Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i
(USC Thesis Other)
Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Using Maxent to Model the Distribution of Prehistoric Agricultural Features in a Portion of the Hōkūli‘a Subdivision in Kona, Hawai‘i by Solomon Ha‘aheo Kailihiwa, III A Thesis Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree MASTER OF SCIENCE (GEOGRAPHIC INFORMATION SCIENCE AND TECHNOLOGY) August 2015 Copyright 2015 Solomon Ha‘aheo Kailihiwa, III ii DEDICATION For Juliana, my loving and supporting wife, and Savia, my reason for enduring. iii ACKNOWLEDGMENTS This endeavor would not have been possible without a number of individuals that it has been my honor to work with and to know. Dr. Patrick V. Kirch introduced me to the world of Hawaiian archaeology, without Uncle Pat, I would not be where I am today. Dr. Paul L. Cleghorn provided me the opportunity to live and to do archaeology in Hawai‘i, he never gave up hope that I would someday attend graduate school. Dr. Alan E. Haun provided the opportunity and support for me to work on this project. Thank you also goes to Dr. Thomas Garrison and Dr. Su Jin Lee for being on my committee and offering valuable input and advice. Dr. Karen Kemp provided invaluable guidance throughout the thesis process. To my parents, Solomon and Audreen Kailihiwa, for building the foundation of my successes. Thank you to Joseph and Carol Vivona whose support helped to keep the torch alive. Last, and definitely not least, immeasurable gratitude to Juliana Kailihiwa for her support and for putting up with me during this research. iv TABLE OF CONTENTS DEDICATION .............................................................................................................................. ii ACKNOWLEDGMENTS ........................................................................................................... iii LIST OF TABLES .........................................................................................................................v LIST OF FIGURES ..................................................................................................................... vi LIST OF ABBREVIATIONS ..................................................................................................... vii ABSTRACT .............................................................................................................................. viii CHAPTER 1: INTRODUCTION ..................................................................................................1 1.1 Motivation ............................................................................................................................2 1.2 Research Hypotheses ...........................................................................................................4 CHAPTER 2: BACKGROUND AND LITERATURE REVIEW .................................................7 2.1 The Kona Field System ........................................................................................................7 2.2 Habitat Suitability Modeling ................................................................................................8 2.3 Ecological Models and Archaeology ...................................................................................9 2.4 Maxent for Modeling Archaeological Site Probability....................................................... 11 CHAPTER 3: METHODOLOGY ............................................................................................... 14 3.1 Study Area ......................................................................................................................... 14 3.2 Agricultural Feature Presence Data .................................................................................... 16 3.2.1 Preparation of Prehistoric Agricultural Presence Data for Maxent 17 3.3 Environmental Variable Data ............................................................................................. 18 3.3.1 Rainfall data 19 3.3.2 Elevation Data 21 3.4 Maxent Parameters............................................................................................................. 23 3.4.1 Computational Features 24 3.4.2 Environmental Variable Assessment in Maxent 26 3.4.4 Maxent Settings 26 3.4.5 Model Validation 29 CHAPTER 4: RESULTS ............................................................................................................. 30 4.1 Maxent Probability Distribution......................................................................................... 30 4.2 Environmental Variable Performance ................................................................................ 31 4.2.1 Response Curves 33 4.2.2 Environmental Variable Contribution and Importance to the Maxent Model 35 4.3 Model Performance ............................................................................................................ 36 CHAPTER 5: DISCUSSION AND CONCLUSIONS ................................................................. 38 5.1 Model Evaluation ............................................................................................................... 38 5.1.1 Archaeological Monitoring and Future Model Evaluation 44 5.2 Conclusions ........................................................................................................................ 44 References ................................................................................................................................... 47 v LIST OF TABLES Table 1 List of Agricultural feature types, and the number of each type observed within the AOI .................................................................................................................................. 17 Table 2 Summary of Environmental Variables. ........................................................................... 19 Table 3 Summary of Maxent Calculation Features and a description of the constraints they impose upon the output model. ........................................................................................ 25 Table .4 Modeling Parameters ..................................................................................................... 27 Table 5 Final Model Parameters .................................................................................................. 28 Table 6 Minimum, maximum, and mean values of the unaltered output model and the smoothed output model .................................................................................................... 31 Table 7 Summary of Environmental Variables and Their Importance ........................................ 35 vi LIST OF FIGURES Figure 1. Research Area of Interest. ...............................................................................................3 Figure 2. Examples of clearing features typically found in the kula and kaluulu zones of the Kona Field System. ............................................................................................................3 Figure 3. Research AOI shown within the elevation zones of the Kona Field System ..................8 Figure 4 Conceptual Model ......................................................................................................... 13 Figure 5 Densely vegetated and mechanically altered areas in and around the AOI .................... 15 Figure 6 AOI and Soil Types ....................................................................................................... 15 Figure 7 Archaeological Feature Map .......................................................................................... 16 Figure 8 Comparison of Raw Rainfall Data (top) and Smoothed Rainfall Data (bottom) ............ 22 Figure 9 Clipped Elevation Raster ............................................................................................... 22 Figure 10 Aspect Raster ............................................................................................................... 23 Figure 11 Final Maxent output ..................................................................................................... 32 Figure 12 Maxent output response curves. ................................................................................... 34 Figure 13 Sample 'Identify' results for a Maxent model pixel ...................................................... 34 Figure 14 Agricultural Clearing Feature Probability Distribution With Elevation Overlay ......... 39 Figure 15 Elevation Response Curve Showing the Average Values and Standard Deviation from the 100 Maxent Replicate Models ........................................................................... 40 Figure 16 Agricultural Clearing Feature Probability Distribution with Summer Rain Overlay (top) and Annual Rainfall Overlay (bottom) .................................................................... 41 Figure 17 Elevation Response Curve of Clearing Feature Probability to Slope Variable ............ 42 vii LIST OF ABBREVIATIONS AIS Archaeological inventory survey AOI Area of Interest ASCII American Standard Code for Information Interchange AUC Area under the receiver operator curve CPU Central Processing Unit DLNR Department of Land and Natural Resources .dxf Drawing Exchange Format ENFA Ecological niche factor analysis GARP Genetic Algorithm for Rule-Set Prediction GPS Global Positioning System GUI Graphical user interface HTML Hypertext markup language NAD83 North American Datum 1983 PRISM Parameter-elevation Regressions on Independent Slopes Model ROC Receiver operator characteristics ROR Relative occurrence rate SHPD State Historic Preservation Division UTM Universal Transverse Mercator WGS1984 World Geodetic System 1984 viii ABSTRACT Archaeological investigations are an integral part of the permitting process for land development in Hawai‘i. The State recognizes that conservation of its historic and cultural heritage is important and that its cultural resources are nonrenewable. A recent archaeological survey for prehistoric and historic agricultural features in the Hōkūli‘a luxury development on Hawai‘i Island afforded an opportunity to test maximum entropy theory as a method for predicting the presence of archaeological features. The Maxent computer program uses a machine learning algorithm that utilizes presence point data and environmental variable rasters to produce a probability distribution of the species of interest. The "species" of interest for this research were agricultural clearing features, associated with sweetpotato (Ipomea batatas) cultivation, identified in a portion of Hōkūli‘a which lies within the Kona Field System. Previous agricultural habitat suitability models for Hawai‘i were used to determine the environmental variables used in this research; the variables included annual rainfall, summer rainfall, elevation, and slope. Maxent produced a probability distribution that matched the expectations of the conceptual model. The model was validated using diagnostic tools included in the Maxent program (1) area under the receiver operator characteristic curve analysis, (2) jack-knife testing, and (3) environmental variable response curve analysis, as well as three research hypotheses. The model does not account for human behavior and may overestimate feature presence in uncultivated, spiritually important areas that are suitable for farming. The results of this research show that Maxent can be used to successfully model certain types archaeological features. 