and n+1 is the number of thresholds. A larger AUC indicates the better agreement between the binary and continuous maps than the agreement that has a lower AUC; the value ranges from 50% (i.e., no agreement) to 100% (i.e., perfect agreement). The ROC analysis is performed using pROC package (Robin et al., 2011). 59 Based on the resulting AUC, the Globcover and VCF may or may not be used together in this dissertation. In other words, if the agreement is too low, for example, around 50%, there is no way the forest-cover maps can be used together because there is no similarity as they measure different forest-covers, and also users should suspect the accuracy of the two datasets. In contrast, if the AUC turns out relatively high, e.g., 80%, it seems reasonable to conclude that the top VCF values are fairly identical to the forest category of the binary forest-cover map, where the AUC shows the certainty of the relationship. Then the trees that have lower VCF values than the threshold can be considered either previously deforested or other wooded lands. One technical issue must be addressed before running the ROC analysis. Since the spatial resolutions of Globcover and VCF are different (Table 3.3), there will be different numbers of total pixels even for the identical study area. However, it is required to have the identical number of pixels to run ROC; therefore, the Globcover layer is resampled to match the spatial resolution of VCF (i.e., 500 by 500 meters), using the nearest neigh- bor algorithm. By fixing the spatial resolution to 500 by 500 meters throughout the dissertation, it is expected to minimize error propagation in estimating forest carbon stock at the landscape level. 3.2.3 Change detection analysis Forest-cover change detection is conducted on the VCF layers at the pixel level to quan- tify the amount of change between 2000 and 2005. VCF’s pixels are constructed in terms of continuous values; hence, the change detection analysis will produce contin- uous values too, so there is a need to determine an objective threshold to show which pixels are forests and which are not, and this is fundamentally why the ROC analysis was necessary in Section 3.2.2. When a pixel contains a value that is equal or larger than the threshold after the change detection analysis, this indicates that the pixel (or the patch of forestland) should not be considered “deforested” and must be considered “degraded” although the forestland might have lost some biomass. 60 3.2.4 Local indicator of spatial association After deforestation and forest degradation are identified at the pixel level in Section 3.2.3, the pixels are aggregated for each county, so that the polygon shapefile of Yun- nan province can contain the amount of deforested areas for each county. The polygon shapefile of counties of Yunnan province (1990) are downloaded from China Historical GIS at Harvard University (CHGIS), where the shapefile is part of China in Time and Space (CITAS) data collection. Anselin’s (1995) Local Indicator of Spatial Association (LISA) is then applied to delineate hotspots of deforestation (and of forest degradation). According to LISA’s logic, there can be four types of hotspots, and such detailed clas- sification provides more information than other clustering techniques. When a county that experiences rapid deforestation is surrounded by other bordering counties that also show high rates of deforestation, then the county becomes a member of High– High (HH) hotspots. The other types of hotspot—namely High–Low (HL), Low–High (LH), and Low–Low (LL)—are specified based on the same logic. For example, when a county turns out to be a member of LH hotspots, REDD stakeholders may want to pay more attention to the county because of its potential leakage issues in the future. If any result were statistically significant at the 95% confidence level, it is considered a deforestation (and forest degradation) hotspot. Such identification of the deforesta- tion hotspots is what is proposed in this dissertation to determine a reference region in setting RELs (VCS, 2011; World Bank BioCarbon Fund, 2008). Statistically, LISA can be explained as a spatial transformation of Pearson’s linear correlation coefficient (i.e., r), and there are three major differences between the LISA and r. The first difference is such that LISA only deals with one variable (i.e., univari- ate), unlike the r coefficient. While an r measures the linear relationship between two (continuous) variables, a LISA measures the linear relationship of one variable in a par- ticular location with the same variable that is situated in different locations. In order to do that, it is mandatory to have a conversion mechanism that can transform the spa- tial information to a pseudo-second variable. The mechanism is the second difference 61 between the LISA and r, and this can be easily done using a spatial weight matrix. The third difference is about the statistical distribution of LISA and r statistics. While r relies on normal distributions when testing to determine if an r coefficient is statistically sig- nificant or not, LISA relies on distributions that are produced non-parametrically (i.e., it does not assume normal distribution). While the LISA dictates a spatial (auto)correlation of the variable that a researcher is interested in with its neighbors, the statistic is calculated based on the following equa- tion: I Y = P i P j6=i w ij (y i y)(y j y) P i P j6=i w ij P i (y i y) 2 (3.2) whereI Y represents the LISA of the variableY that the researcher is interested in (e.g., rate of deforestation); w ij denotes an element of a spatial weight matrixW , while the element shows a type of spatial association between locationsi andj. y i indicates the variable that the researcher is interested in at locationi, andy shows the average value of all y i s for the study area. In this dissertation, the spatial weight matrix is created using Queen’s method and only considers 1st order connectivity, i.e., when county i borders countyj then “1” is assigned tow ij , if not “0.” The LISA analysis is performed using OpenGeoDa, which is downloaded from the GeoDa Center for Geospatial Anal- ysis and Computation at Arizona State University. 3.2.5 Business-as-usual forest-cover change From Section 3.2.2 through Section 3.2.4, “forest,” “deforestation,” and “forest degrada- tion” are characterized solely based on forest-cover information, and then from those terms, historical forest-cover loss (in terms of both area and biomass) is calculated in this section while the changes’ spatial extents are objectively delineated using LISA. On the one hand, these can be used as inputs to project a business-as-usual deforestation (and/or forest degradation) via GEOMOD. On the other hand, the process of specifying those inputs is already embedded in the GEB, and this is what makes this new approach 62 unique. In the current section, projecting business-as-usual deforestation of the GEB is explained with respect to that of the GEOMOD. Only business-as-usual deforestation is projected to dictate the forest loss and its forest biomass loss between 2005 and 2010 via both GEB and GEOMOD. 3.2.5.1 Linear extrapolation Once a hotspot (i.e., a group of counties) is identified, the rates of the hotspot’s defor- estation and forest degradation are calculated by another change detection analysis, which is different from those of Sections 3.2.2 and 3.2.3. The change detection analysis conducted at the hotspot level, where the hotspot(s) may be larger or smaller than the Banna prefecture level, dictates the amount of forest-cover loss between 2000 and 2005 as part of the GEB approach. Another set of rates is calculated similarly at the Banna prefecture level as part of GEOMOD modeling. The rates of deforestation and forest degradation at the Yunnan province level are also calculated, so as to see how all rates fluctuate when different sizes of spatial extent are considered. The rates are calculated based on the following equation: R i(2000;2005) = FD i(2000;2005) F i(2000) (3.3) whereR i(2000;2005) indicates the rate of forest disturbed between 2000 and 2005 in region i,FD i(2000;2005) refers to the amount of forest disturbed between 2000 and 2005 in region i (in hectares), andF i(2000) dictates the amount of existing forest in 2000 in regioni (in hectares). The forest disturbed simultaneously indicates both deforestation and forest degradation. In addition, the rates are assumed to be consistent over time (Figure 3.3); therefore, the business-as-usual forest-cover loss between 2005 and 2010 maintains the same rate as it did between 2000 and 2005. This assumption supports the logic of linear extrapolation method of GEOMOD modeling (Pontius Jr. and Chen, 2006). However, the difference is that GEOMOD allows users to input data arbitrarily, so the method 63 Forest Time Observed Estimated Figure 3.3: Logic of linear extrapolation does not control either characterizations of “forest” or scales of study areas, while the GEB does not allow such arbitrariness. Finally, three distinct rates are estimated at different spatial levels, and only business-as-usual projections of deforestation between 2005 and 2010 are generated because this is the only forest-cover transition that has the data for validation; hence, for the same reason, business-as-usual projections of forest degradation are excluded from this research. 3.2.5.2 Spatial allocation The quantities of business-as-usual forest-cover loss due to deforestation have been finally estimated based on the aforementioned methods. Now, it becomes a matter of how to allocate the estimated quantities to specific locations. The GEOMOD module 64 in IDRISI combines many spatial variables (Figures 3.4a–h) to produce the rank map, which is often referred to as a transition potential map (Eastman et al., 2005) show- ing different potentials of forest-cover change (i.e., deforestation). Unlike the GEO- MOD, the GEB simply employs the night-light imagery (Figure 3.5) to replace the GEO- MOD module in IDRISI, so this night-light approach does not require any other spatial variables that represent proximate causes of deforestation. That being said, the GEB assumes that the night-light imagery is a suitable proxy that explicitly measures anthro- pogenic disturbance (i.e., distribution of human species), and the data implicitly factor in the spatial variables used for the GEOMOD run, such as elevation, slope, road, and population. Although the night-light data may include natural wildfire, which is not anthropogenic, it appears to be reasonable to include the fire information as well if one lasts long enough to be stored in the data because the wildfire also emits carbon to the atmosphere by consuming forests. Two technical issues must be addressed when using the night-light imagery for the GEB. First, the spatial resolution of the imagery is resampled from 1,000 by 1,000 meters to 500 by 500 meters so it becomes identical to the pixel resolution of VCF. In this case, it is reasonable to choose the bilinear resampling method. This works as an interpo- lation function to fill the data gap since the resampling from a coarser to finer spatial resolution requires new data points, and it seems reasonable to consider the values of neighboring pixels when estimating and assigning new values to the newly gener- ated pixels (Mitasova et al., 1996). This is also true for other spatial variables that have different pixel resolutions used for the GEOMOD run. Lastly, in some cases, the night- light data might not have enough variation in terms of pixel values, so there may be many pixels that are ranked as a tie. Then it is reasonable to allocate the leftover pixels randomly after all the ranks produced by the night-light imagery are consumed; this random (spatial) allocation is done by sp package (Pebesma and Bivand, 2012). 65 3.2.5.3 Deforestation driver data Both the GEB and GEOMOD are “descriptive” (Castella et al., 2007, 534) and “geo- graphic” (Heistermann et al., 2006, 141) land change models that have little human decision-making (Agarwal et al., 2002). In other words, they are not geared towards explaining deforestation (hence, descriptive); instead they are more geared towards identifying spatial configuration of deforestation than other economic models (hence, geographic) and are geared towards projecting business-as-usual deforestation rather than testing various scenarios that show human decision-making. By design, the mod- els’ input data are limited to proximate causes of tropical deforestation such as road and stream networks, elevation, slope, aspect, previously disturbed areas, and popula- tion. It is important to note that data acquired after 2000 are excluded, such as Land- Scan (ORNL, 2008) or Global Digital Elevation Map ([GDEM], METI and NASA, 2011) because this dissertation is about forecasting a business-as-usual trend of future defor- estation by using data in the past; hence any future data should not be used for model calibration. Only those acquired around early 2000s are included. “Proximate cause” is different from “underlying driving force,” according to Geist and Lambin (2002). The terms both indicate different factors/drivers of tropical defor- estation (and forest degradation) but with different perspectives. The underlying driv- ing force refers to a more fundamental reason of tropical deforestation such as economic factors, whereas the proximate cause is rather a more superficial yet immediate reason of tropical deforestation, such as proximity to roads or deforested areas. These are actu- ally easier to map than the underlying driving force, and it seems reasonable to regard the proximate causes as physical realizations of the associated underlying driving force. Therefore, when the rubber plantations expand in Banna, one can say that it is due to having a denser road network than before, and the underlying driving force of such expansion is the domestic rubber market growth. In this dissertation, only proximate causes are considered, and it is not part of research objectives to identify the associated underlying driving force. 66 (a) Distance from deforestation (meters) 28,214 0 46,878 0 448,448 0 (b) Distance from roads (meters) (c) Distance from railroads (meters) 28,450 0 15,919 0 (d) Distance from streams (meters) (e) Population (number of people) (f) Elevation (meters) (g) Slope (degrees) (h) Aspect (degrees) 6,438 0 62 0 360 0 Figure 3.4: Spatial variables for the GEOMOD run One of the required input data of the GEB is night-light data (DMSP, 2005). The data represent the visible light spectra during night time, and the pixel values are normal- ized by percent length of observation (Figure 3.5). For example, if there is some light that is observed only half of the night, where this observation repeats on a daily basis (for one year), the pixel value would show 50%. Moreover, it is important to note that the data record natural and anthropogenic fires, gas flaring, and city lights (Elvidge et al., 1997). The night-light data are important inputs for gridded population maps, including LandScan (ORNL, 2008; Dobson et al., 2000), Global Rural–Urban Mapping 67 0 50 100 25 Miles 63 0 Nigh-light (Unit: digital numbers) Figure 3.5: Night-light imagery for the Geographic Emission Benchmark (GEB) Project ([GRUMP], CIESIN et al., 2004), and History Database of the Global Environ- ment ([HYDE], Goldewijk et al., 2011; Goldewijk et al., 2010) because it is considered a good proxy of human habitat at the global level. For the GEOMOD run, many spatial variables are used in tandem to produce a tran- sition potential map showing which pixels are more likely to be deforested compared to others. Road, railroad, and stream layers are downloaded from the CHGIS/CITAS of 68 Harvard University, MA, and population maps are downloaded from Center for Inter- national Earth Science Information Network (CIESIN), Centro Internacional de Agricul- tura Tropical (CIAT), and Columbia University, NY; digital elevation models are from NASA’s Shuttle Radar Topography Mission (SRTM). The list of raw vector layers are summarized in Table 3.5. Distance maps are generated based on the road, railroad, and stream layers while slopes and aspects are produced based on the SRTM data. The spa- tial variables are all used as input data of GEOMOD module to calculate transitional potentials of deforestation (Figures 3.4a–h). Table 3.5: Vector data for (spatial) allocation Data type Vector type Temporal/spatial coverage Population Polygon 1990–2000 (every 5 years), Global Road Line 1993, China Railroad Line 1993, China Stream Line 1993, China Other than the forest biomass benchmark map of the pan-tropical regions (Saatchi et al., 2011), the geographical coverage of remotely sensed data sources used in this chapter is global. The level of spatial resolution is often considered “moderate (or medium)” (Achard et al., 2010; DeFries et al., 2007), which means the pixel size of the global products ranges from 90 meters (e.g., elevation) to 1,000 meters (e.g., night-light) as indicated in Table 3.6. For the models’ runs, all of the data are resampled to 500 by 500 meters spatial resolution. Table 3.6: Remotely sensed data for (spatial) allocation Data type Pixel (in meters) Temporal/spatial coverage Elevation 90 by 90 2000, Global Night-light 1,000 by 1,000 1992–2010 (annually), Global Biomass 1,000 by 1,000 early 2000s, Pan-tropical 69 3.2.6 Biomass estimation Once the areal changes of deforestation and forest degradation are measured between the two time periods, biomass density data of the same region are multiplied to those estimates; hence, the forest biomass loss is accounted based on the following equation: Q fb =A df D fb ; (3.4) whereQ fb indicates the quantity forest biomass loss (in megagrams),A df denotes the total areal change due to deforestation and/or forest degradation in the study area (in hectares), and D fb shows the biomass density information of the changed areas (in megagrams per hectare). The estimation of biomass loss is twofold: the first estimation is the observed loss between 2000 and 2005 due to deforestation. The second estimation is the loss due to the observed forest degradation, where this is estimated with three scenarios, namely 10%, 20%, and 30% AGB loss by employing only one category of forest degradation. Even though there must be different levels of forest degradation, the dissertation does not take into account such variation; hence, it employs only one forest degradation category. The AGB loss is, first, calibrated at the pixel level and later is aggregated at the Banna prefecture level. The forest-cover gain portion is not discussed here. Unlike forest degradation, deforestation means 100% of AGB loss. It is crucial to note that the unit of observation of the biomass benchmark map is in megagrams per hectare, while a pixel is approximately 1,000 by 1,000 meters (Saatchi et al., 2011). That being said, a pixel of the biomass map is in fact equivalent to 100 hectares, so a pixel value must be multiplied by 100 to yield the actual biomass quantity for that specific pixel. Likewise, as the pixel resolution employed in this research is 500 by 500 meters (i.e., 1 pixel = 25 hectares) technically, the aggregated outcomes must be multiplied by25 in order to estimate the actual quantity of AGB loss. However, actually20:75 is used instead of 25, and the process is explained in Appendix B. 70 3.2.7 Biomass data In Section 2.1.2, it was explained how the benchmark biomass data could be potentially erroneous. However, given that this dissertation is geared towards inventing a trans- parent and unified methodology that may work at the pan-tropical level, it appears to be more important to use the data that cover the region indiscriminately although the data might be less accurate than the data of smaller areas. Thus, one of the benchmark data is used here, and this is the only pan-tropical dataset that was freely downloadable at the moment of writing. The spatially explicit forest biomass data of the pan-tropical regions in early 2000s (Saatchi et al., 2011) are used to estimate the forest biomass loss due to different specifications of deforestation and forest degradation. The data are twofold in terms of carbon pools: AGB and BGB. The AGB is directly measured by remote sensing and field measurement, and it is processed in tandem, whereas the BGB is statistically estimated based on the measured AGB by a universal allometric equation (Saatchi et al., 2011). Only the AGB portion is considered in this research; the pixel unit is in megagrams (i.e., metric tonnes) per hectare (Figure 3.6). 