University of Southern California Dissertations and Theses

Integrating Landsat and California pesticide exposure estimation at aggregated analysis scales: accuracy assessment of rurality

(USC Thesis Other)

Integrating Landsat and California pesticide exposure estimation at aggregated analysis scales: accuracy assessment of rurality

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

INTEGRATING LANDSAT AND CALIFORNIA PESTICIDE EXPOSURE
ESTIMATION AT AGGREGATED ANALYSIS SCALES:
ACCURACY ASSESSMENT OF RURALITY

by

Trang Minh VoPham

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)

May 2014

Copyright 2014 Trang Minh VoPham

ii

DEDICATION
This thesis is dedicated to my family - especially Matthew David Weaver - for their
relentless support, and to my University of Pittsburgh epidemiology doctoral mentor, Joel
L. Weissfeld, MD, MPH, who engrained in me a great love and appreciation for all things
geospatial.

iii

ACKNOWLEDGEMENTS
I sincerely thank my thesis advisor, John Wilson, PhD, and committee members, Darren
Ruddell, PhD and Tarek Rashed, PhD, for their invaluable expertise and guidance.

iv

TABLE OF CONTENTS
Dedication ii

Acknowledgements iii

List of Tables vii

List of Figures ix

List of Abbreviations xii

Abstract xv

Chapter One: Introduction 1
1.1 Harnessing Epidemiology and the Spatial Sciences 2

Chapter Two: Background 6
2.1 Residential Proximity to Agricultural Pesticide Applications 6
2.2 Evolution of Pesticide Exposure Estimation 8
2.2.1 GIS-Based Pesticide Exposure Metrics 10
2.2.2 Advantages of GIS-Based Metrics in Cancer
Epidemiology 11
2.2.3 Improvements in GIS-Based Pesticide Exposure
Methods over Time 12
2.2.4 Inception of GIS-Based Pesticide Exposure Metrics:
Crop Maps 13
2.2.5 Utilizing the California Pesticide Use Reports (PURs) 15
2.2.6 Enhancement of PUR-Derived Metrics Using Land
Use Surveys 18
2.2.7 Downscaling PUR-Derived Metrics Using Landsat
Satellite Imagery 22
2.3 Surrogate Measures of Pesticide Exposure: Rurality 24
2.3.1 Misclassification of Pesticide Exposure 25
2.3.2 Variation in Rurality Definitions 26
2.3.3 Variation in Analysis Scales 26

Chapter Three: Methods and Data Sources 29
3.1 Research Hypotheses 30
3.2 Study Area: Kern County, California 30
3.3 Data Sources 32
3.3.1 Pesticide Exposure Data 32
3.3.2 Landsat Imagery 36
3.3.3 Rural-Urban Commuting Area Codes 37
3.3.4 U.S. Census Bureau Urban-Rural Classification 39

v

3.4 Pesticide Exposure Estimation 40
3.4.1 Preparation of PUR, PLSS, and Land Use Survey Data 40
3.4.2 Incorporation of Landsat Imagery: Crop Signature
Library (CSL) 43
3.4.3 Classification of 1985 Landsat Imagery 46
3.4.4 Modified Three-Tier Approach to Estimate Pesticide
Exposure 49
3.5 Rurality Metrics 50
3.6 Statistical Analysis 51

Chapter Four: Results 53
4.1 PUR Extraction 53
4.2 Crop Signature Library (CSL) 57
4.2.1 Stratified Random Sampling (SRS) 64
4.3 Classification of 1985 Landsat Imagery 69
4.3.1 Segmentation 69
4.3.2 Principal Component Analysis (PCA) 72
4.3.3 Classification Using Sum of Squared Difference (SSD) 72
4.3.4 Processing CSL-Classified Crop Fields 76
4.4 Modified Three-Tier Approach 78
4.4.1 Contribution of Landsat Imagery to Modified
Three-Tier PUR Matching 82
4.5 Annual Pesticide Application Rates by Areal Aggregation 90
4.6 Descriptive Analysis: Areal Aggregation and Pesticide Exposure 95
4.7 Kern County Rurality 97
4.7.1 Accuracy Assessment of Rurality 102

Chapter Five: Discussion and Conclusions 112
5.1 Critical Assessment of Methods and Results: Strengths and
Limitations 113
5.1.1 PUR Processing 113
5.1.2 Crop Signature Library (CSL) 113
5.1.3 Classification of 1985 Landsat Imagery 116
5.1.4 Segmentation 118
5.1.5 Modified Three-Tier Approach: Pesticide Exposure 119
5.1.6 Impact of Areal Aggregation on Annual
Pesticide Application Rates 122
5.1.7 Accuracy Assessment of Rurality 122
5.2 Feasibility and Informational Gain of Landsat Remote Sensing 129
5.3 Alternative Approaches to Integrating Landsat in Pesticide
Exposure Estimation 130
5.4 Significance of Results 132
5.5 Future Directions 133
5.6 Summary 134

References 136

vi

Appendices
Appendix A: Pesticide Database 144
Appendix B: Pesticide Use Report Processing 149
Appendix C: Landsat Mosaics, 1990 152
Appendix D: Crop Signature Library 154
Appendix E: Segmentation and Classification 186
Appendix F: Applied Pesticides and Rurality 191

vii

LIST OF TABLES
Table 1: Common pesticide-treated crops in Kern County, 2011 32
Table 2: Data sources 34
Table 3: Landsat 4 and 5: Thematic Mapper (TM) sensor 36
Table 4: Landsat 4 and 5 remote sensing characteristics 37
Table 5: Kern County agricultural use and chemical class PUR
extractions 54

Table 6: Pesticide-treated crops by chemical class,
Kern County (1974-1990) 56

Table 7: Common pesticides by chemical class,
Kern County (1974-1990) 57

Table 8: Landsat images from 1990 used for crop signature library 58

Table 9: Eligibility criteria for SRS 65

Table 10: Land use classes excluded from SRS due to multiuse 65

Table 11: Landsat images from 1985 used for classification 70

Table 12: Principal component analysis of Landsat 1985 NDVI images 72

Table 13: CSL classification approaches for segmented crop layer 75

Table 14: Classification: minimum sum of squared differences (SSD) 75

Table 15: Organochlorines: Tiers 1 and 2A matched crops 82

Table 16: Organophosphates: Tiers 1 and 2A matched crops 83

Table 17: Carbamates: Tiers 1 and 2A matched crops 84

Table 18: Pesticide-treated crop fields and sections intersecting areal
units 96

Table 19: Annual pesticide application rates according to areal
Aggregation 96

viii

Table 20: RUCA and U.S. Census Bureau metric designations by
areal aggregation 99

Table 21: ZCTA vs. census tract rurality designations 101

Table 22: Pesticide rates stratified by rurality: ZCTAs 102

Table 23: Pesticide rates stratified by rurality: census tracts 103

Table 24: ZCTA-level accuracy of RUCA codes: organochlorines 104

Table 25: ZCTA-level accuracy of U.S. Census Bureau urban-rural
classification: organochlorines 104

Table 26: ZCTA-Level accuracy of RUCA codes: organophosphates 105

Table 27: ZCTA-level accuracy of U.S. Census Bureau urban-rural
classification: organophosphates 105

Table 28: ZCTA-level accuracy of RUCA codes: carbamates 106

Table 29: ZCTA-level accuracy of U.S. Census Bureau urban-rural
classification: carbamates 106

Table 30: Census tract-level accuracy of RUCA codes:
organochlorines 108

Table 31: Census tract-level accuracy of U.S. Census Bureau
urban-rural classification: organochlorines 108

Table 32: Census tract-level accuracy of RUCA codes:
organophosphates 109

Table 33: Census tract-level accuracy of U.S. Census Bureau
urban-rural classification: organophosphates 109

Table 34: Census tract-level accuracy of RUCA codes: carbamates 110

Table 35: Census tract-level accuracy of U.S. Census Bureau
urban-rural classification: carbamates 110

ix

LIST OF FIGURES
Figure 1: Kern County, California, study area of interest 31

Figure 2: PLSS sections in Kern County 35

Figure 3: Kern County land use survey, 1990 35

Figure 4: Urbanized Areas and Urban Clusters across California, 2000 38

Figure 5: Kern County UAs and UCs, 2000 39

Figure 6 Methodological workflow: PUR, land use survey, and PLSS
processing 40

Figure 7: Methodological workflow: Landsat remote sensing crop
signature library 43

Figure 8: Landsat Path-Row scenes intersecting Kern County 44

Figure 9: Methodological workflow: classification of Landsat images 46

Figure 10: Modified three-tier pesticide exposure method 49

Figure 11: Pounds of agricultural pesticide usage in Kern County by
chemical class (1974-1990) 55

Figure 12: Agricultural PUR pesticide applications in Kern County by
chemical class (1974-1990) 55

Figure 13: Landsat mosaic (band 3), Paths 41-42 and Rows 35-36,
from October 1990 cropped to Kern County 59

Figure 14: Inset of Landsat mosaic (band 3) from October 1990,
showing crop fields in Kern County 60

Figure 15: Inset of Landsat mosaic (band 4) from October 1990,
showing crop fields in Kern County 60

Figure 16: Landsat mosaic (band 4), Paths 41-42 and Rows 35-36,
from October 1990 cropped to Kern County 61

Figure 17: Inset of NDVI image created from red and near infrared
Landsat bands, October 1990 62

Figure 18: NDVI image cropped to Kern County, October 1990 63

x

Figure 19: Cloud-free zone of 1990 Landsat images available for CSL 66

Figure 20: Land use survey polygons sampled via SRS,
Kern County, 1990 67

Figure 21: Median NDVI values for select SRS-sampled land use
survey polygons, October 1990 69

Figure 22: Segmentation-eligible zone vs. cloud-free CSL zone,
overlaying Landsat mosaic (band 3) from September 1985 71

Figure 23: Segments of spectrally homogeneous pixels, basis of crop
field boundaries for classifying 1985 Landsat NDVI images 73

Figure 24: Segments overlaying color-infrared Landsat image from
August 1985 74

Figure 25: Classification 2-derived segments prior to processing 77

Figure 26: Finalized classification 2-derived segments subsequent to
processing 79

Figure 27: Organochlorine PUR tier matches, Kern County
(1974-1990) 81

Figure 28: Organophosphate PUR tier matches, Kern County
(1974-1990) 82

Figure 29: Carbamate PUR tier matches, Kern County (1974-1990) 82

Figure 30: Tier2A match provided by Landsat, organophosphate
PUR applications, 1974-1990 87

Figure 31: Organochlorines: applied pesticides on crop fields and
sections, Kern County (1974-1990) 88

Figure 32: Organophosphates: applied pesticides on crop fields and
sections, Kern County (1974-1990) 89

Figure 33: Carbamates: applied pesticides on crop fields and sections,
Kern County (1974-1990) 90

Figure 34: Kern County ZCTAs 91

xi

Figure 35: Organochlorines: ZCTA-level annual pesticide application
rates, Kern County (1974-1990) 92

Figure 36: Organophosphates: ZCTA-level annual pesticide application
rates, Kern County (1974-1990) 92

Figure 37: Carbamates: ZCTA-level annual pesticide application rates,
Kern County (1974-1990) 93

Figure 38: Kern County census tracts 94

Figure 39: Organochlorines: census tract-level annual pesticide
application rates, Kern County (1974-1990) 94

Figure 40: Organophosphates: census tract-level annual pesticide
application rates, Kern County (1974-1990) 95

Figure 41: Carbamates: census tract-level annual pesticide application
rates, Kern County (1974-1990) 95

Figure 42: ZCTA-level rurality 99

Figure 43: Census tract-level rurality 101

xii

LIST OF ABBREVIATIONS
AI Active ingredient
CA California
CDPR California Department of Pesticide Regulation
CDWR California Department of Water Resources
CIR Color-infrared
CO-MTRS County, meridian, township, range, and section
CSL Crop signature library
DDE dichlorodiphenyldichloroethylene
DDT dichlorodiphenyltrichloroethane
DEM Digital Elevation Model
EPA Environmental Protection Agency
ETM+ Enhanced Thematic Mapper Plus
FIPS Federal Information Processing Standard
FSA Farm Service Agency
GCP Ground control point
GIS Geographic information system
GloVis Global Visualization
GPS Global Positioning System
ISODATA Iterative Self-Organizing Data Analysis Technique
L1T Level 1 Standard Terrain Correction product
LPGS Level 1 Product Generation System
MAUP modifiable areal unit problem

xiii

MODIS Moderate Resolution Imaging Spectroradiometer
MS Multispectral
MSS Multispectral Scanner
NAD83 North American Datum (1983)
NASA National Aeronautics and Space Administration
NDVI Normalized Difference Vegetation Index
NIR Near infrared band
NLAPS National Land Archive Production System
OCP Organochlorine
OP Organophosphate
PCA Principal component analysis
PEG Parkinson’s Environment and Gene study
PLSS Public Land Survey System
PUR Pesticide Use Report
R Red band
RGB red-green-blue
RHRC Rural Health Research Center
RUCA Rural-Urban Commuting Area
SSD Sum of squared difference
SRS Stratified random sampling
TM Thematic Mapper
UA Urbanized Area
UC Urban Cluster

xiv

USGS U.S. Geological Survey
ZCTA ZIP Code Tabulation Area
ZIP code Zone Improvement Plan code
WGS84 World Geodetic System (1984)

xv

ABSTRACT
Pesticide exposure estimation in epidemiologic studies can be constrained to analysis
scales commonly available for cancer data - census tracts and ZIP codes. Research goals
included (1) demonstrating the feasibility of modifying an existing geographic
information system (GIS) pesticide exposure method using California Pesticide Use
Reports (PURs) and land use surveys to incorporate Landsat remote sensing and to
accommodate aggregated analysis scales, and (2) assessing the accuracy of two rurality
metrics (quality of geographic area being rural), Rural-Urban Commuting Area (RUCA)
codes and the U.S. Census Bureau urban-rural system, as surrogates for pesticide
exposure when compared to the GIS gold standard. Segments, derived from 1985 Landsat
NDVI images, were classified using a crop signature library (CSL) created from 1990
Landsat NDVI images via a sum of squared differences (SSD) measure. Organochlorine,
organophosphate, and carbamate Kern County PUR applications (1974-1990) were
matched to crop fields using a modified three-tier approach. Annual pesticide application
rates (lb/ac), and sensitivity and specificity of each rurality metric were calculated. The
CSL (75 land use classes) classified 19,752 segments [median SSD 0.06 NDVI]. Of the
148,671 PUR records included in the analysis, Landsat contributed 3,750 (2.5%)
additional tier matches. ZIP Code Tabulation Area (ZCTA) rates ranged between 0 and
1.36 lb/ac and census tract rates between 0 and 1.57 lb/ac. Rurality was a mediocre
pesticide exposure surrogate; higher rates were observed among urban areal units. ZCTA-
level RUCA codes offered greater specificity (39.1-60%) and sensitivity (25-42.9%). The
U.S. Census Bureau metric offered greater specificity (92.9-97.5%) at the census tract
level; sensitivity was low (≤6%). The feasibility of incorporating Landsat into a modified

xvi

three-tier GIS approach was demonstrated. Rurality accuracy is affected by rurality
metric, areal aggregation, pesticide chemical class, and pesticide exposure cutoff. Future
research should explore integrating Landsat for higher spatial resolution pesticide
exposure estimation.

1

CHAPTER ONE: INTRODUCTION
The goals of this research were: (1) to demonstrate the feasibility of using Landsat remote
sensing in improving the spatiotemporal resolution of an established GIS-based pesticide
exposure method modified to accommodate cancer data analysis scales; and (2) to
evaluate the validity of two different measures of rurality as indicators of ZIP Code
Tabulation Area (ZCTA)- and census tract-level pesticide exposure and determine which
measure of rurality is a more accurate surrogate of pesticide exposure.
The feasibility of integrating Landsat remote sensing and pesticide exposure
estimation is defined as demonstrating that agricultural pesticide application data can be
matched to Landsat-derived crop fields. Specifically, an established GIS-based method
matches California pesticide application data to land use survey agricultural crop fields to
subsequently estimate human pesticide exposure (Rull and Ritz 2003). As land use
surveys are intermittently updated for California counties every seven to 10 years,
Landsat remote sensing, which captures Earth imagery every 16 to 18 days, provides an
opportunity to produce agricultural crop fields during years lacking land use surveys
(USGS 2013b). Demonstrating that pesticide application data - not matching to land use
survey crops fields - can be matched to Landsat imagery classified into agricultural crops
bolsters the feasibility of using remote sensing in pesticide exposure estimation.
An accuracy assessment of rurality addresses potential exposure misclassification
when using rurality to quantify pesticide exposure – which is important to consider in the
context of epidemiologic studies. Epidemiologic studies seeking to elucidate the
relationship between pesticide exposure and human health outcomes have employed
rurality as a surrogate measure of pesticide exposure, which takes advantage of the high

2

prevalence of agricultural pesticide applications occurring in rural geographic areas
(Franklin and Worgan 2005). Usage of rurality will inevitably misclassify some
geographic areas as pesticide-exposed and vice versa, and the exact quantification of such
misclassification has never before been addressed.
Location is the fundamental focus of the spatial sciences - a multifaceted force
that influences human society (Waller and Gotway 2004). One avenue through which
location impacts society is through the environment playing a direct role in human health.
Where an individual lives and where an individual works has a direct effect on future
health outcomes (Pickle et al. 2005). The environment is defined as exogenous factors
nonessential to the normal functioning of human beings, and includes physical, chemical,
and biological agents, in addition to the social, cultural, and political factors interacting
with these agents (Rothman et al. 2008). Especially for individuals living in rural areas,
the environment is associated with potential residential exposure to chemicals (Ward et
al. 2000; Franklin and Worgan 2005). One group of chemicals that has played a
prominent role in adversely affecting human health is pesticides (Alavanja et al. 2004).
Pesticides are chemicals used to treat pests, such as insects (EPA 2012). In both 2006 and
2007, approximately 5.1 billion lb of total pesticides were used in the U.S. (EPA 2011).
Pesticides have been specifically linked to the development of many types of cancers
(Alavanja et al. 2004), which pose a large public health burden as the second leading
cause of death in the U.S. (CDC 2012). In 2012, there were an estimated 1,638,910 cases
of cancer diagnosed in the U.S. (NCI 2012). An approach that combines methods and
concepts from epidemiology and the spatial sciences will be able to better understand the

3

exact role of pesticides in cancer development and to confront the significant public
health burden of cancer.

1.1 Harnessing Epidemiology and the Spatial Sciences
Epidemiology is a branch of science that seeks to elucidate the role that an exposure, such
as pesticides, may play in the development of disease (Szklo and Nieto 2007).
Epidemiology can involve carrying out research studies designed to provide an unbiased
measure of association between a purported exposure and risk of a particular disease.
Results from epidemiologic studies serve many segments of society, from informing
policy makers to being a platform upon which other researchers build their research. The
capacity to provide an unbiased measure of how an exposure is truly associated with a
disease partly hinges on the validity of the measure used to indicate the exposure.
Validity is the extent to which a measure is an indicator of what it is intended to measure
(Szklo and Nieto 2007). Compared to a gold standard, the validity of a surrogate measure
can be quantified through determining its sensitivity and specificity. Sensitivity refers to
the capacity of a measure (e.g. rurality) to correctly identify features with a characteristic
of interest [e.g. ZCTA- or census tract-level agricultural application of pesticides] (Szklo
and Nieto 2007). Specificity refers to the capacity of a measure to correctly identify
features without a characteristic of interest. A surrogate measure with low sensitivity will
produce more false negatives, while a surrogate measure with low specificity will
produce more false positives. Depending on how frequently the exposure occurs in the
general population and whether sensitivity and/or specificity is affected, an
epidemiologic study using a particular surrogate measure may report an exposure

4

conferring less risk for an outcome than it truly does, or greater risk (i.e. biasing results
towards or away from the null hypothesis of no association) (Szklo and Nieto 2007).
However, research endeavors are often limited in resources and time, and
surrogate measures are frequently employed. Specifically, the main limitation of
determining human exposure to pesticides has been inadequate methods of ascertaining
past exposure (Franklin and Worgan 2005). Methods of determining pesticide exposure
are either qualitative (e.g. self-reported pesticide exposure, occupation, etc.) or
quantitative (e.g. direct biological measurement). A frequently used surrogate measure of
pesticide exposure has been rurality, defined as the quality of a geographic area being
rural (Alavanja et al. 2004; Rural Assistance Center 2012). Rurality may be self-reported
(i.e. qualitative) or determined using various objective criteria, such as population density
(i.e. quantitative). Rural areas are typically associated with agricultural activities, which
are a primary source of pesticide exposure (Franklin and Worgan 2005). Pesticides are
applied on agricultural lands and residential proximity to applications may contribute to
pesticide exposure by way of applied pesticides drifting from intended sites (Ward et al.
2000; Rull and Ritz 2003). However, rurality is an imperfect surrogate of pesticide
exposure likely misclassifying some rural areas as pesticide-exposed when they are truly
not and vice versa.
On the other hand, GIS-based pesticide exposure methods are quantitative
approaches that are increasing in usage (Ritz and Rull 2008; Maxwell et al. 2010a),
which offer an objective alternative to qualitative measures, capable of using data from
multiple databases containing relevant pesticide information and determining historical
exposure through incorporation of many years of data (Alavanja et al. 2004). GIS-based

5

methods represent a potentially cost-effective approach to ascertaining pesticide exposure
compared to other methods, such as measurement and determination of serum levels of a
pesticide. A GIS-based approach is also superior to qualitative measures, such as self-
reported pesticide exposure, which are subject to recall bias (Alavanja et al. 2004;
Franklin and Worgan 2005). GIS-based pesticide exposure methods have expanded to
incorporate remote sensing, such as satellite-borne imagery, due to the spatiotemporal
resolution offered by these data types (Maxwell et al. 2010b; Maxwell 2011). However,
there exists a paucity of literature using remotely sensed data in cancer research
examining environmental exposures, which stands to gain from the high spatiotemporal
resolution of remote sensing data in reconstructing historical exposures (Maxwell et al.
2010a).
Pesticides pose potentially harmful human health effects and have been associated
with the development of chronic diseases, such as cancer (Dich et al. 1997; Alavanja et
al. 2004). In order to adequately determine whether or not pesticide exposure is truly
associated with a health outcome in epidemiologic studies, whether or not a surrogate
measure used to indicate pesticide exposure is truly valid must first be addressed. GIS,
guided by fundamental principles underlying epidemiology and the spatial sciences,
provides a powerful way to combine relevant spatial and non-spatial data to address this
research question in ways that would be not otherwise possible. This research addresses
two gaps in the literature regarding: (1) the lack of cancer epidemiology studies using
remote sensing in environmental exposure assessment observed in Maxwell, Meliker, et
al. (2010); and (2) the absence of research examining the validity of rurality as an
indicator of pesticide exposure.