1 CHAPTER 1: INTRODUCTION Archaeological inventory surveys are the primary method used to identify historic and prehistoric resources throughout the State of Hawai‘i on lands that are slated for development. This research focuses on an archaeological investigation conducted within Hōkūli‘a, a luxury subdivision that spans 10 ahupua‘a (traditional land divisions) located along the western shore of Hawai‘i Island in the North Kona and South Kona Districts within an area known as the Kona Field System, an area of intense prehistoric agricultural production (Allen 2001a; Allen 2004). Archaeological investigations have been conducted within Hōkūli‘a beginning in 1930 and resuming in 1988 through the present (Tomonari-Tuggle and Tuggle 2008). An archaeological survey of prehistoric agricultural features undertaken in 2014-15, requested by the Hawai‘i State Historic Preservation Division (SHPD) of the Department of Land and Natural Resources (DLNR), in the Hōkūli‘a subdivision, provided an opportunity to test the use of the maximum entropy model as a means for creating a probability distribution of archaeological features in areas where it is not possible to conduct a comprehensive survey. The area of interest (AOI) is an approximate 312-acre portion of the Hōkūli‘a Subdivision. Figure 1 shows the AOI and its location within the larger extent of the Kona Field System. The archaeological survey crew was unable to record agricultural features in portions of the AOI covered by dense, impenetrable vegetation and areas altered by heavy machinery. Stands of dense guinea grass (Megathyrsus maximus) are located throughout the property obscuring the ground surface and impeding pedestrian survey. Much of the area has also been mechanically altered by heavy equipment and prehistoric agricultural features have been removed from the landscape. The objective of this research is to use Maxent to model a probability distribution of agricultural clearing features. 2 1.1 Motivation The need to provide statistical support for conclusions derived from the data collected during the ongoing fieldwork motivated this research. Maxent has been used successfully in species distribution modeling (Merow, Smith, and Silander 2013) and is able to create reliable models with a limited number of presence-only samples (Phillips, Anderson, and Schapire 2005; Elith et al. 2011; Merow, Smith, and Silander 2013). The Maxent modeling program was used to produce a probability distribution of prehistoric agricultural features in mechanically disturbed and densely vegetated areas within the AOI based on the observed presence data and environmental variables. Habitat suitability modeling and maximum entropy theory were used to model the probability distribution of the agricultural features. Habitat suitability modeling considers the relationship between environmental variables and the probability that a given species will occur (Hirzel and Le Lay 2008). Maximum entropy uses only observed occurrences of a species to create a probability distribution (Roxburgh and Mokany 2010). The Maxent program developed by Phillips, Dudík, and Schapire (2004) creates a probability distribution using presence-only data and environmental variables. The “species” of interest for this study were the agricultural clearing features, rock mounds and modified outcrops, most commonly found within the kula and kaluulu zones of the Kona Field System. Figure 2 shows examples of the clearing features typically found within the AOI. The Kona Field System is characterized by four distinct planting zones defined by elevation (Newman 1971; Kelly 1983): kula (coast— 500 ft), kaluulu (500—1,000 ft), ‘āpa‘a (1,000—2,500 ft), and ‘ama‘u (2,000—3,000 ft). The habitat for the agricultural features was defined by the environmental variables discussed in previous archaeological reports and historical accounts: elevation, slope, and rainfall (Newman 1971; Kelly 1983; Allen 2001b). 3 Figure 1. Research Area of Interest. An estimated extent of the Kona Field System is shown in brown in the inset map. Figure 2. Examples of clearing features typically found in the kula and kaluulu zones of the Kona Field System. A modified outcrop is shown on the left, a rock mound on the right. 4 The general conceptual model for this research is that the density of the rock mounds and modified outcrops will be relatively sparse along the coastal area and feature density will increase with the rise in elevation up to the ‘āpa‘a zone where the Kona Field System is a more formal collection of walled fields (Allen 2001b). The validity of this model was determined through statistical tests that are included as part of the Maxent program and ground truthing. These tests include binomial tests of omission and the area under the receiver operator curve value (AUC) analysis, which were used to determine the goodness of fit for the Maxent outputs. Jack-knife testing, which assesses the contribution of each environmental variable to the model, and response curves, which show how the model responds to the environmental variables, are also available within Maxent. These tests and three research hypotheses were used to evaluate the model produced by Maxent. 1.2 Research Hypotheses Three research hypotheses were developed to evaluate the Maxent output model. The first hypothesis states that clearing feature (rock mounds and modified outcrops) density increases as one ascends from the shoreline to the upper elevations of the kula zone and the lower elevation of the kaluulu zone. Clearing feature density should decrease starting in the mid- elevations of the kaluulu zone. This hypothesis is based on the results of an archaeological inventory survey (AIS) conducted approximately three kilometers to the north of the current AOI by Haun and Henry (2010). Their findings indicated that the feature density was greatest between the elevations of 360 ft and 680 ft which corresponds to the upper elevations of the kula zone and the lower elevations of the kaluulu zone. Clearing feature density began to decrease at elevations higher 5 than 680 ft and dropped even more at the upper elevations of the kaluulu zone as agricultural features and methods become more formalized in the āpa‘a zone. The second hypothesis states that as precipitation levels approach acceptable levels for dryland taro farming, there is a decrease in the density of informal clearing features that are used in sweetpotato 1 cultivation. Sweetpotato requires much less water than taro (Colocasia esculenta) which is why the sweetpotato was cultivated so intensely in the dry lowlands of the Kona Field System (Handy, Handy, and Pukui 1991). There should be a decrease in the informal agricultural features as precipitation increases to levels where formal, rain-fed, dryland agriculture is sustainable. Ladefoged et al. (2009) stated that the annual precipitation required for optimal sweetpotato cultivation was 750—1000 mm. The third hypothesis states that agricultural practices were limited to slopes less than 35 percent; most agriculture took place on slopes less than 20 percent. Newman (1971) conducted an analysis of the environmental characteristics of the agricultural areas on Hawai‘i Island that were observed by Reverend William Ellis during his tour of the island in 1823. Newman created a number of generalizations of the environmental variables which included annual mean rainfall, elevation, slope, drainage, soil parent material, and soil type. Newman generalized that the agricultural zones for Hawai‘i Island incorporated areas that had a slope less than 35 percent with the majority having a slope less than 20 percent. The following chapters describe the work and findings of this research. Chapter 2 provides background on habitat suitability modeling, the Maxent modeling program, and ecological models used in an archaeological context. Chapter 3 discusses the preparation of the presence data and environmental variable rasters, and the parameters used in Maxent. Chapter 4 1 Valenzuela, Fukuda and Arakaki (2000) explicitly state, "Sweetpotato is, in fact, one word..." 6 discusses the model, its response to the environmental variables, and the methods used to evaluate the model. Chapter 5 discusses the conclusions drawn from the results, improvements that could be made to the model, and future uses of Maxent in the Kona Field System. 7 CHAPTER 2: BACKGROUND AND LITERATURE REVIEW Ecological models have been applied to Hawai‘i in efforts to identify areas suitable for agricultural production. This chapter discusses the Kona Field System, habitat suitability modeling, the Maxent program, and the implications of using environmental variables to model archaeological resources which are the result of human behavior which is not determined by environmental variables. 2.1 The Kona Field System The Kona Field System is a rain-fed agricultural system located in the Kona District on the leeward, west, side of Hawai‘i Island (Allen 2001a; Allen 2004; Cordy 2000; Kelly 1983; Lincoln and Ladefoged 2014). The Kona Field System was described by early visitors to Hawai‘i Island (Ellis 1917; Ledyard 1963; Menzies 1920) and Māhele land claim testimonies (Kelly 1983; Lincoln and Ladefoged 2014; Cordy 2000). The main crops cultivated within the Kona Field System included wauke (paper mulberry [Broussonetia papyrifera]), ‘uala (sweetpotato [Ipomea batatas]), ‘ulu (breadfruit [Autocarpus altilis]), kalo (taro [Colocasia esculenta]), kō (sugar cane [Saccharum officinarum]), kī (ti [Cordyline terminalis]), mai‘a (banana [Musa spp.]), and ‘ama‘u (endemic tree ferns [Sadleria spp.]) (Allen 2001a; Kelly 1983; Lincoln and Ladefoged 2014; Cordy 2000). The Kona Field System was divided into four different zones which were determined by elevation and rainfall (Handy, Handy, and Pukui 1991; Kelly 1983). Kelly defines these zones as kula (0—500 ft), kaluulu (500—1,000 ft), ‘āpa‘a (1,000—2,500 ft), and ‘ama‘u (2,000— 3,000 ft). Figure 3 shows that the AOI for this research lies within the lower three elevation zones with the majority of the area in the kula and kaluulu zones. ‘Uala (sweetpotato) was the 8 main crop of the kula zone, ‘ulu (breadfruit) was the main crop within the kaluulu zone (Newman 1971; Kelly 1983; Lincoln and Ladefoged 2014). Figure 3. Research AOI shown within the elevation zones of the Kona Field System 2.2 Habitat Suitability Modeling Habitat suitability modeling considers the relationship between environmental variables and the probability that a given species will occur (Hirzel and Le Lay 2008). Habitat suitability modeling provides a method of predicting the species occurrence where there is little observational data available on the species of interest (Pearson et al. 2007). Pearson et al. (2007) do caution that modeling species occurrence with limited presence data identifies regions that have characteristics similar to those where the species occur and not the limits of that species' habitat range. Phillips, Dudík, and Schapire (2004) propose the use of maximum entropy to model species geographic distributions. The maximum entropy method for modeling species distribution involves using a collection of known species localities as sample data along with relevant environmental factors to model the distribution of that species within a known 9 geographic extent (Phillips, Dudík, and Schapire 2004). Phillips, Anderson, and Schapire (2005) describe the use and testing of the Maxent program, a widely used maximum entropy modeling toolbox, to model species geographic distributions using habitat suitability, and presence-only data. They assess their resulting models' validity using binomial tests of omission and receiver operator characteristics (ROC) analysis. Elith, et al. (2011) describes how to use the Maxent modeling program and makes suggestions for which parameters should be used. Anderson, Lew, and Peterson (2003) deal with evaluating the validity of predictive models of species' distributions. While they use the Genetic Algorithm for Rule-Set Prediction (GARP), their article is pertinent because their study applies to stochastic models that use presence-only data. They propose a method for determining an optimal underestimation/overestimation of presence (omission/commission) relationship to select models that reasonably predict the distribution of a species. Haegeman and Loreau (2008) caution against the blind use of maximum entropy theory in ecology. They show that the restrictive constraints may dictate the outputs of maximum entropy modeling and cause the models to over-fit the probability distributions to the presence- only data. In order for maximum entropy to be effective in ecological modeling, the environmental variables must be understood and carefully chosen (Haegeman and Loreau 2008). 