3.2.8 Validation The validation process is twofold in this dissertation: first, the business-as-usual projec- tions made by the GEB and GEOMOD are compared to a reference map that is regarded as the true observation. Since the projections are about losing forest-cover over time, the corresponding reference map also must include the forest-cover information of the year of validation. Using map comparison statistics, agreements and disagreements of the business-as-usual projections and the reference map can be assessed and summa- rized in numbers. This validation procedure is to validate the forest-cover area change in terms of area and also to validate the associated forest biomass change in terms of mass. Second, another type of validation process is conducted where it has little to do with the map comparison statistics but much to do with the actual ground-truthing 71 0 50 100 25 Miles 390 0 Above-ground living biomass (Unit: megagrams per hectare) Figure 3.6: Spatially explicit above-ground living biomass (AGB) estimates in the field. It was originally planned to collect primary ground-truth data from the study area to confirm the business-as-usual projections (i.e., how accurate are the pro- jections?). Instead, the forest-use map produced by Kunming Institute of Botany (KIB) of the Chinese Academy of Sciences (CAS) is used. The forest-use map turns out to be as accurate as the primary ground-truth data collected, and, better yet, the map covers 72 the entire study area while the ground-truth data cannot. Thus, the forest-use map is compared with the result of the change detection analysis in Section 3.2.3. 3.2.8.1 Figure of merit The two business-as-usual projections are compared to each other and validated with the observed forest-cover map, while a Figure of Merit ([FoM], Pontius Jr. et al., 2008) is used here as this statistic is required by VCS (2011). The FoM assesses the pre- dictive accuracy of a projected business-as-usual deforestation by simply overlaying the observed forest-cover map of 2005, the predicted forest-cover map of 2010, and the observed forest-cover map of 2009. Here, it is assumed that forest-cover does not change substantially within a year, thus the difference in terms of forest-cover between 2009 and 2010 is negligible. The FoM has both visual and numeric expressions. In this dissertation, only the numeric expression is employed, and it is expressed mathemati- cally as follows: Figureof Merit=B=(A+B+C) (3.5) whereA is a number of pixels for “error due to observed change predicted as persis- tence” (or misses),B is a number of pixels for “correct due to observed change predicted as change” (or hits), andC is a number of pixels for “error due to observed persistence predicted as change” (or false alarms). The FoM ranges from 0 to 100 percent, where 100 percent indicates perfect prediction (Pontius Jr. et al., 2008). The validation procedure is accompanied by the AGB estimates produced by Saatchi et al. (2011) as Kim (2010) argues that predictive accuracies of land change models for REDD purposes cannot be sufficiently validated solely with a reference map of forest- cover, so additional biomass information is indispensable. The recent progress in land change science is still not sufficient to statistically test predictive accuracies because the statistical distributions have not been found, according to Pontius Jr. and Millones (2011). 73 3.2.8.2 Forest-use data Banna’s forest-use maps of 1988, 1992, 2002, and 2006 are provided by KIB CAS, where the maps are produced using a cutting-edge technique that not only characterizes pixels of different forest-covers but also groups the pixels into distinct polygons based on their geometrical characteristics, that is, object-oriented classification method (Figure 3.7). This advanced method usually requires high-resolution remotely sensed data so the tree species of rubber plantations can be differentiated from those of tropical rainforests. Their forest-use information and the accuracy are validated through interviews and ground-truth data collected in the field (Zhang et al., 2006). Consequently, the forest-use maps include numerous polygons; for example, regarding the map from 2006, 4,818 of polygons indicate tropical rainforests, while another 2,215 indicate rubber plantations. The majority of tropical rainforests are legally protected by the government, and some are used for local tourism and/or academic research. The rubber plantations, however, produce rubber, timber, and other non-timber products so that the local people can make economic profits for their living (Z. Yi, pers. comm.). By overlaying the change detection map generated in Section 3.2.3 with the forest- use map provided by KIB CAS, the second validation process shows how accurately the GEB can depict the actual forest-use change, but only in terms of the observed change— not the projected one. Further, distinct types of biomass loss are also estimated as char- acterized in Table 3.2. That is, two types of “deforestation” are employed here individ- ually to see how the estimations of carbon loss differ from each other. It is worth noting that only the gross loss is considered in this calculation, not the net loss; that is, the regional forest re-growth is not considered. 3.2.9 Reference emission level VCS (2011) mandates constructing an REL by multiplying the business-as-usual forest- land loss in the future (in hectares) and the associated biomass loss of the forestland 74 Rubber plantations in 1992 Expansion of rubber plantations between 1992 and 2002 Expansion of rubber plantations between 2002 and 2006 Xishuangbanna Prefecture 0 10 20 5 Miles Figure 3.7: Progressive expansion of rubber plantations in Banna prefecture (in tonnes per hectare). Only AGB is required, so other carbon pools, such as BGB or organic matter in soil, are excluded from this dissertation. The estimated forest biomass (in tonnes) are then converted to the dry weight (multiply by:4, where this values is specifically for Banna (J. Tang, pers. comm.)) and later to the weight of carbon dioxide (multiply by44=12). Finally, the carbon emissions are measured in tCO 2 e (Newell and Vos, 2012, 2011), where the average price per tCO 2 e is 4.9 US Dollars (Diaz et al., 2011). Also required is to document forest carbon stored in wood products, such as furniture made out of local rubber trees (VCS, 2011). This is because forest carbon stocks stored in 75 the wood products have different life-cycles, hence resulting in different carbon emis- sion patterns, so it seems fair to distinguish them from carbon emissions due to slash- and-burn, for instance, which releases carbon rather immediately. This dissertation, however, does not deal with any wood products because there are no available timber yield data in the region. Moreover, the consumption of fossil fuel used for harvest- ing and processing timbers is excluded from this research. Other non-carbon dioxide GHGs, such as nitrogen dioxide, are excluded as well. Finally, the observed and projected carbon emissions (in tCO 2 e) are produced. The more accurately projected outcome is then selected between the GEB and GEOMOD to dictate the business-as-usual carbon emissions between 2010 and 2040. However, another land change modeling component is not employed again here to simulate the forest-cover change between 2010 and 2040 as there are already too many associated errors. Therefore, it does not seem to affect the quality of business-as-usual carbon emission outcome if the observed and projected trends between 2000 and 2010 (ten years) are multiplied by 3 to characterize a range of business-as-usual carbon emissions between 2010 and 2040 (thirty years). 76 Chapter 4 Case Study: Xishuangbanna, China 4.1 Study Area Xishuangbanna Dai Autonomous Prefecture (i.e., Banna) is located in southwest Yun- nan (Figure 1.4) and composed of two counties (i.e., Menghai and Mengla) and one city (i.e., Jinghong City). Dai minority is the most populous ethnic group in Banna, which makes the region Dai-autonomous. The area is about 2 million hectares, where its asl elevation ranges from0 to1;919 meters (mean elevation =655:37 meters). Latitudes and longitudes of the lower left and upper right of the study area are99:9432E,21:1410N and 101:8382E,22:5915N, respectively. Banna is one of the few tropical areas in China, and its climatic and geographical conditions are more similar to those of Southeast Asian countries than the other parts of China. At the continental level, it is part of the Indo- Burma biodiversity hotspot (Myers et al., 2000) and a member of Greater Mekong Sub- region (Xi, 2009). Banna contains 21.7% of the mammal species and 36.2% of the bird species of China (Zhang and Cao, 1995), although the region represents only 0.2% of China in terms of area. Due to these merits, the region has been taken care of by many national and inter- national organizations. Xishuangbanna Tropical Botanic Garden (XTBG)—hosted by CAS and collaborating with the Royal Botanic Gardens, U.K., and the Smithsonian Insti- tution, U.S.A.—exemplifies such conservation efforts (http://english.xtbg.cas.cn). The region, however, also has continuous development issues mainly because the Mekong River flows in the middle of the region (Xi, 2009), which borders Myanmar and Laos. These geographical borders and the river access basically make the region one of the most important international hubs for both ground and hydro transportations in China 77 (Figure 1.4). In short, Banna has experienced and still is experiencing consistent con- flicts between conservation and development. Despite many national forestry and land-use policies driven by the Chinese govern- ment in the past few decades (FAO, 2010, 2006, 2001; Li et al., 2008, 2007; Information Office of the State Council of the People’s Republic of China, 2008; Xu et al., 2006; Mur- ray and Cook, 2004; RFF and CIFOR, 2003), Banna has not much benefited from those conservation efforts and is still vulnerable to rapid deforestation and/or forest degrada- tion. In the past, the study area used to sustain almost 100% of tropical rainforests, but less than 50% is left by 2003, including only 3.6% of old growth tropical rainforests (Li et al., 2009, 2008, 2007). This is mainly due to its proven suitability for producing rubber (i.e., natural latex), where 20% of the region has been converted to the rubber planta- tions already, which is about 400,000 hectares (Qiu, 2009). Between 1988 and 2003, the rubber plantations have expanded at the rate of 324% (Liu et al., 2006), while the rubber production has increased by 8.4-fold between 1988 and 2010 (J. Xu, pers. comm.). The situation is equivalent to losing about 6 million tonnes of biomass, which is essentially losing about 14,000 hectares of tropical rainforests every year since 1976 (Li et al., 2007). The expansion of rubber plantation is the most critical cause of deforestation and/or forest degradation in Banna (J. Xu, pers. comm.). It is partly because the China’s Reform and Innovation Policy has encouraged 1.3 billion Chinese people to be equipped with their own private automobiles and has facilitated the use of industrial trucks, which has led to a tremendous demand for domestic rubber (Li et al., 2007). Given China can fulfill only half of the rubber demand, its market price is way higher than other cash crops, such as rice, bananas, tea, etc. (J. Xu, pers. comm.), so the rapid expansion of rubber plantations in Banna does not seem to be shocking at all (Figures 4.1 and 4.2). In addition to this situation, the Chinese government had initiated a plan to expand the domestic rubber production 780,000 tonnes per year by 2010 (Qiu, 2009). The Natural Forest Protection Plan has also provided a perverse incentive to the rubber plantations 78 in Banna because according to the policy, the rubber plantations are considered “natu- ral forest” and “carbon neutral,” so the policy does not penalize those who replace old growth tropical rainforests to rubber plantations (Li et al., 2007). Another forestry pol- icy, Slope Land Conversion Program, has been promoted by the Chinese government since 1990. This policy regards rubber, eucalyptus, and pine trees as “ecological trees”; that is, those who plant these species will be subsidized by the Chinese government for eight years (D. Zhai, pers. comm.). Such circumstances have been motivating local farmers in Banna to produce more rubber instead of other cash crops (Fu et al., 2010, 2009; Qiu, 2009). Although rice used to be the most important crop in the region, many stopped growing it, partly because its production is more labor intensive than rubber production (Z. Yi, pers. comm.). Rubber plantations in general consume a lot of freshwater to grow the trees and pro- duce natural latex, so for the sake of the landscape’s sustainability, producing rubber is harmful. More freshwater is used when the harvested rubber is processed in factories. However, the local farmers and those who process rubber in factories pay nothing for consuming freshwater, which is an additional incentive to further exploit the regional freshwater (Z. Yi, pers. comm.). Driven by such rubber expansion, Banna has been losing for decades its rich agro/biodiversity, food security, and ecosystem services (Fu et al., 2010, 2009; Qiu, 2009; Li et al., 2009, 2008, 2007; Xi, 2009; Hu et al., 2008; Xu et al., 2007, 2006, 2005; Liu et al., 2006; He and Zhang, 2005; Zhu et al., 2004; Cao et al., 1996; Zhang and Cao, 1995; Chapman, 1991). The rubber plantations are particularly replac- ing habitats of tropical seasonal rainforests because their habitats overlap with those of rubber plantations; it is known the range between 500 meters and 800 meters is most suitable for producing rubber (Li et al., 2007). The associated loss of regional ecosystem services is estimated to be 11.4 billion US Dollars (Qiu, 2009; Hu et al., 2008). At the earlier stage of the rubber expansion, often shrubs were converted to rubber plantations so it is likely that Banna had gained extra carbon in the past. However, at the later stage, degraded tropical rainforests were started to be replaced by rubber 79 100,000 t 150,000 t 200,000 t 250,000 t 300,000 t 350,000 t 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 118,993 123,557 125,020 135,783 137,635 136,206 154,875 175,176 169,479 190,743 204,886 218,065 198,751 228,957 255,272 146,938 153,966 154,982 167,638 171,650 173,142 198,812 223,354 219,156 240,341 264,238 282,233 257,180 298,414 330,635 Yunnan Banna Figure 4.1: Rubber yield in Banna prefecture and Yunnan province plantations (R. Harrison, pers. comm.). Therefore, it is difficult to argue whether the rubber plantations in Banna are actually carbon positive or negative. It might be true there could have been more carbon sequestration if the tropical rainforests were not converted to rubber plantations. However, if tropical rainforests in poor condition were replaced by well-managed rubber plantations, it is also likely the rubber plantations could have stored more carbon than the unhealthy tropical rainforests (R. Harrison, 80 0 ha 125,000 ha 250,000 ha 375,000 ha 500,000 ha 1995 1998 2001 2004 2007 2010 Total Area Harvested Area Figure 4.2: Rubber plantations in Yunnan province pers. comm.). Other than the rubber plantations in Banna, bananas may also have sequestered a fair amount of carbon stocks compared to other cash crops (J. Tang, pers. comm.). Six villages in Banna have already agreed to prevent their rubber expansion when 7,000 US Dollars is guaranteed for each village (Qiu, 2009). Through another survey, however, it was found that many villages still prefer to sustain their business-as-usual 81 (a) Early stage (b) Mid stage (c) Final stage Figure 4.3: Life-cycle of rubber plantations in Banna prefecture, photographed by the author in December 2011 intensive rubber production (Z. Yi, pers. comm.). Producing rubber is more productive than any other cash crops in the region, and its productivity has to do with its life-cycle. The rubber plantations produce natural latex throughout the wet seasons (Figures 4.3a– b); their seeds are used to extract oil and the waste is fed to livestock. Finally, when the trees become too old to produce rubber, the trunks are harvested as timber (Figure 4.3c). Many smallholders plant pineapples and corn for extra cash when their rubber trees are too young to be tapped (Z. Yi, pers. comm.). 82 4.2 Results The section shows a series of outcomes where some of them are the unique products of the GEB while the other outcomes are comparative results between it and the GEO- MOD. More specifically, Sections 4.2.1, 4.2.2, and 4.2.3 are the exclusive outcomes of the GEB, whereas Sections 4.2.4, 4.2.5, and 4.2.6 include the results of the business-as-usual projections made by both the GEB and GEOMOD and their comparison in terms of predictive accuracy. The two approaches use the identical forest-cover data as inputs, which are in fact the results of Section 4.2.2, but with different sizes of study areas and different sets of deforestation driver data. 4.2.1 Receiver operating characteristic The resulting AUC is 81.0% (Figure 4.4). That being said, the top 31% of the VCF’s pixels can be considered “forests” with the 81% certainty in terms of AUC, according to the more conservative definition of Table 3.1, and the threshold is 54% of canopy-cover. In other words, when a pixel contains a value that is equal or larger than 54, the pixel is considered “forest (or forestland),” while others are regarded as either deforested or other wooded lands. 4.2.2 Change detection analysis The forest-cover change detection analysis produces a map that shows deforestation, forest degradation, and forest re-growth in the study area (Figure 4.5). The areal infor- mation of the deforestation and forest degradation is summarized in Table 4.1. When pixels happen to lose some biomass, while the loss is estimated in terms of canopy- cover as proxy, the loss is categorized either deforestation or forest degradation. That said, after the change detection analysis, if the value of the pixels are still equal to or larger than 54, the areas should not be regarded as experiencing “deforestation,” but instead as experiencing “forest degradation.” In contrast, if the values are smaller than 83 100 80 60 40 20 0 0 20 40 60 80 100 Specificity (%) Sensitivity (%) AUC: 81.0% Figure 4.4: Area Under the Curve (AUC) of the Globcover and Vegetation Continuous Field (VCF) 54, the area should be considered as undergoing “deforestation.” In addition, the pix- els that were initially lower than 54 in 2000 but later exceeded the 54 threshold in 2005 should be considered “forest re-growth.” 84 Table 4.1: Amounts of observed forest-cover loss between 2000 and 2005 at the prefec- ture level Forest-cover loss Deforestation 57,257.55 ha Forest degradation 199,299.60 ha -26 -25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 Forest degradation (%) Xishuangbanna Prefecture Deforestation Forest re-growth 0 10 20 5 Miles Figure 4.5: Observed deforestation, forest degradation, and forest re-growth character- ized by forest-cover between 2000 and 2005 in Banna prefecture 85 4.2.3 Local indicator of spatial association The four types of hotspots that LISA has delineated for both deforestation and forest degradation have occurred between 2000 and 2005 (Figures 4.6a–b). The red colors indicate HH hotspots where the darker red shows lower p-values (i.e., higher statistical significance) compared to the brighter red. The other hotspots are drawn in the same manner, where the greens indicate LL, the yellows indicate LH, and the yellowish green indicates HL. Lastly, the gray shows statistically insignificant relationships at 95% con- fidence interval. In short, Banna is a member of both HH hotspots both in terms of deforestation and forest degradation. 86 (a) Deforestation hotspots (b) Forest degradation hotspots 0 50 100 25 Miles 0 50 100 25 Miles HH LL LH HL Not significant p <0.01 p <0.05 Figure 4.6: Hotspots of observed deforestation and forest degradation between 2000 and 2005 in Yunnan province 87 4.2.4 Business-as-usual forest-cover change Multiple rates of deforestation and forest degradation are calculated at three different spatial levels (hence, having different sizes of spatial extents) using the linear extrapo- lation method explained in Section 3.2.5.1 (Table 4.2). The larger red hotspot of defor- estation situated in southern Yunnan is the region that is used as an input spatial extent for the GEB to estimate the rate of deforestation between 2000 and 2005 because the color indicates the hotspot is statistically significant at 95% confidence interval (Figure 4.6a). The rate estimated at the hotspot level dictates the quantity of business-as-usual deforestation between 2005 and 2010 for the GEB, while the rate estimated within the Banna prefecture dictates the quantity of deforestation during the same time period for the GEOMOD run. Although projecting the business-as-usual forest degradation is excluded from this dissertation, when the regional forest degradation is to be pro- jected in the same manner to dictate the loss between 2005 and 2010, the observed rates between 2000 and 2005 show that there may be more degraded forestlands than defor- ested lands (in terms of area not necessarily in terms of forest biomass). The observed rates of deforestation and forest degradation are the highest within the HH hotspots. Table 4.2: Amounts and rates of observed forest-cover loss between 2000 and 2005 at different spatial levels Sub-province level Province level Hotspot level Prefecture level Deforestation (observed) 228,051.63 ha 57,257.55 ha 582,398.55 ha Total forest in 2000 1,169,181.99 ha 363,609.72 ha 3,639,533.40 ha Deforestation rate 19.51 % 15.75 % 16.00 % Degradation (observed) 550,486.71 ha 199,299.60 ha 1,828,275.03 ha Total forest in 2000 964,175.31 ha 363,609.72 ha 3,639,533.40 ha Degradation rate 57.10 % 54.81 % 50.23 % The estimated quantities are all spatially allocated based on the transition potential maps produced by the GEB and GEOMOD (Figures 4.7a–b). Figures 4.8a–b show the business-as-usual projections of deforestation by both the GEB and GEOMOD. 