6

CHAPTER TWO: BACKGROUND
Pesticides are pervasive chemicals designed to be toxic to organisms, such as insects,
herbs, and fungi (Blair 1988; EPA 2012), and are grouped into functional classes
according to which organisms they control, such as insecticides, herbicides, and
fungicides. Pesticides are also further categorized into chemical classes according to their
chemical structure and biological mechanisms of action, such as organochlorine
pesticides (OCPs) (e.g. DDT and endosulfan) (Alavanja et al. 2004). Pesticides are
composed of their active ingredient, in addition to other ingredients, such as solvents,
which comprise the pesticide products available in the market. Pesticides are used most
frequently in agriculture, horticulture, and vector control (e.g. antimalarial), followed by
forestry and livestock production (Dich et al. 1997). Exposure to pesticides occurs
through direct, higher-level routes, such as through occupation, and indirect, lower-level
but more frequent routes, such as through drinking water, food, air, and dust. In the U.S.,
the primary source of exposure is from consumption of dairy, fish, meat, and poultry
products (Ritz and Costello 2006). Food is potentially contaminated with pesticides
during production, storage, and/or transport processes (Oates and Cohen 2011).

2.1 Residential Proximity to Agricultural Pesticide Applications
An important source of pesticide exposure occurs through residential proximity to
pesticide applications on agricultural lands, as pesticides may drift from intended sites
through the ground and the air via spray drift and post-application volatilization to far
locations (i.e. pesticide residues change from liquid to gas/vapor form in various climatic
conditions) (Rull and Ritz 2003; Alavanja et al. 2007; Ritz and Rull 2008; EPA 2009).

7

The most vulnerable populations are those residing in rural areas and farming families
(Ward et al. 2000), who are potentially exposed through dermal contact and ingestion of
pesticides in household dust and in groundwater (Gunier et al. 2001). Farm families
frequently reside within 100 yd (91 m) of crop fields (Ward et al. 2000). Aerial pesticide
applications can drift between 500 and 1,000 m and boom-type sprayers [application via
spray nozzles at regular intervals (Ministry of Agriculture 2013)] can drift between 300
and 800 m (Ward et al. 2000). Pesticides can enter homes through drift and dust on shoes
and farmers’ clothing (Ritz and Rull 2008). Residential proximity, as a route of exposure,
poses a potentially large threat to human health, as pesticides are less likely to degrade
and volatilize in homes due to the absence of moisture, sunlight, and microorganisms,
and are able to persist over time (Ritz and Rull 2008; Gunier et al. 2011).
Although residential proximity to agricultural pesticide applications is itself a
surrogate of actual human exposure, it has been directly tied to levels of pesticides in
carpet dust samples (Gunier et al. 2011). Carpet dust samples from 89 residences in
agricultural areas in California were collected in a determinants of exposure study
(Franklin and Worgan 2005). Gunier et al. (2011) demonstrated that residential proximity
to agricultural pesticides, measured as the application of six pesticides within a 1,250 m
residential buffer [geographic coordinates captured using Global Positioning System
(GPS) device] using pesticide application and land use data 730 days before dust
collection was significantly correlated (p<0.05) with concentrations of pesticides in
carpet dust (ng/g). Spearman rank correlation coefficients ranged between 0.23 and 0.50.
Therefore, measurement of residential proximity to agricultural pesticide applications,

8

captured using a GIS incorporating spatial information, is a potentially meaningful way
to capture human exposure.

2.2 Evolution of Pesticide Exposure Estimation
Substantial progress has been achieved in pesticide exposure assessment by moving from
self-reported measures towards more sophisticated, objective metrics. However, the
majority of epidemiologic studies investigating the association between pesticide
exposure and human health have employed interview-administered methods, which are
associated with various limitations - most prominently recall bias (Franklin and Worgan
2005). Most exposure metrics have also relied on surrogates of true human exposure
(Nuckols et al. 2004). In the context of pesticides, environmental concentration is the
presence of the pesticide in a carrier medium, such as in the air. Exposure concentration
is the presence of the pesticide at the point of contact, such as in the zone of breathing.
Dose is the amount of the pesticide that enters the human body (i.e. absorbed). In
epidemiologic studies, measures of environmental concentration have typically served as
surrogates for exposure concentration and dose.
Pesticide exposure metrics are categorized as qualitative or quantitative (Franklin
and Worgan 2005). Qualitative metrics are derived from questionnaires and interviews,
such as self-reported occupational history, residential locations, and exposure to
pesticides (e.g. garden and residential use). Specific examples include occupation as a
farmer (ever and duration), type of crop raised (duration and acres), and application of
any pesticide (ever and duration) (Alavanja et al. 2004). However, a limitation of
qualitative metrics is the absence of identification of specific pesticides. Recall also

9

varies according to subpopulations, such as farmers likely having better recall as they
may directly participate in the purchase and/or application of pesticides. In comparison,
migrant farmworkers and those occupying pesticide-treated residences may not be able to
recall, or may be unaware of, specific pesticide names (Alavanja et al. 2004). To address
the limitation of recall bias, self-reported measures can be supplemented with a review
from experts, such as occupational hygienists.
Quantitative metrics are derived through direct measurement of external exposure
from the air (i.e. environmental monitoring), or from biological markers in serum, urine,
fat, etc. (i.e. biological monitoring/biomonitoring) (Franklin and Worgan 2005; Alavanja
et al. 2007). Measuring pesticide levels in carpet dust is an example of environmental
monitoring. Biomonitoring is considered the gold standard approach that provides a
measure of human pesticide exposure from all pathways and routes, and is advantageous
when the chemical of interest has a long biological half-life and when its concentration is
not affected by disease. Quantitative metrics also include usage of exposure databases,
such as those collecting information for pesticide regulation purposes. Integrated
pesticide exposure metrics have also been developed, which combine self-reported
information with other relevant data, such as personal protective equipment, to better
estimate exposure (Alavanja et al. 2004). For example, job-exposure matrices (JEMs) are
integrated metrics that are typically region-based (e.g. British JEM) and incorporate
information regarding job title, tasks, and industry to estimate exposure intensity.
Exposure intensity algorithms are an extension of JEMs, which weight cumulative
pesticide exposure by chemical- and applicator-specific information, such as work
practices.

10

2.2.1 GIS-Based Pesticide Exposure Metrics
Built upon the concern of high and persistent pesticide exposure among rural residents,
inadequacies of frequently employed qualitative metrics, and the potentially harmful
effect of pesticides on human health, there has been a burgeoning body of research that
focuses specifically on utilizing the concepts and techniques of the spatial sciences to
improve pesticide exposure assessment. Geographic information system (GIS)-based
approaches represent a quantitative method of pesticide exposure ascertainment
integrating different sources of spatial and non-spatial information, such as the California
Department of Water Resources (CDWR) land use surveys and the California
Department of Pesticide Regulation (CDPR) Pesticide Use Reports (PUR) database
(Alavanja et al. 2004; Nuckols et al. 2004; Franklin and Worgan 2005). GIS-based
metrics improve upon the limitations of existing methodologies. Specifically, recall bias,
prominent in qualitative methods, is minimized through combining objectively acquired
information, such as remotely sensed data. GIS-based metrics can be used to determine
pesticide exposure levels for the general population, as individuals are likely unaware of
agricultural pesticides close to their residence (Ward et al. 2000; Alavanja et al. 2007).
Using GIS also represents a cost-effective and time-efficient approach to assessing
exposure, compared to collecting and measuring biological samples in a large enough
study sample with adequate statistical power to detect meaningful differences.
Furthermore, many pesticides have short biological half-lives, and their biologic
measurement, though useful in assessing recent exposure, may be irrelevant in attempting
to determine past exposure that may have precipitated a chronic disease (Franklin and
Worgan 2005).

11

2.2.2 Advantages of GIS-Based Metrics in Cancer Epidemiology
GIS-based pesticide exposure metrics are especially powerful tools in the context of
cancer epidemiology. Methods of measuring pesticide exposure should consider many
important underlying issues in relation to the study of chronic diseases, such as cancer
(Franklin and Worgan 2005). Cancer is frequently associated with long latency periods
(i.e. time between first exposure and clinical diagnosis of disease), typically 20 years or
more (Blair 1988; Rothman et al. 2008). Historical reconstruction of past exposure is
important in capturing the potential effect of a latency period. When possible, exposure
assessment should precede the onset of disease to showcase a temporal relationship,
which is important evidence of a causal relationship (Alavanja et al. 2004; Szklo and
Nieto 2007). Multiple routes of exposure exist, such as dermal, inhalational, and oral.
Furthermore, individuals are potentially exposed to a variety of pesticides. Specifically,
agricultural workers are likely exposed to multiple pesticides over the crop-growing
season. Depending on the crop type, pesticides can be applied in combination via tank
mixes and over multiple time points during the growing season, which pose difficulties in
determining the impact of a specific pesticide on disease (Franklin and Worgan 2005).
GIS-based pesticide exposure metrics are able to address all of the
aforementioned issues. Through incorporating multiple data sources with locational
information and specific chemicals, often spanning long time periods, meaningful and
relevant measures of human exposure to pesticides can be derived. The following are
potential sources of pesticide exposure that are addressed in using a GIS-based metric:
inhalation of ambient air, persistence in household dust, “take-home” of pesticides from
occupations, soil drift, groundwater contamination, dermal contact in fields, and direct

12

ingestion of contaminated produce (Gunier et al. 2001). Most, if not all, of these potential
sources of exposure are associated with residential proximity to agricultural pesticide
applications.

2.2.3 Improvements in GIS-Based Pesticide Exposure Methods over Time
GIS-based pesticide exposure metrics have improved over time through addressing
fundamental concepts underlying the spatial sciences. GIS allows for the capacity to
combine many spatial data sources, which are often associated with different data
representations (i.e. data models), scales, and levels of accuracy (Nuckols et al. 2004).
Spatial data can be represented as vector data models, which represent entities as points,
lines, and polygons, and are typically associated with real-world phenomena with
discrete, unambiguous boundaries, and as raster data models, such as satellite imagery,
representing data through pixels, or cells, which are better-suited for continuous
phenomena (Waller and Gotway 2004). Data sources are available at different scales, or
spatial resolutions/granularities (i.e. smallest distinguishable and/or mappable unit). Scale
can also refer to analysis scale (i.e. how phenomena are measured/aggregated) and
operational/phenomenon scale (i.e. scale at which a process of interest operates)
(Montello 2001; Nuckols et al. 2004). A data source may be aggregated to a particular
analysis scale that is not relevant to the underlying geographic process of interest it is
attempting to represent. Lastly, error can emanate from positional error (i.e. inaccuracies
in locational information), attribute error (i.e. inaccuracies in data describing specific
locations), and temporal error (i.e. mismatches in temporal currency of data).

13

Geospatial pesticide exposure methods have taken these fundamental issues into
account. Different data models necessitate different types of analytic tools, and advances
in technology and technical knowledge have facilitated the use and analysis of different
data types in GIS environments. Multiple data sources, often times collected for purposes
unrelated to research, have been combined to ultimately provide improved spatiotemporal
resolution. Specifically, the field of GIS-based pesticide exposure assessment has grown
to include high-resolution remote sensing technology, such as aerial photographs and
Landsat satellite imagery, land use surveys, and pesticide exposure databases, to improve
the spatiotemporal resolution of capturing individual-level residential exposure to
agricultural pesticides.

2.2.4 Inception of GIS-Based Pesticide Exposure Metrics: Crop Maps
Pioneering the use of GIS in pesticide exposure assessment,Ward et al. (2000) conducted
a feasibility study to determine the extent to which Landsat satellite imagery could be
used to reconstruct historical crop patterns. Validated using Nebraska Farm Service
Agency (FSA) aerial photographs with annotated crop information, a historical land
cover map of Adams, Buffalo, and Hall Counties in Nebraska was created using a
Landsat multispectral (MS) image from 1984. Six agricultural land cover types (i.e. corn,
sorghum, soybeans, alfalfa, rangeland, and bare soil) were screen-digitized using the FSA
records. Crop-specific probabilities of pesticide use were determined using information
from surveyed farmers of the University of Nebraska Agricultural Extension Service and
usual number of applications of each pesticide from the U.S. Environmental Protection
Agency (EPA) Biological and Economic Analysis Division of the Office of Pesticide

14

Programs. After creating 500 m buffers around geocoded [i.e. assigning a geographic
location to an address record (Waller and Gotway 2004)] residences of study subjects
from a non-Hodgkin lymphoma study, exposure to crop pesticides applied to one or more
major crop types and the (average) distance from each residence to crop field centroid(s)
within the buffer were calculated.
Ward et al. (2000) demonstrated that Landsat remote sensing could provide useful
information relevant to studies seeking to quantify pesticide exposure. For example, the
authors showed rural residences (N=10; outside of town boundaries) had a greater
number of crop fields proximate to their homes and were closer in distance to crop fields
compared to community residences (N=97; located within a town boundary).
Specifically, 100% of rural residences vs. 15% of community residences had at least one
crop field within a 500 m residential buffer. The median distance to crop field centroids
was 378.3 m for rural residences vs. 419.9 m for community residences.
Ward et al. (2006) extended the previous work to determine if there is an
association between residential proximity to agricultural fields and indoor pesticide
concentrations that could adversely affect human health. The authors evaluated the utility
of crop maps to predict crop herbicide levels from residential carpet dust samples. Using
collected vacuum cleaner dust from study subjects of a non-Hodgkin lymphoma study in
Iowa, 14 herbicides were measured. Landsat MS images, validated using FSA records,
were used to create land cover maps to identify corn and soybean fields between 1998
and 2000. Among 112 residences with locations recorded using GPS devices, 58% had
crops within 500 m of their home. Sixty-one percent of rural residences had detectable
levels of herbicides in carpet dust, compared to 15% of in-town residences. The odds of

15

detecting at least one agricultural herbicide was 7.4 [95% confidence interval (CI) 1.3-
41.3] times greater among residences with more than 300 ac of corn and soybean fields
within 750 m compared to no crops within 750 m, adjusted for agricultural jobs.
There was also a significant increase in concentration of agricultural herbicides in
carpet dust (ng/g) per ac increase of crops within 500 to 750 m of a residence [β
(regression coefficient) 1.01; 95% CI 1.00-1.02], adjusted for other buffer distances. The
geometric mean of agricultural herbicides measured in house dust increased with recent
agricultural work, with levels of approximately 366 ng/g [geometric standard deviation
(SD) 4.6] in homes with current agricultural workers, 121.9 ng/g (geometric SD 2.5) with
former agricultural workers, and 111.5 ng/g (geometric SD 2.5) with no agricultural
workers. The authors interpreted their findings as confirmation of the “take-home”
pathway of exposure for families living with an agricultural worker who potentially
exposes family members to pesticides from clothing, shoes, etc.

2.2.5 Utilizing the California Pesticide Use Reports (PURs)
Bell et al. (2001) and Gunier et al. (2001) forged more direct approaches to estimating
potential residential pesticide exposure by utilizing the comprehensive California
Pesticide Use Reports (PUR) database (CDPR 2013). Rather than assuming all crop
fields are treated similarly with respect to pesticide applications, these authors shifted
their focus towards California, which is both agriculturally productive and has legally
required protocols for growers and applicators to report use of all restricted-use pesticides
since 1974 and all pesticides since 1990 (Bell et al. 2001; Gunier et al. 2001; CDPR
2013). PURs include information regarding specific applications of pesticides, such as

16

the name and pounds of pesticide active ingredient applied, acres treated, date of
application, and Public Land Survey System (PLSS) section location of application. The
PLSS system is used to divide and describe U.S. lands for surveying purposes, and
imposes a grid of square sections measuring 1 mi on a side spanning the entire U.S.
(National Atlas 2013).
Bell et al. (2001) examined the relationship between maternal residential
proximity to pesticide applications and fetal death due to congenital anomalies across
Fresno, Kern, Kings, Madero, Merced, Monterey, Riverside, San Joaquin, Stanislaus, and
Tulare Counties in California. Restricted-use pesticides from five pesticide chemical
classes between 1983 and 1984 were assessed. Using PUR-derived pesticide application
locations at the PLSS section-level and maternal addresses located using county maps,
broad and narrow geographic definitions of pesticide exposure were calculated.
According to the broad definition, a mother was exposed to pesticides if a PUR
application was within her section of residence, or the eight adjacent sections. A narrow
definition only considered the section of residence. Despite the advantages of
investigating the effects of specific pesticides and pesticide chemical classes, the authors
did not geocode the exact locations of maternal residences. PLSS sections are 1 mi
2
in
resolution, and distances between residences and pesticide applications could not be
determined.
Gunier et al. (2001) demonstrated an improved usage of PUR data through
calculating average annual pesticide application rates (lb of pesticide active ingredient
per mi
2
) between 1991 and 1994 for each PLSS section. The authors examined pesticide
use based on groupings related to chemical classes and toxicological evidence (i.e.

17

probable carcinogens, possible carcinogens, genotoxic compounds, and reproductive or
developmental toxicants) for all census block groups across California. Pesticide use at
the PLSS section level was allocated to census block groups based on section area within
each census block group. Census block group-level pesticide use density was calculated
after dividing by the census block group area. Gunier et al. (2001) also developed a
methodology to weight annual pesticide application rates according to the pesticide’s
potential to cause cancer (i.e. carcinogenicity) using U.S. EPA classifications and
exposure potential via volatilization and environmental persistence. Using this GIS-based
approach, the authors found that most census block groups in California (57-99%)
averaged less than 1 lb per mi
2
of average annual use for each pesticide group and
individual pesticide evaluated.
Several studies have since adopted the approach detailed in Gunier et al. (2001).
Reynolds et al. (2002) conducted an ecologic study of childhood cancer cases between
1988 and 1994 in California in relation to census block group-level pesticide exposure.
Reynolds et al. (2005) subsequently improved pesticide exposure assessment by
incorporating a residential buffer around each of the geocoded addresses. Investigating
the relationship between maternal residential proximity to pesticide applications and early
childhood cancer, Reynolds et al. (2005) estimated pesticide exposure within a half-mile
of geocoded maternal residences. Pounds of pesticide use were assigned to each study
subject’s residence based on the percentage area of each PLSS section within each half-
mile buffer. Pounds were summed across the relevant time period of interest and divided
by the buffer area (approximately 0.79 mi
2
) for each pesticide toxicological group,
chemical class, and individual pesticide. Despite the utility of using specific PUR

18

pesticide application data in relation to a geocoded residential buffer, methods of
improving the spatial resolution of the PUR data reported beyond the 1 mi
2
PLSS section
level were needed. The authors noted that usage of PLSS section-level data is more
sensitive in capturing potential pesticide applications, but potentially at the cost of
specificity.

2.2.6 Enhancement of PUR-Derived Metrics Using Land Use Surveys
The next advance in GIS-based pesticide exposure methodologies using PUR data
attempted to increase the specificity (i.e. minimization of false positives) of estimates
through finding a relevant buffer distance around geocoded residences to capture
pesticide drift, and through increasing the spatiotemporal resolution of determining which
agricultural lands were applied with pesticides. Rull and Ritz (2003) laid the foundation
for usage of the CDWR land use surveys to make use of the PUR attributes regarding
crop type and field acreage associated with pesticide applications. County-based CDWR
land use surveys are conducted every seven to 10 years to describe land use and crop
cover, with a minimum mapping unit of 0.81 ha (0.003 mi
2
) (Nuckols et al. 2007). Rather
than basing estimates on residence within a 1 mi
2
PLSS section, resolution was improved
to 1:24,000 or 1 in to 2,000 ft by incorporating land use surveys (Rull and Ritz 2003).
Specifically, the authors determined the likely locations of crop fields (using CDWR land
use surveys) near residences (within 500 or 1,000 m) upon which PUR pesticide
applications took place. The authors accounted for the seasonal rotation of crops (i.e.
same fields used for different crops) by collapsing seasonal field crops (e.g. cotton,
grains, potatoes, tomatoes, and alfalfa) into a single field crop class. Rull and Ritz (2003)

19

created a three-tier approach to assign pounds of applied pesticides to agricultural lands
based on the certainty of a PUR crop type matching land use survey data (Goldberg et al.
2007). Annual application rates (lb/ac) were calculated by summing the applied pounds
of pesticide (from PUR) divided by treated crop acres (from land use survey) intersecting
a 500 or 1,000 m residential buffer.
Rull and Ritz (2003) compared their approach to the Bell et al. (2001) broad vs.
narrow approach by generating 1,000 randomly selected samples of 200 addresses from
residential parcel centroids in Kern County, California. The Rull and Ritz (2003) PUR
and land use survey approach was designated as the gold standard, and a residence was
considered exposed if a pesticide-treated field was within a 500 or 1,000 m buffer. The
Bell et al. (2001) approach (broad: exposed if PUR application reported within section of
residence or adjacent sections; narrow: exposed if PUR application reported within
section of residence) and a land use survey-only approach (residential proximity to crop
fields; exposed if crops grown within 500 m of residence) was compared to the gold
standard. The authors demonstrated that measures of association between pesticide
exposure and a health outcome would be attenuated, or biased towards the null
hypothesis, if using lower resolution metrics [i.e. Bell et al. (2001) approach and land use
survey-only approach] that did not combine both PUR and land use survey data. For
various pesticides when compared to the gold standard, the Bell et al. (2001) broad
definition was associated with perfect sensitivity (100%) but poor to good specificity (62-
93.9%). Specificity was improved with the Bell et al. (2001) narrow definition (98.7-
99.4%), though sensitivity decreased (35.3-54.8%). The land use survey-only model was
associated with decreasing specificity and increasing sensitivity with increasing buffer

20

size (500 m buffer: 60.1% sensitivity and 94% specificity; 1,000 m buffer: 72.2%
sensitivity and 87% specificity).
Rull and Ritz (2003) also touched on the impact of residential mobility.
Attenuation of measures of association becomes more pronounced with increasing
exposure prevalence and increasing mobility rate. This issue is particularly problematic if
individuals move to urban areas, which would decrease specificity and increase the
number of false positives. Taken as a whole, Rull and Ritz (2003) demonstrated that
usage of a higher resolution metric increases specificity and decreases the extent to which
measures of association are attenuated, particularly when the true exposure prevalence is
low in the population. Many epidemiologic studies have since adopted this approach for
studying cancers and Parkinson’s disease (Marusek et al. 2006; Rull et al. 2006a; Rull et
al. 2006b; Roberts et al. 2007; Costello et al. 2009; Gatto et al. 2009; Ritz et al. 2009;
Rull et al. 2009; Manthripragada et al. 2010; Cockburn et al. 2011; Wang et al. 2011; Lee
et al. 2012, 2013). In practice, these studies have weighted pesticide application rates
(lb/ac) by the proportion of the area of pesticide-treated acres intersecting a 500 m
residential buffer.
The accuracy of the Rull and Ritz (2003) GIS-based approach of estimating
individual-level residential exposure to agricultural pesticides was demonstrated in a
validation study among participants of the Parkinson’s Environment and Gene (PEG)
study in Central California (Fresno, Kern, and Tulare Counties). For 22 Parkinson’s
disease cases and 22 age- and gender-matched Medicare controls and randomly selected
residential parcels, exposure was defined as the weighted average of organochlorine
pesticide applications (lb/ac) within a 1,000 m residential buffer between 1974 and 1999.