2.3 Ecological Models and Archaeology The Kona Field System has previously been spatially modeled as either general models of agriculture in Hawai‘i (Ladefoged et al. 2009) or concentrated on the extent of the kaluulu zone (Lincoln and Ladefoged 2014). Ladefoged et al (2009) created a model of the two main different types of agricultural systems throughout the state of Hawai‘i: irrigated pondfields and rain-fed dryland systems. They used three variables important in sweetpotato cultivation, 10 rainfall, elevation, and soil fertility, to model the rain-fed dryland agriculture systems. This model of rain-fed dryland agricultural systems successfully predicted the leeward Kohala field system on Hawai‘i Island and the Kalaupapa field system on Moloka‘i. It also under-predicted the presence of the Kona Field System. Discrepancies between the model and observed agricultural development were explained by the presence of variables that were difficult to quantify and could not be accounted for in an ecological model (Ladefoged et al. 2009). Lincoln and Ladefoged (2014) modeled the extent and potential yield of the kaluulu zone using historic descriptions and archaeological survey. Rainfall and elevation rasters were used to determine the environmental boundaries of the kaluulu zone in their model. Lincoln and Ladefoged’s model increased the previously estimated extent and size of the kaluulu zone. Such previous studies on Hawai‘i Island have modeled the extent and yield of the Kona Field System, but they have not looked at the potential distribution of features within the field system. They have utilized habitat suitability modeling, but not maximum entropy. Two more examples of ecological models in archaeology that have used presence-only data with environmental variables are described below. The first was a project in Southern Italy modeling probability distributions of Bronze Age sites and the second was a project using Maxent in Cyprus to predict the presence of ancient and modern terraces. Arnese (2008) used ecological niche factor analysis (ENFA) to create a probability distribution of Bronze Age sites in Kaulonia located in Calabria, Southern Italy. ENFA utilizes presence data and "eco-geographical" variables to model animal distributions (Arnese 2008). The eco-geographical variables used in Arnese's model were elevation, distance from the sea, slope, aspect, and distance from rivers. The ENFA model ranked elevation as the most important variable followed by distance from the sea. Arnese was able to ground truth the model with 11 success. Areas with a high Habitat Suitability value contained Bronze Age sites and areas with a low Habitat Suitability value did not contain any Bronze Age Sites. Galletti et al. (2013) used Maxent successfully to model predictive distributions for ancient and modern terraces in the Troodos foothills on the island of Cyprus. They used 10 environmental variables in their terrace probability distribution modeling. These variables included, elevation, spring vegetation, fall vegetation, distance to nearest calcareous sediments, distance to nearest major stream, slope, albedo, radiation, distance to nearest pillow lava, and temperature. The AUC analysis, model gain, performance gain, and the Kvamme gain statistic were used to evaluate the performance of their Maxent outputs. Galletti et al. were able to show that the distributions of the ancient and modern terraces were determined differently by the environmental variables. 2.4 Maxent for Modeling Archaeological Site Probability The Maxent program 2 uses maximum entropy theory to create a probability distribution of a given species related to a series of environmental variables (Phillips, Dudík, and Schapire 2004; Phillips, Anderson, and Schapire 2005; Phillips and Dudík 2008). Maximum entropy theory uses partial information, presence-only data, to create a probability distribution that best conforms to a set of constraints, environmental variables, without considering any other constraints (Jaynes 1957; Phillips, Dudík, and Schapire 2004). The Maxent program developed by Phillips, Dudík, and Schapire (2004) is a machine learning process that uses multiple iterations to train the model into creating an acceptable model. Due to the stochastic nature of this process, multiple replicates of the model must be run in order to bring the average of the outputs to a suitable result (Kemp 2012). 2 Available for download at http://www.cs.princeton.edu/~schapire/maxent/ 12 Maxent requires a point dataset that contains the observed species occurrence, presence- only, data and environmental raster datasets. In the case of the model described here, the observed species data contains the coordinate information for each agricultural feature recorded during the archaeological agriculture feature survey. The environmental variables for this study were determined by previous studies that modeled rain-fed agricultural systems on Hawai‘i Island (Newman 1971; Ladefoged et al. 2009; Lincoln and Ladefoged 2014). Newman (1971), Ladefoged et al. (2009) and Lincoln and Ladefoged (2014) identify the environmental variables used to determine the extent of rain-fed agriculture in the Kona Field System. These variables include rainfall, elevation, slope, and soil fertility. Within the Kona Field System, precipitation increases with elevation. The conceptual model for this research proposes that agricultural clearing feature density increases with increases in elevation and rainfall until taro cultivation is viable. At that point, farming methods were more formalized in well defined fields and the less structured agricultural clearing feature density declines. The conceptual model is shown in Figure 4 illustrating the increase and decrease in rainfall and agricultural clearing feature density. However, it is important to remember that agriculture is a human activity, and that human behavior may not be dictated completely by the environment (Freilich 1967). Ecological models can be used to identify the potential probability of archaeological feature presence based on environmental variables but the models must be evaluated with empirical observations (Ladefoged et al. 2009). The variables that define a spiritually important place cannot always be quantified in a tangible manner (King 2003). An ecological model's inability to account for intangible variables may cause the model to overestimate agricultural feature presence. 13 Figure 4 Conceptual Model The next chapter discusses the preparation of the data and the Maxent model and the tests used to evaluate the model. 14 CHAPTER 3: METHODOLOGY Maxent creates a probability distribution model by using presence point data and environmental variable raster grids. This chapter discusses the geographic extent of the model and the AOI, the preparation of the agricultural presence data and environmental variable rasters, the Maxent model setup, and finally the diagnostic tools built within the Maxent program that were used to evaluate the final model. 3.1 Study Area As previously described in Chapter 1, the AOI is an approximately 312-acre portion of the Hōkūli‘a luxury development as shown in Figure 1. The AOI is approximately 3,235 m long, and 1,430 m at its widest. Unfortunately, heavy machinery has altered portions of the AOI to an extent where agricultural features were either completely removed from the landscape or altered to such an extent that the features were not identifiable; thus, it was observed that agricultural features were recorded with less frequency in mechanically altered areas. Dense stands of guinea grass also prevented the crew from being able to survey portions of the AOI. Therefore, predicting the distribution of agricultural features in these incompletely surveyed areas is the main objective of this research. Figure 5 shows the AOI as well as the mechanically altered portions and the densely vegetated areas. The study area is situated on a recent pahoehoe flow, thus only one soil type covers the majority of the AOI as shown in Figure 6. This led to the exclusion of soil attributes in this model, as discussed below in Section 3.3 Environmental Variable Data. 15 Figure 5 Densely vegetated and mechanically altered areas in and around the AOI Figure 6 AOI and Soil Types 16 3.2 Agricultural Feature Presence Data Prehistoric and historic agricultural feature spatial and attribute data were collected between the months of August 2014 and March 2015 (Kailihiwa 2015). The agricultural feature data points are shown in Figure 7. The agricultural feature types included clearing features (rock mounds and modified outcrops), agricultural terraces, grow pits, linear rock mounds, agricultural enclosures, retaining walls, and swales; these are summarized in Table 1. Linear rock mounds, agricultural enclosures, retaining walls, and swales were not used in this test of Maxent because of the small number of observed occurrences for each feature. The clearing features, grow pits, and agricultural terraces were each modeled separately. Linear mounds, even though they were the second most numerous feature recorded, were not used to create a model because they are line features and Maxent requires the presence data to be point features. Figure 7 Archaeological Feature Map 17 Table 1 List of Agricultural feature types, and the number of each type observed within the AOI (Kailihiwa 2015) Type Count Clearing Feature 3743 Linear Mounds 139 Grow Pit 104 Terrace 89 Enclosure 42 Swale 2 Retaining Wall 1 The data was collected by a crew of archaeologists systematically surveying and recording agricultural features within the AOI. The crew members were spaced five to ten meters apart (spacing was dependent upon ground visibility) during the survey; spatial data was recorded with handheld Global Positioning System (GPS) devices that provided an accuracy of five to ten meters. The central area of the AOI contained the densest concentration of agricultural feature observed presences. This portion of the AOI was the least vegetated and the least impacted by heavy machinery and provided the most accurate picture of agricultural feature distribution in the entire AOI. The presence samples were not reduced in the area for modeling purposes, even though they undoubtedly influenced the output probability distribution, because this region did represent an unaltered view of agricultural feature intensity in the Kona Field System. 