88 County border 0 20 40 10 Miles (a) Geographic Emission Benchmark (b) GEOMOD modeling 38.7 0 63 0 Figure 4.7: Transition potential maps of the (a) Geographic Emission Benchmark (GEB) and (b) GEOMOD modeling 4.2.5 Biomass estimation The AGB loss is estimated for the observed loss due to deforestation and forest degra- dation between 2000 and 2005 (Table 4.3). When it is assumed that forest degradation would lose 50% of the existing biomass, while deforestation indicates 100% loss, the loss due to forest degradation between 2000 and 2005 is larger than the AGB loss due to deforestation for the same time period. The AGB loss estimates between 2005 and 2010 projected by the GEB and GEOMOD are summarized in Table 4.4. The two biomass estimates by the GEB and GEOMOD overestimate the AGB loss when compared to the observed value (Table 4.4). 89 County border Other land-covers Forest persistance between 2000 and 2005 Forest re-growth between 2000 and 2005 Deforestation between 2000 and 2005 Deforestation predicted between 2005 and 2010 0 20 40 10 Miles (a) Geographic Emission Benchmark (b) GEOMOD modeling Figure 4.8: Business-as-usual loss of forestlands projected by the (a) Geographic Emis- sion Benchmark (GEB) and (b) GEOMOD modeling 4.2.6 Figure of merit Finally, measured by FoMs, the GEB turns out to have a somewhat higher predictive accuracy than the GEOMOD (Table 4.5) at the Banna prefecture level. 4.2.7 Forest-use versus forest-cover The progressive expansion of rubber plantations in the study area is mapped in Figure 3.7, and according to the regional experts of KIB CAS and XTBG CAS, the expansion between 2002 and 2006 could be considered mainly forest-use change—conversion from tropical rainforests to rubber plantations—while the older rubber plantations are mix- ture of afforestation, reforestation, and forest-use change because many of them were 90 Table 4.3: Observed above-ground living biomass (AGB) loss due to deforestation and forest degradation between 2000 and 2005 at the prefecture and province levels Type Prefecture Province Deforestation 38.79 Tg 310.74 Tg 10% Degradation 14.09 Tg 110.66 Tg 20% Degradation 28.45 Tg 223.88 Tg 50% Degradation 71.55 Tg 562.49 Tg Table 4.4: Projected and observed above-ground living biomass (AGB) loss due to defor- estation between 2005 and 2010 at the prefecture level Method Prefecture GEB 51.22 Tg GEOMOD 76.62 Tg Observed 4.44 Tg planted on abandoned rice fields, for instance (R. Harrison, pers. comm.; J. Xu, pers. comm.). Assuming this forest-use change as the other type of deforestation, the results of the two change detection analyses are summarized in Table 4.6. In sum, the areal changes do not look identical in terms of both quantity (Table 4.6) and spatial alloca- tion (Figure 4.9). In other words, the forest biomass loss due to the regional expansion of rubber plantations overwhelms the loss due to the other type of deforestation that is specified by the definition similar to the one by FAO (2007) (Table 4.7), and the for- mer’s spatial distribution is less sparsely allocated compared to the other result that is based on forest-cover (Figure 4.9). When the loss due to forest degradation is taken into account, however, the loss measured based on forest-cover may result in a larger number (Tables 4.3 and 4.6). If deforestation is characterized by forest-cover, the charac- terization treats timber harvest within the rubber plantations as deforestation, whereas the other characterization does not treat the timber harvest from the rubber plantations identical to the timber harvest from local tropical rainforests (Figure 4.9). 91 Table 4.5: Figure of Merits (FoMs) of the Geographic Emission Benchmark (GEB) and GEOMOD modeling at the prefecture level Method Prefecture GEB 30.16 % GEOMOD 26.50 % Table 4.6: Amounts of observed forest-cover and forest-use changes at the prefecture level Forest-cover (2000–2005) Forest-use (2002–2006) Deforestation 57,257.55 ha 237,004.01 ha Forest degradation 199,299.60 ha N/A 4.2.8 Reference emission level The two projection methods in fact overestimate the business-as-usual AGB loss when their estimates are compared to the observed one (Table 4.8). Banna released 63.41 mil- lion tCO 2 e to the atmosphere between 2000 and 2010, while the estimates projected by the GEB and GEOMOD result in 132.02 million and 169.27 million tCO 2 e, respectively. The observed carbon emissions (due to forest-cover based deforestation) may be worth at least 310.72 million US Dollars over the past 10 years. The figures extrapolated to dic- tate the business-as-usual carbon emissions between 2010 and 2040 are shown in Table 4.9. The estimations vary from 103.76 to 396.07 million tCO 2 e in Banna. 4.3 Conclusion To answer the research questions in Section 3.1.1, areal estimates of forest disturbed and rates of deforestation fluctuate substantially when distinct characterizations of “forest” and different sizes of spatial extent are employed. It is possible to characterize “defor- estation” and “forest degradation” based on the remotely sensed forest-cover data, and the outcomes are different from the other type of “deforestation” characterized by the local forest-use information. The differently defined terms produce different biomass 92 Table 4.7: Observed above-ground living biomass (AGB) loss due to distinct deforesta- tions between 2000 and 2005 at the prefecture level (1 Tg = 1 terragram = 1 million tonnes) Type AGB loss Deforestation (forest-use, 2002–2006) 62.41 Tg Deforestation (forest-cover, 2000–2005) 38.79 Tg Table 4.8: Projected and observed carbon emissions due to deforestation between 2000 and 2010 at the prefecture level Method Biomass Carbon emission US Dollar GEB 90.01 MtCO 2 e 132.02 MtCO 2 e 646.92 MtCO 2 e GEOMOD 115.41 MtCO 2 e 169.27 MtCO 2 e 829.44 MtCO 2 e Observed 43.23 MtCO 2 e 63.41 MtCO 2 e 310.72 MtCO 2 e loss estimates, where the amount of disturbed forestlands and their spatial patterns do not look identical either. It is possible to objectively delineate hotspots of deforestation and forest degradation, and they do fix the spatial extents when calibrating the rates. The GEB predicts a more accurate forest-cover change than the GEOMOD run at the prefecture level, and the more accurate business-as-usual forest-cover projection (e.g., in hectares) results in a more accurate measure of business-as-usual forest carbon stock loss (e.g., in tonnes) at least for this case study. The forest carbon stock loss due to deforestation and forest degradation in Banna turns out to be substantial. It is logical to conclude that Banna may emit about 400 million tCO 2 e of carbon for the next 30 years, which is equivalent to losing about a billion US Dollars. 4.4 Discussion Overall, the findings from this dissertation are quite conservative although the wood products originated from Banna’s rubber plantations have not been yet included and analyzed. Other forest carbon pools and biomass losses due to forest degradation are omitted. It employs a stricter definition of “forest” than that of FAO; that is, there 93 -26 -25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 Forest degradation (%) Xishuangbanna Prefecture Deforestation Forest re-growth 0 10 20 5 Miles Rubber plantation Figure 4.9: Observed forest-cover change between 2000 and 2005 overlaid with the rub- ber plantations in 2006 may be many forestlands that are being automatically excluded because they are being defined as other wooded lands. Such exclusion somehow underestimates the forest car- bon stock loss. Lastly, it is common sense in the study area that the only way to expand rubber plantations is to cut down the local tropical rainforests because other lands that are suitable for producing rubber have been mostly exploited (Z. Yi, pers. comm.). Thus, Banna’s business-as-usual loss of forestlands will result in rather more intensive loss of forest carbon by consuming tropical rainforests (not only degraded ones but also 94 Table 4.9: Projected and observed carbon emissions due to deforestation between 2010 and 2040 at the prefecture level Method Prefecture GEB 396.07 MtCO 2 e Observed 103.76 MtCO 2 e healthy ones), which is why many conservation communities have been so concerned about the loss of regional forestlands. The GEB outcomes provide novel insights in estimating forest carbon loss driven by land-use and land-cover change in the study area, and there are, at least, three impor- tant messages that need further discussion. First and foremost, the relationship between areal changes of forestlands and their carbon stock changes must be thoroughly ana- lyzed with respect to errors embedded in forest biomass data. Secondly, there must be a clear purpose for employing a land change model for REDD projects. This is primar- ily because land change models might not always be helpful in terms of dictating the most accurate business-as-usual carbon emissions, but they could be useful for other purposes in the context of REDD, such as identifying drivers of deforestation. Lastly, land change modelers must re-think what they mean by “business-as-usual.” The term should be clarified further as it can be used with different meanings. 4.4.1 Forest-cover and forest carbon stock changes Sloan and Pelletier (2012) assume when one has an accurate measure of business-as- usual forest-cover change in terms of area, the accurate areal outcome would also result in the precise measure of corresponding forest carbon stock change. This is pri- marily why the authors conclude their spatially explicit approach may not be useful for Panama’s REDD because their modeling results turn out to have low predictive accuracy. Contrarily, Guti´ errez-V´ elez and Pontius Jr. (2012) argue that the relationship 95 between forest-use change and carbon stock is not linear, and they are rather indepen- dent of each other. In other words, their research implies that a more accurate forest- cover change projection does not necessarily generate a more accurate measure of forest carbon change. That is mainly because the spatial distribution of existing forest carbon stocks has little to no relationship with the area of forestlands that is to be soon dis- turbed in the near future; thus, knowing a precise location of any deforestation does not always dictate the corresponding carbon emissions in a precise manner. There- fore, according to Guti´ errez-V´ elez and Pontius Jr. (2012), it does not seem logical for Sloan and Pelletier (2012) to reach their conclusion because of the strong assumption. If Guti´ errez-V´ elez and Pontius Jr. (2012) are correct, then the comparative experiment of the GEB and GEOMOD should be considered inconclusive mainly because errors embedded in the biomass data are substantial, although for this case study, the assump- tion made by Sloan and Pelletier (2012) seems to work. 4.4.2 Purpose of land change model The purpose of using land change models have to be specified explicitly. This is impor- tant in terms of both maximizing utilities of the models. As pointed out by Castella et al. (2007), land change models have different objectives, and hence there are different ways to validate the models. For instance, both the GEB and GEOMOD are “descrip- tive” (534) tools, according to Castella et al. (2007), and their goal is to forecast business- as-usual trends of forest-cover loss as accurate as possible, and this makes them subject to thorough statistical validation procedures. On the other hand, a different kind of land change model can be used to simply find out which proximate causes are more influential to deforestation and/or forest degradation so REDD stakeholders can specif- ically target the causes to reduce deforestation and/or forest degradation, which is the purpose of the following chapter. In the case of GEB, the model cannot tell which envi- ronmental or socio-economic factors are more critical in terms of driving deforestation 96 and forest degradation. In contrast, weight tables produced by GEOMOD or beta coef- ficients by logistic regression may give some sense of which driver needs to be targeted to reduce deforestation and/or forest degradation. Both the GEB and GEOMOD are subject to further sensitivity analyses, as are all other land change models. This is par- ticularly more true for land change scientists who are mainly interested in producing the most accurate business-as-usual scenarios. 4.4.3 What do we mean by “business-as-usual?” Finally, it is crucial to point out the term “business-as-usual” is used unclearly. In land change modeling, “business-as-usual” often refers to a linear and one-way transition of land-cover changes (Sloan and Pelletier, 2012; Sangermano et al., 2012; Kim, 2010; Pontius Jr. and Chen, 2006), but does not refer to how the associated agents (e.g., local residents) actually utilize lands in their business-as-usual ways. The two meanings might sound similar but are in fact totally different. The difference is verified from Figures 4.9 and 4.10, where the figures visualize the limitation of the usual business- as-usual approach in predictive land change modeling. Illustrated in Figure 4.9, there are different types and degrees of forest carbon stock loss and gain within the rubber plantations, and it seems reasonable to differentiate such forest carbon dynamics to the ones that occur outside the rubber plantations. In other words, it is to classify the fol- lowing two forest transitions: (1) replacing tropical rainforests with young rubber trees, where rubber trees are considered young only if their age does not exceed five years old (Li and Fox, 2012) and (2) culling mature rubber trees and replacing with young rub- ber trees. The AGB difference between tropical rainforests and matured rubber trees is enormous (Figure 4.11). Apparently, the first transition occurs only once when local res- idents develop rubber plantations by clearing out the existing tropical rainforests, while the second transition occurs every 25 or 35 years depending on the ownerships. In this context, it is not surprising at all that the forest-cover dynamics between 2000 and 2005 (Figure 4.9) look very different from those between 2005 and 2010 (Figure 4.10), and it 97 Xishuangbanna Prefecture Deforestation Forest re-growth 0 10 20 5 Miles Rubber plantation Figure 4.10: Observed forest-cover change between 2005 and 2010 overlaid with the rubber plantations in 2006 does not seem feasible to simulate the internal carbon dynamics of rubber plantations if the term “business-as-usual” is used in a business-as-usual way among land change modelers. A more accurate meaning of “business-as-usual” can be defined and modeled, at least for this study area, if there are additional data about the rubber plantations’ age distribution as well as their ownerships. This kind of data is something that can be 98 achieved by high-resolution remote sensing techniques and surveys. Unlike the pre- vious business-as-usual, the newly defined “business-as-usual” is twofold: (1) expan- sion of rubber plantations, which is generally a simple one-way transition from tropical rainforests to rubber plantations, and (2) internal biomass dynamics (i.e., harvesting and re-growth) that occur within the rubber plantations. They both consider how the rubber plantations have been expanded and operated by the local residents and how the very practice will sustain over time. The former can be modeled, for example, via the GEB or GEOMOD as the land change models are designed to forecast such a sim- ple one-way forest-cover/-use transition. However, the latter is rather an independent system dynamic modeling that uses spatially explicit datasets, namely age, ownership, and carbon stock of the rubber plantations. In this way, the rubber trees’ harvesting and re-growth cycles can be modeled as the cycles are known for the region—every 25 or 35 years for small holders and state farms, respectively. In this manner, the real business-as-usual forest-cover/-use change can be modeled transparently, and the find- ings will provide a fundamental scientific basis that is useful to understanding future deforestation and forest degradation with respect to plantation operation. 99 Rubber plantations Mandan village Tropical raintree (the village’s holy tree) Figure 4.11: Visual comparison of rubber plantations and a tropical raintree in Mandan village of Banna, photographed by the author in December 2011 100 Chapter 5 A Step Forward 5.1 Research Objective By design, the GEB is not geared toward enhancing predictive accuracy of land change models nor is it geared toward theoretically explaining any forest-cover and forest-use change. It attempts to provide an “emission reference” for RELs, so any future effort of land change modelers can further contribute to production of accurate RELs in a more targeted manner. The next step is to invent a new land change model that can actually explain the regional forest carbon dynamics, and then once it is explained, one can then accurately estimate the associated carbon emissions. Since expansion of the monocultural rubber plantations is the most dominant land-use change in Banna, it appears to be reasonable and meaningful to start with modeling the expansion because this may be the most crucial factor determining the regional forest carbon dynamics. Productivity of rubber could function as a proxy in explaining the rubber expansion because if a patch of land does not guarantee a certain level of productivity, then local people may not plant rubber trees to that land. The research questions are as follows: What factors affect rubber productivity in Banna? Do contributions of those factors vary over space? 101 5.2 Methods Geographically Weighted Regression (GWR) is designed to explore spatial non- stationarity of any given spatial datasets by running numerous local regressions at dif- ferent spatial scales, instead of running one global regression at one (usually the most macroscopic) spatial scale (Fotheringham et al., 2002). The spatial scale is determined by size of a “spatial window,” which basically determines a number of observations to be used for fitting each local regression model. After a spatial window is specified, it glides over the study area. For each move, one local regression model is fitted, and observa- tions that only fall within the spatial window are used to construct a local regression model. As the spatial window does not stop moving until the observations are fully exploited, numerous regression models are therefore generated. In addition to this idea of moving spatial windows, GWR includes a geographical weighting scheme— observations that are closer to the center of a spatial window are assigned more weights as opposed to the other observations that are farther from the center. For the GWR in this dissertation, an adaptive window is selected, where this automatically adjusts size of the spatial window based on a fixed number of observations (Fotheringham et al., 2002). Lastly, a bisquare weights function is used for the geographic weighting scheme. The bisquare weights function ik is expressed as follows: ik = 8 < : 1(d ik =h) 2 2 ; ifd ik <r 0; otherwise ; (5.1) whered ik is the Euclidean distance between observationsi (denoting the center of the spatial window) andk (denoting the rest of observations within the spatial window); r indicates the radius of the spatial window; and h smoothens the weight decaying. Based on this geographic weighting scheme, the weight matrix W i is constructed as follows: 102 W i = 2 6 6 6 6 6 6 6 4 i1 0 ::: 0 0 i2 ::: 0 . . . . . . . . . . . . 0 0 ::: iN 3 7 7 7 7 7 7 7 5 ; (5.2) where N denotes the total number of observations (Brunsdon et al., 1998). Finally, GWR’s coefficient ^ i is expressed as follows: ^ i = X T W i X 1 X T W i Y; (5.3) where a nonparametric approach is used to estimate the coefficients. As a result, GWR produces numerous ^ i s that vary over space, and their significance levels are indicated by pseudot statistics (Fotheringham et al., 2002). 5.3 Data The rubber yield of 2010 is used as the dependent variable, while the others are used as the independent variables. Daily rubber yield and age of rubber tree data were collected from 2008 to 2010, where the data collection only occurred during the wet seasons. After 1,173 georeferenced points were collected at the plot level through a stratified sampling for the entire study area, the daily yield data are processed to dictate the annual yield of 2010. One can estimate the age of rubber trees by measuring length of tapping scars on the barks (Figure 4.3a)—approximately 20 centimeters is equivalent to one year (Z. Yi, pers. comm.). For more information about the rubber yield and age data, find Yi et al. (2013). The other spatial variables include the following: percent of clay in soil, soil pH (Yi et al., 2013), monthly precipitation for the wet season, annual precipitation, monthly mean temperature for the wet season, annual mean temperature (Hijmans et al., 2005, Worldclim), elevation, slope, aspect (METI and NASA, 2011), population (ORNL, 2008), distances from road and stream networks, and protected areas (Yi et al., 2013). While 103 other variables range from lower to higher values in a linear fashion, the aspect ranges from 0 to 359.9 non-linearly (clockwise from north), expressed in degrees. 