21

Using lipid-adjusted dichlorodiphenyldichloroethylene (DDE) levels measured in serum
as the gold standard, the GIS-based metric was associated with 87% specificity and 38%
sensitivity. The GIS-based metric, body mass index, age, gender, mixing and loading
pesticides by hand (derived from occupational questionnaire data), and residential
pesticide use explained 47% of the variance in DDE serum levels.
Building off of the Rull and Ritz (2003) approach, Nuckols et al. (2007) also
utilized PUR and land use survey data to derive pesticide exposure, but at the crop level
by not collapsing field crops. Collapsing crops may introduce additional issues in
interpretation when the collapsed category does not include the crop type in the PUR
database (Rull and Ritz 2003). For each PLSS section intersecting a 500 m residential
buffer, the authors calculated an annual crop-specific pesticide application rate (lb/ac)
between 1988 and 1994 by dividing the total amount of pesticides applied to each crop of
interest during the time period by the total area of the crop field within the PLSS section.
This was weighted according to the crop area within a 500 m buffer, and then divided by
the area of the buffer to obtain a pesticide use density measure in lb per mi
2
. Six
pesticides were evaluated and residences of participants from a California Department of
Health Services childhood cancer study were geocoded. Using this metric as the gold
standard, the performance of a PUR-only metric (pesticide application rate weighted by
the area of the PLSS sections within a 500 m buffer) was assessed. A residence was
considered exposed if there was any pesticide use on a crop field within 500 m of a
residence (gold standard), or pesticide use in a section within 500 m (PUR-only model).
The authors also used an additional exposure cutoff of greater than the 25
th
percentile of
pesticide use. By using a coarser-scale, PUR-only metric, sensitivity is 100%. Nuckols et

22

al. (2007) demonstrated good specificity with various pesticides [e.g. 96% specificity
with dicofol (5% prevalence) compared to 86% specificity with propargite (15%
prevalence)] using a cutoff of any pesticide use within 500 m. However, specificity
decreased to between 29 and 45% when excluding residences with no pesticide use
within 500 m and using a 25
th
percentile cutoff of exposure. The results of overall
agreement mirrored that of specificity, where overall agreement between the gold
standard and PUR-only metric was high (88-98%), but decreased when excluding
residences with no pesticide use within 500 m (35-58%).

2.2.7 Downscaling PUR-Derived Metrics Using Landsat Satellite Imagery
Approaches to ascertaining pesticide exposure using PUR data have been further refined
with remotely sensed data, which are data captured from a distance, such as aerial
photographs and satellite imagery (Waller and Gotway 2004). Despite being a rich
resource for large-scale data, few cancer epidemiologic studies have utilized remotely
sensed data for environmental exposure assessment (Maxwell et al. 2010a). Remote
sensing is particularly relevant to cancer epidemiology because it can be used to
reconstruct environmental exposures and characterize environmental change. Landsat
satellite imagery, for example, provides moderate to high resolution data spanning 39
years (USGS 2013b). One primary advantage is the multispectral and multitemporal
features of Landsat data, which allow for landscape features to be distinguished based on
spectral and phenological (i.e. seasonal changes in vegetation) characteristics (USGS
2011).

23

Maxwell et al. (2010b) demonstrated the potential use of Landsat imagery in
improving pesticide exposure assessment in California. The authors showed how the
spatiotemporal uncertainty regarding crop field-level changes due to the infrequent
CDWR land use surveys could be addressed by using temporally varying Landsat
imagery. Specifically, CDWR land use surveys represent a snapshot of the agricultural
landscape at one point in time. During the interim time between land use surveys, if there
is more than one survey, the shapes, sizes, and existence of crop fields may have
changed. Furthermore, during a given crop-growing season, fields may be used for more
than one crop (i.e. multi-cropped) and only a portion of a field in the CDWR land use
survey may be utilized for growing a particular crop.
A time series of 24 Landsat 5 and 7 images in 2000 was collected for Fresno
County, California. The time series of Landsat imagery was intersected with CDWR land
use surveys to determine crop field locations. Normalized Difference Vegetation Index
(NDVI) values, which measure vegetative growth, were derived from the Landsat images
for 17 crop types. Maxwell et al. (2010b) demonstrated that variation in NDVI values
across the year can showcase evidence of multi-cropped fields and potential
misclassification of fields from the land use surveys (e.g. absence of NDVI-based
evidence for vegetation within a land use survey-labeled crop field). Maxwell (2011)
subsequently presented a case study of the exact methods used to execute a Landsat
imagery-based approach to downscale, or improve the spatial resolution of, PUR data
reported at the PLSS section level. PUR applications of the pesticide paraquat were
selected for Fresno County, California, in 1994. Crop field boundaries based on similar
phenological characteristics (i.e. NDVI values) were derived from Landsat imagery

24

across different dates throughout 1994. Crop types were determined by comparing pixels
from these delineated crop field boundaries to the crop signature library established in
Maxwell et al. (2010b). It was shown that Landsat data could be used to identify PUR
errors (e.g. PUR data indicates pesticide applications on non-existent crop fields) and to
determine which exact area of a field was used for a particular crop type associated with a
PUR pesticide application.

2.3 Surrogate Measures of Pesticide Exposure: Rurality
Despite the burgeoning research into geospatial methods of pesticide exposure
ascertainment, data limitations, lack of technical knowledge, etc., may necessitate the use
of crude surrogates, or proxy indicators, of pesticide exposure. Rurality, or the extent to
which a geographic area is rural, has been used as an indicator of pesticide exposure
(Alavanja et al. 2004). The rationale for using rurality stems from agricultural lands
associated with pesticide applications being more common in rural areas (Ward et al.
2000; Alavanja et al. 2004; Franklin and Worgan 2005). Although some measures of
rurality fall under the umbrella of qualitative pesticide exposure metrics (e.g. self-
reported residence in a rural area), some are quantitative metrics that offer an objective
alternative, such as through incorporating existing information regarding population
density. In the context of epidemiology and investigating the association between
pesticide exposure and one or more health outcomes, there are three primary issues
regarding the use of rurality to indicate pesticide exposure that should be considered: (1)
potential misclassification of pesticide exposure, (2) different definitions of rurality, and
(3) variation in analysis scales.

25

2.3.1 Misclassification of Pesticide Exposure
Validity, or accuracy, is the extent to which a measure is an indicator of what it intends to
measure (Szklo and Nieto 2007). Validity and exposure misclassification are related
concepts, where the impact of validity is manifest in exposure misclassification, or
information bias. Specifically, the validity of a pesticide exposure metric not only
influences the accuracy of the metric in truly indicating pesticide exposure, but may
inflate or obscure exposure-disease relationships (Franklin and Worgan 2005). These
relationships can be quantified as measures of association in epidemiologic studies, such
as odds ratios [i.e. odds of disease among individuals exposed to the purported exposure
of interest compared to those not exposed (Szklo and Nieto 2007)]. The exact effect of
using an inaccurate exposure metric on measures of associations depends on the extent of
exposure misclassification and the prevalence of the exposure in the study population of
interest.
Exposure misclassification can be understood in terms of the classic error model,
where an exposure metric is measured with error and is an imperfect surrogate for the
true exposure (Nuckols et al. 2004). The degree of misclassification can be measured
using sensitivity (i.e. capacity of a measure to correctly identify features with a
characteristic of interest) and specificity (i.e. capacity of a measure to correctly identify
features without a characteristic of interest) (Szklo and Nieto 2007). For example, in the
context of a comparative, epidemiologic study with two study groups - cancer cases and
non-cancer controls - evaluating the association between pesticide exposure and cancer,
nondifferential misclassification (i.e. extent of exposure misclassification does not differ
between the study groups) of pesticide exposure will bias the measure of association

26

towards the null hypothesis of no association. Furthermore, if the prevalence of pesticide
exposure is low in a study population (<10%), then decreases in specificity will
substantially attenuate the measure of association. However, for more frequent exposures,
reductions in sensitivity is associated with greater bias (Szklo and Nieto 2007). Rurality
likely misclassifies some geographic areas as exposed to pesticides when they are truly
not and vice versa; however, the exact extent to which rurality may misclassify pesticide-
exposed geographic areas remains unknown.

2.3.2 Variation in Rurality Definitions
Different definitions of how to define rurality exist, and different exposure-disease
relationships may be observed depending on which definition is used (Rural Assistance
Center 2012). For example, a rurality metric may only consider population information in
delineating a rural geographic area, or may consider multiple factors, such as population
and work commuting information. Although an analysis may compare and contrast
results using different definitions, this approach may not meaningfully contribute to
determining which rurality definition most adequately reflects pesticide exposure, or the
processes underlying pesticide exposure.

2.3.3 Variation in Analysis Scales
Furthermore, variation in analysis scales of rurality may also influence study results.
Analysis scales, or how data is measured/aggregated, may vary due to compulsory
aggregation of cancer data, general data availability, ecologic study designs, and
incorporation of contextual information in studying individual-level phenomena

27

(Montello 2001). Cancer data, such as that derived from cancer registries, are often
aggregated to areal units (e.g. census tracts and ZIP codes) for the purposes of patient
confidentiality (Boscoe et al. 2004; Waller and Gotway 2004). To avoid scale-translation
issues, exposure data can be aggregated to identical areal units, such as to evaluate the
potential association between census tract-level rurality (exposure) and census tract-level
cancer incidence rates (outcome) (Boscoe et al. 2004). These aggregations form the
fundamental units of analysis in the majority of studies employing geospatial techniques
(i.e. ecologic studies) (Nuckols et al. 2004), where the unit of analysis is not the
individual, but an aggregated unit (Szklo and Nieto 2007). Ecologic studies can be
important in generating hypotheses; however, depending on how the data are aggregated,
different results can be observed [i.e. modifiable areal unit problem (MAUP)] (O'Sullivan
and Unwin 2010). Studies have also examined individual-level phenomena while
incorporating both individual-level and ecologic, contextual variables to perform multi-
level analyses (Jacquez 2004). This approach is meaningful in attempting to capture the
effect of a variable that may operate at a scale beyond the individual. Irrespective of a
study’s unit of analysis as the individual or an ecologic aggregate, the underlying issue is
that usage of different rurality metrics, whether based on different rurality definitions or
aggregated to different scales, may lead to different results.
Taken together, rurality is an intuitive surrogate measure of pesticide exposure in
that applications of agricultural pesticides frequently occur in rural geographic areas.
However, analytic results may vary according to usage of different rurality definitions
and investigations at different analysis scales. Most importantly, a rurality metric is
inevitably associated with inaccuracies, likely misclassifying some geographic areas as

28

pesticide-exposed that are truly not and vice versa. In many study populations, such as
the state of California, the overall frequency of pesticide exposure is low [2.2% of
California population reside in rural areas (USDA 2013b)]. In the context of researching
the relatively infrequent exposure of residential proximity to agricultural applications of
pesticides, the impact an exposure metric with suboptimal specificity (i.e. high number of
false positives) will bias the results of a study towards the null hypothesis and attenuate
the true exposure-disease relationship. Therefore, determining the validity/accuracy of a
rurality-based exposure metric as an indicator of pesticide exposure, as well as
understanding the impact of using different rurality definitions at varying analysis scales,
is important in elucidating its performance and adequacy as a surrogate measure of true
pesticide exposure.

29

CHAPTER THREE: METHODS AND DATA SOURCES
In order to focus the analysis, the study area of interest was Kern County, California and
historical pesticide exposure between 1974 and 1990 was calculated. The study area and
time period of interest were constrained by the California Department of Pesticide
Regulation (CDPR) Pesticide Use Report (PUR) database, as pesticide reporting to the
CDPR began in 1974 (CDPR 2000b). As there are a wide array of pesticides in use
throughout California, the analysis focused on three pesticide chemical classes previously
associated with primary liver cancer - organochlorine pesticides (OCPs),
organophosphates (OPs), and carbamates (Cordier et al. 1993; Ezzat et al. 2005; Persson
et al. 2012).
For each census tract and ZCTA in Kern County, annual pesticide chemical class-
specific application rates (lb/ac) were calculated using a GIS. Pounds of applied
pesticides were derived from the PUR database and crop field acreage from land use
surveys, classified Landsat imagery, and PLSS sections. A new GIS-based pesticide
exposure methodology is presented, which modifies the Rull and Ritz (2003) three-tier
approach combining PURs, CDWR land use surveys, and PLSS sections to estimate
census tract- and ZCTA-level pesticide exposure and to incorporate Landsat imagery. A
crop signature library (CSL) of Normalized Difference Vegetation Index (NDVI) values
was created using Landsat imagery in 1990, which was used to classify segments derived
from Landsat NDVI images in 1985 into agricultural crop fields.
Rurality was measured for each ZCTA and census tract using two common
metrics, Rural-Urban Commuting Area (RUCA) codes and the U.S. Census Bureau
urban-rural classification system. A statistical analysis, including calculating measures of

30

validity (i.e. sensitivity and specificity), was performed to formally evaluate the extent to
which ZCTA- and census tract-level rurality metrics are valid indicators of pesticide
exposure, as well as to determine which surrogate measure offers greater accuracy. All
GIS-related geoprocessing and visualization was performed in ArcGIS 10.1 and IDRISI
Selva; statistical analyses was performed in SAS 9.3 .

3.1 Research Hypotheses
Rural designations using RUCA codes and the U.S. Census Bureau urban-rural
classification system were hypothesized to be less sensitive and less specific compared to
the GIS-based pesticide exposure metric (i.e. gold standard) in assigning ZCTA- and
census tract-level pesticide exposure. The RUCA code system was hypothesized to be a
more accurate surrogate measure of pesticide exposure compared to the U.S. Census
Bureau urban-rural classification system due to its incorporation of both population and
work commuting information. The U.S. Census Bureau urban-rural classification system
only incorporates population information. In other words, RUCA codes, by virtue of their
definition, were hypothesized to better reflect areas truly rural where agricultural
applications of pesticides are more likely to occur.

3.2 Study Area: Kern County, California
California is the third largest state in the U.S., 158,706 mi
2
in size with 58 counties
(CA.gov 2013). The most populous cities are Los Angeles, San Diego, San Jose, San
Francisco, and Fresno. In 2012, there were 38,041,430 individuals residing in California,
2.2% (836,441 individuals) of whom were rural residents (USDA 2013b). In 2007, over

31

25% of the total statewide land area was devoted to farmland. California is the most
agriculturally productive state in the U.S. in terms of farm output and productivity
(Economic Research Service 2012).
In 2007, approximately 3% of all California farms were located in Kern County
(N=2,117) (USDA 2007a). Kern County is one of 19 counties nestled in California’s
agriculturally intensive Central Valley, which produces 25% of the food Americans
consume (Figure 1) (NPR 2002). Over 2 million ac of Kern County were devoted to
farmland, with an average size of 1,116 ac per farm. Over 68% of Kern County farms
were used for cropland (N=1,449; 942,827 c), of which 81% (N=1,169; 764,929 ac) were
devoted to harvested cropland and 15% (N=222; 41,081 ac) to pasture grazing (USDA
2007c). A total of 836 farms were used as orchards (407,208 ac).

Figure 1 Kern County, California, study area of interest
(Data from U.S. Census Bureau 2013)

32

In 2011, 191 million lb of pesticide active ingredients were used in California
(CDPR 2011a). Over 28 million lb were used in Kern County, ranking it as the second
highest county in the state for pesticide usage after Fresno County. A total of 945 Kern
County farms reported using chemicals to control insects, followed by 735 for
weed/grass/brush control, 120 for nematode control, and 471 for disease in crops and
orchards (USDA 2007b). The most frequently pesticide-treated agricultural commodities
in 2011 included grapes, carrots, oranges, and pistachios (Table 1) (CDPR 2011b).

Table 1 Common pesticide-treated crops in Kern County, 2011
1

Crop Rank
Applied pesticides
(lb)
Pesticide
applications (N)
Treated
land (ac)
Almonds 1 7,996,450 31,468 2,990,073
Grapes 2 4,227,137 40,740 2,096,170
Carrots 3 3,170,438 4,203 243,140
Oranges 4 2,476,753 12,092 639,149
Pistachios 5 2,032,358 7,248 699,843

1
Data adapted from CDPR (2011b)

3.3 Data Sources
3.3.1 Pesticide Exposure Data
Table 2 lists the data sources that were used to execute the research methodology. The
CDPR PUR database is the most comprehensive pesticide reporting system in the world,
collecting data regarding agricultural pesticide use throughout California (CDPR 2013).
Between 1974 and 1989, commercial pest control operators (e.g. structural applicators)
were required to report all pesticide use and farmers were required to report restricted
pesticide use, or pesticides with high potential to cause public health harm. Since 1990, a

33

full-use reporting system has been adopted. The PUR database focuses on agricultural
pesticide applications, but also includes applications to parks, golf courses, etc. PUR
information includes the name and pounds of pesticide active ingredient applied, field
and crop acreage treated, date of application, and PLSS section of application. The PUR
database contains 45,000 pesticide products; 1,000 new products are added each year and
1,000 are inactivated due to nonrenewal, suspension, or cancellation.
PUR data are reported at the PLSS section level (Rull and Ritz 2003). The PLSS
system was introduced earlier and divides the country into townships measuring six miles
on a side and these, in turn, are subdivided into 36 1 mi
2
sections (National Atlas 2013).
PLSS surveys start at an initial point from which townships are surveyed north, south,
east, and west. The north-to-south line running through the initial survey point is called
the principal meridian for that PLSS survey. The east-to-west line running through the
initial point is called the baseline, and is perpendicular to the principal meridian.
Townships are identified through a township designation (i.e. north or south of baseline)
and range designation (i.e. east or west of principal meridian). In California, PLSS
sections are uniquely identified according to their county, principal meridian, township,
range, and section. There are 8,455 PLSS sections intersecting Kern County (Figure 2).
The CDWR conducts land use surveys of agricultural lands for California
counties focusing on over 70 crop types. Aerial photographs, satellite imagery, and GPS
devices are used to delineate crop field boundaries. County-based surveys are conducted
every seven to 10 years (CDWR 2013). The earliest land use survey conducted in Kern
County occurred in 1990 (Figure 3).

34
Table 2 Data sources
1

Dataset Description
Geographic
extent Data type Spatial resolution
Temporal
currency
CDPR PURs Agricultural
pesticide use
database
California,
U.S.
Text files Reported at PLSS (1 mi
2
)
section level
1974 to present
PLSS sections Cadastral dataset California Vector data model
(polygon)
PLSS polylines defined by
survey points accurate to ≥40
ac level
Updated in 2011
CDWR land use
surveys
Surveys of
agricultural lands
California Vector data model
(polygon)
Minimum mapping unit of
0.003 mi
2

1976 to present;
updated every 7-
10 years
RHRC RUCA
codes Version 2.0
Census tract and
ZIP code
rural/urban
designations
U.S. Text files Reported at ZCTA and census
tract level
2000
U.S. Census
Bureau
TIGER/Line
shapefiles
Administrative
delineations:
census tracts and
ZCTAs
U.S. Vector data model
(polygon)
200 ft resolution 2000
U.S. Census
urban-rural
classification
Urbanized Areas
(UAs) and Urban
Cluster (UCs)
U.S. Vector data model
(polygon)
200 ft resolution 2000
USGS and NASA
Landsat imagery
Remotely sensed
satellite imagery
Global Raster data model 30 m for red and near infrared
spectral bands (Landsat 4 and
5 Thematic Mapper sensors)
1972 to present

1
Data from RHRC (2000); U.S. Census Bureau (2000, 2013); Cal-Atlas Geospatial Clearinghouse (2013); CDPR (2013); CDWR (2013),
National Atlas (2013); and USGS (2013b)

35

Figure 2 PLSS sections in Kern County
(Data from Cal-Atlas Geospatial Clearinghouse 2013; and U.S. Census Bureau 2013)

Figure 3 Map showing geographic extent of 1990 Kern County land use survey
(Data from CDWR 2013; and U.S. Census Bureau 2013)

36

3.3.2 Landsat Imagery
The Landsat program was started by the U.S. Geological Survey (USGS) and the
National Aeronautics and Space Administration (NASA) to continuously collect Earth
imagery (USGS 2013b). The first Landsat satellite, Landsat 1 or the Earth Resources
Technology Satellite (ERTS), was launched in 1972. The Return Beam Vidicon (RBV)
and Multispectral Scanner (MSS) sensors were onboard Landsat 1 through 3. Landsat 4
and 5 launched in 1982 and 1984, respectively, included the Thematic Mapper (TM) and
MSS sensors (Table 3). Landsat 7 included the Enhanced Thematic Mapper Plus (ETM+)
sensor, and Landsat 8, launched in 2013, carries the Operational Land Imager (OLI) and
Thermal Infrared Sensor (TIRS) sensors (Maxwell et al. 2010a). The Landsat 4 and 5
remote sensing system characteristics, relevant to this analysis, are listed in Table 4.

Table 3 Landsat 4 and 5: Thematic Mapper (TM) sensor
1

Band
number
Spectral
range ( m) Wavelength
Spatial
resolution (m)
1 0.45 - 0.52 Blue-green 30
2 0.52 - 0.60 Green 30
3 0.63 - 0.69 Red 30
4 0.76 - 0.90 Near infrared 30
5 1.55 - 1.75 Mid infrared 30
6 10.4 - 12.5 Far infrared 120
7 2.08 - 2.35 Mid infrared 30

1
Data adapted from Campbell and Wynne (2011)

The NDVI is the most commonly used vegetation index to characterize vegetative
growth, or greenness (i.e. relative vegetative density and health) (USGS 2011). NDVI
values can be derived from Landsat satellite imagery using the red (R) and near infrared

37

Table 4 Landsat 4 and 5 remote sensing characteristics
1

Characteristic Description
Inception
Landsat 4: 1982
Landsat 5: 1984
Revisit frequency 16 days
Orbit Near-polar, sun-synchronous
Swath width 185 km
Geographic extent Global
Sensors Multispectral Scanner (MSS), Thematic Mapper (TM)
Geographic reference UTM coordinate system, WGS84 datum
Applications Earth observation (EO)

1
Data adapted from USGS (2013b)

(NIR) bands: NIR-R/NIR+R. The NDVI system uses the wavelengths of light absorbed
and reflected by green plants captured in Landsat satellite sensors; reflectance properties
change as the growing season progresses. NDVI values range between -1.0 and +1.0.
NDVI values less than 0.1 indicate areas with barren rock, sand, or snow (i.e. sparse
vegetation). NDVI values between 0.2 and 0.6 indicate moderate vegetation, and values
beyond 0.6 indicate dense vegetation.