3.2.1 Preparation of Prehistoric Agricultural Presence Data for Maxent Maxent requires that the presence data is in a comma delimited value text file that contains three fields, at a minimum, that identify the "species," the X-coordinates and Y- coordinates. The X and Y-coordinate values for each datum point were added to the attribute 18 table of the original presence point dataset. The values for each environmental variable were also added to the attribute table for each sample point. The addition of the environmental variable values ensured that Maxent would run more efficiently because it would not have to go through each environmental raster to determine the values for each sample presence point (Phillips n.d.). Once the pertinent data was added to the attribute table, it was converted to a comma delimited text file. 3.3 Environmental Variable Data Maxent requires that each of the environmental variables is stored in ASCII grid format with all rasters sharing the same exact cell size and the same exact geographic extent. The presence-only data must be in a comma delimited value table with a minimum of three fields which denote species, X-coordinates and Y-coordinates. It is possible to set up Maxent once the presence-only data and the environmental variable rasters are in the proper format. As discussed in Chapter 2, Ladefoged et al. (2009), Lincoln and Ladefoged (2014), and Newman (1971) identify the environmental variables that help to define the extent of the rain-fed agriculture of the Kona Field System. These variables, which include rainfall, elevation, slope, and soil fertility, are summarized in Table 2. The archaeological survey that generated the presence-only data was conducted on a single soil type, pahoehoe bedrock (see Figure 6). Test runs with soil fertility variables in Maxent showed that the model over-fit the probability distribution to the soil polygon. Haun and Henry (2010) demonstrated that the soil in the region is rocky enough that agricultural clearing features do occur on soil types other than pahoehoe flows. Therefore, soil data were not used in the final model so that the probability distribution created by Maxent would extend beyond the immediate boundaries of the archaeological survey 19 area. This section discusses each of the environmental datasets and their preparation for use in Maxent. Table 2 Summary of Environmental Variables. These variabels were identified in previous models of prehistoric agriculture in Hawai‘i and were used during test model runs. Environmental Variable Content Format Comments Elevation Elevation data Raster in ASCII grid format Two foot elevation contour dataset converted to a raster grid with a cell size of 10 m. Source: Belt Collins, 2002, aerial survey Slope Slope data derived from elevation data Raster in ASCII grid format Slope in percent rise with a cell size of 10 m. Derived from the elevation raster Aspect Aspect raster derived from elevation data Raster in ASCII grid format Derived from the elevation raster Rainfall Data: Annual Average Summer Average Average totals for June, July, August, and September Average yearly rainfall totals and average totals for summer months Raster in ASCII grid format 30 year average rain totals. Original Cell size is 250 m. Raster grids Resampled to 10 m cell size to match the elevation and slope data. Projected from WGS84 to NAD83 UTM zone 5N; Source: Giambelluca et al. 2013 3.3.1 Rainfall data The timing of the rainfall and the rainfall totals are important factors in sweetpotato cultivation. Handy, Handy, and Pukui (1991) state that the best time to plant sweet potato in Kona is during the summer months. Valenzuela, Fukuda, and Arakaki (2000) define the tolerable rainfall range for sweetpotato during its growth cycle to be 500 to 1300 mm. While rainfall datasets that cover the period of occupation for the Kona Field System are not readily available, long-term modern rainfall data was available from Giambelluca et al. (2013). The rainfall data 3 available from Giambelluca et al. (2013) covers seven of the eight principal Hawaiian Islands, Ni‘ihau is excluded from this dataset. The datasets are rasters with a cell size of approximately 250 m x 250 m in the World Geodetic System 1984 (WGS1984) geographic coordinate system. Giambelluca et al. (2013) used raingage data and expert 3 Available for download from http://rainfall.geography.hawaii.edu/downloads.html 20 knowledge supplemented by five predictors to create this data set. The five predictors included updated raingage data with values averaged over 30 years from 1978--2007, Parameter-elevation Regressions on Independent Slopes Model (PRISM) data for the years 1971—2000, National Weather Service mean radar rainfall estimates, mean rainfall estimates created by experimental model forecasts, and vegetation distribution (Giambelluca et al. 2013). Thirteen grids are included in the dataset, an annual grid and a grid for each month. The rainfall data from Giambelluca et al. (2013) that were selected for use in this model included annual, June, July, August, and September totals averaged over a 30 year time period from 1978—2007. This data is not contemporaneous with the Kona Field System occupation, but it does show the expected pattern of rainfall increasing with elevation as well as rainfall totals during the summer months in Kona, which, as mentioned above, were important for the cultivation of sweetpotato in the region. 3.3.1.1 Rainfall Data Preparation The rainfall datasets were downloaded from the Rainfall Atlas of Hawai‘i site in the Esri grid format. The original data had a cell size of 250 m in the World Geodetic System 1984 (WGS1984) geographic coordinate system. The dataset was projected into the North American Datum 1983 (NAD83) Universal Transverse Mercator (UTM) zone 5 North (5N) projected coordinate system. As can be seen in Figure 8 top, these large cells produce a blocky appearance at the scale of the study area. Since rainfall is a continuous phenomenon, it was dec ided that it was appropriate to smooth this data as they were projected and resized to the model's computational scale of 10 m. The rainfall grids were smoothed using bilinear interpolation which averages the four nearest input cells weighted by their distance to the center of the input raster cell (Esri 2009). This resulted in a surface with gradually increasing rainfall amounts over 21 smaller-sized cells rather than the sudden jumps in values with the coarser dataset (see Figure 8 bottom). Once the rainfall data was projected and resized it was clipped and converted to an ASCII grid, ready for use in Maxent. 3.3.2 Elevation Data Since a fine resolution elevation grid was not available, an AutoCAD Drawing Exchange Format (.dxf) file containing 2 ft contour lines of the Hōkūli‘a property provided the elevation data. This data was derived from an aerial photogrammetry survey flown in 2002 and has an expected vertical accuracy of two meters. The vertical accuracy was affected by the height of the guinea grass covering the ground in the survey area. After converting the lines into a shapefile containing a line feature class, non-elevation lines were removed and elevation values were assigned to each contour line. The .dxf file was in the Old Hawaiian State Plane Hawaii 1 projected coordinate system and was reprojected to the NAD83 UTM zone 5N projected coordinate system. Once the file was reprojected it was then georeferenced to the World Imagery basemap available in ArcGIS 10.2.2, which has an horizontal accuracy of 10 m. With the conversion from CAD to GIS data complete, it was possible to create the elevation raster. 3.3.2.1 Elevation and Elevation Derived Rasters The elevation contours were interpolated into a raster using the assigned values, in feet, of each line to determine the value of each cell. The cell size was set to 10 m by 10 m. A mask that followed the outline of the elevation contour shapefile was used to clip the elevation raster to limit the model's calculation to the area in which the elevation was known. This same mask was used to clip all of the environmental variables to a common spatial extent. Figure 9 shows the elevation raster clipped to the model's spatial extent. 22 Figure 8 Comparison of Raw Rainfall Data (top) and Smoothed Rainfall Data (bottom) Figure 9 Clipped Elevation Raster 23 The elevation raster was used to create a slope raster. The cell size and extent of the slope raster were set to the same values as the elevation raster. The output measurement for the slope raster was set to percent rise due to the slope variables (percentage) discussed in Newman (1971). An aspect raster was also created from the elevation raster to determine whether or not the amount of sun exposure for sweetpotato contributed significantly to the model. Cell size and extent were matched to the elevation raster. Figure 10 shows the aspect raster derived from the elevation raster. Once the rasters were clipped to the proper extent they were converted to ASCII grid format and ready to use in Maxent. Figure 10 Aspect Raster 3.4 Maxent Parameters The Maxent program has four different areas in which model parameters can be selected. The first area to setup the modeling parameters are in the main graphical user interface (GUI). The main GUI is where the presence data and the environmental variables are selected. The output model format, file type, and output directory are determined on the main GUI. The main GUI is also where the computational features and the jack-knife testing and response curves are 24 selected. The 'Settings' button is also on the main GUI and contains three tabs of parameters that can be set for each model run. This section first discusses the parameters set on the main GUI and then discusses the tabs in Settings. 