5.4 Results According to the Box-Cox transformation, the dependent variable should be square- rooted before fitting an Ordinary Least Squares regression (OLS) because« is around (or a bit higher than):5. Moran’sIs are estimated, first, from the square-rooted dependent variable before fitting an OLS (=:7850***) and second, from the residuals after the OLS fitting (=:4786***), where only one nearest neighbor is considered (i.e., 1st order spatial autocorrelation). The result of Durbin-Watson test shows 1:9028 (p-value =:068). The fitted OLS model is shown in Table 5.1, while the independent variables are selected in a stepwise way. Through cross validating, an adaptive bandwidth is found optimal (CV score = 1596.66) when 28:9% of observations are included for each local regression. With the bandwidth, a GWR is constructed based on the same variables used for the OLS, and its spatially varying coefficients are summarized in Table 5.2. Unlike the signs of OLS’ coefficients, those of GWR’s coefficients range from negative to positive for each inde- pendent variable. The reds show negative relations, while the blues show positive ones; the vector points are interpolated via Inverse Distance Weighting (IDW) to generate raster surfaces (Figures 5.2–5.12). 104 Table 5.1: Coefficients of Ordinary Least Squares regression (OLS) Variable Estimate Significance (Intercept) -259.6 *** Age of rubber tree .06637 *** Percent of clay in soil -.1586 ** Soil pH -.1031 * Elevation -.04608 *** Aspect .007721 *** Slope .04405 * Population Distance from road networks Distance from stream networks .0001191 ** Annual precipitation .1205 *** Annual mean temperature 1.978 *** Nature reserve (dummy) Squared age of rubber tree Squared elevation .00002157 *** Squared aspect -.00002062 *** Squared slope -.0009900 * Squared annual precipitation -.00004304 *** Squared annual mean temperature -.004721 *** Interaction (percent of clay in soil .0001108 ** and annual precipitation) Interaction (population and slope) Interaction (population and distance from road networks) 105 Table 5.2: Coefficients of Geographically Weighted Regression (GWR) Variable Minimum Median Maximum (Intercept) -1,125. -54.19 479.4 Age of rubber tree -.08929 .02609 .1421 Percent of clay in soil -1.557 -.03468 2.655 Soil pH -.8945 -.1058 1.834 Elevation -.04393 .002763 .03057 Aspect -.008468 .005215 .01441 Slope -.07818 -.02220 .1647 Distance from stream networks -.0003201 .00007448 .0008648 Annual precipitation -1.290 -.01888 .9685 Annual mean temperature -4.922 .2192 8.596 Squared elevation -.00002350 -.000003319 .00003390 Squared aspect -.00003428 -.00001672 .00001832 Squared slope -.003698 .0003348 .002464 Squared annual precipitation -.0003474 .000006229 .0004574 Squared annual mean temperature -.01907 -.0005882 .01074 Interaction (percent of clay in soil -.001896 .00002492 .001084 and annual precipitation) 106 5.5 Discussion The OLS seems to have little to no 1st order serial autocorrelation, according to the Durbin-Watson statistic. The Moran’s I statistics indicate positive spatial autocorrela- tion, implying there may be spatial structures unexplained by the independent vari- ables. The results justify the use of GWR or other spatial regression models. According to the OLS model, age of rubber tree appears to have a significant posi- tive correlation with rubber yield, which is reasonable—older trees produce more rub- ber. The squared term of age of rubber tree is excluded from the OLS, indicating there is no non-linear relationship between age of rubber tree and rubber yield, where the non-linear relationship may indicate the situation—if rubber trees get too old then the aging may result in decrease of rubber yield. Since there is no non-linear relationship, it seems reasonable to conclude that rubber plantations managed by smallholders and the local government are re-planted every 25 years and 35 years, respectively (Yi et al., 2013). That is, unproductive old rubber plantations are not dominant in the region, and the re-planting is actually happening. Percent of clay in soil is negatively correlated with rubber yield, implying that if more clay resides in soil, then it is harder for rubber trees to absorb water from the ground. The negative correlation between soil pH and rubber yield is expected as it is common sense given that in the study area rubber trees do not grow on limestone, which has a relatively high soil pH. There are, however, a limited number of rubber trees growing along slopes in limestone region because the topsoil is deep enough (J. Tang, pers. comm.). Elevation is negatively correlated with rubber yield, and this indicates there are more rubber plantations in lower lands. Its squared term shows there is also a non-linear relationship between elevation and rub- ber yield. In other words, if elevation gets lower, after a certain point, fewer rubber plantations are detected. In Banna, banana plantations are usually at the lowest slope; rubber plantations are at the medium slope (Figure 5.1); and tea/coffee plantations are at the top slope. Generally, the lower the rubber plantations are located, the older they 107 are (D. Schmidt-Vogt, pers. comm.). It is reasonable to verify that aspect has a non- linear relationship with rubber yield. More specifically, 0 means perfect north, while 180 indicates perfect south. Within the range of 0 and 180 (or from facing north to south), a higher aspect value shows more solar radiation. However, this relationship is reversed between 180 and 360 (360 is technically identical to 0), so a lower aspect value indicates more exposure to solar radiation. The similar pattern is found for slope, annual precipitation, and annual mean temperature. Distance from stream networks has a positive correlation with rubber yield, and this may be because planting rubber trees near the streams is forbidden by the government as rubber plantations facilitate topsoil loss. The interaction of percent of clay in soil and annual precipitation turns out significant, indicating more rainfall may generate more clay in soil. All of human- related terms, such as population (including the interaction terms), distance from road networks, and nature reserve, turn out as insignificant. Unlike the OLS model, the GWR provides more localized information in a spatially disaggregated yet exploratory fashion (Figures 5.2–5.12), which makes the two models supplementary to each other. From the OLS perspective, the GWR provides different perspectives to explore spatial drift of parameters (Fotheringham, 1997) or potential reason(s) for the spatial autocorrelation embedded in the residuals of the OLS. From the GWR perspective, the OLS provides outcomes that are spatially aggregated and are helpful when analyzing multiscalar relationships of the dependent and independent variables. In the case of age of rubber tree, although the OLS shows older trees tend to produce more rubber than younger ones at the aggregated level, some part of the study area seems to have the opposite relationship if that is to be explored in a spa- tially disaggregated manner (Figure 5.3). Similar patterns are found from the rest of the independent variables (Figures 5.4–5.11). When the maps of elevation (Figure 5.6) and squared elevation (Figure 5.13) are compared, it becomes clear that only the areas where 108 Figure 5.1: Visual comparison of rubber and banana plantations in Banna prefecture, photographed by the author in December 2011 slope is negatively correlated to rubber yield are also having non-linearity. Similar pat- terns are found from all of the squared terms (Figures 5.14–5.17), which supplements the interpretation of the OLS. To conclude, age of rubber tree and in particular the geographic and climatic con- ditions such as percent of clay in soil, soil pH, elevation, aspect, slope, annual precip- itation, annual mean temperature, as well as the interaction between percent of clay in soil and annual precipitation, all significantly affect the rubber productivity, and hence rubber expansion in Banna, while their contributions vary over space. In con- trast, the human-related terms have little to do with the rubber productivity. This chap- ter demonstrates that it may be methodologically “wrong” in land change modeling 109 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles I n t e r c e p t < - 1 , 0 0 0 - 1 , 0 0 0 – - 8 0 0 - 8 0 0 – - 6 0 0 - 6 0 0 – - 4 0 0 - 4 0 0 – - 2 0 0 - 2 0 0 – 0 0 – 2 0 0 2 0 0 – 4 0 0 > 4 0 0 Figure 5.2: Spatial drift of intercept to employ an aspatial regression approach on spatial data. Thus, a future land change model should incorporate spatial structures into the model. 110 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles A g e o f r u b b e r t r e e ( y e a r s ) < - 0 . 0 6 - 0 . 0 6 – - 0 . 0 3 - 0 . 0 3 – 0 0 – 0 . 0 3 0 . 0 3 – 0 . 0 6 0 . 0 6 – 0 . 0 9 0 . 0 9 – 0 . 1 2 > 0 . 1 2 Figure 5.3: Spatial drift of age of rubber tree 111 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles P e r c e n t o f c l a y i n s o i l < - 1 - 0 . 1 – - 0 . 5 - 0 . 5 – 0 0 – 0 . 5 0 . 5 – 1 1 – 1 . 5 1 . 5 – 2 > 2 Figure 5.4: Spatial drift of percent of clay in soil 112 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S o i l p H < - 0 . 6 - 0 . 6 – - 0 . 3 - 0 . 3 – 0 0 – 0 . 3 0 . 3 – 0 . 6 0 . 6 – 0 . 9 0 . 9 – 1 . 2 1 . 2 – 1 . 5 1 . 5 – 1 . 8 > 1 . 8 Figure 5.5: Spatial drift of soil pH 113 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles E l e v a t i o n ( m e t e r s ) < - 0 . 0 4 - 0 . 0 4 – - 0 . 0 3 - 0 . 0 3 – - 0 . 0 2 - 0 . 0 2 – - 0 . 0 1 - 0 . 0 1 – 0 0 – 0 . 0 1 0 . 0 1 – 0 . 0 2 > 0 . 0 2 Figure 5.6: Spatial drift of elevation 114 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles A s p e c t ( d e g r e e s ) < - 0 . 0 0 7 5 - 0 . 0 0 7 5 – - 0 . 0 0 5 - 0 . 0 0 5 – - 0 . 0 0 2 5 - 0 . 0 0 2 5 – 0 0 – 0 . 0 0 2 5 0 . 0 0 2 5 – 0 . 0 0 5 0 . 0 0 5 – 0 . 0 0 7 5 0 . 0 0 7 5 – 0 . 0 1 0 . 0 1 – 0 . 0 1 2 5 > 0 . 0 1 2 5 Figure 5.7: Spatial drift of aspect 115 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S l o p e ( d e g r e e s ) < - 0 . 0 6 - 0 . 0 6 – - 0 . 0 3 - 0 . 0 3 – 0 0 – 0 . 0 3 0 . 0 3 – 0 . 0 6 0 . 0 6 – 0 . 0 9 0 . 0 9 – 0 . 1 2 0 . 1 2 – 0 . 1 5 > 0 . 1 5 Figure 5.8: Spatial drift of slope 116 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles D i s t a n c e f r o m s t r e a m n e t w o r k s ( m e t e r s ) < - 0 . 0 0 0 2 - 0 . 0 0 0 2 – 0 0 – 0 . 0 0 0 2 0 . 0 0 0 2 – 0 . 0 0 0 4 0 . 0 0 0 4 – 0 . 0 0 0 6 0 . 0 0 0 6 – 0 . 0 0 0 8 > 0 . 0 0 0 8 Figure 5.9: Spatial drift of distance from stream networks 117 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles A n n u a l p r e c i p i t a t i o n ( m i l l i m e t e r s ) < - 1 . 2 5 - 1 . 2 5 – - 1 - 1 – - 0 . 7 5 - 0 . 7 5 – - 0 . 5 - 0 . 5 – - 0 . 2 5 - 0 . 2 5 – 0 0 – 0 . 2 5 0 . 2 5 – 0 . 5 0 . 5 – 0 . 7 5 > 0 . 7 5 Figure 5.10: Spatial drift of annual precipitation 118 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles A n n u a l m e a n t e m p e r a t u r e ( C e l s i u s * 1 0 ) < - 4 . 5 - 4 . 5 – - 3 - 3 – - 1 . 5 - 1 . 5 – 0 0 – 1 . 5 1 . 5 – 3 3 – 4 . 5 4 . 5 – 6 6 – 7 . 5 > 7 . 5 Figure 5.11: Spatial drift of annual mean temperature 119 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles I n t e r a c t i o n ( p e r c e n t o f c l a y i n s o i l & a n n u a l p r e c i p i t a t i o n ) < - 0 . 0 0 1 5 - 0 . 0 0 1 5 – - 0 . 0 0 1 2 - 0 . 0 0 1 2 – - 0 . 0 0 0 9 - 0 . 0 0 0 9 – - 0 . 0 0 0 6 - 0 . 0 0 0 6 – - 0 . 0 0 0 3 - 0 . 0 0 0 3 – 0 0 – 0 . 0 0 0 3 0 . 0 0 0 3 – 0 . 0 0 0 6 0 . 0 0 0 6 – 0 . 0 0 0 9 > 0 . 0 0 0 9 Figure 5.12: Spatial drift of interaction (percent of clay in soil and annual precipitation) 120 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S q u a r e d e l e v a t i o n < - 0 . 0 0 0 0 1 8 - 0 . 0 0 0 0 1 8 – - 0 . 0 0 0 0 1 2 - 0 . 0 0 0 0 1 2 – - 0 . 0 0 0 0 0 6 - 0 . 0 0 0 0 0 6 – 0 0 – 0 . 0 0 0 0 0 6 0 . 0 0 0 0 0 6 – 0 . 0 0 0 0 1 2 0 . 0 0 0 0 1 2 – 0 . 0 0 0 0 1 8 0 . 0 0 0 0 1 8 – 0 . 0 0 0 0 2 4 0 . 0 0 0 0 2 4 – 0 . 0 0 0 0 3 > 0 . 0 0 0 0 3 Figure 5.13: Spatial drift of squared elevation 121 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S q u a r e d a s p e c t < - 0 . 0 0 0 0 3 - 0 . 0 0 0 0 3 – - 0 . 0 0 0 0 2 5 - 0 . 0 0 0 0 2 5 – - 0 . 0 0 0 0 2 - 0 . 0 0 0 0 2 – - 0 . 0 0 0 0 1 5 - 0 . 0 0 0 0 1 5 – - 0 . 0 0 0 0 1 - 0 . 0 0 0 0 1 – - 0 . 0 0 0 0 0 5 - 0 . 0 0 0 0 0 5 – 0 0 – 0 . 0 0 0 0 0 5 0 . 0 0 0 0 0 5 – 0 . 0 0 0 0 1 0 . 0 0 0 0 1 – 0 . 0 0 0 0 1 5 > 0 . 0 0 0 0 1 5 Figure 5.14: Spatial drift of squared aspect 122 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S q u a r e d s l o p e < - 0 . 0 0 3 - 0 . 0 0 3 – - 0 . 0 0 2 - 0 . 0 0 2 – - 0 . 0 0 1 - 0 . 0 0 1 – 0 0 – 0 . 0 0 1 0 . 0 0 1 – 0 . 0 0 2 > 0 . 0 0 2 Figure 5.15: Spatial drift of squared slope 123 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S q u a r e d a n n u a l p r e c i p i t a t i o n < - 0 . 0 0 0 3 - 0 . 0 0 0 3 – - 0 . 0 0 0 2 - 0 . 0 0 0 2 – - 0 . 0 0 0 1 - 0 . 0 0 0 1 – 0 0 – 0 . 0 0 0 1 0 . 0 0 0 1 – 0 . 0 0 0 2 0 . 0 0 0 2 – 0 . 0 0 0 3 0 . 0 0 0 3 – 0 . 0 0 0 4 > 0 . 0 0 0 4 Figure 5.16: Spatial drift of squared annual precipitation 124 X i s h u a n g b a n n a P r e f e c t u r e 0 10 20 5 Miles S q u a r e d a n n u a l m e a n t e m p e r a t u r e < - 0 . 0 1 8 - 0 . 0 1 8 – - 0 . 0 1 5 - 0 . 0 1 5 – - 0 . 0 1 2 - 0 . 0 1 2 – - 0 . 0 0 9 - 0 . 0 0 9 – - 0 . 0 0 6 - 0 . 0 0 6 – - 0 . 0 0 3 - 0 . 0 0 3 – 0 0 – 0 . 0 0 3 0 . 0 0 3 – 0 . 0 0 6 0 . 0 0 6 – 0 . 0 0 9 > 0 . 0 0 9 Figure 5.17: Spatial drift of squared annual mean temperature 125 Chapter 6 Concluding Remarks 6.1 Intellectual Merit of Empirical Results This dissertation provides a sound scientific foundation for REDD implementation by macroscopically connecting various research modules in distinct fields, and the con- nection demonstrates a few hindrances that REDD must tackle for its successful imple- mentation. From Forest Science’s viewpoint, the research offers a new spatial method- ology to objectively characterize “forest,” “deforestation,” and “forest degradation” in a systematic manner using remote sensing. From Land Change Science’s viewpoint, the research points out the negative consequences of common interchangeable use of the terms “land-use” and “land-cover” when explaining business-as-usual land-use and land-cover change. From Industrial Ecology’s viewpoint, the research provides a way to factor in a “direct land-use” (2) component to the life-cycle assessment of wood or paper products (Newell and Vos, 2012) based on the peer-reviewed and freely avail- able forest-cover and forest biomass data. Lastly, from Spatial Science’s viewpoint, the research shows there are potential research modules that spatial methods can contribute to (indirectly) mitigate climate change. In short, all of the aforementioned must be addressed to precisely measure the climate change benefit of REDD. 6.2 Broader Implications of Findings By design, REDD has a substantial outreach component because its ultimate objective is to establish a sustainable income source for those local people who successfully protect the forests on their own lands. The research outcomes shall provide novel insights into 126 the procedures relevant to a future REDD project in Banna because it offers a baseline for future RELs. The findings of this dissertation shall influence key national and inter- national forest policy debates as they reveal the limitations of current REDD protocols. Lastly, the use of freely available global datasets and free and open source statistical computer programs that support spatial analyses will facilitate the dissemination of the findings, especially to those REDD stakeholders who do not have the resources suffi- cient for conducting costly pilot studies. 6.3 Future Research This doctoral dissertation research will be expanded into three distinct yet related research foci in the future. The first focus is about conducting comparative REL studies over the pan-tropical regions, while the second focus is geared towards understanding and explaining why and how forestlands are removed or converted to other land-covers and land-uses where the Forest Transition Theory (FTT) may be playing a key role in this regard. Lastly, the third focus is about measuring carbon emissions and footprints of post-deforestation phases using Life-Cycle Assessment (LCA) by taking into consid- eration both biological and industrial carbon cycles (Gower, 2003). The future find- ings will benefit international GHG accounting protocols in general as they can inform whether REDD actually mitigates anthropogenic climate change or not. 6.3.1 Comparative RELs One of the main purposes of this dissertation is to invent the unified REL methodology that can be applicable for the pan-tropical regions. Therefore, it seems to be logical that the GEB should be tested over different areas with varying spatial scales to understand the multiscalar nature of prospective and spatially explicit RELs. The future research will include numerous ROC and LISA runs and change detection analyses to dictate the forest-cover change between 2000 and 2010, and how such change is related to REDD. 127 6.3.2 Forest transition theory The FTT (Mather, 1992) is geared towards explaining why forests are destined to be first removed from a landscape for other land-uses then eventually to re-grow later on, and the theory has been studied by many researchers (Lambin and Meyfroidt, 2011; Meyfroidt and Lambin, 2010, 2009; Angelsen, 2007; Rudel et al., 2005; Geist and Lambin, 2002; Angelsen and Kaimowitz, 1999; Mather and Needle, 1998; Mather, 1992). There are two distinct ways to comprehend forest transition, and they appear to be comple- mentary to each other. On the one hand, one might want to approach this topic rather theoretically by systematically controlling different causes that presumably trigger for- est transition while such causes could be too simple to represent a reality, but sim- ple enough for researchers to actually model the changes and interpret the outcomes (Angelsen, 2007; Mather and Needle, 1998). On the other hand, others may want to understand such forest transition rather empirically. That is, through understanding spatio-temporal dynamics, or a series of spatial snapshots over time, of deforestation at multiple spatial scales, researchers attempt to invent a grand theory that explains seemingly different forest transitions at the global level. As this dissertation visually portrays the forest transition in Chinese tropics, the research can be useful for the latter, especially given the support of the global datasets employed. 6.3.3 The carbon footprint of forest transition The LCA is designed to evaluate environmental burdens associated with a product by quantifying energy, material usage, and environmental release (e.g., carbon emissions), to assess the impact of those, for instance, to climate change; it is a main tool in Indus- trial Ecology (Graedel and Allenby, 2003). International GHG accounting protocols are in fact a collection of LCA-based primers for different types of carbon offsetting projects, while many of them still assume biogenic emissions have no harm for climate change (ICLEI, 2010, 2009; The Climate Registry, 2008). However, such an assumption has been 128 proven wrong by recent research, and numerous attempts have been made in LCA to include biogenic emissions due to land change (e.g., fiber acquisition) when estimating carbon footprints of wood products and biomass fuels (McKechnie et al., 2011; Newell and Vos, 2011; Geyer et al., 2010; Miller, 2010; Miller et al., 2007; Manomet Center for Conservation Sciences, 2010; Perez-Garcia et al., 2005; White et al., 2005). Fiber acqui- sition activities sometimes turn out to be the largest emission source above all other emissions, such as fossil fuel used for transportation (Newell and Vos, 2011), while dif- ferent harvest practices, forest managements, and forest ownerships are also important factors that determine the carbon footprints (Perez-Garcia et al., 2005; White et al., 2005). An REL is mainly calculated by multiplying business-as-usual forestlands loss and degradation (in hectares) and carbon stocks of living biomass (in tonnes per hectare) to estimate business-as-usual carbon emissions (in tCO 2 e), according to REDD proto- cols (Calmel et al., 2010; VCS, 2010c,d; World Bank BioCarbon Fund, 2008). This is incomplete and unfair given that many other GHG accounting protocols in different industries include the emissions of different life-cycles (ICLEI, 2010, 2009; The Climate Registry, 2008; Boston, 2007; WRI, 2006; WBCSD and WRI, 2005). For instance, if har- vested fibers are used as fuelwoods, this activity will release carbon more immediately to the atmosphere as opposed to the cases where fibers are used as wood products, such as paper or furniture, which might actually store carbon for a couple decades. In sum, land change and different life-cycles must be considered in tandem to produce an ideal REL. Carbon stored in wood products must be measured as well as the other forest carbon pools such as BGB. Accounting the wood products’ carbon stock is not only required by VCS (VCS, 2011), but also will inform many academics in Forest Science and Indus- trial Ecology how forest carbon stocks are actually stored in wood products and how they may influence outcomes of REDD projects. The associated BGB can be taken into account fairly easily because Saatchi et al. (2011) also provide the BGB data in a raster format, and the data (along with the findings of this dissertation) can be modeled via 129 spatially explicit book-keeping models, for instance (Carlson et al., 2012). These future directions appear to be reasonable and feasible in terms of evaluating and comprehend- ing how present and future REDD projects are and will be actually contributing to cli- mate change mitigation. 