3.3.3 Rural-Urban Commuting Area Codes
The Rural Health Research Center (RHRC) RUCA codes (Version 2.0) classify ZCTAs
and census tracts using a 33-code scheme that incorporates 2000 U.S. Census Bureau
daily work commuting, population density, and urbanization information. The RUCA
system is comprised of a two-level classification system (RHRC 2000). The first level
consists of whole numbers between 1 and 10 corresponding to metropolitan,
micropolitan, small town, and rural area categories (USDA 2012). Second-level

38

subgroups reflect secondary commuting flows associated with the first-level categories.
Different RUCA code groupings exist to classify geographic units as urban or rural.

Figure 4 Urbanized Areas and Urban Clusters across California, 2000
(Data from U.S. Census Bureau 2013)

39

3.3.4 U.S. Census Bureau Urban-Rural Classification
The 2000 U.S. Census Bureau urban-rural classification system categorizes geographic
areas across the U.S. as urban or rural. Specifically, Urbanized Areas (UAs) are
geographic areas with a densely settled core of census block groups or census blocks with
a population density of at least 1,000 individuals per mi
2
, surrounding census blocks with
a population density of at least 500 individuals per mi
2
, and with a total population of
50,000 or more (Figures 4 and 5). Urban Clusters (UCs) are geographic areas with a
densely settled core of census block groups or census blocks with a population density of
at least 1,000 individuals per mi
2
, surrounding census blocks with a population density of
at least 500 individuals per mi
2
, and with a total population of at least 2,500 individuals,
but less than 50,000 (U.S. Census Bureau 2000). Any geographic areas outside of UAs
and UCs are considered rural. Data from the year 2000 were chosen to compare with the
RUCA code assignments.

Figure 5 Kern County UAs and UCs, 2000 (Data from U.S. Census Bureau 2013)

40

3.4 Pesticide Exposure Estimation
3.4.1 Preparation of PUR, PLSS, and Land Use Survey Data
All PUR files between 1974 and 1990 were downloaded. Each PUR record contains
information regarding an individual active ingredient used in a pesticide application.
Since a pesticide product may contain multiple chemicals, there may be multiple PUR
records for a single pesticide application (CDPR 2000b).
The workflow for processing PUR, land use survey, and PLSS data for use in the
GIS environment is illustrated in Figure 6. Using agricultural pesticide references (Dich
et al. 1997; Gunier et al. 2001; Alavanja et al. 2004; Greene and Pohanish 2005; Rull et
al. 2006a, 2009; Wood 2010; AgroPages 2013), a database of pesticides belonging to the
organochlorine (OCP), organophosphate (OP), and carbamate chemicals classes was
compiled - including pesticide name and CDPR chemical code. Unique identifiers

Figure 6 Methodological workflow: PUR, land use survey, and PLSS processing

41

comprised of the CDPR county code, principal meridian, township, range, and section
(CO-MTRS) were created to combine PUR data with PLSS and land use survey data.
Per CDPR data quality standards (CDPR 2000a), the following PUR logic checks
were performed separately for each year between 1974 and 1989 (Appendix B; Table
B1): (1) duplicates and (2) spatially inconsistent county (using CO-MTRS) outside
county boundary. The following logic checks were performed for 1990 PUR data: (1)
duplicates, (2) spatially inconsistent county, (3) inconsistent county code, (4) missing
agricultural field location identifiers, (5) inconsistent CO-MTRS for a location, (6)
inconsistent acres planted, and (7) treated acres greater than planted acres. The CDWR
land use survey logic check was omitted due incorporating land use survey information in
the tiered PUR matching methodology. Not all CDPR logic checks were performed on
PUR data between 1974 and 1989 because the logic checks were created for PUR data
from 1990 onward. Therefore, some of the variables required for the logic checks, such
as the grower identification number, were available starting in 1990. Depending on the
logic check definition, PUR records identified using a logic check were either excluded
from the analysis or the first record was retained.
To increase comparability between PUR and CDWR land use survey data, in
addition to addressing the uncertainty regarding the seasonal rotation of crops (e.g.
double-cropping) and intercropping, the following crop types in both datasets were
collapsed into a single field crop category: (1) grain and hay crops; (2) field crops; (3)
pasture; and (4) truck, nursery, and berry crops (Rull and Ritz 2003; Rull et al. 2006a). A
crosswalk between PUR commodity codes (different codes were used between 1974 and
1989 compared to 1990 onward) and CDWR land use survey crop codes (1981-1992 and

42

1993-1997 legends) was created. Executive decisions regarding CDPR commodity codes
assigned to CDWR land use survey crops that were not exact name matches were
documented [e.g. CDPR tangerines (pre-1990 code 2104; 1990 code 2008) assigned to
CDWR land use survey oranges (code C3)].
Between 1974 and 1989, outliers were identified as pesticide application rates
meeting two of the three CDPR outlier flag definitions created for PUR data beginning in
1990 (CDPR 2002): (1) pesticide application rates [lb active ingredient (AI)/treated ac]
greater than 200 lb/ac (greater than 1,000 lb/ac if fumigation) (only considers PUR
records reported in acres), and (2) pesticide application rates (lb AI/treated unit)
(considers all PUR records) greater than 50 times the median rate for all uses of that
pesticide product [identified using manufacturing firm number, label sequence number,
revision number, and registration firm number between 1974 and 1983, and as EPA
registration number between 1984 and 1989], commodity code, unit type, and record type
(production agriculture vs. monthly report). Outliers were identified in 1990 PUR records
using three CDPR-created outlier flags (the two aforementioned outlier definitions in
addition to neural network). All outliers were imputed with the statewide median rate for
the pesticide AI in that year; pounds of AI were recalculated.
Between 1974 and 1989, agricultural, non-summary PUR records were extracted;
in 1990, all daily and monthly production agriculture PUR records were extracted. The
following selections were made from PUR records: those associated with an OCP, OP, or
carbamate in the compiled pesticide database and applied in Kern County. As a result of
the logic checks, only PUR records with a valid CO-MTRS identifier were included.

43

Pounds of applied AI were then summed according to crop type and CO-MTRS (Figure
6).

Figure 7 Methodological workflow: Landsat remote sensing and crop signature library

3.4.2 Incorporation of Landsat Imagery: Crop Signature Library (CSL)
A crop signature library (CSL) was compiled using a time series of Landsat 4 and
Landsat 5 Thematic Mapper (TM) imagery acquired between January and October 1990
for Kern County - 1990 is the year in which the earliest Kern County CDWR land use
survey is available as a ground truth (Figure 7) (Maxwell et al. 2010b; Maxwell 2011;
CDPR 2013). Landsat images for November and December were not available in 1990.
Images from Paths 41 and 42 and Rows 35 and 36 were requested from USGS Global
Visualization (GloVis) Viewer as they cover the geographic extent of Kern County
(Figure 8). Images with excessive cloud cover were excluded. All images were Standard
Terrain Correction [Level 1T (L1T)] products, which are radiometrically and
geometrically processed images using ground control points (GCPs) and a Digital
Elevation Model (DEM) (USGS 2013a). All images were associated with a common

44

coordinate system: Universal Transverse Mercator (UTM) Zone 11N (WGS84 datum;
meter).

Figure 8 Landsat Path-Row scenes intersecting Kern County
(Data from U.S. Census Bureau 2013; and USGS 2013c)

Using IDRISI Selva, TM images for band 3 (red) and band 4 (near infrared) were
radiometrically corrected to at-sensor reflectance using published radiometric calibration
coefficients and image metadata [MTL files for Level 1 Product Generation System
(LPGS)-processed images] (Chander et al. 2009). Atmospheric correction was executed
using the Chavez cosine estimation of atmospheric transmittance (COST) model, taking
into account date, time of day, band center wavelength, gain, bias, cosine of the solar
zenith angle (90-solar elevation angle), and assuming the downwelling spectral
irradiance, path radiance due to haze (i.e. digital number of objects with zero reflectance,
such as deep clear lakes), and spectral diffuse sky irradiance is zero (Campbell and

45

Wynne 2011). Chavez (1996) developed the COST, or Cos(t), model to integrate the
Dark Object Subtraction (DOS) model for haze removal, in addition to estimating
absorption from Rayleigh scattering and atmospheric gases. For each month in 1990, all
four Path/Row scenes (where available) were mosaicked one Path at a time (cover
overlap method matching on grey level using non-background values). Clouds near the
Path 41-to-42 overlapping region were masked out before a mosaic to join the two Paths.
Subsequent to mosaicking, negative reflectance values, potentially associated with
random error related to water and/or shadows, were recoded to a reflectance value of 0
(YCEO 2013). A median spatial filter (3x3 kernel) was applied to each mosaic to
minimize random noise (Vassiliou et al. 1988; Mather and Koch 2011), and the mosaic
was cropped to a smaller geographic area enclosing Kern County. NDVI values were
calculated using bands 3 and 4. NDVI images were re-projected to the California Teale
Albers (NAD83 datum; meter) coordinate system (30 m spatial resolution; nearest
neighbor resampling to not alter pixels).
Guided by a natural color [red-green-blue (RGB) band combination] multispectral
(MS) Landsat 5 image of California provided by Cal-Atlas (Cal-Atlas Geospatial
Clearinghouse 2013), polygons representing the geographic extent of each monthly
NDVI image not affected by clouds or shadows were digitized and intersected to create a
cloud-free zone. Using the 1990 Kern County CDWR land use survey, stratified random
sampling (SRS) eligibility criteria for land use survey polygons included (1) single-use
(i.e. not double- or triple-cropped, intercropped, or mixed); (2) at least 4 ha in area
(Maxwell et al. 2010b); and (3) intersecting the cloud-free zone. SRS using strata defined
by land use classes (e.g. C1=grapefruit) was performed to select at most 30 eligible

46

polygons per stratum. Non-agricultural land use classes, such as native vegetation (NV),
were included in the CSL to facilitate the discrimination between as many land use
classes as possible during subsequent classification.
Negative NDVI values, indicative of an absence of green vegetation, were
recoded to 0 (Beck et al. 2006). Separately for each month between January and October
1990, NDVI values for each pixel were extracted using a mask defined by the SRS-
sampled land use survey polygons. NDVI values for points straddling multiple crop field
boundaries were deleted. The median NDVI value for each land use survey polygon was
calculated; the median NDVI value for each land use class for each month was retained
in the CSL for subsequent classification (Figure 9).

Figure 9 Methodological workflow: classification of Landsat images

3.4.3 Classification of 1985 Landsat Imagery
Using the CSL, NDVI Landsat images from 1985 for Paths 41 and 42 and Rows 35 and
36 were radiometrically and atmospherically processed and digitally enhanced to

47

facilitate classification according to land use class using a sum of squared differences
(SSD) measure (Figure 9). An earlier time period of 1985 was chosen to address the PUR
records occurring before 1990. All images were acquired using the Landsat 5 TM sensor,
processed with either the LPGS system (MTL metadata) or the National Land Archive
Production System (NLAPS) (WO metadata) system (Chander et al. 2009) and associated
with the UTM Zone 11N coordinate system (WGS84 datum; meter). Landsat images
between January and October 1985 were requested from GloVis to parallel the CSL. Due
to the absence of images for Path 42 (majority of Kern County agricultural fields in this
Path) in February 1985, this month was excluded from classification.
Images were corrected to at-sensor reflectance (Chander et al. 2009), cloud-
masked, mosaicked, smoothed using a median filter (3x3 kernel), and cropped to a
geographic area enclosing Kern County. NDVI values were derived using the red and
near infrared bands 3 and 4, respectively. Cloud- and shadow-free geographic areas
within the NDVI images available for all months (except February) in 1985 were
digitized and intersected to create a segmentation-eligible zone.
A principal component analysis (PCA) was performed on the nine monthly NDVI
images as a data compression method (Maxwell 2011; Lippitt et al. 2012). Principal
components 1, 2, and 3 (Maxwell 2011) were used as inputs for an object-based
segmentation to delineate crop field boundaries - and other land features in the selected
geographic extent. Segmentation groups adjacent pixels into segments according to
spectral homogeneity (Campbell and Wynne 2011). Segmentation was performed
iteratively using different input parameters, comparing resultant segmentation products to
crop field boundaries according to a color composite of the first three principal

48

components (PC1: red, PC2: green, PC3: blue) and an August 1985 Landsat image
displayed using a color-infrared (CIR) band combination (USDA 2013a). CIR imagery is
useful when examining crop field boundaries and irrigated vegetation (CDOC 2013). The
following parameters were used for the final segmentation output: window of 3, tolerance
of 80, weight mean factor of 0.5, and weight variance factor of 0.5. All datasets were re-
projected to the California Teale Albers coordinate system (NAD83 datum; meter).
Using all NDVI pixel values for each segment, the resultant segmentation vector
shapefile was classified using the 1990 CSL according to a distance measure: the smallest
sum of squared differences (Maxwell 2011) using the median NDVI value for each
segment compared to the median value of all CSL land use classes for each available
month in 1985 (January, March to October). A sensitivity analysis was performed
comparing resultant classified segments when using a CSL with (1) all land use classes;
(2) land use classes except broad groupings (e.g. F for field crop, no subclass given); or
(3) land use classes except broad groupings and SRS land use strata with less than 30
samples. Inclusion of land use classes without a subclass may have obscured differences
in spectrally heterogeneous polygons. Inclusion of land use classes not meeting SRS-
stratum sample sizes may have resulted in selection of samples not truly representative of
the land use class. The classified segments (selected from one of three aforementioned
approaches) were processed to exclude non-agricultural land use classes and segments
corresponding to known areas without vegetation (using CIR image and land use survey).
The processed, CSL-classified segments were used as 1985 crop field boundaries in the
modified three-tier matching.

49

3.4.4 Modified Three-Tier Approach to Estimate Pesticide Exposure
The Rull and Ritz (2003) three-tier approach was modified to assign PUR-derived pounds
of applied pesticides for each year between 1974 and 1990 to crop fields within PLSS
sections derived from the 1990 Kern County land use survey and the 1985 Landsat-
classified layer (Figure 10).

Figure 10 Modified three-tier pesticide exposure method

Land use survey and Landsat crop fields were dissolved by crop type to the PLSS
section level and intersected with sections to facilitate tiered matching. Sliver polygons
resulting from the intersection [area ≤0.11 ac (smallest 1990 land use survey polygon) or
area ≤12.36 ac and length ≥200 m] were excluded from matching.
Tier 1 match: PUR-derived data were matched to a land use survey crop field
when the crop type and PLSS section matched.

50

Tier 2A match: PUR data were matched to a Landsat-derived crop field when the
crop type and PLSS section matched.
Tier 2B match: PUR data were matched to the other land use survey crop fields
within the PLSS section.
Tier 2C match: PUR data were matched to the other Landsat-derived crop fields
within the PLSS section.
Tier 3: If no land use survey and Landsat-derived crop fields were present within
a PLSS section, PUR data were matched to the entire PLSS section.
The primary difference between the modified three-tier approach and the existing
Rull and Ritz (2003) approach is the use of Landsat imagery to derive additional crop
field information to determine likely locations of PUR applications (Tiers 2A and 2C).
For each ZCTA and census tract in Kern County, organochlorine-, organophosphate-, and
carbamate-specific annual pesticide application rates (lb/ac) were calculated by weighting
land use survey and Landsat crop field- and section-specific application rates by the
proportion of each aerial unit comprised of each crop field or section. ZCTA boundaries
were clipped to the Kern County geographic extent. The weighted average of the
pesticide application rates for each aerial unit was divided by 17 years to calculate an
annual rate (1974 to 1990).

3.5 Rurality Metrics
The 2000 U.S. Census Bureau ZCTAs were used to approximate ZIP code boundaries;
the 2000 U.S. Census Bureau census tracts were used as boundaries (U.S. Census Bureau
2013). Per Grubesic and Matisziw (2006), ZCTAs were checked for water features (HH)

51

and large tracts of land with no mailing addresses/ZIP codes (XX). As ZCTAs may be
spatially discontiguous, ZCTAs were dissolved using 5-digit ZCTA codes.
RUCA codes (2006 ZIP Code Version 2.0 and 2000 Census Tract Version) were
joined to the 2000 U.S. Census Bureau ZCTA and census tract boundaries. Each ZCTA
and census tract was assigned a single RUCA code. Using the recommended
Categorization C, census tracts and ZCTAs with values of 1.0, 1.1, 2.0, 2.1, 3.0, 4.1, 5.1,
7.1, 8.1, or 10.1 were coded as urban; all other values were coded as rural (RHRC 2000).
The 2000 version of UAs and UCs were chosen to be comparable to the RUCA codes
created using 2000 U.S. Census Bureau information. The 2000 U.S. Census Bureau
urban-rural classification system (U.S. Census Bureau 2000) was implemented by coding
any ZCTA and census tract intersecting a UA or UC as urban. All other geographic areas
were coded as rural.

3.6 Statistical Analysis
The following measures and tests quantified the accuracy of RUCA codes and the U.S.
Census Bureau urban-rural system compared to the modified three-tier gold standard
approach. Using 5-digit ZIP codes and census tract Federal Information Processing
Standard (FIPS) codes as identifiers, the following statistical analysis was executed: For
each rurality metric and pesticide chemical class, sensitivity was calculated as the number
of ZCTAs, or census tracts, classified as rural divided by the number of ZCTAs, or
census tracts, truly exposed to pesticides (gold standard). Specificity was calculated as
the number of ZCTAs, or census tracts, classified as not rural (i.e. urban) divided by the
number of ZCTAs, or census tracts, truly not exposed to pesticides (gold standard). For

52

the gold standard, the following pesticide exposure cutoffs were evaluated: >0 lb/ac,
≥50th percentile, and ≥75
th
percentile of annual application rates.
Wilcoxon rank-sum tests determined if pesticide chemical class-specific annual
application rates differed according to geographic aggregation (ZCTAs and census tracts)
as well as if rates differed according to rurality. Separately for each areal aggregation, the
kappa statistic was calculated as the proportion of the observed agreement in rurality
according to each rurality metric not due to chance: (proportion of observed agreement-
proportion of expected agreement due to chance) / (1-proportion of expected agreement
due to chance). Separately for each rurality metric, chi-square and Fisher’s exact tests
determined if the proportion of ZCTAs compared to census tracts categorized as rural
was different, and if the proportion of pesticide-exposed areal units differed according to
the gold standard and each rurality metric. All statistical tests were two-sided (α=0.05).
The data analysis was generated using the SAS System for Windows software, Version
9.3 (Copyright © 2013 SAS Institute Inc. SAS and all other SAS Institute Inc. product or
service names are registered trademarks or trademarks of SAS Institute Inc., Cary, NC,
USA).

53

CHAPTER FOUR: RESULTS
The compiled pesticide database included 157 pesticide active ingredients from three
mutually exclusive pesticide chemical classes: organochlorines, organophosphates, and
carbamates (Appendix A; Tables A1-A3).

4.1 PUR Extraction
There were between 476,981 and 1,305,573 PUR records each year from 1974 to 1989.
Agricultural use PUR records comprised between 45.9 and 78.5% of all PUR records,
representing the majority of PUR records for earlier time periods (Appendix B; Table
B2). Subsequent to logic checks 1 and 2, the majority of agricultural use PUR records
remained eligible for inclusion in the analysis (between 82.1% and 92.3%). As
anticipated, the number of PUR records (N=2,657,840) available in 1990 far exceeded
previous years with the adoption of full-use reporting (Appendix B; Table B3). After
applying logic checks 1 through 7 to PUR records in 1990, 71.4% of agricultural use
PUR records remained (Appendix B; Tables B1-B3).
Table B4 (Appendix B) shows PUR records remaining after extracting
applications associated with the pesticide chemical classes of interest. Organochlorines,
organophosphates, and carbamates together comprised a large proportion of agricultural
use PURs - ranging between 23.2 and 48.8% between 1974 and 1990. Very few outliers
were present (between 0.002 and 0.7%). Between 1974 and 1989, the majority of outliers
met definition 2 (refer to Section 3.4.1). In 1990, the majority of the outliers met
definition 3.

54

Table 5 Kern County agricultural use and chemical class PUR extractions
1

Year Kern County (N)
2
Total (N)
3

1974 11,743 11,715
1975 8,122 8,116
1976 7,228 7,219
1977 6,989 6,985
1978 8,021 8,017
1979 6,393 6,378
1980 6,423 6,419
1981 6,393 6,389
1982 6,455 6,442
1983 7,187 7,181
1984 8,207 8,195
1985 7,226 7,219
1986 9,492 9,486
1987 9,161 9,160
1988 10,898 10,894
1989 9,580 9,568
1990 19,849 19,288
Total 149,367 148,671

1
Data from CDPR (2013)
2
This number reflects PUR records after logic checks were applied and agricultural
use, Kern County location, and pesticide chemical classes were extracted.
3
This number reflects PUR records additionally excluding non-agricultural and/or
ambiguous PUR-derived commodities and applications reporting 0 lb.

Of the 149,367 agricultural use and chemical class-specific PUR records
occurring in Kern County between 1974 and 1990, 148,671 were included in the analysis
after excluding non-agricultural, ambiguous commodities (e.g. outdoor plants in
containers) and 0 lb of active ingredient reported (Table 5). There were N=95,621
organophosphate PUR records, followed by N=38,436 carbamate PUR records, and
N=14,614 organochlorine PUR records. Organophosphates were consistently the most
frequently used type of chemical, followed by carbamates and organochlorines (Figures
11-12). This reflects the ban of organochlorines in the U.S. starting in 1972 - slightly

55

Figure 11 Pounds of agricultural pesticide usage in Kern County by
chemical class (1974-1990) (Data from CDPR 2013)

Figure 12 Agricultural PUR pesticide applications in Kern County by
chemical class (1974-1990) (Data from CDPR 2013)

56

before PUR data became available in 1974 (CDC 2009). As demonstrated in Figures 11
and 12, the number of PUR applications mirrors the quantity of applied pesticides (lb).
PUR application numbers include instances where an application was comprised of
multiple active ingredients applied on a single crop type. The spike in applied pesticides
and PUR applications in 1990 reflects full-use reporting beginning in 1990, where
farmers were required to report all pesticide use, irrespective of restricted-use status (i.e.
the pre-1990 protocol).
The most frequently pesticide-treated crops in Kern County (using PUR
commodity codes) were cotton, onion, and alfalfa for organochlorines; cotton, almond,
and alfalfa for organophosphates; and alfalfa, cotton, and lettuce for carbamates (Table
6). The most frequently used organochlorine, organophosphate, and carbamate pesticide
active ingredient was dicofol (1.4 million lb), dimethoate (1.3 million lb), and methomyl
(1.1 million lb), respectively (Table 7).