3.4.1 Computational Features Phillips, Dudík, and Schapire (2011), Merow, Smith and Silander (2013), and Elith et al. (2011) provide descriptions of the computational feature types (Table 3) and explain the appropriate use for each feature type. The computational features are used to constrain the Maxent output based on the environmental variable values (Phillips, Dudík, and Schapire 2011). Linear features constrain the output model to the same values for each continuous environmental variable as the observed locations for that species. Quadratic features constrain the output model to the same environmental variable variance as the presence-only data. Product features constrain the covariance of one variable with other environmental variables by multiplying two of the environmental variables. Threshold features create a probability grid where the environmental variable value is greater than a threshold value. Hinge features also create a probability grid for values above a threshold, but use a linear function to create a smoother model. Discrete features are automatically created for each variable that is categorical rather than continuous. Maxent allows users to choose which type of calculation features to use based on how the model should respond to the environmental variables. The number of presence samples determines which features are used by default. Linear features are automatically chosen when the sample number is below 10, Maxent automatically adds quadratic features when there are more than 10 samples, hinge features are added when there are more than 15 samples, and threshold and product features are added when there are more than 80 samples. Elith et al. (2011) 25 recommend simplifying Maxent models by using only hinge features to create an easily interpreted model. It is hard to determine the nature of environmental variable interaction with a small number of presence samples (Elith et al. 2011). Hinge features, like linear features are based on a single environmental variable, but their values are a constant below a threshold value. Hinge features are a special case of linear and threshold features and using them in the model makes using linear and threshold features redundant (Elith et al. 2011; Merow, Smith, and Silander 2013). Therefore, hinge, quadratic, and product features were utilized in the final model of this project in order to create a distribution that showed the probability of where agricultural clearing features may be present dependent upon the relationship and interaction of the environmental variables. Table 3 Summary of Maxent Calculation Features and a description of the constraints they impose upon the output model. Feature Type Description Linear Matches the output value of the environm ental variables to the same values as the presence-only data. Quadratic Constrains the environmental variable variance to match the presence-only data. Product Constrains the covariance between two environmental variables. Useful for modeling complex relationships between variables. Threshold Creates a binary probability grid using a step function for variable values above a certain threshold value Hinge Creates a binary probability grid using a linear function for variable values above a certain threshold value. 26 3.4.2 Environmental Variable Assessment in Maxent The main Maxent GUI provides two options for assessing the importance of the environmental variables to the model: (1) response curves and (2) jack-knife testing. Response curves demonstrate the influence each environmental variable has on the Maxent model (Phillips, Dudík, and Schapire 2011). Jack-knife testing creates models first omitting each variable and then using only one variable to determine the importance of an environmental variable to the predictive distribution of a species (Phillips, Dudík, and Schapire 2011). Both of these options were selected to determine the importance of the environmental variables and to observe whether or not the predictive distribution responded to each variable as expected. 3.4.3 Maxent Output Format Maxent has three different output formats (1) raw, (2) logistic (default), and (3) cumulative. The raw output shows a relative occurrence rate (ROR) of the target species (Merow, Smith, and Silander 2013). The logistic output is a post processing of the raw output and is used to show an approximate probability that the species will occur within a specific grid cell (Phillips, Dudík, and Schapire 2011; Merow, Smith, and Silander 2013). The cumulative output shows the suitability of conditions at a given location (Phillips, Dudík, and Schapire 2011). The logistic output format was chosen for this project because of its ease of interpretation for species' probability of presence. 3.4.4 Maxent Settings The Maxent settings GUI contains three different tabs, (1) Basic, (2) Advanced, and (3) Experimental, where parameters are set before running Maxent. There are a total of 46 parameters among the three tabs. The parameters which can be set in Maxent are summarized in Table 4. 27 Table 4 Modeling Parameters. These parameters affect the final output model produced by Maxent. Parameter Tab Description Random seed Basic Random sets of sample points are used during each model run for testing and training the model Random test percentage Basic Percentage of presence samples set aside as test points used in evaluating the model Regularization multiplier Basic Multiply the regularization parameters by this number. The lower the number the tighter the distribution is to the presence-only data occurrences. A higher number creates a more diffuse distribution Max number of background points Basic The number of cells chosen at random for background points if the number of grid cells is larger than specified number Replicates Basic Number of times to run the model. Replicated run type Basic Testing and training method to use when running more than one replicate. Three types Cross Validation (divides samples into groups and each group takes a turn being used as test samples), Bootstrap (test samples chosen randomly by sampling with replacement), and Subsampling (test samples chosen randomly by sampling without replacement). Maximum iterations Advanced The maximum number of iterations to run before model learning is terminated. Adjust sample radius Advanced Adjust the size of the sample dots on the pictures of predictions by adding the specified number of pixels. Negative values reduce the size of the radius Default prevalence Advanced Probability that an individual is observed at a suitable location Threads Experiment- al Number of processor threads in a computer's central processing unit (CPU) to use. Test runs were conducted using different parameter settings, environmental variables, and presence point data in order to determine the parameter settings, environmental variables, and agricultural feature types that would be used in the final model. More than 30 different test runs were conducted and evaluated based on their goodness-of-fit, stability, environmental variable performance, and whether or not the final model met expectations. Models for agricultural clearing features, grow pits, and terraces were created. The paucity of presence data and environmental variables for the grow pits and terraces created unreliable models. The abundance of clearing feature presences (n=3743) appeared to compensate for the small number of 28 environmental variables and appeared to create reliable probability distributions which are discussed in the next chapter. The parameters used in the final model are summarized in Table 5. Table 5 Final Model Parameters. These are the values, and the justification for those values, that the model parameters were set to during the final model run. Parameter Value Justification Random seed True Random seeds are used to keep the bootstrapping runs from using replicate test and training samples. Random test percentage 25 Each model replicate used 25 percent of the samples for testing, training and statistics calculation Regularization multiplier 8 Model runs indicated that this number was sufficiently large enough to create smooth response curves and spread-out the probability distribution from the presence-only data while maintaining a stable model. Max number of background points 10000 Background points are used as pseudo-absences for model training. Presence probability is assumed to be unknown for each background point. The higher the number, the better the fit of the model. This number balanced speed and accuracy for modeling. Replicates 100 Number of replicates sufficient enough to converge the average and median values of the Maxent output. Replicated run type Bootstrap Bootstrapping created the models with the best goodness of fit and stability. Each of the 100 replicates run was able to use 25 percent of the sample points selected at random. Maximum iterations 1000 The higher the number the more training that Maxent can do per replicate run resulting in a more stable model. Adjust sample radius 0 Not pertinent to the final output of this model. Default prevalence 0.8 Probability that an individual is observed at a suitable location Threads 8 CPU used for the model runs was an Intel Core i7 I7-4770 3.4 GHz Quad-Core Processor. It has eight threads. 29 3.4.5 Model Validation Maxent contains statistical tools to help in the validation of the output predictive distributions. The main tool used to assess the goodness of fit for a Maxent model is the area under the curve (AUC) which is determined by the receiver operating curve (ROC). The ROC, in the case of Maxent, is a measure of the relative occurrence rate (ROR) because the model is created using presence-only data (Merow, Smith, and Silander 2013). The response curves and results from the jack-knife testing are additional tools created by Maxent to assess its outputs. The final method of model validation, to be conducted in the future, will be ground truthing the areas of dense vegetation. 30 CHAPTER 4: RESULTS A prehistoric agricultural clearing feature probability distribution was produced using presence point data and environmental variable raster data in the Maxent program. The output distribution was produced as an ASCII grid. Maxent provides diagnostic tools that were used to assess the reliability of the model. These tools include jack-knife testing to determine environmental variable importance, response curves to demonstrate how the model responded to the environmental variables, and area under the receiver operator curve (AUC) to assess the goodness-of-fit of the model. This chapter begins with a discussion of the output model generated by Maxent followed by a discussion of model validity using the diagnostic tools available in Maxent. 4.1 Maxent Probability Distribution The Maxent program uses a machine learning algorithm that trains itself by contrasting a specified number of the available presence samples to a specified number of background sample points. The background sample points are used as "pseudo absences" and the values are contrasted with the values of the presence data. The final probability distribution here is a product of 100 bootstrap replicate runs. Each run used 75 percent of the presence samples (n = 2808) and 10,000 background samples to train each model and 25 percent of the presence samples (n = 935) to test each replicate model. A smoothing filter was used in ArcMap 10.2.2 on the final output to smooth out the values and create a more realistic model. The smoothing filter is a 9 cell (3x3) window that passes over each cell in the raster dataset and assigns the average values of the 9 cells to the center cell, reducing anomalies in the model. The higher values of the raster were reduced slightly (95 to 92), but not enough to affect the overall predicted distribution of the agricultural 31 clearing features. Table 6 summarizes the change between the smoothed and unsmoothed raster. The only value that changed is the max value. Table 6 Minimum, maximum, and mean values of the unaltered output model and the smoothed output model. Model Output Minimum Value Maximum Value Mean Value Unaltered Model 0 95 51 Smoothed Model 0 92 51 Figure 11 shows the final, smoothed average of the probability distributions of the 100 replicate runs, and the standard deviation of those runs. The values for the distribution are the probability that the agricultural clearing features are present in each grid cell using breaks of 10 percent. The standard deviation is depicted using one percent breaks in order to better show the variation in the standard deviation which ranges from close to zero to approximately 10.3 percent and has the greatest values near the westernmost and at the easternmost extent of the model. The modeled area extends the probability distribution of agricultural clearing features beyond the boundaries of the AOI. Soil data variables were too homogenous and constrained the model too tightly to the soil polygons, as a result, the soil data were not used in this model. Elevation-derived data and rainfall datasets were used to create a more generalized model that extends beyond the boundary of the AOI. The next section of this chapter discusses the environmental variables and their performance in the model. 