130 Appendix A: Glossary of Acronyms AGB Above-Ground living Biomass AUC Area Under the Curve BAU Business-As-Usual BGB Below-Ground living Biomass BSI British Standards Institution CAS Chinese Academy of Sciences CCBA Climate, Community and Biodiversity Alliance CH 4 Methane CHGIS China Historical GIS CIAT Centro Internacional de Agricultura Tropical CIESIN Center for International Earth Science Information Network CIFOR Center for International Forestry Research CITAS China In Time And Space CO 2 Carbon dioxide CV Cross Validation DBH Diameter at Breath Height DMSP Defense Meteorological Satellite Program EPA United States Environmental Protection Agency ESA European Space Agency FAO Food and Agriculture Organization of the United Nations FoM Figure of Merit FTT Forest Transition Theory GCM General Circulation Model GCS Geographic Coordinate System GDEM Global Digital Elevation Map GEB Geographic Emission Benchmark GHG GreenHouse Gas GIS Geographic Information Science/Systems GLCF Global Land Cover Facility GOFC-GOLD Global Observation of Forest and Land Cover Dynamics GRUMP Global Rural–Urban Mapping Project GWR Geographically Weighted Regression HWP HardWood Product HYDE HistorY Database of the global Environment ICLEI International Council for Local Environmental Initiatives ICRAF World Agroforestry Centre IDW Inverse Distance Weighting IGBP International Geosphere–Biosphere Programme IPCC Intergovernmental Panel on Climate Change KIB Kunming Institute of Botany 131 LCA Life-Cycle Assessment LEAF Lowering Emissions in Asia’s Forests LISA Local Indicator of Spatial Association MRV Measurement, Reporting and Verification NASA National Aeronautics and Space Administration OLS Ordinary Least Squares regression model ORNL Oak Ridge National Laboratory pAUC partial Area Under the Curve PES Payment for Ecosystem Services REDD Reducing Emissions from Deforestation and forest Degradation REDD+ REDD, plus the role of conservation, sustainable management of forests and enhancement of forest carbon stocks REL Reference Emission Level RFF Resources For the Future ROC Receiver Operating Characteristic SRTM Shuttle Radar Topography Mission tCO 2 e tonnes of Carbon Dioxide equivalent TEEB The Economics of Ecosystems and Biodiversity UN-REDD United Nations Collaborative Programme on Reducing Emis- sions from Deforestation and Forest Degradation in Developing Countries UNEP United Nations Environmental Programme UNFCCC United Nations Framework Convention on Climate Change VCF Vegetation Continuous Field VCS Verified Carbon Standard WBCSD World Business Council for Sustainable Development WRI World Resources Institute XTBG Xishuangbanna Tropical Botanic Garden 132 Appendix B: Geographic Coordinate System What makes spatial data useful and unique compared to any other aggregated data in a tabular form is due to the fact that the spatial data explicitly show information in a spatially disaggregated manner. The spatial data of Banna’s forest biomass portray the information using thousands of pixels where each pixel contains information of a particular forest biomass density, rather than using a single aggregated number for the entire region. Such spatially explicit attributes must be preserved as much as possible so as to precisely estimate the forest biomass of the study area. Choosing a proper Geo- graphic Coordinate System (GCS) is crucial in determining an accurate forest biomass estimation because any estimation would have to rely on a measure of area, which is GCS-dependent. Technically speaking, one cannot directly calculate an area based on a map that uses latitudes and longitudes to assign coordinates since its GCS does not support the metric system. In other words, the map would have no measurements for length (e.g., meter) and therefore, of course, no measurements for area (e.g., square meters or hectare) but would only contain latitudes and longitudes where these are actually measurements for angles (e.g., degree). In this situation, it is typical to project the map to convert its set of coordinates to another set that supports the metric system, so that any areal information can be accurately measured. The dissertation, however, does not follow this typical way, and the process is explained as follows. Table 2: Areal information by different sources Yearbook (2010) CITAS (1990) Banna 2,925,800 ha 1,963,024 ha Yunnan 39,400,000 ha 38,690,385 ha In Table 2, the first column is from Yunnan Yearbook published in 2010 and accessed through the China Data Center at University of Michigan, Ann Arbor, MI, and the upper number looks substantially different from the other upper number of the sec- ond column. It seems impossible to have such a different number unless the study area 133 happens to include an additional county or two, which has never happened, so the dif- ference might be due to the typo (“2” instead of “1”). The second column is the value documented as part of the administrative GIS layer used, and it says the value was calculated with the Lambert Conformal Conic Projection. The latter value is chosen for this dissertation because the way it is produced seems more transparent than the other value. A pixel ends up being equivalent to 20:75 hectares (= 1;963;024=94;589) as Banna prefecture is composed of 94,589 pixels, and this value is the areal factor that is used throughout the dissertation. 134 Appendix C: Sensitivity of ROC and AUC The ROC result in Section 4.2.1 can be different if another ROC analysis were con- ducted at different spatial levels that are different from the Yunnan province. In the case of AUC calibrated at the Banna prefecture level, where the prefecture is nested within the Yunnan province, the value turns out to be significantly different from the AUC calibrated at the province level when the two AUCs are tested by the bootstrap method with 2,000 iteration (Table 3). It is crucial to note that a direct comparison of AUCs with- out any statistical tests is irrelevant (Peterson et al., 2008; Pontius Jr. and Millones, 2011; Robin et al., 2011). The difference might be due to the fact that Banna’s landscape is too complicated given that much deforestation, forest degradation, and forest re-growth are all happening simultaneously compared to the other parts of Yunnan province. In addition, the choice of spatial resolution (e.g., either 300-meter or 500-meter spatial res- olution) can significantly affect the resulting AUCs as well (Table 3). Although finding an optimal spatial scale to define “forest” was not one of the aims of this dissertation, at least these two spatial factors must be kept in mind when defining deforestation and forest degradation based on forest-cover. Table 3: Comparison of Area Under the Curves (AUCs) Spatial scale Pixel resolution AUC Bootstrap Test Yunnan province 300 by 300 meters .8098 significant Banna prefecture 300 by 300 meters .6911 (p<:001) significant Banna prefecture 500 by 500 meters .6986 (p<:001) It is worth noting that there are different ways to analyze AUCs, and one popular way is to employ a partial AUC ([pAUC], Robin et al., 2011). A pAUC is a subset of an (ordinary) AUC, and it does not consider the entire area under the curve. It is produced by limiting the ranges of sensitivity and/or specificity and utilizing that area specifi- cally to assess predictive accuracy. This is a common practice in predictive approaches, 135 such as ecological niche modeling (Peterson et al., 2008) or land change modeling (East- man et al., 2005; Kim, 2010; Pontius Jr. and Schneider, 2001). Employing a pAUC can be useful here if one has additional forest-cover information of the study area because the known portion of forest-cover can simply be excluded here, allowing one to pay more attention to unknown forest-cover. 136 Appendix D: Specifications of GEOMOD Parameter Input Beginning land-use map Forest-cover map of 2005 Mask or strata image Unchecked Neighborhood search mode Unconstrained Beginning time 2005 Ending time 2010 Time step 5 Suitability map Create from driver images Number of simulation 1 Number of drivers 8 Equal weight analysis Checked Environmental impact Unchecked Use validation image Unchecked Number of pixel Total 156,046 Number of pixel State 1 BGN 41,027 Number of pixel State 1 END 34,566 Number of pixel State 2 BGN 115,019 Number of pixel State 2 END 121,480 Output interim time image Unchecked 137 Appendix E: Sensitivity of OLS Variable Estimate Significance (Intercept) -5.328e+01 *** Age of rubber tree Percent of clay in soil Soil pH -1.131e01 * Elevation -1.255e02 *** Aspect 5.078e03 *** Slope Population Distance from road networks Distance from stream networks 9.731e05 * Precipitation (March) 2.590e01 *** Precipitation (April) Precipitation (May) 3.960e01 *** Precipitation (June) 2.114e01 *** Precipitation (July) Precipitation (August) Precipitation (September) 1.217e01 *** Precipitation (October) -7.280e02 * Precipitation (November) 2.456e01 *** Annual precipitation 4.107e02 ** Mean temperature (March) -9.715e02 . Mean temperature (April) Mean temperature (May) Mean temperature (June) -1.842e01 *** Mean temperature (July) Mean temperature (August) Mean temperature (September) 2.299e01 *** Mean temperature (October) Mean temperature (November) -1.164e01 * Annual mean temperature Nature reserve (dummy) -8.607e01 *** Squared age of rubber tree 9.113e04 *** Squared elevation 5.274e06 *** Squared aspect -1.479e05 *** Squared slope Squared annual precipitation -4.690e05 *** Squared annual mean temperature Interaction (percent of clay in soil and annual precipitation) 138 Interaction (population and slope) Interaction (population and distance from road networks) Variable Estimate Significance (Intercept) 2.806e+02 *** Age of rubber tree Percent of clay in soil Soil pH -1.492e01 *** Elevation -3.529e03 *** Aspect 5.228e03 *** Slope -7.030e03 Population Distance from road networks Distance from stream networks 6.730e05 . Precipitation (May–Jul.) -3.513e01 *** Precipitation (Aug.–Oct.) -2.083e02 ** Annual precipitation 1.940e01 *** Mean temperature (May–Jul.) -6.021e02 . Mean temperature (Aug.–Oct.) -3.478e+00 *** Annual mean temperature 8.994e+00 *** Nature reserve (dummy) -7.173e01 ** Squared age of rubber tree 9.463e04 *** Squared elevation Squared aspect -1.494e05 *** Squared slope Squared precipitation (May–Jul.) 2.450e04 *** Squared precipitation (Aug.–Oct.) Squared annual precipitation -5.868e05 *** Squared annual mean temperature -2.046e02 *** Interaction (percent of clay in soil and annual precipitation) Interaction (population and slope) Interaction (population and distance from road networks) 139 Bibliography Aceh (2008). Reducing Carbon Emissions from Deforestation in the Ulu Masen Ecosystem, Aceh, Indonesia – A Triple-Benefit Project Design Note for CCBA Audit. Aceh, Indonesia: The Provincial Government of Nanggroe Aceh Darussalam (Aceh). Achard, F., R. S. DeFries, H. D. Eva, M. C. Hansen, P . Mayaux, and H.-J. Stibig (2007). Pan-tropical monitoring of deforestation. Environmental Research Letters 2(4), 045022. Achard, F., H. D. Eva, P . Mayaux, H.-J. Stibig, and A. Belward (2004). Improved esti- mates of net carbon emissions from land cover change in the tropics for the 1990s. Global Biogeochemical Cycles 18(2), 1–12. Achard, F., H. D. Eva, H.-J. Stibig, P . Mayaux, J. Gallego, T. Richards, and J.-P . Mal- ingreau (2002). Determination of deforestation rates of the world’s humid tropical forests. Science 297, 999–1002. Achard, F., H.-J. Stibig, H. D. Eva, E. J. Lindquist, A. Bouvet, O. Arino, and P . Mayaux (2010). Estimating tropical deforestation from earth observation data. Carbon Man- agement 1(2), 271–87. Agarwal, C., G. M. Green, J. M. Grove, T. P . Evans, and C. M. Schweik (2002). A Review and Assessment of Land-Use Change Models: Dynamics of Space, Time, and Human Choice. Technical report, United States Department of Agriculture Forest Service Northeastern Research Station, Newtown Square, PA. Agrawal, A., A. Chhatre, and R. Hardin (2008). Changing governance of the world’s forests. Science 320(5882), 1460–2. Anderson, L. O., Y. E. Shimabukuro, R. S. Defries, and D. Morton (2005). Assessment of Deforestation in Near Real Time Over the Brazilian Amazon Using Multitemporal Fraction Images Derived From Terra MODIS. IEEE Geoscience and Remote Sensing Letters 2(3), 315–18. Angelsen, A. (2007). Forest Cover Change in Space and Time: Combining the von Th ¨ unen and Forest Transition Theories (WPS4117). World Bank Policy Research Work- ing Paper, 1–43. Angelsen, A. and D. Kaimowitz (1999). Rethinking Causes Lessons from of the Models Deforestation: Economic Models. World Bank Research Observer 14(1), 73–98. Anselin, L. (1995). Local Indicators of Spatial Association—LISA. Geographical Analy- sis 27(2), 93–115. 140 Asner, G. P . (2009). Tropical forest carbon assessment: Integrating satellite and airborne mapping approaches. Environmental Research Letters 4(3), 034009. Asner, G. P ., D. E. Knapp, J. Boardman, R. O. Green, T. Kennedy-Bowdoin, M. Eastwood, R. E. Martin, C. Anderson, and C. B. Field (2012). Carnegie Airborne Observatory-2: Increasing science data dimensionality via high-fidelity multi-sensor fusion. Remote Sensing of Environment 124, 454–65. Asner, G. P ., G. V . N. Powell, J. Mascaro, D. E. Knapp, J. K. Clark, J. Jacobson, T. Kennedy-Bowdoin, A. Balaji, G. Paez-Acosta, E. Victoria, L. Secada, M. Valqui, and R. F. Hughes (2010). High-resolution forest carbon stocks and emissions in the Amazon. Proceedings of the National Academy of Sciences 107(38), 1–5. Asner, G. P ., T. K. Rudel, T. M. Aide, R. S. DeFries, and R. Emerson (2009). A contem- porary assessment of change in humid tropical forests. Conservation Biology 23(6), 1386–95. Aspinall, R. (2006). Editorial. Journal of Land Use Science 1(1), 1–4. Baccini, A., S. J. Goetz, W. S. Walker, N. T. Laporte, M. Sun, D. Sulla-Menashe, J. Hackler, P . S. A. Beck, R. Dubayah, M. A. Friedl, S. Samanta, and R. A. Houghton (2012). Estimated carbon dioxide emissions from tropical deforestation improved by carbon- density maps. Nature Climate Change 2(3), 182–5. Baker, D. J., G. Richards, A. Grainger, P . Gonzalez, S. Brown, R. S. DeFries, A. Held, J. Kellndorfer, P . Ndunda, and D. Ojima (2010). Achieving forest carbon information with higher certainty: A five-part plan. Environmental Science and Policy 13(3), 249–60. Baker, T. R., O. L. Phillips, Y. Malhi, S. Almeida, L. Arroyo, A. Di Fiore, T. Erwin, N. Higuchi, T. J. Killeen, S. G. Laurance, W. F. Laurance, S. L. Lewis, A. Monteagudo, D. A. Neill, P . N. Vargas, N. C. A. Pitman, J. N. M. Silva, and R. V . Mart´ ınez (2004). Increasing biomass in Amazonian forest plots. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 359(1443), 353–65. Benito, d. P . B. and d. G. J. Pe˜ nas (2011). Greenhouses, land use change, and predic- tive models: MaxEnt and Geomod working together. In M. Paegelow and M. T. C. Olmedo (Eds.), Modelling Environmental Dynamics: Advances in Geomatic Solutions, Chapter 11, pp. 297–317. Berlin, Heidelberg: Springer-Verlag. Bicheron, P ., P . Defourny, C. Brockmann, L. Schouten, C. Vancutsem, M. Huc, S. Bon- temps, M. Leroy, F. Achard, M. Herold, F. Ranera, and O. Arino (2008). GLOBCOVER: Products Description and Validation Report. Ramonville-Saint-Agne, France: MEDIAS- France/POSTEL. Bonan, G. B., R. S. DeFries, M. T. Coe, and D. S. Ojima (2004). Land use and climate. In G. Gutman, A. C. Janetos, C. O. Justice, E. F. Moran, J. F. Mustard, R. R. Rindfuss, D. Skole, B. L. Turner, and M. A. Cochrane (Eds.), Land Change Science: Observing, Monitoring and Understanding Trajectories of Change on the Earths Surface, Volume 6 of Remote Sensing and Digital Image Processing, Chapter 17, pp. 307–20. Dordrecht, Netherlands: Springer Netherlands. 141 Bontemps, S., P . Defourny, E. V . Bogaert, O. Arino, V . Kalogirou, and J. R. Perez (2011). GLOBCOVER 2009: Products Description and Validation Report. Paris, France: Univer- sit´ e catholique de Louvain (UCL) and European Space Agency (ESA). Boston, A. (2007). Best Practices and Better Protocols. British Columbia, Canada: British Columbia Ministry of Environment, Community Energy and Emissions Inventory Working Group, and Holland Barrs Planning Group. Broich, M., M. C. Hansen, F. Stolle, P . V . Potapov, B. A. Margono, and B. Adusei (2011). Remotely sensed forest cover loss shows high spatial and temporal variation across Sumatera and Kalimantan, Indonesia 2000–2008. Environmental Research Letters 6(1), 014010. Brown, D. G. and J.-D. Duh (2004). Spatial simulation for translating from land use to land cover. International Journal of Geographical Information Science 18(1), 35–60. Brown, S. (1997). Estimating Biomass and Biomass Change of Tropical Forests: A Primer (FAO Forestry Paper 134). Rome, Italy: Food and Agriculture Organization (FAO). Brown, S. (2002a). Measuring, monitoring, and verification of carbon benefits for forest- based projects. Philosophical Transactions of the Royal Society of London. Series A, Mathe- matical, Physical, and Engineering Sciences 360(1797), 1669–83. Brown, S. (2002b). Product 5: Report of the application of GEOMOD to the Rio Bravo, Belize project area. Arlington, VA: Winrock International. Brown, S. (2005). Product 1: Description of GEOMOD and application to the Noel Kempff, Bolivia project area. Arlington, VA: Winrock International. Brown, S., M. Hall, K. Andrasko, F. Ruiz, W. Marzoli, G. Guerrero, O. Masera, A. Dushku, B. DeJong, and J. Cornell (2007). Baselines for land-use change in the trop- ics: Application to avoided deforestation projects. Mitigation and Adaptation Strategies for Global Change 12(6), 1001–26. Brown, S. and A. E. Lugo (1992). Aboveground biomass estimates for tropical moist forests of the Brazilian Amazon. Interciencia 17(1), 8–18. Brunsdon, C., A. S. Fotheringham, and M. E. Charlton (1998). Geographically weighted regression-modelling spatial non-stationarity. Journal of the Royal Statistical Society: Series D (The Statistician) 47(3), 431–43. BSI British Standards, Carbon Trust, and Department for Environment Food and Rural Affairs (2011). P AS 2050:2011 – Specification for the assessment of the life cycle greenhouse gas emissions of goods and services. London, UK: BSI British Standards. Buermann, W., S. Saatchi, T. B. Smith, B. R. Zutta, J. a. Chaves, B. Mil´ a, and C. H. Graham (2008). Predicting species distributions across the Amazonian and Andean regions using remote sensing data. Journal of Biogeography 35(7), 1160–76. Busch, J., F. Godoy, W. R. Turner, and C. A. Harvey (2011). Biodiversity co-benefits of reducing emissions from deforestation under alternative reference levels and levels of finance. Conservation Letters 4(2), 101–15. 142 Calmel, M., A. Martinet, N. Grondard, T. Dufour, M. Rageade, A. Fert´ e-Devin, and ONF International (2010). REDD+ at project scale: Evaluation and development guide. Paris, France: ONF International. Cao, M., J. Zhang, Z. Feng, J. Deng, and X. Deng (1996). Tree species composition of a seasonal rain forest in Xishuangbanna, Southwest China. Tropical Ecology 37(2), 183– 92. Carlson, K. M., L. M. Curran, D. Ratnasari, A. M. Pittman, B. S. Soares-Filho, G. P . Asner, S. N. Trigg, D. a. Gaveau, D. Lawrence, and H. O. Rodrigues (2012). Committed car- bon emissions, deforestation, and community land conversion from oil palm planta- tion expansion in West Kalimantan, Indonesia. Proceedings of the National Academy of Sciences 109(19), 7559–64. Castella, J.