Table 6 Pesticide-treated crops by chemical class, Kern County (1974-1990)
1

Pesticide chemical class Crop
2
PUR applications (N)
Organochlorines
Cotton 7,927
Onion 1,711
Alfalfa 1,511
Organophosphates
Cotton 29,302
Almond 12,768
Alfalfa 9,638
Carbamates
Alfalfa 7,514
Cotton 8,083
Lettuce 5,491

1
Data from CDPR (2013)
2
These figures use PUR commodity codes and not land use survey crop codes.

57

Table 7 Common pesticides by chemical class, Kern County (1974-1990)
1

Pesticide chemical
class
Pesticide active
ingredient
PUR
applications (N)
Applied
chemical (lb)
Organochlorines
Dicofol 8,205 1,425,049.0
Dacthal (DCPA) 2,023 877,093.5
Endosulfan 1,547 169,378.7
Methoxychlor 1,284 122,234.0
Quintozene 550 41,285.6
Organophosphates
Dimethoate 9,658 1,263,884.6
Tribufos 9,516 1,950,829.0
Parathion 9,432 1,495,975.6
Azinphos methyl 8,008 1,535,406.1
Diazinon 6,695 803,246.7
Carbamates
Methomyl 21,041 1,061,297.3
Aldicarb 8,001 1,046,430.0
Carbaryl 2,406 1,072,007.0
Carbofuran 2,314 140,738.8
Benomyl 1,593 122,668.1

1
Data from CDPR (2013)

4.2 Crop Signature Library (CSL)
Table 8 lists the Landsat images that were processed for inclusion into the CSL. Figure
13 is a mosaic of Paths 41 and 42 and Rows 35 and 36 created from radiometrically and
atmospherically corrected October 1990 band 3 (red) images subsequent to applying a
median spatial filter, reclassifying negative reflectance values to 0, and cropping to a
geographic extent enclosing Kern County. Figure 14 allows for a larger scale
examination of the crop fields captured by Landsat imagery. Figures 15 and 16 show the
aforementioned processing, but using band 4 (near infrared) Landsat images from
October 1990. Original mosaics for the bands 3 and 4 images are shown in Appendix C
(Figures C1-C2).

58

Table 8 Landsat images from 1990 used for crop signature library
1,2

Month

Path 41 Path 42
Row 35 Row 36 Row 35 Row 36
January
Cloud cover 10% 0% 0% 0%
Acquisition 1/22/1990 1/22/1990 1/29/1990 1/29/1990
February
Cloud cover 0% 0% 0% 0%
Acquisition 2/15/1990 2/15/1990 2/14/1990 2/14/1990
March
Cloud cover 10%
Excluded
0% 10%
Acquisition 3/27/1990 3/18/1990 3/18/1990
April
Cloud cover 40% 20% 0% 40%
Acquisition 4/28/1990 4/28/1990 4/3/1990 4/3/1990
May
Cloud cover 10% 10% 0% 10%
Acquisition 5/30/1990 5/30/1990 5/5/1990 5/5/1990
June
Cloud cover
Excluded Excluded
0% 10%
Acquisition 6/6/1990 6/6/1990
July
Cloud cover None
available
None
available
0% 50%
Acquisition 7/8/1990 7/8/1990
August
Cloud cover 10% 10% 0% 0%
Acquisition 8/18/1990 8/18/1990 8/25/1990 8/25/1990
September
Cloud cover 0% 0% 0% 20%
Acquisition 9/3/1990 9/3/1990 9/10/1990 9/10/1990
October
Cloud cover 0% 0% 0% 0%
Acquisition 10/5/1990 10/5/1990 10/12/1990 10/28/1990

1
Data from USGS (2013b)
2
Landsat images were not available for November and December 1990. Images were excluded
due to excessive cloud cover overlapping the Kern County geographic extent.

59

Figure 13 Landsat mosaic (band 3), Paths 41-42 and Rows 35-36, from October 1990 cropped to Kern County
(Data from U.S. Census Bureau 2013; and USGS 2013b)

60

Figure 14 Inset of Landsat mosaic (band 3) from October 1990, showing crop fields in
Kern County (Data from USGS 2013b)

Figure 15 Inset of Landsat mosaic (band 4) from October 1990, showing crop fields in
Kern County (Data from USGS 2013b)

61

Figure 16 Landsat mosaic (band 4), Paths 41-42 and Rows 35-36, from October 1990 cropped to Kern County
(Data from U.S. Census Bureau 2013; and USGS 2013b)

62

NDVI images were created using the red and near infrared bands for each month
between January and October 1990. For example, in October 1990, NDVI values ranged
between -0.52 and 1; negative values are indicative of non-vegetation (e.g. barren rock)
and positive values closer to 1 are indicative of dense vegetation (Figures 17 and 18).

Figure 17 Inset of NDVI image created from red and near infrared Landsat bands,
October 1990 (Data from USGS 2013b)

63

Figure 18 NDVI image cropped to Kern County, October 1990
(Data from U.S. Census Bureau 2013; USGS 2013b)

64

4.2.1 Stratified Random Sampling (SRS)
Stratified random sampling (SRS) was applied to the 1990 Kern County land use survey
in geographic areas meeting three eligibility criteria. SRS-eligible land use survey
polygons must have been: (1) single-use; (2) at least 4 ha in area; and (3) within the
cloud-free zone (Table 9). A total of 16,635 of the 16,769 land use survey polygons
included in the 1990 Kern County dataset satisfied the single-use criterion, followed by
12,197 satisfying the single-use and area criteria, and 11,832 satisfying all three criteria.
The majority of crop types excluded from 1990 Kern County land use survey due to
double-cropping or intercropping/mixed (i.e. not single-use) were grain and hay crops
(N=55; 41% non-single-use), almonds (N=33; 24.6%), and potatoes (N=29; 21.6%)
(Table 10). Figure 19 shows the cloud-free zone representing the geographic area where
cloud- and shadow-free NDVI images were available for all months in 1990. As July
1990 was missing Path 41, this portion of the study area was excluded.
Out of the 11,832 land use survey polygons eligible for SRS, 1,423 were
randomly selected within 81 land use strata (excluded Z: outside of study area)
(Appendix D; Table D1). At most 30 samples were randomly selected within each
stratum (Figure 20). However, 49 land use classes had samples sizes less than 30 due to a
low prevalence of such classes in Kern County subsequent to applying the
aforementioned eligibility criteria. SRS samples for these land use classes may not be
representative of the strata.

65

Table 9 Eligibility criteria for SRS
1

Eligibility criteria Sample (n)
a

Single-use 16,635
≥4 ha 12,197
In cloud-free zone 11,832

1
Data from CDWR (2013)

Table 10 Land use classes excluded from SRS due to multiuse
1

Land use class N (%)
Grain and hay crop 55 (41.0%)
Almond 33 (24.6%)
Potato 29 (21.6%)
Onion and garlic 7 (5.2%)
Corn 3 (2.2%)
Carrot 2 (1.5%)
Melon, squash, cucumber 2 (1.5%)
Cotton 1 (0.8%)
Pistachio 1 (0.8%)
Tomato 1 (0.8%)

1
Data from CDWR (2013)

NDVI values for each pixel of each NDVI image between January and October
1990 intersecting any of the SRS-sampled land use survey polygons were extracted. A
total of 645,127 NDVI values were extracted for each month - from a total of 6,451,270
NDVI values contributing to the 1990 CSL (Appendix D; Table D2). This was
subsequent to removing one pixel with an NDVI value associated with a point that
straddled the line between two land use survey polygons. For any given month between
January and October 1990, the native vegetation land use class contributed the most
NDVI values to the CSL (N=149,648) across all of its 30 SRS-sampled polygons. This
reflects the typically large size of its polygons (median 71.75 ac vs. 43.44 ac for all other

66

Figure 19 Cloud-free zone of 1990 Landsat images available for CSL
(Data from U.S. Census Bureau 2013; and USGS 2013)

67

Figure 20 Land use survey polygons sampled via SRS, Kern County, 1990
(Data from CDWR 2013; and U.S. Census Bureau 2013)

68

land use classes). Commercial (motel) contributed the fewest NDVI values (n=1 SRS;
N=52 NDVI). Out of the agricultural land use classes of interest to the analysis, pistachio
contributed the most NDVI values (n=30 SRS; N=26,878 NDVI) and avocado
contributed the fewest NDVI values (n=1 SRS; N=61 NDVI).
Figure 21 shows select SRS-sampled land use survey polygons and their extracted
NDVI pixel values for October 1990. Median NDVI values for the selected peach and

Figure 21 Median NDVI values for select SRS-sampled land use survey polygons,
October 1990. The polygons are located near PLSS section 15M32S29E14.
(Data from CDWR 2013)

69

nectarine land use survey polygons are similar (ranging from 0.4923 to 0.5116).
Variability between different land use survey polygons is manifest in the low median
NDVI value of the vineyard land use survey polygon (0.1404) relative to the highest
median NDVI value of the peach and nectarine land use survey polygons (0.5116).
As an objective improvement over the Maxwell (2011) approach, all NDVI values
of each SRS-sampled polygon were used to compute a median NDVI value for that
specific polygon - harnessing all available spectral data from the NDVI images. This is in
contrast to the Maxwell (2011) method, which selects one pixel per polygon. The final
CSL contained median NDVI values for all SRS-sampled polygons from each land use
class (using the median NDVI value for each polygon) from January and October 1990.
All negative NDVI values were recoded to 0 (no vegetation) (Beck et al. 2006). Refer to
Appendix D (Figures D1-D56) for figures showing NDVI values of all agricultural land
use classes included in the CSL.

4.3 Classification of 1985 Landsat Imagery
4.3.1 Segmentation
Table 11 lists the Landsat images used to create a classified 1985 Kern County crop field
layer using the CSL. February was excluded due to the absence of the majority of
agricultural crop fields (Path 42) and November and December were not considered as
the CSL only extended into October. A Landsat Multispectral Scanner (MSS) image was
available for July 1985, but not used due to a different spatial resolution compared to
Thematic Mapper (TM) images. Subsequent to radiometric and atmospheric processing,
mosaicking, reclassification, spatial filtering, cropping to Kern County, and NDVI

70

Table 11 Landsat images from 1985 used for classification
1,2

Month

Path 41 Path 42
Row 35 Row 36 Row 35 Row 36
January
Cloud cover
Excluded Excluded
20% 0%
Acquisition 1/31/1985 1/31/1985
February
Cloud cover 10% 10% None
available
None
available Acquisition 2/25/1985 2/25/1985
March
Cloud cover
Excluded Excluded
0% 0%
Acquisition 3/20/1985 3/20/1985
April
Cloud cover 10% 10% 0% 50%
Acquisition 4/14/1985 4/14/1985 4/5/1985 4/5/1985
May
Cloud cover 10% 10% 0% 20%
Acquisition 5/16/1985 5/16/1985 5/23/1985 5/23/1985
June
Cloud cover 10% 1% 0% 18%
Acquisition 6/1/1985 6/17/1985 6/8/1985 6/8/1985
July
Cloud cover 0% 0% 1% None
available Acquisition 7/3/1985 7/3/1985 7/26/1985
August
Cloud cover None
available
10% 0% 50%
Acquisition 8/20/1985 8/11/1985 8/11/1985
September
Cloud cover 0% 0% 1% 0%
Acquisition 9/21/1985 9/21/1985 9/12/1985 9/12/1985
October
Cloud cover
Excluded Excluded
0% 0%
Acquisition 10/14/1985 10/14/1985

1
Data from USGS (2013b)
2
Images were excluded due to excessive cloud cover overlapping the Kern County geographic
extent

calculations, all NDVI images were clipped to the geographic extent containing NDVI
values for all months in January and March through October and containing no clouds.
Figure 22 shows the geographic extent of the segmentation-eligible zone (no clouds and
available 1985 data) compared to the cloud-free zone used to determine eligible SRS land
use survey polygons. Note that some reflectance values exceed 1, which is physically
acceptable and may arise from bright surfaces (e.g. clouds) (YCEO 2013).

71

Figure 22 Segmentation-eligible zone vs. cloud-free CSL zone, overlaying Landsat mosaic (band 3)
from September 1985 (Data from USGS 2013b; and U.S. Census Bureau 2013)

72

4.3.2 Principal Component Analysis (PCA)
A principal component analysis (PCA) was performed using the nine NDVI images
(January, March to October 1985) (Table 12). Over 82% of the overall variance is
explained by the first three components. These three principal components were used to
create a crop field boundary layer via segmentation (Appendix E; Figure E1). The
segmented polygon feature class consisted of 19,752 segments, each representing a
spectrally homogeneous grouping of pixels derived from the three input principal
components (Figures 23 and 24).

Table 12 Principal component analysis of Landsat 1985 NDVI images
Principal
component Variance (%)
Principal
component Variance (%)
1 56.06 6 3.15
2 16.84 7 2.82
3 9.42 8 2.40
4 4.63
9 1.13
5 3.54

4.3.3 Classification Using Sum of Squared Difference (SSD)
NDVI values for each segment were extracted from monthly NDVI images in 1985.
There were a total of 7,825,045 NDVI values across all segments for each month. There
was an average of 396 NDVI values intersecting each segment (median 258, minimum
19, maximum 7,742), and thus contributing to the classification of each segment. Median
NDVI values for each segment were compared to median NDVI values from each land
use class in the CSL using a sum of squared differences (SSD) measure. Each segment
was classified by assigning it to one land use class based on the smallest SSD. Three

73

Figure 23 Segments of spectrally homogeneous pixels, basis of crop field boundaries for
classifying 1985 Landsat NDVI images

different CSL classification approaches were executed, each more conservative than the
previous (Table 13). Minimum SSDs did not differ between classifications 1 and 2
(median of 0.06) (Table 14). The minimum SSD of classification 3 (median of 0.08) was
slightly different from classifications 1 and 2 - indicating a larger difference in NDVI
values for land use matches using classification 3.

74

Figure 24 Segments overlaying color-infrared Landsat image
from August 1985 (Data from USGS 2013b)

75

Table 13 CSL classification approaches for segmented crop layer
Classification Description
Land use
classes (N)
1: Standard
Excluded Z land use class
(outside of study area)
81
2: Subclass-
required
Excluded land use classes without specified
subclass
75
3: Strict
Excluded land use classes without specified
subclass and with SRS samples <30
28

Table 14 Classification: minimum sum of squared differences (SSD)
N Mean  SD Median (IQR) Min. Max.
Classification 1 19,752 0.09  0.09 0.06 (0.10) 0.001 0.84
Classification 2
1
19,752 0.09  0.09 0.06 (0.11) 0.001 0.84
Classification 3 19,752 0.12  0.12 0.08 (0.13) 0.002 0.87

1
Classification 2 (subclass-required) was selected to classify 1985 Landsat images.

However, the crop types with the largest SSDs did vary according to classification
method (using 99
th
percentile cutoff). For classification 1 among segments with SSDs
≥0.39 NDVI, apricots comprised the majority (N=21; 10.6%) (data not shown). For
classification 2 among segments with SSDs ≥0.40 NDVI, cole crops comprised the
majority (N=30; 15.2%). For classification 3 among segments with SSDs ≥0.54 NDVI,
oranges comprised the majority (N=54; 27.4%). Results may differ if using land use
class-specific SSD percentiles.
In terms of the 20 most frequently classified land use classes (Appendix E;
Figures E2-E4), classifications 1 and 2 were similar, as a large number of segments were
classified as jojoba (classification 1: N=2,308; classification 2: N=2,618) and cotton
(classification 1: N=1,879; classification 2: N=1,879). Classification 3 was dramatically

76

different from classifications 1 and 2, as most segments were classified as feed lots
(N=3,335), a semi-agricultural land use class.
After comparing and contrasting the three classified crop field layers,
classification 2 (subclass-required) was chosen for its comparability to classification 1 in
land use classification frequencies, to address the potential heterogeneity in including
land use classes without a specified subclass, and the decision to not break the SRS
randomization used in creating the CSL. Figure 25 is a map showing the preprocessed
classified crop field layer (using classification 2), which is dominated by jojoba crop
fields along the periphery. Cotton fields displayed in blue are interspersed throughout the
layer.

4.3.4 Processing CSL-Classified Crop Fields
The original classified layer (Figure 25) was iteratively processed to exclude non-
agricultural segments using the 1990 Kern County land use survey and a color-infrared
(CIR) band combination of Landsat images in August 1985. Using the CSL-classified
land use classes assigned to the segments, non-agricultural classes were deleted (e.g.
urban). There was a potential misclassification of segments assigned to the jojoba land
use class. A large number of segments were classified as jojoba (N=1,572 after
aforementioned processing). However, there were few jojoba land use survey polygons
according to the 1990 Kern County land use survey (N=9). It is conceivable that jojoba
constituted some of the native vegetation land use classes, commonly present in the same
peripheral geographic regions of the land use survey. The discrepancy between the
relatively few jojoba crop fields in Kern County in 1990, a short time after the 1985

77

Figure 25 Classification 2-derived segments prior to processing

78

Landsat images were acquired (used for CSL classification), lends support to likely
misclassification. These jojoba-classified segments potentially share a similar spectral
profile between the months of January and March to October with some other land use
class either included in or absent from the CSL (Appendix D; Figure D26). All jojoba
segments were excluded from the analysis.
Figure 26 shows the final CSL-classified 1985 Kern County crop field boundaries
including 10,008 crop fields. The percentages of agricultural use polygons in the 1990
land use survey (Appendix E; Table E1) slightly differ from the CSL-classified layer
(Appendix E; Table E2). Note that the land use survey geographic extent is larger than
that of the CSL-classified layer. There was a total of 49 land use classes in the 1985
classified layer - predominantly cotton (N=1,878; 18.76%). This number overestimates
the actual number of cotton fields as single cotton fields may have been represented as
multiple adjacent segments by virtue of the segmentation process.

4.4 Modified Three-Tier Approach
Processed PUR pesticide records (lb of applied AI) for every year between 1974 and
1990 were matched to the 1990 Kern County land use survey, the newly created CSL-
classified crop field layer derived from 1985 Landsat data, or PLSS section data (Figures
27-29). Across all three chemical classes, the majority of PUR records matched to Tier 1:
84.5% of organochlorine records, 85.5% of organophosphate records, and 83% of
carbamate records. The contribution of the CSL-classified Landsat layer was modest,
ranging from 2.1 to 2.4% for Tier 2A and from 0.1 to 0.2% for Tier 2C. Had the land use
survey not been used as the first tier, more PUR records would likely match to the

79

Figure 26 Finalized classification 2-derived segments subsequent to processing

80

Landsat-derived crop fields - particularly PUR records of applied pesticides close in time
to 1985. Percentage tier matches may also differ when examining individual years. Note
that the number of PUR records reflects the application of different individual pesticide
active ingredients. Some records may also be a part of the same pesticide application,
comprised of multiple pesticides ingredients applied on crop.
The PUR records for each chemical class between 1974 and 1990 are associated
with 2,952,761.16 lb (N=14,614 PUR records) for organochlorines, 12,367,594.88 lb
(N=95,621) for organophosphates, and 3,777,562.21 lb (N=38,436) for carbamates. Note
that these totals reflect rounding error inherent in distributing PUR pounds to all crop
fields in a PLSS section via Tiers 2B and 2C.

Figure 27 Organochlorine PUR tier matches, Kern County (1974-1990)

81

Figure 28 Organophosphate PUR tier matches, Kern County (1974-1990)

Figure 29 Carbamate PUR tier matches, Kern County (1974-1990)

82

4.4.1 Contribution of Landsat Imagery to Modified Three-Tier PUR Matching
According to the most specific crop-matching Tiers 1 and 2A where PUR records were
matched exactly to crop type, collapsed field crops (grain and hay crops, field crops,
pasture, and truck nursery and berry crops) consistently represented the majority of
matches (Tables 15-17). More specifically, the Landsat-derived crop field boundaries
were useful in matching collapsed field crops, in addition to the subtropical fruits
(lemons, grapefruits) and deciduous fruits and nuts (walnuts, almonds, peaches and
nectarines, plums) agricultural land use classes.

Table 15 Organochlorines: Tiers 1 and 2A matched crops
Tier Crop type N (%)
Tier 1 (N=12,355)
Collapsed field crop 12,021 (97.3%)
Orange 261 (2.1%)
Almond 21 (0.2%)
Lemon 16 (0.1%)
Plum 14 (0.1%)
Apple 8 (0.1%)
Peach and nectarine 6 (0.05%)
Walnut 5 (0.04%)
Grapefruit 3 (0.02%)
Tier 2A (N=304)
Collapsed field crops 270 (88.8%)
Lemon 20 (6.6%)
Grapefruit 9 (3.0%)
Pear 3 (1.0%)
Apple 1 (0.3%)
Plum 1 (0.3%)

83

Table 16 Organophosphates: Tiers 1 and 2A matched crops
Tier Crop type N (%)
Tier 1 (N=81,784)
Collapsed field crops 60,067 (73.4%)
Almond 11,626 (14.2%)
Orange 6,238 (7.6%)
Peach and nectarine 1,230 (1.5%)
Apple 781 (1.0%)
Pistachio 614 (0.8%)
Plum 562 (0.7%)
Lemon 214 (0.3%)
Walnut 183 (0.2%)
Kiwi 76 (0.1%)
Olive 67 (0.1%)
Apricot 43 (0.1%)
Grapefruit 32 (0.04%)
Rice 21 (0.03%)
Fig 14 (0.02%)
Cherry 10 (0.01%)
Prune 4 (0.005%)
Avocado 1 (0.001%)
Pear 1 (0.001%)
Tier 2A
(N=2,319)
Collapsed field crops 1,539 (66.4%)
Lemon 167 (7.2%)
Peach and nectarine 164 (7.1%)
Almond 146 (6.3%)
Grapefruit 94 (4.1%)
Plum 87 (3.8%)
Apple 29 (1.3%)
Kiwi 27 (1.2%)
Orange 24 (1.0%)
Prune 12 (0.5%)
Walnut 10 (0.4%)
Apricot 8 (0.3%)
Pear 5 (0.2%)
Miscellaneous deciduous 3 (0.1%)
Miscellaneous subtropical fruit 2 (0.1%)
Olive 1 (0.04%)
Rice 1 (0.04%)

84

Table 17 Carbamates: Tiers 1 and 2A matched crops
Tier Crop type N (%)
Tier 1 (N=31,900)
Collapsed field crops 26,295 (82.4%)
Orange 2,808 (8.8%)
Almond 1,300 (4.1%)
Peach and nectarine 973 (3.1%)
Plum 143 (0.4%)
Apple 135 (0.4%)
Lemon 84 (0.3%)
Pistachio 65 (0.2%)
Miscellaneous deciduous 28 (0.1%)
Grapefruit 23 (0.1%)
Olive 20 (0.1%)
Apricot 19 (0.1%)
Cherry 7 (0.02%)
Tier 2A (N=912)
Collapsed field crops 573 (62.8%)
Peach and nectarine 129 (14.1%)
Grapefruit 58 (6.4%)
Miscellaneous deciduous 48 (5.3%)
Lemon 47 (5.2%)
Almond 20 (2.2%)
Plum 14 (1.5%)
Apricot 8 (0.9%)
Apple 7 (0.8%)
Olive 3 (0.3%)
Orange 2 (0.2%)
Pear 2 (0.2%)
Prune 1 (0.1%)

Figure 30 shows a specific example of a Tier 2A match, where Landsat NDVI
images identified an almond field treated with organophosphate pesticides between 1974
and 1990 not otherwise found in the land use survey. An orange orchard (from land use
survey) was present in section 15M27S27E06, which provided a Tier 1 match according
to specific crop type and section. However, had a Tier 2A not been implemented, the

85

14,565.55 lb applied on an almond field would have been matched to the orange orchard
using the standard Rull and Ritz (2003) approach (i.e. the 2
nd
tier).
Figures 31 through 33 illustrate the pounds of applied active ingredient pesticide
(1974-1990) matched to any of the tiers following a spatial union of the land use survey,
CSL-classified, and PLSS section layers. The spatial union overestimates the actual
number of tier-matched pounds of AI due to the intersecting land use survey and Landsat
crop fields. However, the union is informative in illustrating the overall pattern of
pesticide applications across Kern County. As reflected in the consistently higher
pesticide use with organophosphates during any given year (Figure 11), organophosphate
pesticide use was more dispersed throughout Kern County compared to organochlorines
and carbamates. Higher organophosphate pesticide usage was also more concentrated and
more frequently occurring (≥5,000 lb) in the central and northwestern portions of the
county. This geographic pattern of concentrated pesticide usage was also observed with
organochlorines and carbamates, which is where a large portion of the Central Valley
agricultural fields are located. Crop fields derived from land use survey, Landsat, and
PLSS section data not treated with pesticides are not shown.