4.2 Environmental Variable Performance Maxent creates a hypertext markup language (html) file that summarizes the model results. It creates a file for each replicate model run and one file for the summary of all the replicates. The results for the different diagnostic tests conducted by Maxent are included in each of the html files. The environmental variable performance is evaluated using response curves, contribution and permutation importance, and jack-knife testing. 32 Figure 11 Final Maxent output 33 4.2.1 Response Curves Maxent creates two sets of response curves for the environmental variables. The first set of curves are called marginal curves and they demonstrate how the model prediction changes as the values of each environmental variable changes slightly while the rest of the variables remain at their average values. Phillips, Dudík, and Schapire (2011) warn that the marginal curves may be difficult to interpret if the environmental variables are correlated. The annual precipitation (rf_ann) and summer precipitation (rf_summer) are highly correlated with the monthly rain datasets; the values for annual and summer precipitation cannot be changed without changing the values of the monthly precipitation datasets. As a result, The marginal response curves for this model are not appropriate for determining how the Maxent output is affected by each environmental variable. The second set of response curves shows that the Maxent prediction reaches a peak and then decreases as the values go up for each environmental variable. Figure 12 shows the response curves created to represent a Maxent prediction that uses only that environmental variable. The Y-axis for each of the response curves is the probability of agricultural feature presence. The X-axis for the elevation response curve (elev) is elevation in feet, and the X-axis for the rainfall response curves is rainfall in millimeters. The standard deviation for the 100 replicate runs, depicted in blue, increases as the number of presence samples represented for those values of the environmental variables decrease. Figure 12 shows the result of the "Identify" tool used in ArcMap 10.2.2. This tool has been configured to show the cell values for each of the visible layers. The two results displayed in Figure 10 show the values of the Maxent model, Agricultural Clearing Feature Average, along with the values of the eight environmental variables that intersect with that Maxent model pixel. 34 Figure 12 Maxent output response curves. Figure 13 Sample 'Identify' results for a Maxent model pixel with a low probability value (left) and a high probability value (right). Note that the variable values in the high probability pixel are near to the peak values for each variable, while the reverse is true for the low probability pixel. 35 4.2.2 Environmental Variable Contribution and Importance to the Maxent Model Maxent uses two different methods to estimate variable importance. The first method implemented by Maxent creates a table using data gathered during the training of the model that summarizes the environmental variable contribution to the model and the permutation importance, or stability, of the variable. Variable contribution is determined by the amount of increase or decrease of the model fit, called gain, caused by an environmental variable for each iteration of the Maxent algorithm. The permutation importance is calculated by randomly changing the value of an environmental variable among the model training points. The lower this value is the more stable the variable's contribution to the model. The value of the decrease in the training AUC is normalized so that the data can be represented as percentages for both the percent contribution and the permutation importance (Phillips n.d.). Table 7 lists the environmental variables used in this model and ranks them according to their contribution to the model. Table 7 Summary of Environmental Variables and Their Importance. The variables are ranked according to their percent contribution with September Rainfall having the most contribution and July Rainfall the least contribution. Variable Percent contribution Permutation importance September Rainfall 48.8 19.9 Elevation 22.5 17.9 Summer Rainfall 19.4 17.8 August Rainfall 5 24.6 June Rainfall 1.5 13.7 Slope 1.1 0.7 Annual Rainfall 1.1 0.8 July Rainfall 0.6 4.6 36 Another method that Maxent uses to determine variable importance is the jack-knife test. The jack-knife test trains the model removing each environmental variable to calculate which variable causes the largest decrease in the model's gain. This variable contains the most information not found in the other environmental variables. The second part of the jack-knife test is training the model using each environmental variable by itself. The environmental variable with the highest gain is considered to have the most useful information by itself. The summer rainfall variable (rf_summer) caused the largest decrease in the gain averaged over the 100 replicate runs of this model. The September rainfall variable (rf_sep) had the highest gain when used alone for training. The Maxent output, however, warns that caution must be used when interpreting the results of jack-knife testing when the environmental variables are highly correlated, as the rainfall variables are in this case. 4.3 Model Performance The area under the receiver operator curve (AUC) is the most commonly used diagnostic to evaluate a Maxent model (Merow, Smith, and Silander 2013). The AUC is a comparison of the true positive rate and the false positive rate, or how well the model is able to predict presence and absence. Maxent uses presence-only data. The AUC created for Maxent models shows how well the model is able to distinguish presence from random (Phillips, Anderson, and Schapire 2005). The value for the AUC ranges from 0 to 1, the closer the value of the AUC is to 1 the better the fit of the model. An AUC value of 0.5 equals random prediction (Phillips, Anderson, and Schapire 2005). The average AUC for the 100 replicates of this model was 0.819 with a standard deviation of 0.009. According to the diagnostic tools provided by Maxent, the model produced for this project performs well. The next chapter discusses the results of the model with respect to the 37 expectations presented by the conceptual model in Chapter 2. Chapter 5 also discusses the applicability of Maxent in archaeology, future work, and improvements that can be made to the current model. 38 CHAPTER 5: DISCUSSION AND CONCLUSIONS The purpose of this thesis was to assess the usefulness of Maxent for predicting the presence of archaeological features. A recent archaeological survey of agricultural features within a portion of the Hōkūli‘a luxury development on Hawai‘i Island provided the opportunity to test Maxent (See Figure 1 in Chapter 1). This chapter discusses the results of the Maxent model, how they compare to the conceptual model and expectations of agricultural feature distribution in the Kona Field System, how human behavior may affect the interpretation of the results, limitations of this model, and the usefulness and applicability of using Maxent in an archaeological context. 5.1 Model Evaluation The diagnostic tools included as part of the Maxent package indicated that the model produced for this project is a strong, reliable model. Three research hypotheses listed in Chapter 1 were also used to evaluate the model. The first hypothesis states that agricultural clearing feature density increases with elevation from the shoreline to the upper elevations of the kula zone and the lower elevations of the kaluulu zone (approximately 0—680 ft elevation), and that the feature density should start to decrease starting in the middle elevations of the kaluulu zone (approximately 680 ft elevation). This distribution pattern was noted during an archaeological inventory survey conducted three kilometers to the north of the current AOI by Haun and Henry (2010). The model shows that the region within the AOI that has the highest probability of agricultural clearing features ranges in elevation from approximately 175 ft to 600 ft. Figure 14 shows the modeled probability distribution overlaid with elevation contours to illustrate this. Cells within the elevation range of 175 ft—600 ft have an 80 percent or better probability of containing an agricultural clearing feature. Figure 15 shows the response curve for an elevation- 39 only model created within Maxent. The response curve shows that there is a 70 percent or better probability of agricultural clearing features present in the 50 ft—680 ft elevation range. The results for the distribution of agricultural features with respect to elevation do not exactly match the observations made by Haun and Henry (2010), but that does not invalidate the model. The model does show a general trend of increasing presence with elevation before peaking and presence probability starts to decline above 350 ft in elevation. This apparent lack of correspondence may be due to the fact that elevation is not the only variable involved in this model. The second hypothesis states that as precipitation levels approach acceptable levels for dryland taro farming, there is a decrease in the density of informal clearing features that are used in sweetpotato cultivation. As noted in Chapter 2, Valenzuela, Fukuda, and Arakaki (2000) Figure 14 Agricultural Clearing Feature Probability Distribution With Elevation Overlay 40 Figure 15 Elevation Response Curve Showing the Average Values and Standard Deviation from the 100 Maxent Replicate Models identify the rainfall tolerance for sweetpotato to be 500—1300 mm per growth cycle and the optimal rainfall to be 900—1300 mm. Ladefoged et al (2009) used 750—1000 mm as their optimal range for sweetpotato cultivation. Handy, Handy, and Pukui (1991) identified the rainfall during the summer months as the most important for sweetpotato cultivation in the Kona Field System. Figure 16 shows the output model with an overlay of the summer months' rainfall as well as the output model with annual