-C., S. P . Kam, D. D. Quang, P . H. Verburg, and C. T. Hoanh (2007). Com- bining top-down and bottom-up modelling approaches of land use/cover change to support public policies: Application to sustainable management of natural resources in northern Vietnam. Land Use Policy 24(3), 531–45. Chapman, E. C. (1991). The expansion of rubber in southern Yunnan, China. The Geo- graphical Journal 157(1), 36–44. Chhatre, A. and A. Agrawal (2009). Trade-offs and synergies between carbon storage and livelihood benefits from forest commons. Proceedings of the National Academy of Sciences 106(42), 17667–70. CIESIN, IPFRI, World Bank, and CIAT (2004). Global Rural–Urban Mapping Project (GRUMP): Gridded Population of the World, Version 3, with Urban Reallocation (GPW- UR). Palisades, NY: Center for International Earth Science Information Net- work (CIESIN) at Columbia University, International Food Policy Research Institute (IPFRI), the World Bank, and Centro Internacional de Agricultura Tropical (CIAT). CIFOR (2008). Moving Ahead with REDD: Issues, Options and Implications. Bogor, Indone- sia: Center for International Forestry Research (CIFOR). Congalton, R. G. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment 37(1), 35–46. Cramer, W., A. Bondeau, S. Schaphoff, W. Lucht, B. Smith, and S. Sitch (2004). Tropical forests and the global carbon cycle: Impacts of atmospheric carbon dioxide, climate change and rate of deforestation. Philosophical Transactions of the Royal Society of Lon- don. Series B, Biological Sciences 359(1443), 331–43. Damuth, J. (2001). Scaling of growth: plants and animals are not so different. Proceedings of the National Academy of Sciences 98(5), 2113–4. DeFries, R. and C. Rosenzweig (2010). Toward a whole-landscape approach for sus- tainable land use in the tropics. Proceedings of the National Academy of Sciences 107(46), 19627–32. Defries, R. S., F. Achard, S. Brown, M. Herold, D. Murdiyarso, B. Schlamadinger, and C. de Souza Jr (2006). Reducing Greenhouse Gas Emissions from Deforestation 143 in Developing Countries: Considerations for Monitoring and Measuring. Techni- cal report, Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD), Rome, Italy. DeFries, R. S., F. Achard, S. Brown, M. Herold, D. Murdiyarso, B. Schlamadinger, and C. Desouzajr (2007). Earth observations for estimating greenhouse gas emissions from deforestation in developing countries. Environmental Science and Policy 10(4), 385–94. DeFries, R. S., M. C. Hansen, J. R. G. Townshend, A. C. Janetos, and T. R. Loveland (2000). A new global 1-km dataset of percentage tree cover derived from remote sensing. Global Change Biology 6, 247–54. DeFries, R. S., M. C. Hansen, J. R. G. Townshend, and R. Sohlberg (1998). Global land cover classifications at 8 km spatial resolution: the use of training data derived from Landsat imagery in decision tree classifiers. International Journal of Remote Sens- ing 19(16), 3141–68. DeFries, R. S., R. A. Houghton, M. C. Hansen, C. B. Field, D. L. Skole, and J. Town- shend (2002). Carbon emissions from tropical deforestation and regrowth based on satellite observations for the 1980s and 1990s. Proceedings of the National Academy of Sciences 99(22), 14256–61. DeFries, R. S., T. Rudel, M. Uriarte, and M. Hansen (2010). Deforestation driven by urban population growth and agricultural trade in the twenty-first century. Nature Geoscience 3(3), 178–81. Diaz, D., K. Hamilton, and E. Johnson (2011). State of the Forest Carbon Markets 2011: From Canopy to Currency. Washington, DC: Ecosystem Marketplace. DMSP (2005). Version 4 Defense Meteorological Satellite Program (DMSP)-Operational Lines- can System (OLS) Nighttime Lights Time Series. Boulder, Colorado: National Geophys- ical Data Center, National Oceanic and Atmospheric Administration (NOAA). Dobson, J. E., E. A. Bright, P . R. Coleman, and B. A. Worley (2000). LandScan: A Global Population Database for Estimating Populations at Risk. Photogrammetric Engineering and Remote Sensing 66(7), 849–57. Dushku, A. and S. Brown (2003). Spatial Modeling of Baselines for LULUCF Carbon Projects: The GEOMOD modeling approach. Manila, Philippines: Winrock International. Eastman, J. R. (2007). Land Change Modeler Tutorial. Worcester, MA: Clark Labs, Clark University. Eastman, J. R. (2012). IDRISI Selva Manual. Worcester, MA: Clark Labs, Clark University. Eastman, J. R., R. G. Pontius Jr., and S. Crema (2007). Re: Comparison of Land Change Modeling Environments. Technical report, Clark Labs, Clark University, Worcester, MA. Eastman, J. R., M. E. Van Fossen, and L. A. Sol´ orzano (2005). Transition Potential Mod- eling for Land Cover Change. In D. Maguire, M. Batty, and M. F. Goodchild (Eds.), GIS, Spatial Analysis and Modeling, Chapter 17, pp. 339–68. Redlands, CA: ESRI Press. 144 Elith, J., S. J. Phillips, T. Hastie, M. Dud´ ık, Y. E. Chee, and C. J. Yates (2011). A statistical explanation of MaxEnt for ecologists. Diversity and Distributions 17(1), 43–57. Elvidge, C. D., K. E. Baugh, E. A. Kihn, H. W. Kroehl, and E. R. Davis (1997). Map- ping City Lights With Nighttime Data from the DMSP Operational Linescan System. Photogrammetric Engineering and Remote Sensing 63(6), 727–34. Fang, J., A. Chen, C. Peng, S. Zhao, and L. Ci (2001). Changes in forest biomass carbon storage in China between 1949 and 1998. Science 292, 2320–2. FAO (2001). Global Forest Resources Assessment 2000: FAO Forestry Paper 140. Rome, Italy: Food and Agriculture Organization (FAO). FAO (2006). Global Forest Resources Assessment 2005: FAO Forestry Paper 147. Rome, Italy: Food and Agriculture Organization (FAO). FAO (2007). Definitional issues related to reducing emissions from deforestation in developing countries (Forests and Climate Change Working Paper 5). Rome, Italy: Food and Agricul- ture Organization of the United Nations (FAO). FAO (2010). Global Forest Resources Assessment 2010: FAO Forestry Paper 163. Rome, Italy: Food and Agriculture Organization (FAO). FAO, UNDP , and UNEP (2008). UN Collaborative Programme on Reducing Emissions from Deforestation and Forest Degradation in Developing Countries (UN-REDD). Rome, Italy, New York, NY, and Nairobi, Kenya: Food and Agriculture Organization of the United Nations (FAO), United Nations Development Programme (UNDP), and United Nations Environment Programme (UNEP). FAO and UNEP (1999). Terminology for Integrated Resources Planning and Management. Rome, Italy and Nairobi, Kenya: Food and Agriculture Organization (FAO) and United Nations Environment Programme (UNEP). FARGC (2009). Project Design Document – Reduced Emissions from Deforestation and Forest Degradation in Oddar Meanchey Province, Cambodia: A Community Forestry Initiative for Carbon and Biodiversity Conservation and Poverty Reduction. Phnom Penh, Cambodia: Forestry Administration of the Royal Government of Cambodia (FARGC). FAS (2008). The Juma Sustainable Development Reserve Project: Reducing Greenhouse Gas Emissions from Deforestation in the State of Amazonas, Brazil – Project Design Document (PDD) for validation at Climate, Community and Biodiversity Alliance (CCBA). Ama- zonas, Brazil: Amazonas Sustainable Foundation (FAS). FCO (2009). The Kasigau Corridor REDD Project Phase I: Rukinga Sanctuary – Project Design Document (PDD) for validatoin using the Climate, Community and Biodiversity (CCB) Project Design Standards (2nd ed.). Grass Valley, CA: Forester’s Co-Op (FCO). Foley, J. A., G. P . Asner, M. H. Costa, M. T. Coe, H. K. Gibbs, E. A. Howard, S. Olson, J. Patz, N. Ramankutty, R. Defries, and P . Snyder (2007). Amazon revealed: Forest degradation and loss of ecosystem goods and services in the Amazon Basin. Frontiers in Ecology and the Environment 5(1), 25–32. 145 Fotheringham, A. S. (1997). Trends in quantitative methods I: Stressing the local. Progress in Human Geography 21(1), 88–96. Fotheringham, A. S., C. Brunsdon, and M. E. Charlton (2002). Geographically weighted regression: the analysis of spatially varying relationships. New York, NY: John Wiley and Sons. Fu, Y., H. Brookfield, H. Guo, J. Chen, A. Chen, and J. Cui (2009). Smallholder rubber plantation expansion and its impact on local livelihoods, land use and agrobiodi- versity, a case study from Daka, Xishuangbanna, southwestern China. International Journal of Sustainable Development and World Ecology 16(1), 22–9. Fu, Y., J. Chen, H. Guo, H. Hu, A. Chen, and J. Cui (2010). Agrobiodiversity loss and livelihood vulnerability as a consequence of converting from subsistence farming systems to commercial plantation-dominated systems in Xishuangbanna, Yunnan, China: A household level analysis. Land Degradation and Development 284, 274–84. Gaveau, D. L. A., S. Wich, J. Epting, D. Juhn, M. Kanninen, and N. Leader-Williams (2009). The future of forests and orangutans (Pongo abelii) in Sumatra: Predicting impacts of oil palm plantations, road construction, and mechanisms for reducing car- bon emissions from deforestation. Environmental Research Letters 4(3), 034013. Geist, H. J. and E. F. Lambin (2002). Proximate Causes and Underlying Driving Forces of Tropical Deforestation. BioScience 52(2), 143–50. Geoghegan, J., S. C. Villar, P . Klepeis, P . M. Mendoza, Y. Ogneva-Himmelberger, R. R. Chowdhury, B. L. Turner Jr., and C. Vance (2001). Modeling tropical deforestation in the southern Yucat´ an peninsular region: comparing survey and satellite data. Agri- culture, Ecosystems and Environment 85(1-3), 25–46. Geyer, R., J. P . Lindner, D. M. Stoms, F. W. Davis, and B. Wittstock (2010). Coupling GIS and LCA for biodiversity assessments of land use, Part 2: Impact assessment. International Journal of Life Cycle Assessment 15(7), 692–703. Gibson, L., T. M. Lee, L. P . Koh, B. W. Brook, T. A. Gardner, J. Barlow, C. A. Peres, C. J. A. Bradshaw, W. F. Laurance, T. E. Lovejoy, and N. S. Sodhi (2011). Primary forests are irreplaceable for sustaining tropical biodiversity. Nature 478, 378–81. Gillespie, T. W., G. M. Foody, D. Rocchini, A. P . Giorgi, and S. S. Saatchi (2008). Mea- suring and modelling biodiversity from space. Progress in Physical Geography 32(2), 203–21. Gillespie, T. W., S. S. Saatchi, S. Pau, S. Bohlman, A. P . Giorgi, and S. Lewis (2009). Towards quantifying tropical tree species richness in tropical forests. International Journal of Remote Sensing 30(6), 1629–34. GOFC-GOLD (2010). A sourcebook of methods and procedures for monitoring and report- ing anthropogenic greenhouse gas emissions and removals caused by deforestation, gains and losses of carbon stocks in forests remaining forests, and forestation. Alberta, Canada: Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD). 146 Goldewijk, K. K., A. Beusen, and P . Janssen (2010). Long-term dynamic modeling of global population and built-up area in a spatially explicit way: HYDE 3.1. The Holocene 20(4), 565–73. Goldewijk, K. K., A. Beusen, G. Van Drecht, and M. De Vos (2011). The HYDE 3.1 spatially explicit database of human-induced global land-use change over the past 12,000 years. Global Ecology and Biogeography 20(1), 73–86. Goodchild, M. F. (2003). Geographic Information Science and Systems for Environmen- tal Management. Annual Review of Environment and Resources 28(1), 493–519. Gower, S. T. (2003). Patterns and Mechanisms of the Forest Carbon Cycle. Annual Review of Environment and Resources 28(1), 169–204. Gower, S. T., A. McKeon-Ruediger, A. Reitter, M. Bradley, D. J. Refkin, T. Tollefson, F. J. Souba Jr, A. Taup, L. Embury-Williams, S. Schiavone, J. Weinbauer, A. C. Janetos, and R. Jarvis (2006). Following the paper trail: The impact of magazine and dimensional lumber production on greenhouse gas emissions. Technical report, The H John Heinz III Center for Science, Economics and the Environment, Washington, DC. Graedel, T. E. and B. R. Allenby (2003). Industrial ecology (2nd ed.). Upper Saddle River, NJ: Prentice Hall. Greenoxx (2009). Project Design Document: Madre de Dios Amazon REDD Project. Monte- video, Uruguay: Greenoxx. Guti´ errez-V´ elez, V . H. and R. G. Pontius Jr. (2012). Influence of carbon mapping and land change modelling on the prediction of carbon emissions from deforestation. Environmental Conservation, 1–12. Hall, C. A. S., H. Tian, Y. Qi, R. G. Pontius Jr., and J. Cornell (1995). Modelling Spatial and Temporal Patterns of Tropical Land Use Change. Journal of Biogeography 22(4/5), 753. Hansen, M., R. DeFries, C. Dimiceli, C. Huang, R. Sohiberg, X. Zhan, and J. Townshend (1998). Red and infrared space partitioning for detecting land cover change. In 1998 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp. 2512–14 vol.5. Institute of Electrical and Electronics Engineers (IEEE). Hansen, M. C., R. S. DeFries, J. R. G. Townshend, M. Carroll, C. Dimiceli, and R. A. Sohlberg (2003). Global Percent Tree Cover at a Spatial Resolution of 500 Meters: First Results of the MODIS Vegetation Continuous Fields Algorithm. Earth Interac- tions 7(10), 1–15. Hansen, M. C., R. S. DeFries, J. R. G. Townshend, R. Sohlberg, C. DiMiceli, and M. Car- roll (2002). Towards an operational MODIS continuous field of percent tree cover algorithm: examples using AVHRR and MODIS data. Remote Sensing of Environ- ment 83, 303–19. Hansen, M. C., S. V . Stehman, and P . V . Potapov (2010). Quantification of global gross forest cover loss. Proceedings of the National Academy of Sciences 107(19), 8650–5. 147 Hansen, M. C., S. V . Stehman, P . V . Potapov, B. Arunarwati, F. Stolle, and K. Pittman (2009). Quantifying changes in the rates of forest clearing in Indonesia from 1990 to 2005 using remotely sensed data sets. Environmental Research Letters 4(3), 034001. Hansen, M. C., S. V . Stehman, P . V . Potapov, T. R. Loveland, J. R. G. Townshend, R. S. Defries, K. W. Pittman, B. Arunarwati, F. Stolle, M. K. Steininger, M. Carroll, and C. Dimiceli (2008). Humid tropical forest clearing from 2000 to 2005 quantified by using multitemporal and multiresolution remotely sensed data. Proceedings of the National Academy of Sciences 105(27), 9439–44. Harris, N. L., S. Brown, S. C. Hagen, S. S. Saatchi, S. Petrova, W. Salas, M. C. Hansen, P . V . Potapov, and A. Lotsch (2012). Baseline map of carbon emissions from defor- estation in tropical regions. Science 336, 1573–6. Harris, N. L., S. Petrova, F. Stolle, and S. Brown (2008). Identifying optimal areas for REDD intervention: East Kalimantan, Indonesia as a case study. Environmental Research Letters 3(3), 035006. Harvey, C. A., O. Zerbock, S. Papageorgiou, and A. Parra (2010). What is needed to make REDD+ work on the ground? Lessons learned from pilot forest carbon initiatives. Arlington, VA: Conservation International. He, Y. and Y. Zhang (2005). Climate Change from 1960 to 2000 in the Langcang River Valley, China. Mountain Research and Development 25(4), 341–48. Heath, L. S., R. A. Birdsey, C. Row, and A. J. Plantinga (1996). Carbon pools and flux in U.S. forest projects. In M. J. Apps and D. T. Price (Eds.), Forest Ecosystems, Forest Management, and the Global Carbon Cycle, pp. 271–8. New York, NY: Springer-Verlag. Heistermann, M., C. M ¨ uller, and K. Ronneberger (2006). Land in sight? Achievements, deficits and potentials of continental to global scale land-use modeling. Agriculture, Ecosystems and Environment 114(2-4), 141–58. Herold, M., F. Achard, R. S. DeFries, D. L. Skole, S. Brown, and J. Townshend (2006). Report of the Workshop on Monitoroing Tropical Deforestation for Compensated Reductions. Technical report, Global Observation of Forest and Land Cover Dynam- ics (GOFC-GOLD) and the Global Terrestrial Observing System of the United Nations (GTOS), Jena, Germany. Hijmans, R. J., S. E. Cameron, J. L. Parra, P . G. Jones, and A. Jarvis (2005). Very high resolution interpolated climate surfaces for global land areas. International Journal of Climatology 25, 1965–78. Hong, B., K. E. Limburg, M. H. Hall, G. Mountrakis, P . M. Groffman, K. Hyde, L. Luo, V . R. Kelly, and S. J. Myers (2012). An integrated monitoring/modeling framework for assessing human-nature interactions in urbanizing watersheds: Wappinger and Onondaga Creek watersheds, New York, USA. Environmental Modelling and Soft- ware 32, 1–15. Houghton, R. A. (1996). Converting Terrestrial Ecosystems from Sources to Sinks of Carbon. Ambio 25(4), 267–72. 148 Houghton, R. A. (1999). The annual net flux of carbon to the atmosphere from changes in land use 1850-1990. Tellus B 51(2), 298–313. Houghton, R. A. (2005). Aboveground Forest Biomass and the Global Carbon Balance. Global Change Biology 11(6), 945–58. Houghton, R. A., R. D. Boone, J. M. Melillo, C. A. Palm, G. M. Woodwell, N. Myers, B. Moore III, and D. L. Skole (1985). Net flux of carbon dioxide from tropical forests in 1980. Nature 316, 617–20. Houghton, R. A., J. E. Hobbie, J. M. Melillo, B. Moore, B. J. Peterson, G. R. Shaver, and G. M. Woodwell (1983). Changes in the Carbon Content of Terrestrial Biota and Soils between 1860 and 1980: A Net Release of CO2 to the Atmosphere. Ecological Monographs 53(3), 235–62. Hu, H., W. Liu, and M. Cao (2008). Impact of land use and land cover changes on ecosystem services in Menglun, Xishuangbanna, Southwest China. Environmental Monitoring and Assessment 146(1-3), 147–56. Huettner, M., R. Leemans, K. Kok, and J. Ebeling (2009). A comparison of baseline methodologies for ‘Reducing Emissions from Deforestation and Degradation’. Carbon Balance and Management 4(4). ICLEI (2009). International Local Government GHG Emissions Analysis Protocol. Bonn, Germany: International Council for Local Environmental Initiatives (ICLEI) – Local Governments for Sustainability. ICLEI (2010). Local Government Operations Protocol for the quantification and reporting of greenhouse gas emissions inventories. Bonn, Germany: International Council for Local Environmental Initiatives (ICLEI) – Local Governments for Sustainability. ICRAF (2012). Reference Emission Levels (REL) in the context of REDD and land-based NAMAs: Forest transition stages can inform nested negotiations. Technical report, World Agroforestry Centre (ICRAF), Gigiri, Nairobi. IE (2008). Genesis Forest Project: Reducing Greenhouse Gas Emissions from Deforestation and Degradation in the State of Tocantins, Brazil – Project Design Document (PDD) for valida- tion at Climate, Community and Biodiversity Standard (CCBS). Palmas, Brazil: Ecologica Institute (IE). INBio, FONAFIFO, and FUNDECOR (2009). Programmatic project for the payment for environmental services to mitigate greenhouse gas emissions through avoided deforestation of privately owned tropical rainforests in high conservation value areas in the Central Volcanic Range of Costa Rica. Heredia, Costa Rica, San Jos´ e, Costa Rica, and Centroam´ erica, Costa Rica: National Biodiversity Institute (INBio), National Forestry Financing Fund (FONAFIFO), and Foundation for the Development of the Central Volcanic Range (FUNDECOR). Information Office of the State Council of the People’s Republic of China (2008). China’s Policies and Actions for Addressing Climate Change (1st ed.). Beijing, China: Foreign Languages Press. 149 IPCC (2000). Land Use, Land-Use Change and Forestry: A Special Report of the Intergovern- mental Panel on Climate Change. Cambridge, UK: Cambridge University Press. IPCC (2006). 2006 IPCC Guidelines for National Greenhouse Gas Inventories. Kanagawa, Japan: Institute for Global Environmental Strategies for the Intergovernmental Panel on Climate Change (IPCC). IPCC Working Group I (2007). Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, 2007. Cambridge, UK, and New York, NY: Cambridge University Press. IPCC Working Group III (2007). Contribution of Working Group III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, 2007. Cambridge, UK, and New York, NY: Cambridge University Press. Jasinski, E., D. Morton, R. S. DeFries, Y. Shimabukuro, L. Anderson, and M. C. Hansen (2005). Physical Landscape Correlates of the Expansion of Mechanized Agriculture in Mato Grosso, Brazil. Earth Interactions 9(16), 1–18. Johnson, D. L. and L. A. Lewis (2007). Land degradation (2nd ed.). Lanham, MD: Row- man and Littlefield Publishers Inc. Kim, O. S. (2010). An Assessment of Deforestation Models for Reducing Emissions from Deforestation and Forest Degradation (REDD). Transactions in GIS 14(5), 631–54. Kim, O. S. (2013). In Pursuit of Low Carbon Cities: Understanding Limitations of ICLEI’s International Local Government Greenhouse Gas Emissions Protocol. Journal of the Korean Geographical Society 48(1), 151–65. Laar, A. V . and A. Akc ¸a (2007). Tree height. In Forest Mensuration, pp. 47–56. Dordrecht, Netherlands: Springer. Lal, R. (2008). Carbon sequestration. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 363(1492), 815–30. Lambin, E. F. (1997). Modelling and monitoring land-cover change processes in tropical regions. Progress in Physical Geography 21(3), 375–93. Lambin, E. F. and P . Meyfroidt (2010). Global Forest Transition: Balance of Evidence for a Coming End to Deforestation. Annual Review of Environment and Resources 36(1), 343–71. Lambin, E. F. and P . Meyfroidt (2011). Global land use change, economic globalization, and the looming land scarcity. Proceedings of the National Academy of Sciences 108(9), 3465–72. Lawrence, W., S. Saatchi, R. S. Defries, J. Dietz, R. Rice, L. A. Dietz, M. S. D. Araujo, and K. Alger (1995). Utilization of SAR and Optical Remote Sensing Data for Habitat Conservation in the Tropical Forest of Brazil. In 1995 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp. 1480–2. Institute of Electrical and Electronics Engineers (IEEE). 