86

Figure 30 Tier2A match provided by Landsat, organophosphate PUR applications,
1974-1990 (Data from CDPR 2013; and CDWR 2013)

87

87

Figure 31 Organochlorines: applied pesticides on crop fields and sections, Kern County (1974-1990)
(Data from CDPR 2013; and U.S. Census Bureau 2013)

88

88

Figure 32 Organophosphates: applied pesticides on crop fields and sections, Kern County (1974-1990)
(Data from CDPR 2013; and U.S. Census Bureau 2013)

89

89

Figure 33 Carbamates: applied pesticides on crop fields and sections, Kern County (1974-1990)
(Data from CDPR 2013; and U.S. Census Bureau 2013)

90

4.5 Annual Pesticide Application Rates by Areal Aggregation
Figure 34 shows the 47 ZCTAs intersecting some portion of Kern County. The absence
of ZCTAs in some areas of the county reflects the absence of ZIP codes, and thus mail
delivery, in these regions (Grubesic and Matisziw 2006). At the ZCTA level, annual
pesticide application rates differed according to pesticide chemical class (Figures 35-37).
Consistent across all classes are higher application rates in the central and northwestern
portions of Kern County. Pesticide application rates were highest for organophosphates,
ranging between 0 and 1.36 lb/ac (Figure 36). This was followed by carbamates, ranging
between 0 and 0.39 lb/ac (Figure 37), and organochlorines ranging between 0 and 0.25
lb/ac (Figure 35). Organochlorine usage was absent in nine ZCTAs, followed by five
ZCTAs absent of carbamate usage, and two ZCTAs absent of organophosphate usage.

Figure 34 Kern County ZCTAs (Data from U.S. Census Bureau 2013)

91

Figure 35 Organochlorines: ZCTA-level annual pesticide application rates, Kern County
(1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

Figure 36 Organophosphates: ZCTA-level annual pesticide application rates, Kern
County (1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

92

Figure 37 Carbamates: ZCTA-level annual pesticide application rates, Kern County
(1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

Figure 38 shows the 140 census tracts located in Kern County. Annual pesticide
application rates were highest for organochlorines (maximum of 1.57 lb/ac) (Figure 39).
Organophosphate usage was relatively high, upwards of 1.41 lb/ac (Figure 40).
Carbamate-specific annual application rates ranged between 0 and 0.55 lb/ac (Figure 41).
Looking across all three chemical classes, there was an absence of pesticide applications
in central Kern County near the city of Bakersfield, but higher rates in northwestern Kern
County that decreased as one moved eastward towards the Sierra Nevada Mountains.
Forty census tracts were absent of organochlorine usage, followed by 29 census tracts
absent of carbamate usage, and 16 census tracts absent of organophosphate usage.

93

Figure 38 Kern County census tracts (Data from U.S. Census Bureau 2013)

Figure 39 Organochlorines: census tract-level annual pesticide application rates, Kern
County (1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

94

Figure 40 Organophosphates: census tract-level annual pesticide application rates, Kern
County (1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

Figure 41 Carbamates: census tract-level annual pesticide application rates, Kern County
(1974-1990) (Data from CDPR 2013; and U.S. Census Bureau 2013)

95

As a result of the choice of areal aggregations under study, some PUR tier
matches were not included in calculating ZCTA-level application rates due to the absence
of ZCTAs in those geographic areas. The affected PUR records matched to six land use
survey crop fields and 15 PLSS sections for organochlorines, 15 land use survey crop
fields, three Landsat-derived crop fields, and 80 PLSS sections for organophosphates, and
12 land use survey crop fields, two Landsat-derived crop fields, and 48 PLSS sections for
carbamates. Census tract rates were not affected due to the entire extent of Kern County
being covered by these areal units.
Table 18 shows the number of pesticide-treated crop fields (from land use survey
or Landsat) and PLSS sections stratified by areal aggregation and pesticide chemical
class. ZCTAs were more frequently intersected with organophosphate-treated crop fields
and sections (median 15 fields/sections; maximum 658), which is reflected in the fewer
number of ZCTAs absent of organophosphate pesticide applications (N=2). This pattern
persists at the census tract level (median 4 fields/sections; maximum 828), where again,
relatively few census tracts were absent of organophosphate applications (N=16). Note
that a single crop field or section could have intersected multiple ZCTAs, crop fields
belonging to the collapsed field crop group are represented as a single multipart polygon
within a section, and that a Landsat-derived crop field may intersect a land use survey
crop field.

4.6 Descriptive Analysis: Areal Aggregation and Pesticide Exposure
Despite the variation in shapes and sizes of the areal units under study, ZCTA- and
census tract-level annual application rates were comparable across all three pesticide

96

Table 18 Pesticide-treated crop fields and sections intersecting areal units
ZCTAs (N=47)
Mean ± SD Median (IQR) Min. Max.
Organochlorines 62.1 ± 105.8 6 (133) 0 457
Organophosphates 100.1 ± 157.7 15 (167) 0 658
Carbamates 82.4 ± 135.0 12 (145) 0 577
Census tracts (N=140)
Organochlorines 22.2 ± 69.7 1 (7) 0 486
Organophosphates 36.1 ± 107.6 4 (11) 0 828
Carbamates 29.8 ± 90.2 3 (10) 0 657

chemical classes (Table 19). Census tract-level organochlorine rates were comparable to
ZCTA-level rates (median 0.001 lb/ac) (p=0.5705). Census tract-level organophosphate
rates were slightly higher (median 0.02 lb/ac) than ZCTA-level rates (median 0.01 lb/ac)
(p=0.6104). Census tract-level carbamate rates were also comparable to ZCTA-level rates
(median 0.01 lb/ac) (p=0.9801). Higher maximum application rates were typically
observed when aggregated at the census tract level.

Table 19 Annual pesticide application rates according to areal aggregation

Organochlorines
N Mean ± SD Median (IQR) Min. Max. p
1

ZCTA 47 0.03 ± 0.06 0.001 (0.08) 0 0.25
0.5705
Census tract 140 0.04 ± 0.15 0.001 (0.03) 0 1.57
Organophosphates
ZCTA 47 0.16 ± 0.29 0.01 (0.30) 0 1.36
0.6104
Census tract 140 0.13 ± 0.24 0.02 (0.14) 0 1.41
Carbamates
ZCTA 47 0.05 ± 0.08 0.01 (0.06) 0 0.39
0.9801
Census tract 140 0.05 ± 0.09 0.01 (0.05) 0 0.55

1
Wilcoxon rank-sum test

97

4.7 Kern County Rurality
The geographic pattern of rurality varied according to aggregation and rurality metric.
When evaluating ZCTA-level rurality, rural geographic areas were predominantly located
in the western and eastern portions of Kern County for both metrics (Figure 42).
Agreement in ZCTA-level rurality between RUCA codes and the U.S. Census Bureau
metric was poor beyond chance (kappa=0.03; Table 20). Among rural ZCTAs
categorized using the U.S. Census Bureau metric, 55.6% were categorized as urban using
RUCA codes. Among rural ZCTAs categorized using RUCA codes, 60% were
categorized as urban using the U.S. Census Bureau metric.
At the census tract level, geographic patterns of rurality were also dissimilar
compared to ZCTA-level patterns. A larger proportion of Kern County was designated as
rural according to RUCA codes compared to the U.S. Census Bureau metric (Figure 43).
Agreement in census tract-level rurality was also poor between RUCA codes and the U.S.
Census Bureau metric (kappa=0.04; Table 20). Among rural census tracts categorized
using the U.S. Census Bureau metric, 71.4% were categorized as urban using RUCA
codes. Among rural census tracts categorized using RUCA codes, 92.9% were
categorized as urban using the U.S. Census Bureau metric.
Rurality designations significantly differed according to areal aggregation across
both rurality metrics. When using the RUCA metric, a larger proportion of ZCTAs was
categorized as rural (42.6%) compared to census tracts (20%) (p=0.0022; Table 21).
When using the U.S. Census Bureau metric, a larger proportion of ZCTAs was also
categorized as rural (38.3%) compared to census tracts (5%) (p<0.0001; Table 21). Given
the larger number of census tracts compared to ZCTAs in Kern County, census tracts, by

98

Figure 42 ZCTA-level rurality (Data from RHRC 2000; and U.S. Census Bureau 2013)

99

Table 20 RUCA and U.S. Census Bureau metric designations by areal aggregation

ZCTA
RUCA
Rural Urban Total Kappa
1

U.S. Census
Bureau
Rural 8 10 18
0.03 Urban 12 17 29
Total 20 27 47
Census tract
U.S. Census
Bureau
Rural 2 5 7
0.04 Urban 26 107 133
Total 28 112 140

1
Kappa values range between -1 and +1. Kappa values <0.4 indicate poor agreement beyond
chance. Kappa values between 0.4 and 0.75 indicate fair to good agreement beyond chance.
Kappa values >0.75 indicate excellent agreement beyond chance.

design, may be more homogeneous aggregations. In other words, since ZCTAs are
relatively larger in area (median area: 56,288.28 ac), their structure may mask urban/rural
differences within the ZCTA that is better captured when Kern County is partitioned
according to smaller census tracts (median area: 554.5 ac), which may explain some of
the differences in rurality designations by aggregation.
The extent to which rural areal units are representative of pesticide application
practices is highlighted in Tables 23 and 24. Urban ZCTAs and census tracts, whether
defined according to RUCA codes or the U.S. Census Bureau metric, were consistently
characterized by higher application rates between 1974 and 1990 across all three
pesticide chemical classes. Interestingly, rates were significantly different between urban
and rural ZCTAs using the U.S. Census Bureau metric across all pesticide chemical
classes. For example, carbamate application rates among urban ZCTAs (using the U.S.
Census Bureau metric) were significantly higher (median 0.01 lb/ac) compared to rural
ZCTAs (median 0.0005 lb/ac) (p=0.0011). Similar results were observed among census

100

Figure 43 Census tract-level rurality.
Sources: Data from RHRC (2000); U.S. Census Bureau (2013).

101

Table 21 ZCTA vs. census tract rurality designations

RUCA
Rural Urban Total p
1,2

ZCTA 20 27 47
0.0022**
Census tract 28 112 140
U.S. Census Bureau urban-rural classification
ZCTA 18 29 47
<0.0001***
Census tract 7 133 140

1
Chi-square test
2
**p<0.01; *** p<0.001

tracts; however, rates were significantly different according to RUCA code designations.
For example, carbamate application rates among urban census tracts (using RUCA codes)
were significantly higher (median 0.01 lb/ac) compared to rural census tracts (median
0.0003 lb/ac) (p=0.0074). Maximum application rates were consistently higher among
urban census tracts.
The median number of pesticide-treated crop fields and sections was typically
higher among urban ZCTAs (median 7-32 fields) compared to rural ZCTAs (median 1.5-
10.5 fields), as reflected in the relatively higher application rates (Appendix F; Table F1).
However, the median number of treated fields and sections was generally higher among
rural census tracts (median 1-38 fields) vs. urban census tracts (median 1-4 fields),
though the application rates do not reflect this pattern (Appendix F; Table F2). Larger
urban-rural differences were observed when examining the U.S. Census Bureau metric.
Overlaying rural and urban ZCTAs and census tracts and the tiered PUR matches
also mirror the results shown in Tables 22 and 23 (Appendix F; Figures F2-F7).
Pesticide-treated crop fields and sections across all three chemical classes frequently
intersect urban ZCTAs and census tracts along the central portions of Kern County.

102

Table 22 Pesticide rates stratified by rurality: ZCTAs
Organochlorines

N
Mean ±
SD
Median (IQR) Min. Max. p
1,2

RUCA
Rural 20 0.02 ± 0.04 0.0002 (0.020) 0 0.13
0.0533
Urban 27 0.04 ± 0.07 0.005 (0.100) 0 0.25
U.S.
Census
Bureau
Rural 18 0.01 ± 0.03 0.0001 (0.002) 0 0.11
0.0093**
Urban 29 0.05 ± 0.07 0.005 (0.100) 0 0.25
Organophosphates
RUCA
Rural 20 0.20 ± 0.38 0.001 (0.22) 0 1.36
0.1822
Urban 27 0.13 ± 0.19 0.02 (0.30) 0 0.68
U.S.
Census
Bureau
Rural 18 0.04 ± 0.09 0.001 (0.01) 0 0.31
0.0056**
Urban 29 0.24 ± 0.34 0.02 (0.37) 0.001 1.36
Carbamates
RUCA
Rural 20 0.05 ± 0.10 0.001 (0.040) 0 0.39
0.1186
Urban 27 0.04 ± 0.06 0.01 (0.090) 0 0.19
U.S.
Census
Bureau
Rural 18 0.01 ± 0.02 0.0005 (0.002) 0 0.06
0.0011**
Urban 29 0.07 ± 0.10 0.01 (0.100) 0.001 0.39

1
Wilcoxon rank-sum test
2
**p<0.01

4.7.1 Accuracy Assessment of Rurality
The accuracy of using ZCTA-level rurality metrics as a surrogate measure for pesticide
exposure varied according to rurality metric, pesticide chemical class, and GIS metric
(gold standard) pesticide exposure cutoff (Tables 24-29). Specificity was consistently
higher than sensitivity. When evaluating RUCA codes, sensitivity was generally highest
when using a 0 lb/ac cutoff, decreasing when using a 50
th
percentile cutoff, and

103

Table 23 Pesticide rates stratified by rurality: census tracts
Organochlorines

N
Mean ±
SD
Median (IQR) Min. Max. p
1,2

RUCA
Rural 28 0.01 ± 0.02 0.00004 (0.010) 0 0.10
0.0398*
Urban 112 0.05 ± 0.17 0.003 (0.050) 0 1.57
U.S.
Census
Bureau
Rural 7 0.02 ± 0.04 0.0004 (0.002) 0 0.10
0.5390
Urban 133 0.05 ± 0.15 0.001 (0.040) 0 1.57
Organophosphates
RUCA
Rural 28 0.05 ± 0.12 0.003 (0.05) 0 0.55
0.0078**
Urban 112 0.15 ± 0.25 0.03 (0.19) 0 1.41
U.S.
Census
Bureau
Rural 7 0.06 ± 0.12 0.01 (0.06) 0 0.32
0.5822
Urban 133 0.14 ± 0.24 0.02 (0.14) 0 1.41
Carbamates
RUCA
Rural 28 0.01 ± 0.03 0.0003 (0.01) 0 0.09
0.0074**
Urban 112 0.06 ± 0.10 0.01 (0.08) 0 0.55
U.S.
Census
Bureau
Rural 7 0.01 ± 0.02 0.003 (0.02) 0 0.06
0.4538
Urban 133 0.05 ± 0.10 0.01 (0.05) 0 0.55

1
Wilcoxon rank-sum test
2
*p<0.05; **p<0.01

increasing when using a 75
th
percentile cutoff. Specificity when using RUCA codes
followed a similar pattern, where specificity decreased when using a 50
th
percentile cutoff
compared to a 0 lb/ac cutoff, but was highest when using a 75
th
percentile cutoff. On the
other hand, when examining the U.S. Census Bureau metric, sensitivity decreased and
specificity increased as the pesticide exposure cutoffs became more conservative.
Sensitivity ranged between 25 and 42.9% for RUCA codes. In other words, the
probability of a ZCTA being classified as rural given the ZCTA was pesticide-exposed
ranged between 0.25 and 0.429 - where a probability of 1 is perfect sensitivity. Another
to express this result is to state that RUCA codes correctly identified between 25 and

104

Table 24 ZCTA-level accuracy of RUCA codes: organochlorines

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

RUCA
Exposed 15 5
39.5% 44.4% 0.4653 Not
exposed
23 4
Pesticide exposure cutoff: ≥50
th
percentile (0.001 lb/ac)
RUCA
Exposed 6 14
25.0% 39.1% 0.0129
*
Not
exposed
18 9
Pesticide exposure cutoff: ≥75
th
percentile (0.08 lb/ac)
RUCA
Exposed 4 16
33.3% 54.3% 0.4541 Not
exposed
8 19

1
Chi-square test or Fisher’s exact test
2
*p<0.05

Table 25 ZCTA-level accuracy of U.S. Census Bureau urban-rural classification:
organochlorines

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

U.S.
Census
Bureau
Exposed 12 6
31.6% 33.3% 0.0676 Not
exposed
26 3
Pesticide exposure cutoff: ≥50
th
percentile (0.001 lb/ac)
U.S.
Census
Bureau
Exposed 5 13
20.8% 43.5% 0.0119* Not
exposed
19 10
Pesticide exposure cutoff: ≥75
th
percentile (0.08 lb/ac)
U.S.
Census
Bureau
Exposed 1 17
8.3% 51.4% 0.0167* Not
exposed
11 18

1
Chi-square test or Fisher’s exact test
2
*p<0.05

105

Table 26 ZCTA-Level accuracy of RUCA codes: organophosphates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

RUCA
Exposed 19 1
42.2% 50.0% >0.99 Not
exposed
26 1
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
RUCA
Exposed 7 13
29.2% 43.5% 0.0579 Not
exposed
17 10
Pesticide exposure cutoff: ≥75
th
percentile (0.30 lb/ac)
RUCA
Exposed 4 16
33.3% 54.3% 0.4541 Not
exposed
8 19

1
Chi-square test or Fisher’s exact test

Table 27 ZCTA-level accuracy of U.S. Census Bureau urban-rural classification:
organophosphates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

U.S.
Census
Bureau
Exposed 16 2
35.6% 0% 0.1415 Not
exposed
29 0
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
U.S.
Census
Bureau
Exposed 5 13
20.8% 43.5% 0.0119* Not
exposed
19 10
Pesticide exposure cutoff: ≥75
th
percentile (0.30 lb/ac)
U.S.
Census
Bureau
Exposed 1 17
8.3% 51.4% 0.0167* Not
exposed
11 18

1
Chi-square test or Fisher’s exact test
2
*p<0.05

106

Table 28 ZCTA-level accuracy of RUCA codes: carbamates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

RUCA
Exposed 18 2
42.9% 60.0% >0.99 Not
exposed
24 3
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
RUCA
Exposed 6 14
26.1% 41.7% 0.0254* Not
exposed
17 10
Pesticide exposure cutoff: ≥75
th
percentile (0.06 lb/ac)
RUCA
Exposed 5 15
41.7% 57.1% 0.9426 Not
exposed
7 20

1
Chi-square test or Fisher’s exact test
2
*p<0.05

Table 29 ZCTA-level accuracy of U.S. Census Bureau urban-rural classification:
carbamates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

U.S.
Census
Bureau
Exposed 13 5
31.0% 0% 0.0056** Not
exposed
29 0
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
U.S.
Census
Bureau
Exposed 3 15
13.0% 37.5% 0.0005** Not
exposed
20 9
Pesticide exposure cutoff: ≥75
th
percentile (0.06 lb/ac)
U.S.
Census
Bureau
Exposed 1 17
8.3% 51.4% 0.0167* Not
exposed
11 18

1
Chi-square test or Fisher’s exact test
2
*p<0.05; **p<0.01

107

way 42.9% of all truly pesticide-exposed ZCTAs. The remaining 57.1 to 75% of ZCTAs
represent false negatives, or ZCTAs that were incorrectly classified as urban. Sensitivity
ranged between 8.3 and 35.6% for the U.S. Census Bureau metric.
Specificity ranged between 39.1 and 60% for RUCA codes. In other words,
between 39.1 and 60% of all truly unexposed ZCTAs were classified as urban. The
remaining 40 to 60.9% of ZCTAs represent false positives, or ZCTAs incorrectly
classified as rural. Specificity ranged between 0 and 51.4% for the U.S. Census Bureau
metric.
Significant differences were observed when comparing the GIS gold standard to
the U.S. Census Bureau metric - depending on chemical class - across all pesticide
exposure cutoffs. For example, a larger proportion of pesticide-exposed ZCTAs (50
th

percentile) were false negatives compared to true positives. In other words, a substantial
proportion of pesticide-exposed ZCTAs were misclassified as urban. Eighty-seven
percent of carbamate-exposed ZCTAs were false negatives (using 50
th
percentile), while
13% were true positives (p=0.0005; Table 29). Fewer significant differences were
observed when comparing RUCA codes to the GIS gold standard.
The accuracy of census tract-level rurality metrics also differed according to
rurality metric, pesticide chemical class, and pesticide exposure cutoff. Specificity was
consistently high, upwards of 77.1% for RUCA codes, and upwards of 97.5% for the U.S.
Census Bureau metric (Tables 30-35). In other words, RUCA codes classified at most
77.1% of census tracts not exposed to pesticides as urban, and the U.S. Census Bureau
metric classified at most 97.5% of census tracts not exposed to pesticides as urban.