rainfall. The figure shows that there is a good probability for agricultural clearing feature presence between 340 mm—460 mm of summer rain, which correlates to approximately 810 mm—1,050 mm of annual rainfall (See Figure 16). The highest probability of feature presence (greater than 80 percent) does occur within the annual rainfall range of 850 mm—990 mm. This is relatively close to the range used by Ladefoged et al (2009). The rainfall data used in this model is modern and attention should not be focused on the actual 41 values of rainfall data. These data were used to show the general pattern that rainfall totals increase with elevation (Kirch et al. 2012). Figure 16 Agricultural Clearing Feature Probability Distribution with Summer Rain Overlay (top) and Annual Rainfall Overlay (bottom) The third hypothesis is based on Newman's (1971) analysis of agricultural practices in Hawai‘i. The hypothesis is that agricultural practices were limited to slopes less than 35 percent, and that most agricultural activities took place on slopes less than 20 percent. Figure 17 shows 42 the response curve for the average of 100 models using only the slope variable within Maxent. The X-axis values for the graph are percent slope. The response curve shows that the probability of presence for agricultural clearing features peaks at approximately 30 percent slope and decreases as the percent slope value rises above 30 percent. This number does not quite match Newman's (1971) observation but it does show that agricultural activity declines above a certain slope. The Maxent probability distribution model shows that the greatest concentration of agricultural clearing features occurs within the kula zone and lower kaluulu zone. This is an ecological model that uses environmental variables to determine the probability that a cell within the model surface will contain an individual of the species of interest, agricultural clearing features in this case. Importantly, however, the model does not take into account human behavior and cultural factors so it is important to consider the human role on the distribution of agricultural practices. Figure 17 Elevation Response Curve of Clearing Feature Probability to Slope Variable 43 The environmental variables determine, to an extent, how and where crops can be cultivated. People are capable of altering the landscape to increase the cultivability of an area and to grow crops in relatively marginal areas (Kirch et al. 2004; Vitousek et al. 2004). Vitousek et al. (2004) did observe areas in Hawai‘i suitable for dryland agriculture that were not cultivated demonstrating that even though an area was suitable for farming, there is the possibility that it was not cultivated. Historical and cultural factors help in understanding human behavior and cultural development (Freilich 1967). Human behavior must be kept in mind when interpreting this model. The areas modeled as low probability of agricultural clearing feature presence may be an underestimation of the features that may be present. Native Hawaiians developed farming techniques to address the challenges posed by different regions (Handy, Handy, and Pukui 1991). Kirch et al. (2004) noted that Native Hawaiians, at times, were forced to farm in marginal areas to feed growing populations. The actual presence of agricultural clearing features could be greater than predicted by the ecological model due to this behavior. On the other hand, areas that are suitable for agricultural cultivation may be important for spiritual or cultural reasons that are not obvious. Spiritual and cultural importance may be intangible and cannot be measured or evaluated empirically (King 2003). This may cause agriculturally suitable areas to remain uncultivated, in which case the model may overestimate the probability of agricultural clearing feature presence. For example, Pu‘u Ohau, a culturally significant cinder cone, is located in the southwest corner of the model and may be an area where the probability of agricultural clearing feature presence may be overestimated. The model calculates that the eastern slope of the pu‘u has a high probability of agricultural clearing feature presence. However, Pu‘u Ohau contains a number of burial and ceremonial sites (Kaschko 1984; 44 Hammatt et al. 1997; Belt Collins Hawaii 2007) and this Maxent model cannot not account for these sites and the spiritual and cultural importance of Pu‘u Ohau. 5.1.1 Archaeological Monitoring and Future Model Evaluation In the future, model performance will be evaluated in the field as a by-product of archaeological monitoring of ground disturbing activities to be conducted within the AOI. The archaeological monitor will be required to record any features that are encountered during these activities (Tomonari-Tuggle and Tuggle 1999) and the data collected will be used to evaluate the Maxent model. Of course, it will not be possible to evaluate through ground truthing the model performance for the areas that have been completely altered by heavy machinery (e.g. the golf course) as there is no evidence left of the cultural resources that were present within those areas. 5.2 Conclusions The probability distribution of agricultural clearing features modeled by the Maxent program appears to be a reliable and stable model based on the diagnostic tests conducted within Maxent. The mean AUC for the 100 replicates is 0.819 with a standard deviation of 0.009 which is an indicator that the model performs well, the closer the AUC is to 1.0 the better the model performance (Phillips n.d.). The environmental variable response curves show the model responding to individual variables as predicted by the conceptual model. This model could be improved with a better elevation dataset. The elevation data used in this model included recent alterations to the landscape such as house lot grades, the golf course, manmade ponds, and roads. These landscape alterations may have affected the influence of the slope variable on the probability distribution. Maxent's usefulness in regards to modeling archaeological feature distribution is conditional and is dependent upon the number of presence samples that are used to create the 45 model. Test models were produced while the archaeological agricultural feature survey, which provided the presence data for this project, was in progress. Variable contributions and importance changed with the number as well as the distribution of the presence samples. Probability distributions for agricultural terraces and grow pits were also created using Maxent, but the models produced for these feature types were unreliable and the results were not useful due to the small number of presence samples used by the model. As a result of this research, it is concluded that Maxent can be used to model prehistoric agricultural systems by using environmental variables that can define the proper growing conditions for the crops cultivated within that system. Maxent cannot model human behavior and may overestimate feature presence in spiritually important areas that are suitable for crop cultivation where farming does not occur. The variables that define the spiritual importance of a place are often intangible and cannot be empirically recorded (King 2003). As a result, these spiritual variables cannot be used in Maxent. Despite this inability to model human behavior, this research suggests that there is further use for Maxent in the Kona Field System. The program and technique offer a way to synthesize the commercial archaeology data of the region to create a more complete picture of the Kona Field System than the piecemeal study areas do on their own. Soil attribute data could also be added back into the model over the larger geographic region as there is more than one soil type present within the larger Kona Field System. The Maxent program should be able to reliably model the distribution of agricultural features in kula and kaluulu zones as the majority of the agricultural features present within those zones can be represented as points. However, the linear mounds and long, cross-slope terraces observed in the formal areas of the Kona Field System, the ‘āpa‘a zone, present a challenge for Maxent since they are linear features rather than points. 46 This could be resolved by using coarser environmental data at a smaller scale and placing points in the center of the agricultural fields for the presence data as done by Kemp (2012). Using Maxent to model the Kona Field System may expand its extent beyond the currently accepted boundaries. While Maxent is incapable of reliably modeling human habitation sties, burials, and ceremonial sites it can model regions of human activities that are partially dependent upon environmental variables. Modeling the distribution of these activities, e.g. farming, can be used as a proxy in trying to determine areas that may be of cultural importance. Haun and Henry (2010) showed that there were large habitation complexes and substantial ceremonial sites located within areas that were densely covered by agricultural features. The Maxent model can be used to determine which areas warrant further archaeological investigation and provide an idea of the level of effort that an area may require to be properly documented. 47 References Allen, Melinda S. 2004. "Bet-Hedging Strategies, Agricultural Change, and Unpredictable Environments: Historical Development of Dryland Agriculture in Kona, Hawaii." Journal of Anthropological Archaeology 23 (2): 196-224. doi:http://dx.doi.org.libproxy.usc.edu/10.1016/j.jaa.2004.02.001. ———. 2001a. "Cultural Setting and Recent Garden History." In Gardens of Lono: Archaeological Investigations at the Amy B.H. Greenwell Ethnobotanical Garden Kealakekua, Hawai'i, edited by Melinda S. Allen, 1-13. Honolulu, Hawai'i: Bishop Museum Press. ———. 2001b. "The Environmental Context of Traditional Dryland Agriculture in Kona." In Gardens of Lono: Archaeological Investigations at the Amy B.H. Greenwell Ethnobotanical Garden Kealakekua, Hawai'i, edited by Melinda S. Allen, 15-22. Honolulu, Hawai'i: Bishop Museum Press. Anderson, Robert P., Daniel Lew, and A. Townsend Peterson. 2003. "Evaluating Predictive Models of Species’ Distributions: Criteria for Selecting Optimal Models." Ecological Modelling 162 (3): 211-232. doi:http://dx.doi.org.libproxy.usc.edu/10.1016/S0304- 3800(02)00349-6. Arnese, Alessio. 2008. "Applying Ecological Niche Factor Analysis for Predictive Modelling in the Kaulonia Field Survey. Layers of Perception. Proceedings of the 35th International Conference on Computer Applications and Quantitative Methods in Archaeology (CAA). Axel, P, K. Lambers, and I. Herzog Editors.