150 LEAF (2011). Road to REDDiness – Making RELs Work. Bangkok, Thailand: Lowering Emissions in Asia’s Forests (LEAF). Lefsky, M. A. (2010). A global forest canopy height map from the Moderate Resolution Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophysical Research Letters 37(L15401), 1–5. Lepers, E., E. F. Lambin, A. C. Janetos, R. S. DeFries, F. Achard, N. Ramankutty, and R. J. Scholes (2005). A Synthesis of Information on Rapid Land-Cover Change for the Period 1981-2000. BioScience 55(2), 115–24. Li, H., T. M. Aide, Y. Ma, W. Liu, and M. Cao (2007). Demand for rubber is causing the loss of high diversity rain forest in SW China. Biodiversity and Conservation 16, 1731–45. Li, H., Y. Ma, T. M. Aide, and W. Liu (2008). Past, present and future land-use in Xishuangbanna, China and the implications for carbon dynamics. Forest Ecology and Management 255(1), 16–24. Li, H., Y. Ma, W. Liu, and W. Liu (2009). Clearance and fragmentation of tropical rain forest in Xishuangbanna, SW, China. Biodiversity Conservation 18, 3421–40. Li, Z. and J. M. Fox (2012). Mapping rubber tree growth in mainland Southeast Asia using time-series MODIS 250m NDVI and statistical data. Applied Geography 32(2), 420–32. Liu, W., H. Hu, Y. Ma, and H. Li (2006). Environmental and Socioeconomic Impacts of Increasing Rubber Plantations in Menglun Township, Southwest China. Mountain Research and Development 26(3), 245–53. Lubowski, R., A. Plantinga, and R. Stavins (2006). Land-use change and carbon sinks: Econometric estimation of the carbon sequestration supply function. Journal of Envi- ronmental Economics and Management 51(2), 135–52. Macedo, M. N., R. S. DeFries, D. C. Morton, C. M. Stickler, G. L. Galford, and Y. E. Shimabukuro (2012). Decoupling of deforestation and soy production in the southern Amazon during the late 2000s. Proceedings of the National Academy of Sciences 109(4), 1341–6. Malhi, Y., J. T. Roberts, R. A. Betts, T. J. Killeen, W. Li, and C. A. Nobre (2008). Climate change, deforestation, and the fate of the Amazon. Science 319, 169–72. Manomet Center for Conservation Sciences (2010). Massachusetts Biomass Sustainability and Carbon Policy Study: Report to the Commonwealth of Massachusetts Department of Energy Resources. Brunswick, Maine: Natural Capital Initiative Report NCI-2010-03. Mather, A. S. (1992). The Forest Transition. Area 24(4), 367–79. Mather, A. S. and C. L. Needle (1998). The forest transition: a theoretical basis. Area 30(2), 117–24. 151 McKechnie, J., S. Colombo, J. Chen, W. Mabee, and H. L. MacLean (2011). Forest bioen- ergy or forest carbon? Assessing trade-offs in greenhouse gas mitigation with wood- based fuels. Environmental Science and Technology 45(2), 789–95. METI and NASA (2011). Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Map (GDEM). Tokyo, Japan: Ministry of Economy, Trade, and Industry (METI) of Japan and the United States National Aeronautics and Space Administration (NASA). Meyfroidt, P . and E. F. Lambin (2009). Forest transition in Vietnam and displacement of deforestation abroad. Proceedings of the National Academy of Sciences 106(38), 16139–44. Meyfroidt, P . and E. F. Lambin (2010). Forest transition in Vietnam and Bhutan: Causes and environmental impacts. In H. Nagendra and J. Southworth (Eds.), Reforesting Landscapes: Linking Pattern and Process, Landscape Series 10, Volume 10 of Landscape Series, Chapter 14, pp. 315–39. Dordrecht, Netherlands: Springer Netherlands. Miles, L., A. C. Newton, R. S. DeFries, C. Ravilious, I. May, S. Blyth, V . Kapos, and J. E. Gordon (2006). A global overview of the conservation status of tropical dry forests. Journal of Biogeography 33(3), 491–505. Miller, S. A. (2010). Minimizing land use and nitrogen intensity of bioenergy. Environ- mental Science and Technology 44(10), 3932–9. Miller, S. a., A. E. Landis, and T. L. Theis (2007). Environmental trade-offs of biobased production. Environmental Science and Technology 41(15), 5176–82. Mitasova, H., J. Hofierka, M. Zlocha, and L. R. Iverson (1996). Modelling topographic potential for erosion and deposition using GIS. International Journal of Geographical Information Systems 10(5), 629–41. Mollicone, D., A. Freibauer, E. D. Schulze, S. Braatz, G. Grassi, and S. Federici (2007). Elements for the expected mechanisms on ‘reduced emissions from deforestation and degradation, REDD’ under UNFCCC. Environmental Research Letters 2(045024), 1–7. Morton, D. C., R. S. Defries, J. T. Randerson, L. Giglio, W. Schroeder, and G. R. Van Der Werf (2008). Agricultural intensification increases deforestation fire activity in Amazonia. Global Change Biology 14(10), 2262–75. Morton, D. C., R. S. DeFries, Y. E. Shimabukuro, L. O. Anderson, F. Del Bon Esp´ ırito- Santo, M. C. Hansen, and M. Carroll (2005). Rapid Assessment of Annual Deforesta- tion in the Brazilian Amazon Using MODIS Data. Earth Interactions 9(8), 1–22. Moutinho, P . and S. Schwartzman (Eds.) (2005). Tropical Deforestation and Climate Change. Bel´ em, Brazil and Washington DC: Amazon Environmental Research Institue (IPAM) and Environmental Defense Fund. Murray, G. and I. G. Cook (2004). The Greening of China. Beijing, China: China Intercon- tinental Press. Myers, N., R. A. Mittermeier, C. G. Mittermeier, G. A. da Fonseca, and J. Kent (2000). Biodiversity hotspots for conservation priorities. Nature 403(6772), 853–8. 152 Newell, J. P . and R. O. Vos (2011). “Papering” Over Space and Place: Product Carbon Footprint Modeling in the Global Paper Industry. Annals of the Association of American Geographers 101(4), 1–12. Newell, J. P . and R. O. Vos (2012). Accounting for forest carbon pool dynamics in prod- uct carbon footprints: Challenges and opportunities. Environmental Impact Assessment Review 37, 23–36. Olander, L. P ., H. K. Gibbs, M. K. Steininger, J. J. Swenson, and B. C. Murray (2008). Reference scenarios for deforestation and forest degradation in support of REDD: A review of data and methods. Environmental Research Letters 3(025011), 1–12. ORNL (2008). LandScan. Oak Ridge, TN: Oak Ridge National Laboratory (ORNL), University of Tennessee, and Battelle (UT-Battelle). Paladino, L. and R. G. Pontius Jr. (2004). Accuracy Assessment and Uncertainty in Baseline Projections for Land-Use Change Forestry Projects. Pebesma, E. and R. Bivand (2012). Package ‘sp’. Vienna, Austria: The R Project for Statistical Computing. Penman, J., M. Gytarsky, T. Hiraishi, T. Krug, D. Kruger, R. Pipatti, L. Buendia, K. Miwa, T. Ngara, K. Tanabe, and F. Wagner (2003). Definitions and Methodological Options to Inventory Emissions from Direct Human-Induced Degradation of Forests and Deveg- etation of Other Vegetation Types. Technical report, Institute for Global Environ- mental Strategies (IGES) for the Intergovernmental Panel on Climate Change (IPCC), Kanagawa, Japan. Perez-Garcia, J., B. Lippke, D. Briggs, J. B. Wilson, J. Bowyer, and J. Meil (2006). The environmental performance of renewable building materials in the context of resi- dential construction. Wood Science and Technology 37(Corrim Special Issue), 3–17. Perez-Garcia, J., B. Lippke, J. Comnick, and C. Manriquez (2005). An assessment of car- bon pools, storage, and wood products market substitution using life-cycle analysis results. Wood and Fiber Science 37(Corrim Special Issue), 140–8. Peterson, A. T., M. A. Ortega-Huerta, J. Bartley, V . S´ anchez-Cordero, J. Sober´ on, R. H. Buddemeier, and D. R. B. Stockwell (2002). Future projections for Mexican faunas under global climate change scenarios. Nature 416(6881), 626–9. Peterson, A. T., M. Pape, and J. Sober´ on (2008). Rethinking receiver operating charac- teristic analysis applications in ecological niche modeling. Ecological Modelling 213(1), 63–72. Phelps, J., M. C. Guerrero, D. A. Dalabajan, B. Young, and E. L. Webb (2010). What makes a ‘REDD’ country? Global Environmental Change 20(2), 322–32. Piao, S., J. Fang, P . Ciais, P . Peylin, Y. Huang, S. Sitch, and T. Wang (2009). The carbon balance of terrestrial ecosystems in China. Nature 458, 1009–13. Polsky, C. (2004). Putting Space and Time in Ricardian Climate Change Impact Studies: Agriculture in the U.S. Great Plains, 1969–1992. Annals of the Association of American Geographers 94(3), 549–64. 153 Pongratz, J., L. Bounoua, R. S. DeFries, D. C. Morton, L. O. Anderson, W. Mauser, and C. A. Klink (2006). The Impact of Land Cover Change on Surface Energy and Water Balance in Mato Grosso, Brazil. Earth Interactions 10(19), 1–17. Pontius Jr., R. G. (2000). Quantification error versus location error in comparison of categroical maps. Photogrammetric Engineering and Remote Sensing 66(8), 1011–6. Pontius Jr., R. G. (2001). Modeling the spatial pattern of land-use change with GEO- MOD2: application and validation for Costa Rica. Agriculture, Ecosystems and Envi- ronment 85(1-3), 191–203. Pontius Jr., R. G., W. Boersma, J.-C. Castella, K. Clarke, T. Nijs, C. Dietzel, Z. Duan, E. Fotsing, N. Goldstein, K. Kok, E. Koomen, C. D. Lippitt, W. McConnell, A. Mohd Sood, B. Pijanowski, S. Pithadia, S. Sweeney, T. N. Trung, A. T. Veldkamp, and P . H. Verburg (2008). Comparing the input, output, and validation maps for several models of land change. Annals of Regional Science 42, 11–37. Pontius Jr., R. G. and H. Chen (2006). GEOMOD Modeling. Worcester, MA: Clark Labs, Clark University. Pontius Jr., R. G. and M. Millones (2011). Death to Kappa: Birth of quantity disagree- ment and allocation disagreement for accuracy assessment. International Journal of Remote Sensing 32(15), 4407–29. Pontius Jr., R. G. and L. C. Schneider (2001). Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA. Agriculture, Ecosystems and Environment 85, 239–48. Pontius Jr., R. G. and J. Spencer (2005). Uncertainty in extrapolations of predictive land- change models. Environment and Planning B: Planning and Design 32(2), 211–30. Qiu, J. (2009, January). Where the rubber meets the garden. Nature 457(7227), 246–7. Ramankutty, N., H. K. Gibbs, F. Achard, R. S. DeFries, J. A. Foley, and R. A. Houghton (2007). Challenges to estimating carbon emissions from tropical deforestation. Global Change Biology 13(1), 51–66. REDD Forests (2009). Reducing carbon emissions by protecting a native forest in Tasmania – Project Design Document (PDD) for validation at Climate, Community and Biodiversity Alliance (CCBA). St. Andrews, Australia: REDD Forests. RFF and CIFOR (2003). China’s Forest: Global lessons from market reforms. Washington, DC: Resources For The Future (RFF) and Center For International Forestry Research (CIFOR). Rindfuss, R. R., S. J. Walsh, B. L. Turner Jr., J. Fox, and V . Mishra (2004). Developing a science of land change: Challenges and methodological issues. Proceedings of the National Academy of Sciences 101(39), 13976–81. Robbins, P . (2001). Fixed categories in a portable landscape: the causes and conse- quences of land-cover categorization. Environment and Planning A 33(1), 161–79. 154 Robbins, P . (2004). Political Ecology: A Critical Introduction. Oxford, UK: Blackwell Pub- lishing Ltd. Robbins, P . and T. Maddock (2000). Interrogating land cover categories: Metaphor and method in remote sensing. Cartography and Geographic Information Science 27(4), 295– 309. Robin, X., N. Turck, A. Hainard, N. Tiberti, F. Lisacek, J.-C. Sanchez, and M. M ¨ uller (2011). pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 12(1), 77. Rudel, T. K., O. T. Coomes, E. Moran, F. Achard, A. Angelsen, J. Xu, and E. Lambin (2005). Forest transitions: Towards a global understanding of land use change. Global Environmental Change 15(1), 23–31. Rudel, T. K., R. Defries, G. P . Asner, and W. F. Laurancee (2009). Changing Drivers of Deforestation and New Opportunities for Conservation. Conservation Biology 23(6), 1396–1405. Saatchi, S. S., N. L. Harris, S. Brown, M. Lefsky, E. T. A. Mitchard, W. Salas, B. R. Zutta, W. Buermann, S. L. Lewis, S. Hagen, S. Petrova, L. White, M. Silman, and A. Morel (2011). Benchmark map of forest carbon stocks in tropical regions across three conti- nents. Proceedings of the National Academy of Sciences 108(24), 9899–904. Saatchi, S. S., R. A. Houghton, R. C. Dos Santos Alval´ a, J. V . Soares, and Y. Yu (2007). Distribution of aboveground live biomass in the Amazon basin. Global Change Biol- ogy 13(4), 816–37. Sandker, M., S. K. Nyame, J. F¨ orster, N. Collier, G. Shepherd, D. Yeboah, D. E.-d. Blas, M. Machwitz, S. Vaatainen, E. Garedew, G. Etoga, C. Ehringhaus, J. Anati, O. D. K. Quarm, and B. M. Campbell (2010). REDD payments as incentive for reducing forest loss. Conservation Letters 3(2), 114–21. Sangermano, F., J. Toledano, and J. R. Eastman (2012). Land cover change in the Boli- vian Amazon and its implications for REDD+ and endemic biodiversity. Landscape Ecology 27(4), 571–84. Sathaye, J. A. and K. Andrasko (2006). Land use change and forestry climate project regional baselines: A review. Mitigation and Adaptation Strategies for Global Change 12(6), 971–1000. Sathaye, J. A. and K. Andrasko (2007). Special issue on estimation of baselines and leakage in carbon mitigation forestry projects. Mitigation and Adaptation Strategies for Global Change 12, 963–70. SCS (2011). The Rimba Raya Biodiversity Reserve REDD Project – Project Design Document (PDD) submitted fro validation according to CCB Standards of the Climate, Community and Biodiversity Alliance. Emeryville, CA: Scientific Certification Systems (SCS). Sehgal, R. N. M., W. Buermann, R. J. Harrigan, C. Bonneaud, C. Loiseau, A. Chasar, I. Sepil, G. Valkinas, T. Iezhova, S. Saatchi, and T. B. Smith (2011). Spatially explicit 155 predictions of blood parasites in a widely distributed African rainforest bird. Proceed- ings of the Royal Society. Series B, Biological Sciences 278(1708), 1025–33. Shanley, P . and L. Luz (2003). The Impacts of Forest Degradation on Medicinal Plant Use and Implications for Health Care in Eastern Amazonia. BioScience 53(6), 573–84. Skole, D. L., J. H. Samek, and M. J. Smalligan (2011). Implications of allometry. Proceed- ings of the National Academy of Sciences 108(4), E12. Sloan, S. and J. Pelletier (2012). How accurately may we project tropical forest-cover change? A validation of a forward-looking forest baseline for reduced emissions from deforestation and degradation. Global Environmental Change 22(2), 440–53. The Climate Registry (2008). General Reporting Protocol for the Voluntary Reporting Pro- gram, Version 1.1. Los Angeles, CA: The Climate Registry. The Economist (2010). A special report on forests. New York, NY: The Economist. Turner Jr., B. L., E. F. Lambin, and A. Reenberg (2007). The emergence of land change science for global environmental change and sustainability. Proceedings of the National Academy of Sciences 104(52), 20666–71. Turner Jr., B. L. and P . Robbins (2008). Land-Change Science and Political Ecology: Similarities, Differences, and Implications for Sustainability Science. Annual Review of Environment and Resources 33(1), 295–316. UN-REDD (2013). Measurement Reporting and Verification. WWW Document: UN-REDD Programme (www.un-redd.org). UNFCCC (1998). Kyoto Protocol to the United Nations Framework Convention on Climate Change. Kyoto, Japan: United Nations Framework Convention on Climate Change (UNFCCC). UNFCCC COP (2011). Report of the Conferences of the Parties on its sixteenth session, held in Cancun from 29 November to 10 December 2010. van der Werf, G. R., D. C. Morton, R. S. DeFries, J. G. J. Olivier, P . S. Kasibhatla, R. B. Jackson, G. J. Collatz, and J. T. Randerson (2009). CO2 emissions from forest loss. Nature Geoscience 2, 737–8. VCS (2010a). Approved VCS Methodology VM0006 Version 1.0 Methodology for Carbon Accounting in Project Activities that Reduce Emissions from Mosaic Deforestation and Degradation Sectoral Scope 14. Washington, DC: Verified Carbon Standard (VCS). VCS (2010b). Approved VCS Methodology VM0009 Version 1.0 Methodology for Avoided Mosaic Deforestation of Tropical Forests Sectoral Scope 14. Washington, DC: Verified Car- bon Standard (VCS). VCS (2010c). Approved VCS Module VMD0006 Version 1.0 REDD Methodological Mod- ule: Estimation of baseline carbon stock changes and greenhouse gas emissions from planned deforestation (BL-PL) Sectoral Scope 14. Washington, DC: Voluntary Carbon Standard (VCS). 156 VCS (2010d). Approved VCS Module VMD0007 Version 1.0 REDD Methodological Module: Estimation of baseline carbon stock changes and greenhouse gas emissions from unplanned deforestation (BL-UP) Sectoral Scope 14. Washington, DC: Verified Carbon Standard (VCS). VCS (2011). Approved VCS Module VMD0007 Version 2.0 REDD Methodological Module: Estimation of baseline carbon stock changes and greenhouse gas emissions from unplanned deforestation (BL-UP) Sectoral Scope 14. Washington, DC: Verified Carbon Standard (VCS). VCS (2012a). Approved VCS Methodology VM0009 Version 2.1 Methodology for Avoided Deforestation Sectoral Scope 14. Washington, DC: Verified Carbon Standard (VCS). VCS (2012b). Approved VCS Methodology VM0015 Methodology for Avoided Unplanned Deforestation Sectoral Scope 15. Washington, DC: Verified Carbon Standard (VCS). VCS (2012c). VCS Module VMD0007 REDD Methodological Module: Estimation of baseline carbon stock changes and greenhouse gas emissions from unplanned deforestation (BL-UP). Washington, DC: Verified Carbon Standard (VCS). Veldkamp, T. A. and E. F. Lambin (2001). Predicting land-use change. Agriculture, Ecosystems and Environment 85(1-3), 1–6. Venter, O. and L. P . Koh (2011). Reducing emissions from deforestation and forest degra- dation (REDD+): game changer or just another quick fix? Annals of the New York Academy of Sciences, 1–14. WBCSD and WRI (2005). The Greenhouse Gas Protocol: The GHG Protocol for Project Accounting. Conches-Geneva, Switzerland, and Washington, DC: World Busi- ness Council for Sustainable Development (WBCSD) and World Resources Institute (WRI). White, M. K., S. T. Gower, and D. E. Ahl (2005). Life cycle inventories of roundwood production in northern Wisconsin: Inputs into an industrial forest carbon budget. Forest Ecology and Management 219(1), 13–28. Wood, C. H. and D. L. Skole (1998). Linking satellite, census, and survey data to study deforestation in the Brazilian Amazon. In D. Liverman, E. F. Moran, R. R. Rindfuss, and P . C. Stern (Eds.), People and pixels: linking remote sensing and social science, Chap- ter 4, pp. 70–93. Washington, DC: National Academy Press. World Bank BioCarbon Fund (2008). Methodology for Estimating Reductions of GHG Emis- sions from Mosaic Deforestation. Washington, DC: World Bank. WRI (2006). The Greenhouse Gas Protocol: The Land Use, Land-Use Change, and Forestry Guidance for GHG Project Accounting. Washington, DC: World Resources Institute (WRI). WRI and WBCSD (2011). Greenhouse Gas Protocol: Product Life Cycle Accounting and Reporting Standard. Washington, DC, and Conches-Geneva, Switzerland: World Resources Institute (WRI) and World Business Council for Sustainable Development (WBCSD). 157 Xi, J. (2009). Valuation of Ecosystem Services in Xishuangbanna Biodiversity Conservation Corridors Initiative Pilot Site, China. Mandaluyong City, Philippines: Asian Develop- ment Bank. Xu, J. (2011). China’s new forests aren’t as green as they seem. Nature 477, 371. Xu, J., J. Fox, J. B. Vogler, P . Zhang, Y. Fu, L. Yang, J. Qian, and S. Leisz (2005). Land- use and land-cover change and farmer vulnerability in Xishuangbanna prefecture in southwestern China. Environmental Management 36(3), 404–13. Xu, J., E. T. Ma, D. Tashi, Y. Fu, Z. Lu, and D. Melick (2006). Integrating Sacred Knowl- edge for Conservation: Cultures and Landscapes in Southwest China. Ecology and Society 10(2), 1–25. Xu, J., Y. Yang, J. Fox, and X. Yang (2007). Forest transition, its causes and environ- mental consequences: Empirical evidence from Yunnan of Southwest China. Tropical Ecology 48(2), 137–50. Yi, Z., C. H. Cannon, J. Chen, C. Ye, and R. D. Swetnam (2013). Developing indicators of economic value and biodiversity loss for rubber plantations in Xishuangbanna, southwest China: A case study from Menglun township. Ecological Indicators. Zaks, D. P . M., C. C. Barford, N. Ramankutty, and J. A. Foley (2009). Producer and con- sumer responsibility for greenhouse gas emissions from agricultural production—a perspective from the Brazilian Amazon. Environmental Research Letters 4(4), 044010. Zhang, J. and M. Cao (1995). Tropical forest vegetation of Xishuangbanna, SW China and its secondary changes, with special reference to some problems in local nature conservation. Biological Conservation 73(3), 229–38. Zhang, P ., J. Xu, M. Wang, and X. Deng (2006). Spatial and temproal dynamics of rubber plantations and its impacts on tropical forest in Xishuangbanna. Remote Sensing for Land and Resources 3, 51–5. Zhang, Q., M. T. Bennett, K. Kannan, and L. Jin (2010). Payments for Ecological Services and Eco-Compensation: Practices and Innovations in the People’s Republic of China. Man- daluyong City, Philippines: Asian Development Bank. Zhu, H., Z. F. Xu, H. Wang, and B. G. Li (2004). Tropical rain forest fragmentation and its ecological and species diversity changes in southern Yunnan. Biodiversity and Conservation 13(7), 1355–72. Zomer, R. J., A. Trabucco, R. Coe, and F. Place (2009). Trees on Farm: Analysis of Global Extent and Geographical Patterns of Agroforestry. ICRAF Working Paper. Nairobi, Kenya: World Agroforestry Centre (ICRAF). 158