108

Table 30 Census tract-level accuracy of RUCA codes: organochlorines

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

RUCA
Exposed 17 11
17.0% 72.5% 0.1606 Not
exposed
83 29
Pesticide exposure cutoff: ≥50
th
percentile (0.001 lb/ac)
RUCA
Exposed 10 18
14.3% 74.3% 0.0910 Not
exposed
60 52
Pesticide exposure cutoff: ≥75
th
percentile (0.03 lb/ac)
RUCA
Exposed 4 24
11.4% 77.1% 0.1432 Not
exposed
31 81

1
Chi-square test

Table 31 Census tract-level accuracy of U.S. Census Bureau urban-rural classification:
organochlorines

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

U.S.
Census
Bureau
Exposed 6 1
6.0% 97.5% 0.6730 Not
exposed
94 39
Pesticide exposure cutoff: ≥50
th
percentile (0.001 lb/ac)
U.S.
Census
Bureau
Exposed 2 5
2.9% 92.9% 0.4411 Not
exposed
68 65
Pesticide exposure cutoff: ≥75
th
percentile (0.03 lb/ac)
U.S.
Census
Bureau
Exposed 1 6
2.9% 94.3% 0.6801 Not
exposed
34 99

1
Fisher’s exact test

109

Table 32 Census tract-level accuracy of RUCA codes: organophosphates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1,2

RUCA
Exposed 24 4
19.4% 75.0% 0.5269 Not
exposed
100 12
Pesticide exposure cutoff: ≥50
th
percentile (0.02 lb/ac)
RUCA
Exposed 8 20
11.4% 71.4% 0.0112* Not
exposed
62 50
Pesticide exposure cutoff: ≥75
th
percentile (0.14 lb/ac)
RUCA
Exposed 2 26
5.7% 75.2% 0.0147* Not
exposed
33 79

1
Chi-square test or Fisher’s exact test
2
*p<0.05

Table 33 Census tract-level accuracy of U.S. Census Bureau urban-rural classification:
organophosphates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

U.S.
Census
Bureau
Exposed 6 1
4.8% 93.8% 0.5809 Not
exposed
118 15
Pesticide exposure cutoff: ≥50
th
percentile (0.02 lb/ac)
U.S.
Census
Bureau
Exposed 3 4
4.3% 94.3% >0.99 Not
exposed
67 66
Pesticide exposure cutoff: ≥75
th
percentile (0.14 lb/ac)
U.S.
Census
Bureau
Exposed 1 6
2.9% 94.3% 0.6801 Not
exposed
34 99

1
Fisher’s exact test

110

Table 34 Census tract-level accuracy of RUCA codes: carbamates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

RUCA
Exposed 21 7
18.9% 75.9% 0.5316 Not
exposed
90 22
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
RUCA
Exposed 8 20
11.4% 71.4% 0.0112* Not
exposed
62 50
Pesticide exposure cutoff: ≥75
th
percentile (0.05 lb/ac)
RUCA
Exposed 3 25
8.6% 76.2% 0.0510 Not
exposed
32 80

1
Chi-square test
2
*p<0.05

Table 35 Census tract-level accuracy of U.S. Census Bureau urban-rural classification:
carbamates

Pesticide exposure cutoff: >0 lb/ac
GIS metric (gold standard)
Exposed Not exposed Sensitivity Specificity p
1

U.S.
Census
Bureau
Exposed 5 2
4.5% 93.1% 0.6344 Not
exposed
106 27
Pesticide exposure cutoff: ≥50
th
percentile (0.01 lb/ac)
U.S.
Census
Bureau
Exposed 3 4
4.3% 94.3% >0.99 Not
exposed
67 66
Pesticide exposure cutoff: ≥75
th
percentile (0.05 lb/ac)
U.S.
Census
Bureau
Exposed 1 6
2.9% 94.3% 0.6801 Not
exposed
34 99

1
Chi-square test

111

Sensitivity was relatively low across both rurality metrics and all pesticide
chemical classes, ranging from 5.7 to 19.4% for RUCA codes and from 2.9 to 6% for the
U.S. Census Bureau metric. When examining RUCA codes, sensitivity decreased when
using more conservative pesticide exposure cutoffs, while specificity remained relatively
constant across all pesticide exposure cutoffs. When examining the U.S. Census Bureau
metric, sensitivity and specificity remained constant across pesticide exposure cutoffs.
Significant differences were observed when comparing RUCA codes to the GIS
gold standard. Among organophosphate and carbamate usage, a large proportion of
pesticide-exposed census tracts (using the 50
th
or 75
th
percentile) were misclassified as
urban. For example, 5.7% of organophosphate-exposed census tracts (75
th
percentile)
were classified as rural and 94.3% were misclassified as urban (p=0.0147; Table 32).
Similar results were observed for carbamates using a 50
th
percentile cutoff (p=0.0112;
Table 34).

112

CHAPTER FIVE: DISCUSSION AND CONCLUSIONS
Landsat satellite-borne imagery represents an invaluable resource (data available
beginning in 1972) that can be integrated into pesticide exposure methodologies (USGS
2013b). Given its moderate spatial and temporal resolution - the Landsat Thematic
Mapper sensors acquired images with at least 30 m spatial resolution every 16 days -
Landsat has the capacity to contribute information in determining likely crop field
locations of PUR pesticide applications occurring in the past. This research demonstrated
the feasibility of incorporating Landsat imagery into a GIS-based method to improve the
spatiotemporal resolution of pesticide exposure estimation, representing a large-scale
implementation of the Maxwell (2011) Landsat methods across Kern County (Maxwell et
al. 2010b; Maxwell 2011). The presented pesticide exposure approach modified an
existing individual-level, GIS-based three-tier method (Rull and Ritz 2003) that
incorporates pesticide, land use, and cadastral datasets to accommodate the ZIP code and
census tract analysis scales and to integrate Landsat remote sensing. This modified three-
tier approach can be adopted for analysis scales finer than aggregated areal units.
Furthermore, comparing the validity of two commonly used rurality metrics,
RUCA codes and the U.S. Census Bureau urban-rural classification system, provided
clarity with respect to which measure is a superior surrogate of pesticide exposure to use
in terms of accuracy. These rurality metrics may prove valuable in situations (i.e. places)
with limited pesticide information. Hence, the absence of pesticide information, lack of
technical knowledge, etc., may require the use of rurality-based metrics. The results
highlight wide variability in terms of sensitivity and specificity as a function of rurality
metric, areal aggregation, pesticide chemical class, and pesticide exposure cutoff.

113

5.1 Critical Assessment of Methods and Results: Strengths and Limitations
5.1.1 PUR Processing
The pesticide database of organochlorines, organophosphates, and carbamates included
157 pesticide active ingredients (AIs). The pesticide database may not have included all
possible pesticide AIs in each chemical class, potentially underestimating pounds of
pesticide used, and thus ZCTA- and census tract-level annual application rates.
Regarding PUR data processing, although applying logic checks is a conservative PUR
data cleaning approach that has not been documented in the literature, these methods
were used by the CDPR to evaluate PUR data quality (CDPR 2000a). Applying two of
the three 1990-onward PUR outlier definitions to PUR data between 1974 and 1989 has
also never been performed in previous research, but adds consistency to PUR handling.

5.1.2 Crop Signature Library (CSL)
The CSL was a major component of this research, spanning January to October 1990 and
including NDVI values for sampled land use survey polygons acquired via stratified
random sampling (SRS). The CSL formed the basis for the classification of 1985 Landsat
imagery to be incorporated into the modified three-tier pesticide exposure methodology.
Its major strength and improvement upon previous CSL-related research was harnessing
all available spectral information in the form of NDVI values from each SRS-sampled
land use survey polygon. This represented an objective alternative to the Maxwell (2010)
approach, which selects one pixel per polygon at the location of the polygon label -
except when the spectral tone of the pixel is not representative of the polygon.

114

A potential limitation relates to the mixed pixel problem, or when a pixel is not
completely occupied by a single homogeneous category (Campbell and Wynne 2011).
This is a common issue at the edges of large, discrete objects and linear features. A scene
divided into discrete pixel areas averages the brightness values over the entire pixel area.
As the geographic features under study (crop fields; mean area of SRS polygons
407,928.4 m
2
, median area 176,249.8 m
2
) are larger than the pixels of the Landsat images
(30 m
2
), the CSL represents an H-resolution model, characterized by spectral responses
of features mixed together so that composite signatures of the pixels do not match the
pure spectral signature(s) (Strahler et al. 1986). For example, as NDVI values for each
SRS-sampled land use survey polygon were extracted, NDVI values of pixels along the
edges of the polygons may represent NDVI values of multiple crop types by virtue of the
spatial resolution of the sensor. However, this issue may not have been impactful
regarding the CSL if neighboring land use survey polygons were relatively uniform with
respect to NDVI values, despite being associated with different crop types.
Obtaining high quality, cloud-free Landsat images was challenging. Cloud cover
affected the inclusion of entire Path-Row images for March and June 1990. Furthermore,
Landsat images were not available for November and December 1990, which would have
contributed to a more complete CSL and enhanced the discrimination of 1985 Landsat
images that were subsequently used to create an agricultural crop field layer. In addition,
including multiple images within the same month, rather than limiting the CSL to one
image per month, would have better captured intra-month NDVI variability and further
improved classification.

115

Another prominent issue was the shrinking geographic coverage from which a
CSL could be created. Affected by cloud cover and Path-Row data availability, a
modestly sized region of northwestern Kern County was used to execute SRS (Figure
19). Although SRS ensured that all land use classes were sampled irrespective of their
prevalence across Kern County, the potential for random error was introduced as 49 land
use classes included SRS samples less than the a priori specified stratum size of 30. SRS
samples within these 49 strata may not be representative of the land use class, and may
have been characterized by atypical NDVI values. Expansion of the study area beyond
Kern County would increase the population of each land use class strata from which
polygons could be sampled.
Cloud cover and shadows were addressed at three stages of CSL creation and
1985 Landsat image classification: (1) selection of images for inclusion in the analysis;
(2) masking cloud-affected areas before mosaicking Path-Row images together; and (3)
digitizing geographic areas without cloud cover before SRS or PCA. Note that if clouds
were present in an image but were not within a region overlapping Kern County, these
areas were not masked out but were eventually excluded after cropping each
radiometrically and atmospherically processed image to the Kern County extent. Some
Landsat images may have been salvageable, with geographic areas covering some
geographic portions of Kern County. However, cloud cover-related decisions were
ultimately guided by the amount of cloud-free areas intersecting Kern County and
anticipated mosaicking difficulties.
The major limiting factor affecting the CSL was the 1990 Kern County land use
survey used as the ground truth. The CSL is only as informative as its source data - the

116

land use survey guided the extraction of NDVI values for land use classes according to
land use survey-delineated crop field boundaries. Land use surveys focus on agricultural
land use (CDWR 2013). Therefore, other land use classes related to non-agricultural uses
may not be adequately captured in this dataset. This issue is manifest in the large number
of segments classified as jojoba - nine land use survey crop fields in 1990 vs. 1,572
segments were classified as jojoba with Landsat imagery. The external validity, or
generalizability, of the CSL is limited by the study area of Kern County and the eligibility
criteria applied prior to selecting SRS samples: single-use crop fields, at least 4 ha in
area, and within the designated cloud-free zone. Grain and hay crops were the most
frequently occurring multi-use or inter-cropped land use class; their exclusion from the
CSL may have limited the ability of the CSL to classify segments. Results may have been
impacted by sources of positional and attribute error in the data sources. For example,
data entry error may have misclassified land use survey crop field names. The positional
accuracy of the crop fields is also affected by methods from which boundaries were
delineated (e.g. GPS). Furthermore, the extent to which the CSL is consistently
representative of monthly NDVI spectral profiles for the included crop types during time
periods before and after 1990 should be explored.

5.1.3 Classification of 1985 Landsat Imagery
Landsat images from 1985 selected for CSL-based classification were also affected by
cloud cover for the months of January, March, and October. Particular Path-Row scenes
were altogether missing, limiting the geographic extent of classified crop fields. Ideally,
all Path-Row scenes would have been available for all months in 1985 paralleling the

117

1990 CSL (January to October). However, a decision was made to assign more weight to
the contribution of more time points with NDVI values with respect to classification
rather than capturing as much geographic extent as possible. In other words, only
geographic areas with NDVI image data availability for all months (January, March to
October; February was excluded due to absence of Path 42) were selected for subsequent
PCA and segmentation. An alternative would have been to include as much of Kern
County as possible by choosing only to consider months where all Path-Row scenes were
available (April-June, September), which would have produced different classified crop
fields. Although the alternative strategy would have resulted in a classified crop field
layer covering all of Kern County, it may not have been able to adequately discriminate
between land use classes, as only including four months of NDVI data may have
misclassified land use classes with similar spectral profiles during these months.
A key component of this research that should be addressed is the implementation
of a formal accuracy assessment of the classified Landsat imagery. In practice, classified
remote sensing products should be compared to a gold standard to derive an error matrix
(Campbell and Wynne 2011). Measures such as user’s and producer’s accuracy, as well
as kappa, can be calculated to quantify the extent to which the classified product is a
valid representation of the phenomenon it seeks to represent. Future research should
explore the validity of this CSL in terms of classifying crop fields in Kern County at
different points in time and also in different geographic areas around California.

118

5.1.4 Segmentation
The segmentation process is subjective in allowing the end user to specify particular
parameters (e.g. tolerance in IDRISI Selva) to achieve a segmented layer of objects
representing a satisfactory likeness of the geographic phenomenon under study - crop
field boundaries in this analysis. A further challenge related to the need to account for
differences in crop field boundaries over the course of 1985, which was addressed in the
PCA using all NDVI images from January and March to October to output principal
components. Various tolerance parameters ranging between 10 and 100 were used and
segmentation results were examined against a PCA composite and a color-infrared (CIR)
Landsat image from August 1985. This method is in contrast to Maxwell (2011), which
used Definiens eCognition software to specify different parameters (scale, shape, and
compactness) to derive boundaries.
The segmentation process should be further explored to optimally derive
segments truly representative of crop field boundaries. For example, Figure 24 shows
some segments that appear to cross multiple crop field boundaries (likely same crop type)
and also multiple segments are present within the same crop field. It should be noted that
as a part of the modified three-tier pesticide exposure method, all crop fields from each
dataset (land use survey and Landsat) were dissolved, separately, within each PLSS
section - the geographic level of reporting for PUR data. Therefore, although multiple,
adjacent segments of the same crop type did not impact pesticide exposure estimation, it
is still meaningful to produce resultant segments that truly represent real-world features.

119

5.1.5 Modified Three-Tier Approach: Pesticide Exposure
A new modified three-tier pesticide exposure methodology was developed, which
honored the existing Rull and Ritz (2003) three-tier method through utilizing a tiered
approach incorporating land use survey and PLSS section data. However, the modified
method introduced tiers derived from Landsat imagery classified according to agricultural
crop types, which allowed for determining the independent contribution of Landsat
imagery to tiered matching above and beyond land use survey and PLSS data (Tiers 2A
and 2C; Figures 27-29).
Implementing the modified three-tier approach demonstrated that most PUR
records were claimed by Tier 1 using land use survey crop fields. Results may differ if
examining pesticide chemical classes other than organochlorines, organophosphates, and
carbamates, time periods beyond 1990, and individual years. Furthermore, as crop types
are not perfectly comparable between the PUR and land use survey datasets (e.g. some
PUR crop types not found in land use survey), PUR-to-land use survey crosswalk crop
assignments may have affected tiered results.
Another methodological consideration is the way in which overlapping crop fields
from the land use survey and Landsat data were treated. Rates were calculated by
weighting the application rates (pounds of applied pesticides divided by acres of land use
survey- or Landsat-derived crop field or PLSS section within each section) by the
proportion of the ZCTA or census tract comprised of that particular crop field or section.
This weighted average approach takes into account the entire geographic area of each
areal unit, irrespective of pesticide treatment. The tiered approach was implemented in
such a way that treated Landsat-derived crop fields as boundaries independent of land use

120

survey crop fields without the use of a spatial union to sum applied pounds within
overlapping areas. In other words, land use survey and Landsat crop field boundaries
matched to PUR data may have intersected each other. This does not affect PLSS section
boundaries as the design of the modified three-tier method only matches PUR data to a
section when no land use survey or Landsat crop fields are present in a section. It may be
appropriate to treat land use survey and Landsat crop fields independently as a particular
crop field may exist at one point in time, and be replaced with some other crop field at a
later time.
However, these occurrences were infrequent as Landsat modestly contributed
tiered matches. Specifically, 3.42% of organochlorine-treated land use survey crop fields
were intersected by organochlorine-treated Landsat crop fields (20,022.62 ac; data not
shown), followed by 4.82% of carbamate-treated land use survey fields intersecting
Landsat fields (32,750.72 ac), and 6.21% of organophosphate-treated land use survey
fields intersecting Landsat fields (45,156.36 ac). Future research should explore the
optimal way in which to incorporate overlapping crop field boundaries representing
multiple time periods.
Another prominent issue was that of sliver polygons, resulting from the
intersection of PLSS sections and land use survey or Landsat crop fields. The geographic
resolution of PUR data (PLSS section level) necessitated the intersection of these data
layers to identify likely locations of treated crop fields when implementing tiered
matching. Specifically, if multiple fields of a particular crop type exist within a section,
PUR data does not discriminate between which crop field was treated (Goldberg et al.
2007). As crop fields may span multiple sections, sliver polygons were produced as a

121

result of their intersection. One approach to handling sliver polygons is to retain the
acreage of the sliver polygon with its source crop field in calculating application rates.
Another approach, adopted in this research, was to exclude sliver polygons using shape
length and shape area attributes. However, the criteria for exclusion likely excluded
smaller crop fields.
The contribution of Landsat was modest across all pesticide chemical classes at
Tiers 2A (2.1-2.4%) and 2C (0.1-0.2%), which supports the notion that integrating
Landsat remote sensing improved, to a small degree, pesticide exposure assessment
through addressing PUR applications that did not match land use survey crop fields. If
Landsat was implemented as a tier prior to considering land use survey crop fields, its
contribution may have been more pronounced. However, as the Landsat layer was
derived from a CSL that used the 1990 Kern County land use survey as a ground truth, it
was more appropriate to use the land use survey as Tier 1. In addition, if implemented in
a different California county, the contribution of Landsat may have differed due to the
prevalence of different crop types. For example, a California county with a higher
prevalence of rice crop fields (N=1 rice crop field in 1990 Kern County land use survey)
may have observed larger tier contributions from Landsat data. Nevertheless, the
feasibility component of this study was to modify the existing Rull and Ritz (2003) three-
tier approach to evaluate if Landsat could contribute additional crop field location
information beyond land use survey data. Integrating Landsat, by way of creating a CSL
and classification, was demonstrated to be a feasible analytic addition to the pesticide
exposure ascertainment process.

122

5.1.6 Impact of Areal Aggregation on Annual Pesticide Application Rates
Annual pesticide application rates varied according to areal aggregation and pesticide
chemical class. Rates were not significantly different at the ZCTA vs. census tract level,
although maximum rates were typically higher at the census tract level. For example,
organochlorine-specific annual application rates ranged from 0 to 0.26 lb/ac (median
0.001 lb/ac) at the ZCTA level and from 0 to 1.57 lb/ac (median 0.001 lb/ac) at the
census tract level. Differences in application rates according to areal aggregation are a
manifestation of the modifiable areal unit problem (MAUP). The delineation of ZCTA
and census tract boundaries does not necessarily reflect agricultural crop boundaries, let
alone pesticide-treated crop field boundaries and pesticide application practices.
Another important consideration is the geoprocessing of ZCTA boundaries prior
to calculating application rates. The calculated ZCTA-level rates reflect pesticide
exposure specifically associated with residence within Kern County - by virtue of
extracting Kern County PUR records. For example, the full extent of ZCTAs spanning
multiple counties, such as 93527, was not considered. ZCTA boundaries were clipped to
the Kern County extent for use in weighting application rates. Calculated ZCTA rates
represent an ecologic measure of pesticide exposure for individuals residing in both Kern
County and a particular ZCTA.

5.1.7 Accuracy Assessment of Rurality
The performance of each rurality metric as a surrogate for pesticide exposure was largely
a function of the rurality metric, areal aggregation, pesticide chemical class, and pesticide
exposure cutoff. It was hypothesized that RUCA codes, by virtue of incorporating both

123

population and work commuting information, would be both more sensitive and specific
in assigning pesticide exposure compared to the GIS gold standard metric. Rurality is an
intuitive surrogate measure of pesticide exposure as agricultural pesticide applications
occur more frequently in rural areas (Franklin and Worgan 2005).
However, rural ZCTAs and census tracts were typically characterized by lower
median annual pesticide application rates compared to their urban counterparts across all
chemical classes. These patterns run counter to what was expected, i.e. higher pesticide
exposure in rural geographic areas. These results potentially shed light on the distinction
between rural and urban areas being unrelated to pesticide application practices in Kern
County, which is predominantly rural. It is conceivable that urban areas, as demarcated
by RUCA codes and the U.S. Census Bureau metric in Kern County, are actually more
rural as compared to urban areas in other counties outside of the Central Valley - by
virtue of selecting a predominantly rural study area.
Furthermore, a rurality metric that is not binary and has multiple categories
corresponding to different levels of rurality may be more appropriate in trying to capture
pesticide exposure. The RHRC presents additional methods to categorize RUCA codes
(RHRC 2000). For example, a four-category classification discriminates between areal
units that are urban, large rural, small rural, and isolated. In addition, had other pesticide
chemical classes been examined, the expected pattern of higher rates in rural geographic
areas may have been observed.
Another plausible interpretation stems from the imperfect implementation of the
U.S. Census Bureau metric with respect to ZCTAs and census tracts, i.e. handling the
large swaths of ZCTAs and census tracts not intersecting Urbanized Areas and Urban

124

Clusters (Appendix F; Figure F1). Even if a small proportion of a ZCTA or census tract
intersected with an Urbanized Area or Urban Cluster, it was classified as urban.
Therefore, as fewer ZCTAs and census tracts were classified as rural, there was less
opportunity for pesticide-treated crop fields to intersect rural areal units. An area cutoff
could have been applied, where an Urbanized Area or Urban Cluster must have
intersected a particular proportion of the ZCTA or census tract for it to be classified as
urban. However, this approach would have ignored the portion of the areal unit that truly
was urban, even if the proportion of the overall areal unit intersecting the Urbanized Area
or Urban Cluster was small. The U.S. Census Bureau metric may be more useful when
using, for example, individual-level residential locations, rather than areal aggregations.
Results regarding the sensitivity and specificity of the U.S. Census Bureau metric should
be interpreted with caution.
The extent to which the rurality metrics differed in how Kern County was
classified as rural vs. urban was striking. Geographic patterns of rurality were seemingly
similar at the ZCTA level - rural ZCTAs using RUCA codes and the U.S. Census Bureau
metric were observed in the eastern and western portions of Kern County. However,
agreement was poor between the two metrics (kappa=0.03). At the census tract level,
rurality was more widespread when using RUCA codes - in the eastern, western, and
central portions of the county. Rurality, when measured using the U.S. Census Bureau
metric, was observed in the northeastern and eastern portions of Kern County. Agreement
was also poor between the two metrics (kappa=0.04).
In terms of the standard by which to judge satisfactory vs. unsatisfactory
sensitivity and specificity, absolute differences from 100% (perfect sensitivity and

125

specificity) and relative differences across areal aggregations, pesticide chemical classes,
and pesticide exposure cutoffs were considered. Across both areal aggregations,
specificity was superior to sensitivity. This reflects the satisfactory capacity of rurality,
whether measured using RUCA codes or the U.S. Census Bureau metric, to correctly
identify geographic units truly not exposed to pesticides. At the ZCTA level, RUCA
codes were superior to the U.S. Census Bureau metric. The highest specificity for RUCA
codes was observed for carbamates (60%) using a cutoff of 0 lb/ac as pesticide-exposed,
while the highest specificity for the U.S. Census Bureau metric (51.4%) was observed for
all chemical classes using a 75
th
percentile cutoff. RUCA codes were also superior to the
U.S. Census Bureau metric in terms of sensitivity - highest observed for carbamates
(42.9%) using a 0 lb/ac cutoff. The highest sensitivity offered by the U.S. Census Bureau
metric was observed for organophosphates (35.6%) using a 0 lb/ac cutoff. A larger
number of statistically significant differences when comparing the U.S. Census Bureau
metric to the GIS gold standard is also indicative of its mediocre performance at the
ZCTA level.
At the census tract level, specificity was also consistently higher than sensitivity
across all pesticide chemical classes and pesticide exposure cutoffs. Sensitivity and
specificity remained relatively constant across all pesticide exposure cutoffs. In contrast
to the ZCTA level, the U.S. Census Bureau metric offered superior specificity compared
to RUCA codes. The highest specificity using the U.S. Census Bureau metric was
observed for organochlorines (97.5%) using a 0 lb/ac cutoff, compared to 77.1% for
organochlorines using a ≥75
th
percentile cutoff when with RUCA codes. Across all
pesticide chemical classes, sensitivity was poor (≤19.4%). RUCA codes were more

126

sensitive than the U.S. Census Bureau metric- the highest sensitivity was observed for
organophosphates (19.4%) using a 0 lb/ac cutoff. The highest sensitivity offered by the
U.S. Census Bureau metric was observed for organochlorines (6%) using a 0 lb/ac cutoff.
The prevalence of different pesticide chemical classes used on crops across Kern
County impacted the results of the rurality accuracy assessment. RUCA codes were most
sensitive and specific for carbamates usage (0 lb/ac cutoff) at the ZCTA level.
Interestingly, pesticide-treated crop fields and sections were quite prevalent throughout
Kern County, irrespective of chemical class. The largest concentration of treated fields
was towards the northwest, reflecting pervasive agricultural practices in this region of the
Central Valley. Therefore, although organophosphate usage accounted for the majority of
PUR applications and pounds applied in Kern County, carbamate-treated crop fields
happened to intersect rural ZCTAs more frequently, and non-treated fields urban ZCTAs,
when using the RUCA metric.
On the other hand, when evaluating census tracts, the U.S. Census Bureau metric
was most specific for organochlorine usage (0 lb/ac cutoff) and RUCA codes were most
sensitive for organophosphate usage (0 lb/ac cutoff). Organochlorines were associated
with the fewest median number of pesticide-treated fields intersecting census tracts, both
rural and urban, compared to organophosphates and carbamates (Appendix F; Table F2).
This was paired with the fact that the U.S. Census Bureau metric classifies the majority
of Kern County as urban, which served to increase its specificity with respect to
organochlorine usage. The relatively higher prevalence of organophosphates worked to
increase the capacity of RUCA codes to accurately classify census tracts as pesticide-
exposed/rural, ultimately increasing sensitivity.