." Berlin Germany, April 2- 6, 2007. 48 Belt Collins Hawaii, Ltd. 2007. Supplemental Environmental Impact Statement Preparation Notice, Hokuli`a, North Kona and South Kona Island of Hawai`i. Hawaii. Cordy, Ross. 2000. Exalted Sits the Chief: The Ancient History of Hawai`i Island. Honolulu, HI: Mutual Publishing. Elith, Jane, Steven J. Phillips, Trevor Hastie, Miroslav Dudík, Yung En Chee, and Colin J. Yates. 2011. "A Statistical Explanation of Maxent for Ecologists." Diversity and Distributions (17): 43-57. Ellis, William. 1917. A Narrative of a Tour through Hawaii, Or Owhyhee; with Remarks on the History, Traditions, Manners, Customs and Language of the Inhabitants of the Sandwich Islands. Honolulu, HI: Hawaiian Gazette Co., Ltd. Esri. "Georeferencing a Raster Dataset." ArcGIS Desktop 9.3 Help. Esri, last modified November 12, 2009, accessed March 14, 2015, http://webhelp.esri.com/arcgisdesktop/9.3/index.cfm?TopicName=Georeferencing_a_raster _dataset. Freilich, Morris. 1967. "Ecology and Culture: Environmental Determinism and the Ecological Approach in Anthropology." Anthropological Quarterly 40 (1): 26-43. http://www.jstor.org.libproxy.usc.edu/stable/3316830. Galletti, Christopher S., Elizabeth Ridder, Steven E. Falconer, and Patricia L. Fall. 2013. "Maxent Modeling of Ancient and Modern Agricultural Terraces in the Troodos Foothills, Cyprus." Applied Geography (39): 46-56. Giambelluca, Thomas W., Qi Chen, Abby G. Frazier, Jonathan P. Price, Yi-Leng Chen, Pao-Shin Chu, Jon K. Eischeid, and Donna M. Delparte. 2013. "Online Rainfall Atlas of Hawai‘i." 49 Bulletin of the American Meteorological Society 94: 313-316. doi:10.1175/BAMS-D-11- 00228.1. Haegeman, Bart and Michel Loreau. 2008. "Limitations of Entropy Maximization in Ecology." Oikos (117): 1700-1710. Hammatt, Hallet H., Douglas F. Borthwick, Brian L. Colin, Ian Masterson, Jennifer J. Robins, and Helen Wong-Smith. 1997. Archaeological Inventory Survey and Limited Subsurface Testing of a 1,540-Acre Parcel in the Ahupua‘a of Honuaino, Hokukano, Kanaueue, Haleki‘i, Ke‘eke‘e,‘Ilikahi, Kanakau, Kalukalu, and Onouli, Districts of North and South Kona. Vol. I, II, and III. Kailua, Hawaii: Cultural Surveys Hawai‘i. Handy, E. S. Craighill, Elizabeth Green Handy, and Mary Kawena Pukui. 1991. Native Planters in Old Hawaii: Their Life, Lore, & Environment. Honolulu, HI: Bishop Museum Press. Haun, Alan E. and J. David Henry. 2010. Archaeological Inventory Survey TMK: (3) 7-9-05:76, 77 and 78, Lands of Honalo, Ma‘ihi 1-2 and Kuamo‘o 1, North Kona District, Island of Hawaii. Kailua Kona, Hawaii: Haun & Associates. Hirzel, Alexandre H. and Gwenaëlle Le Lay. 2008. "Habitat Suitability Modelling and Niche Theory." Journal of Applied Ecology (45): 1372-1381. Jaynes, Edwin T. 1957. "Information Theory and Statistical Mechanics." Physical Review 106 (4): 620-630. doi:10.1103/PhysRev.106.620. Kailihiwa, Solomon. "Hōkūli‘a Agricultural Feature Survey Data during Monitoring." Archaeological Spatial and Attribute Data, Unpublished Data, . Kaschko, Michael. 1984. Pu‘u Ehu Estates. Hilo, Hawaii: Paul H. Rosendahl, Ph.D., Inc. Kelly, Marion. 1983. Nā Māla O Kona: Gardens of Kona, a History of Land use in Kona, Hawai‘i. Honolulu, HI: Department of Anthropology Bernice P. Bishop Museum. 50 Kemp, Karen K. 2012. The Hawai`i Island Crop Probability Map: An Update of the Crop Growth Parameters for the Hawai`i County Crop Model: Final Report Prepared for the County of Hawai`i Research Adn Development Department. King, Thomas F. 2003. Places that Count, Traditional Cultural Properties in Cultural Resource Management. Heritage Resource Management Series., edited by Don Fowler. Lanham, Maryland: AltaMira Press. Kirch, P. V., G. Asner, O. A. Chadwick, J. Field, T. Ladefoged, C. Lee, C. Puleston, S. Tuljapurkar, and P. M. Vitousek. 2012. "Building and Testing Models of Long-Term Agricultural Intensification and Population Dynamics: A Case Study from the Leeward Kohala Field System, Hawai’i." Ecological Modelling 227 (0): 18-28. doi:http://dx.doi.org.libproxy.usc.edu/10.1016/j.ecolmodel.2011.11.032. Kirch, P. V., A. S. Hartshorn, O. A. Chadwick, P. M. Vitousek, D. R. Sherrod, J. Coil, L. Holm, and W. D. Sharp. 2004. "Environment, Agriculture, and Settlement Patterns in a Marginal Polynesian Landscape." Proceedings of the National Academy of Sciences of the United States of America 101 (26): 9936-9941. doi:10.1073/pnas.0403470101. Ladefoged, Thegn N., Patrick Vinton Kirch, Samuel M. Gon III, Oliver A. Chadwick, Anthony S. Hartshorn, and Peter M. Vitousek. 2009. "Opportunities and Constraints for Intensive Agriculture in the Hawaiian Archipelago Prior to European Contact." Journal of Archaeological Science 36: 2374-2383. Ledyard, John. 1963. John Ledyard's Journal of Captain Cook's Last Voyage, edited by James Kenneth Munford. Corvallis, OR: Oregon State University Press. 51 Lincoln, Noa and Thegn Ladefoged. 2014. "Agroecology of Pre-Contact Hawaiian Dryland Farming: The Spatial Extent, Yield and Social Impact of Hawaiian Breadfruit Groves in Kona, Hawai`i." Journal of Archaeological Science 49: 192-202. Menzies, Archibald. 1920. Hawaii Nei 128 Years Ago, edited by William Frederick Wilson. Honolulu, HI:. Merow, Cory, Matthew J. Smith, and John A. Silander. 2013. "A Practical Guide to MaxEnt for Modeling Species? Distributions: What it does, and Why Inputs and Settings Matter." Ecography 36 (10): 1058-1069. doi:10.1111/j.1600-0587.2013.07872.x. Newman, T. Stell. 1971. "Hawaii Island Agricultural Zones, Circa A. D. 1823: An Ethnohistorical Study." Ethnohistory 18 (4): 335-351. Pearson, Richard G., Christopher J. Raxworthy, Miguel Nakamura, and A. Townsend Peterson. 2007. "ORIGINAL ARTICLE: Predicting Species Distributions from Small Numbers of Occurrence Records: A Test Case using Cryptic Geckos in Madagascar." Journal of Biogeography 34 (1): 102-117. doi:10.1111/j.1365-2699.2006.01594.x. Phillips, Steven J. "A Brief Tutorial on Maxent." Unpublished tutorial distributed with software. Phillips, Steven J., Robert P. Anderson, and Robert E. Schapire. 2005. "Maximum Entropy Modeling of Species Geographic Distributions." Ecological Modelling 190 (2006): 231-259. Phillips, Steven J. and Miroslav Dudík. 2008. "Modeling of Species Distributions with Maxent: New Extensions and a Comprehensive Evaluation." Ecography (31): 161-175. Phillips, Steven J., Miroslav Dudík, and Robert E. Schapire. 2004. "A Maximum Entropy Approach to Species Distribution Modeling." Proceedings of the Twenty-First International Conference on Machine Learning: 655-662. ———. 2011. Maximum Entropy Modeling of Species Geographic Distribution. Vol. 3.3.3k. 52 Roxburgh, Stephen H. and Karel Mokany. 2010. "On Testing Predictions of Species Relative Abundance from Maximum Entropy Optimisation." Oikos (119): 583-590. Tomonari-Tuggle, Myra J. and H. David Tuggle. 2008. Hōkūli‘a Preservation Plan for Non- Burial Sites at Hōkūli‘a, Districts of North and South Kona, Island of Hawai‘i (TMK 7-9- 12:4). Report Prepared for 1250 Oceanside Partners by International Archaeological Research Institute, Inc. ———. 1999. Hokuli‘a: An Integrated Archaeological Mitigation Plan. Honolulu, Hawaii: International Archaeological Research Institute, Inc. Valenzuela, Hector, Steven Fukuda, and Alton Arakaki. 2000. Sweetpotato Production Guidelines for Hawaii. Honolulu, Hawaii: Agricultural Publications Distribution Office, College of Tropical Agriculture and Human Resources. University of Hawai'i at Manoa. Vitousek, P. M., T. N. Ladefoged, P. V. Kirch, A. S. Hartshorn, M. W. Graves, S. C. Hotchkiss, S. Tuljapurkar, and O. A. Chadwick. 2004. "Soils, Agriculture, and Society in Precontact Hawai`i." Science 304 (5677): 1665-1669. doi:10.1126/science.1099619.
Abstract (if available)
Abstract
Archaeological investigations are an integral part of the permitting process for land development in Hawai‘i. The State recognizes that conservation of its historic and cultural heritage is important and that its cultural resources are nonrenewable. A recent archaeological survey for prehistoric and historic agricultural features in the Hōkūli‘a luxury development on Hawai‘i Island afforded an opportunity to test maximum entropy theory as a method for predicting the presence of archaeological features. The Maxent computer program uses a machine learning algorithm that utilizes presence point data and environmental variable rasters to produce a probability distribution of the species of interest. The ""species"" of interest for this research were agricultural clearing features, associated with sweetpotato (Ipomea batatas) cultivation, identified in a portion of Hōkūli‘a which lies within the Kona Field System. Previous agricultural habitat suitability models for Hawai‘i were used to determine the environmental variables used in this research
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model
PDF
Predicting the presence of historic and prehistoric campsites in Virginia’s Chesapeake Bay counties
PDF
A Maxent-based model for identifying local-scale tree species richness patch boundaries in the Lake Tahoe Basin of California and Nevada
PDF
Modeling prehistoric paths in Bronze Age Northeast England
PDF
Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama
PDF
Modeling burn probability: a Maxent approach to estimating California's wildfire potential
PDF
Using pattern oriented modeling to design and validate spatial models: a case study in agent-based modeling
PDF
Demonstrating GIS spatial analysis techniques in a prehistoric mortuary analysis: a case study in the Napa Valley, California
PDF
Effect of spatial patterns on sampling design performance in a vegetation map accuracy assessment
PDF
Creating a water quality geodatabase for the West Hawai‘i Island region
PDF
Using volunteered geographic information to model blue whale foraging habitat, Southern California Bight
PDF
Evaluating glacier movement fluctuations using remote sensing: a case study of the Baird, Patterson, LeConte, and Shakes Glaciers in central southeastern Alaska
PDF
Preparing for immigration reform: a spatial analysis of unauthorized immigrants
PDF
Questioning the cause of calamity: using remotely sensed data to assess successive fire events
PDF
Calculating solar photovoltaic potential on residential rooftops in Kailua Kona, Hawaii
PDF
Using Maxent modeling to predict habitat of mountain pine beetle in response to climate change
PDF
Using Landsat and a Bayesian hard classifier to study forest change in the Salmon Creek Watershed area from 1972–2013
PDF
Exploring urban change using historical maps: the industrialization of Long Island City (LIC), New York
PDF
Residential housing code violation prediction: a study in Victorville, CA using geographically weighted logistic regression
PDF
Applying least cost path analysis to search and rescue data: a case study in Yosemite National Park
Asset Metadata
Creator
Kailihiwa, Solomon Ha‘aheo, III
(author)
Core Title
Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i
School
College of Letters, Arts and Sciences
Degree
Master of Science
Degree Program
Geographic Information Science and Technology
Publication Date
07/01/2015
Defense Date
04/30/2015
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
archaeology,Hawaii,Hawaiian archaeology,Kona,Kona Field System,Maxent,OAI-PMH Harvest,prehistoric agriculture
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Kemp, Karen K. (
committee chair
), Garrison, Thomas G. (
committee member
), Lee, Su Jin (
committee member
)
Creator Email
kailihiw@usc.edu,kailihiwa@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-583369
Unique identifier
UC11301734
Identifier
etd-KailihiwaS-3533.pdf (filename),usctheses-c3-583369 (legacy record id)
Legacy Identifier
etd-KailihiwaS-3533.pdf
Dmrecord
583369
Document Type
Thesis
Format
application/pdf (imt)
Rights
Kailihiwa, Solomon Ha‘aheo, III
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
Tags
archaeology
Hawaiian archaeology
Kona Field System
Maxent
prehistoric agriculture