Asset Metadata

Creator Kim, Oh Seok (author)

Core Title A spatially explicit approach to measuring carbon dynamics for reducing emissions from deforestation and forest degradation: a case study of Chinese forests

Contributor Electronically uploaded by the author (provenance)

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Geography

Publication Date 07/18/2015

Defense Date 04/08/2013

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag carbon emission,China,climate change,Deforestation,land change,oai:digitallibrary.usc.edu:usctheses,OAI-PMH Harvest,REDD,Rubber,spatial regression,Tropics

Language English

Advisor McKenzie, Roderick C. (committee chair), Newell, Joshua P. (committee chair), Curtis, Andrew J. (committee member), Heikkila, Eric J. (committee member), Nugent, Jeffrey B. (committee member)

Creator Email ohseok.kim@gmail.com,oskim@korea.ac.kr

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-292295

Unique identifier UC11287810

Identifier etd-KimOhSeok-1792.pdf (filename),usctheses-c3-292295 (legacy record id)

Legacy Identifier etd-KimOhSeok-1792.pdf

Dmrecord 292295

Document Type Dissertation

Rights Kim, Oh Seok

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Abstract (if available)

Abstract This dissertation proposes a Geographic Emission Benchmark (GEB) for reducing emissions from deforestation and forest degradation (REDD+). The GEB is designed to function as a "baseline" to indicate reference (carbon) emission levels as business-as-usual if REDD+ does not take place. The GEB is also a new land change model that produces prospective and spatially explicit outcomes. Unlike other land change models forecasting business-as-usual deforestation, the GEB internally (1) characterizes "forest" in a general sense based on remotely sensed data and (2) identifies "deforestation hotspots" using a spatial clustering technique in order to transparently estimate regional rates of deforestation. As a byproduct, it also portrays a historic trend of forest degradation. In terms of predictive accuracy, the GEB produces a more accurate projection than GEOMOD, measured by Figure of Merit, at least for this case study of Chinese tropics. Thus, the research concludes the GEB is a reasonably accurate baseline for current and future reference emission levels. If other land change models are to be used to set a prospective and spatially explicit reference emission level for a particular region, their predictive accuracies must outperform that of the regional GEB. Finally, this dissertation demonstrates an application of spatial regression attempting to identify drivers of deforestation in the study area. The result shows that biophysical factors have more influence on the regional deforestation than other factors such as population or infrastructure. Thus, when forecasting business-as-usual forest loss for a future REDD+ project, the biophysical drivers must be considered more seriously than other drivers of deforestation in the Chinese tropics.