127

The results of the accuracy assessment were also affected by the pesticide
chemical classes included the analysis. Different pesticide active ingredients were likely
used more frequently for particular crop types, and evaluating only organochlorines,
organophosphates, and carbamates may not reflect this pesticide usage in Kern County.
For example, excluding other pesticide chemical classes may have misclassified areal
units truly pesticide-exposed as not exposed to pesticides (using gold standard), which
would have decreased the specificity of both rurality metrics (increased false positive
rate). In other words, an areal unit categorized as rural would have been designated as not
exposed to pesticides using the current gold standard GIS metric, but would have been
designated as pesticide-exposed had other chemical classes been considered.
In a real-world scenario implementing a comparative epidemiologic study using
rurality as a surrogate measure of pesticide exposure (assuming low prevalence of
pesticide exposure in study population), usage of ZCTA-level RUCA codes (using 0 lb/ac
cutoff) would result in less attenuation (i.e. bias towards null hypothesis) in study results
due to its superior specificity across all pesticide chemical classes. On the other hand,
although usage of census tract-level RUCA codes is associated with relatively high
specificity, census tract-level U.S. Bureau Census urban-rural designations were
associated with even higher specificity, and would thus result in less attenuation (using 0
lb/ac cutoff). It is important to note that the gain with respect to specificity must be
balanced against the impact of sensitivity. Sensitivity was found to be mediocre at the
ZCTA level and poor at the census tract level, meaning truly pesticide-exposed areal
units were misclassified as not exposed (urban) according to each rurality metric (false
negatives). These results are limited in generalizability as the study area was Kern

128

County, which is predominantly rural; rurality may perform differently as a surrogate
measure in different geographic areas across California. In regions with higher pesticide
exposure (>10%), sensitivity has a greater impact on study results (Szklo and Nieto
2007).
The methods by which ZCTAs and census tracts are delineated also affected the
results of the accuracy assessment. Census tracts, characterized by a larger number of
areal units constituting Kern County (N=140) compared to ZCTAs (N=47), may be more
homogeneous with respect to urban/rural characteristics. This homogeneity may be more
relevant to delineating urban vs. rural areas, reflected in the higher specificity (ability to
correctly identify areal units not pesticide-exposed) across all census tract-level rurality
metrics when compared to the GIS gold standard. The highest census tract-level
specificity was 97.5% compared to 60% at the ZCTA level. This alludes to the notion
that census tract aggregations may be delineated in such a way that better determines
urban processes lacking agricultural pesticide applications. Furthermore, the overall poor
sensitivity (ability to correctly identify areal units exposed to pesticides) observed at the
census tract level may also reflect how census tracts are designed in such a way that is
not conducive to correctly identifying geographic units truly exposed to pesticides.
Choice of pesticide exposure cutoff (>0 lb/ac, ≥50
th
percentile, and ≥75
th

percentile) should be guided by knowledge of pesticide exposures meaningful to the
application at hand. Usage of a 0 lb/ac cutoff is the most liberal, expected to result in the
highest sensitivity, or capacity to capture all pesticide-exposed areal units. On the other
hand, usage of the most conservative cutoff in this study, for example the 75
th
percentile,
would be expected to result in higher specificity, or the capacity to identify all non-

129

exposed areal units. Results may also differ depending on which gold standard pesticide
exposure metric is used - for example, a gold standard without Landsat imagery
integration. Although PUR-derived, GIS-based pesticide exposure metrics do not directly
address all possible routes of pesticide exposure, such as from occupation and diet, they
do reflect residential proximity to agricultural pesticides, which has been demonstrated to
be significantly associated with within-household pesticide levels (Gunier et al. 2011).

5.2 Feasibility and Informational Gain of Landsat Remote Sensing
The feasibility and utility of integrating Landsat remotely sensed imagery into an existing
three-tier pesticide exposure methodology is manifest in the creation of a CSL,
classification of additional Landsat imagery, and PUR records matching to Landsat-
derived tiers. However, given the constraints of computing power and time, the benefit of
incorporating Landsat imagery may be greater at finer spatial resolutions - finer than the
ZCTA and census tract levels. By design, annual pesticide application rates were derived
by weighting rates according to the proportion of the areal unit comprised of the field or
section. Therefore, integrating Landsat imagery may only result in incremental
improvements in enhancing pesticide exposure ascertainment when examining ZCTA-
and census tract-level rates. Although the weighted average approach to calculating rates
represents an average exposure for all individuals residing in a ZCTA or census tract,
evaluating such large areal units masks the heterogeneity that exists in pesticide
application rates at larger scales. When estimating pesticide exposure at aggregated
analysis scales, a PLSS section-only method (Bell et al. 2001; Gunier et al. 2001) or a
land use survey and PLSS section method (Rull and Ritz 2003), may produce similar

130

pesticide application rates. To effectively evaluate if Landsat imagery meaningfully
improves the spatiotemporal resolution of pesticide exposure estimation, the performance
of the modified three-tier method compared to existing methodologies should be assessed
at various spatial resolutions (e.g. individual-level residences vs. aggregated analysis
scales).
Another methodological approach would have been to classify additional Landsat
images from multiple years between 1974 and 1990, not only 1985, which could have
improved the informational gain from Landsat data. The modified three-tier method
could have incorporated the date of PUR applications in the tiered matching. The
availability of Landsat imagery beginning in 1972 allows for this approach. However,
earlier Landsat sensors, such as the Multispectral Scanner (MSS), will differ in such
characteristics as spatial resolution. This would have to be reconciled against using the
1990 CSL to classify Landsat images, which is constrained to 1990 due to land use
survey ground truth data availability. Future research should explore the extent to which
classification is affected by using a CSL that differs in spatial resolution to the classified
images.

5.3 Alternative Approaches to Integrating Landsat in Pesticide Exposure Estimation
Although the proposed modified three-tier method yielded a modest amount of PUR tier
matches using Landsat-derived crop fields, an approach that matches pesticide data to
Landsat-derived crop fields (not considering land use surveys) would better address the
potential utility of Landsat remote sensing in pesticide exposure estimation. Landsat
imagery provides a valuable opportunity to address temporal voids in land use surveys

131

through its moderate temporal resolution (16 to 18 days) in capturing Earth imagery since
1972. Although many agricultural crop types are likely long-standing between years,
some agricultural fields are rotated on an annual basis. Crop rotation is implemented to
maintain soil fertility via alternation of plant species. For example, a crop rotation
schedule may require planting an agricultural parcel of land with a different crop type
each year. Sloping land may be subject to excessive soil loss if row crops, such as corn,
are grown for many consecutive years. Rotation of corn with sod-based forage crops (e.g.
grass) minimizes soil loss (Lerner and Lerner 2008). Therefore, Landsat imagery, when
classified accurately, can provide a temporally accurate snapshot of agricultural lands
with 30 m spatial resolution in delineating crop field boundaries, not otherwise provided
with the intermittently updated land use surveys. A formal comparison between the Rull
and Ritz (2003) method and a three-tier approach using Landsat-derived crop fields
should be implemented to determine the extent to which Landsat imagery can provide a
more temporally accurate agricultural landscape at a particular point in time lacking land
use surveys – ultimately addressing Landsat’s utility in pesticide exposure estimation (i.e.
tier 1 PUR matches).
The proposed classification method using a sum of squared differences measure
harnesses the temporal variability of NDVI to identify crop types. Typical minimum
distance measures classify imagery according to the minimum distance between a pixel
value and the mean value of an informational class (Campbell and Wynne 2011). The
sum of squared differences measure, as applied in this research, determined the minimum
distance (squared) between each segment’s median NDVI value and the median NDVI
value of each land use class – summed across all months in 1985 with available imagery.

132

However, other classification methods could be used that can also take into account the
temporal variability of NDVI that may also yield accurate results. For example, Wardlow
and Egbert (2008) implemented an unsupervised classification approach [Iterative Self-
Organizing Data Analysis Technique (ISODATA)] to classify a time series of Moderate
Resolution Imaging Spectroradiometer (MODIS) NDVI images of Kansas between
March 22 and November 1. Spectral-temporal clusters were generated and assigned to the
crop, non-crop, or confused classes via comparing the distribution of the cluster’s pixels,
cluster means, and visual interpretation of the land cover types using Landsat Enhanced
Thematic Mapper Plus (ETM+) imagery. Future research should explore the optimal and
most accurate way to classify a temporal series of NDVI images into agricultural crop
fields.

5.4 Significance of Results
Demonstrating the feasibility of using remotely sensed data in a GIS-based pesticide
exposure metric at cancer data analysis scales to enhance the spatiotemporal resolution of
identifying pesticide-applied locations is needed. Quantifying the exact extent of
exposure misclassification from using two surrogate measures of rurality compared to a
GIS-based pesticide exposure gold standard, as well as determining which measure is
superior in terms of accuracy, have never before been performed. The results of this
research are specifically relevant to cancer epidemiology - the units of analysis included
census tracts and ZIP codes (ZCTAs used to approximate boundaries), which are typical
geographic aggregations for cancer registry data used to preserve patient confidentiality
(Boscoe et al. 2004; Waller and Gotway 2004). The results shed light on potential

133

exposure misclassification when using rurality-based metrics, and are applicable to
ecologic studies utilizing pesticide data aggregated to areal units and individual-level
studies using contextual, ecologic metrics.
The results are generalizable to epidemiologic literature examining pesticide
exposure in California and other states with similar data. In the absence of data on
pesticide applications and land use, understanding which rurality metric most
meaningfully captures the processes underlying pesticide exposure - both in terms of the
rurality definition and analysis scale - is important to explore. The impact of pesticides on
elevating the risk for certain cancers has been established (Alavanja et al. 2004), and
research into how to accurately measure pesticide exposure is integral to implementing
epidemiologic studies addressing this research topic. Ultimately, this research harnessed
GIS tools in order to directly address how the validity of surrogate measures of exposure
can directly impact the inferences derived from epidemiologic studies investigating
human health outcomes.

5.5 Future Directions
Future research should explore the utility of integrating Landsat into GIS-based pesticide
exposure metrics beyond Kern County and at finer spatial scales [e.g. within the 500 m
residential buffers implemented by Rull and Ritz (2003)]. A formal comparison between
the Rull and Ritz (2003) three-tier method and a Landsat-only pesticide exposure method
would be better able to highlight the contribution of Landsat imagery to locating
agricultural pesticide applications and ultimately to pesticide exposure estimation. A
validity study demonstrating the accuracy of the CSL in discriminating between land use

134

classes when examining imagery from a different geographic area and at a different time
point (e.g. creation of an error matrix comparing Landsat-classified crop fields to a
ground truth) would highlight the accuracy and generalizability of the CSL. Alternative
methods of classification to produce agricultural crop fields using temporal NDVI data
should be explored. Investigating the contribution of Landsat in PUR-matching at
different points in time would shed light on how Landsat could enhance the temporal
resolution of identifying pesticide-treated crop fields. The performance of rurality
compared to different GIS-based gold standard pesticide exposure metrics would also be
informative. Measuring pesticide exposure as a cumulative measure (e.g. total pounds of
applied pesticide) within an areal aggregation as opposed to pesticide exposure density
(rate in lb/ac) presented in this research would be valuable to explore as well.

5.6 Summary
The feasibility of incorporating Landsat remotely sensed imagery into a modified GIS-
based pesticide exposure metric accommodating cancer data analysis scales was
demonstrated. Strengths included the methodological improvement over previous
research via objectively harnessing all NDVI spectral information in creating a crop
signature library for use in subsequent classification of Landsat imagery, and developing
a modified three-tier pesticide exposure method that can be used at other analysis scales.
The accuracy of commonly used rurality metrics as pesticide exposure surrogates was
assessed, which has never before been researched. RUCA codes offer superior specificity
at the ZCTA level while the U.S. Census Bureau urban-rural classification metric offers
superior specificity at the census tract level. Accuracy varies according to rurality metric,

135

areal aggregation, pesticide chemical class, and pesticide exposure cutoffs, which should
be tailored to specific research applications. Future research should explore the
integration of Landsat imagery at finer spatial resolution pesticide exposure
methodologies (i.e. individual-level), examine the contribution of a Landsat-only method
to estimate pesticide exposure compared to the existing Rull and Ritz (2003) three-tier
method, validate the NDVI crop signature library, and evaluate the utility of using
different GIS-based pesticide exposure metrics (e.g. cumulative pounds).

136

REFERENCES

AgroPages. 2013. Crop Protection Database 2013 [cited 2013]. Available from
http://www.agropages.com/AgroData/.
Alavanja, M. C., J. A. Hoppin, and F. Kamel. 2004. Health effects of chronic pesticide
exposure: cancer and neurotoxicity. Annu Rev Public Health 25:155-97.
Alavanja, M. C., M. H. Ward, and P. Reynolds. 2007. Carcinogenicity of agricultural
pesticides in adults and children. J Agromedicine 12 (1):39-56.
Beck, P. S., C. Atzberger, K. A. Høgda, B. Johansen, and A. K. Skidmore. 2006.
Improved monitoring of vegetation dynamics at very high latitudes: A new
method using MODIS NDVI. Remote sensing of environment 100 (3):321-334.
Bell, E. M., I. Hertz-Picciotto, and J. J. Beaumont. 2001. A case-control study of
pesticides and fetal death due to congenital anomalies. Epidemiology 12 (2):148-
56.
Blair, A., Zahm, S.H., Cantor, K.P., Stewart, P.A. 1988. Estimating Exposure to
Pesticides in Epidemiological Studies of Cancer. In Biological Monitoring for
Pesticide Exposure: Measurement, Estimation, and Risk Reduction, 38-46:
American Chemical Society.
Boscoe, F. P., M. H. Ward, and P. Reynolds. 2004. Current practices in spatial analysis of
cancer data: data characteristics and data sources for geographic studies of cancer.
Int J Health Geogr 3 (1):28.
CA.gov. State of California: Facts 2013 [cited 2013. Available from
http://www.ca.gov/about/facts.html.
Cal-Atlas Geospatial Clearinghouse. 2013 [cited 2013. Available from
http://www.atlas.ca.gov/download.html.
Campbell, J. B., and R. H. Wynne. 2011. Introduction to Remote Sensing. Fifth ed. New
York, NY: The Guilford Press.
CDC (Centers for Disease Control and Prevention). Fourth National Report on Human
Exposure to Environmental Chemicals 2009 [cited 2013. Available from
http://www.cdc.gov/exposurereport/pdf/FourthReport.pdf.
———. Cancer 2012 [cited 2013. Available from
http://www.cdc.gov/chronicdisease/resources/publications/AAG/dcpc.htm.
CDOC (California Department of Conservation). FMMP - Mapping Procedures 2013
[cited 2013. Available from
http://www.conservation.ca.gov/dlrp/fmmp/mccu/Pages/making_map.aspx.

137

CDPR (California Department of Pesticide Regulation). Appendix C: California’s
Pesticide Use Report An Assessment of Spatial Data Quality 2000a [cited 2013.
Available from http://www.cdpr.ca.gov/docs/pur/appendix_c_dataq_ldr.pdf.
———. DPR Pesticide Use Reporting: An Overview of California's Unique Full
Reporting System 2000b [cited 2013. Available from
http://www.cdpr.ca.gov/docs/pur/purovrvw/ovr52000.pdf.
———. 2002. Pesticide Use Report Data User Guide & Documentation.
———. Pesticide Use Reporting 2011a [cited 2013. Available from
http://www.cdpr.ca.gov/docs/pur/pur11rep/lbsby_co_11.pdf.
———. Top 5 Sites 2011b [cited 2013. Available from
http://www.cdpr.ca.gov/docs/pur/pur11rep/top_5_sites_ais_lbs11.pdf.
———. Pesticide Use Reporting (PUR) 2013 [cited 2013. Available from
http://www.cdpr.ca.gov/docs/pur/purmain.htm.
CDWR (California Department of Water Resources). Land Use Survey Overview 2013
[cited 2013. Available from
http://www.water.ca.gov/landwateruse/lusrvymain.cfm.
Chander, G., B. L. Markham, and D. L. Helder. 2009. Summary of current radiometric
calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors.
Remote sensing of environment 113 (5):893-903.
Chavez, P. S. 1996. Image-based atmospheric corrections-revisited and improved.
Photogrammetric engineering and remote sensing 62 (9):1025-1035.
Clark Labs. 2013. IDRISI Selva. Worcester, Massachusetts, United States.
Cockburn, M., P. Mills, X. Zhang, J. Zadnick, D. Goldberg, and B. Ritz. 2011. Prostate
cancer and ambient pesticide exposure in agriculturally intensive areas in
California. Am J Epidemiol 173 (11):1280-8.
Cordier, S., T. B. Le, P. Verger, D. Bard, C. D. Le, B. Larouze, M. C. Dazza, T. Q.
Hoang, and L. Abenhaim. 1993. Viral infections and chemical exposures as risk
factors for hepatocellular carcinoma in Vietnam. Int J Cancer 55 (2):196-201.
Costello, S., M. Cockburn, J. Bronstein, X. Zhang, and B. Ritz. 2009. Parkinson's disease
and residential exposure to maneb and paraquat from agricultural applications in
the central valley of California. Am J Epidemiol 169 (8):919-26.
Dich, J., S. H. Zahm, A. Hanberg, and H. O. Adami. 1997. Pesticides and cancer. Cancer
Causes Control 8 (3):420-43.

138

Economic Research Service. Overview 2012 [cited 2013. Available from
http://www.ers.usda.gov/data-products/agricultural-productivity-in-the-us.aspx.
EPA (Environmental Protection Agency). Pesticide issues in the works: pesticide
volatilization 2009 [cited 2013. Available from
http://www.epa.gov/pesticides/about/intheworks/volatilization.htm.
———. Pesticides Industry Sales and Usage: 2006 and 2007 Market Estimates 2011
[cited 2013. Available from
http://www.epa.gov/opp00001/pestsales/07pestsales/market_estimates2007.pdf.
———. Pesticides 2012 [cited 2013. Available from http://www.epa.gov/pesticides/.
Esri. 2013. ArcGIS 10.1. Redlands, California, United States.
Ezzat, S., M. Abdel-Hamid, S. A. Eissa, N. Mokhtar, N. A. Labib, L. El-Ghorory, N. N.
Mikhail, A. Abdel-Hamid, T. Hifnawy, G. T. Strickland, and C. A. Loffredo.
2005. Associations of pesticides, HCV, HBV, and hepatocellular carcinoma in
Egypt. Int J Hyg Environ Health 208 (5):329-39.
Franklin, C., and J. Worgan. 2005. Occupational and Residential Exposure Assessment
for Pesticides. Hoboken, NJ: Wiley.
Gatto, N. M., M. Cockburn, J. Bronstein, A. D. Manthripragada, and B. Ritz. 2009. Well-
water consumption and Parkinson's disease in rural California. Environ Health
Perspect 117 (12):1912-8.
Goldberg, D. W., X. Zhang, J. C. Marusek, J. P. Wilson, B. Ritz, and M. G. Cockburn.
2007. Development of an automated pesticide exposure analyst for California’s
central valley. Paper read at Proc Urban Regional Info Syst Assoc GIS Public
Health Conf, New Orleans.
Greene, S. A., and R. P. Pohanish. 2005. Sittig's handbook of pesticides and agricultural
chemicals. Norwich, N.Y.: William Andrew Pub.
Grubesic, T. H., and T. C. Matisziw. 2006. On the use of ZIP codes and ZIP code
tabulation areas (ZCTAs) for the spatial analysis of epidemiological data. Int J
Health Geogr 5:58.
Gunier, R. B., M. E. Harnly, P. Reynolds, A. Hertz, and J. Von Behren. 2001.
Agricultural pesticide use in California: pesticide prioritization, use densities, and
population distributions for a childhood cancer study. Environ Health Perspect
109 (10):1071-8.
Gunier, R. B., M. H. Ward, M. Airola, E. M. Bell, J. Colt, M. Nishioka, P. A. Buffler, P.
Reynolds, R. P. Rull, A. Hertz, C. Metayer, and J. R. Nuckols. 2011.
Determinants of agricultural pesticide concentrations in carpet dust. Environ
Health Perspect 119 (7):970-6.

139

Jacquez, G. M. 2004. Current practices in the spatial analysis of cancer: flies in the
ointment. Int J Health Geogr 3 (1):22.
Lee, P. C., Y. Bordelon, J. Bronstein, and B. Ritz. 2012. Traumatic brain injury, paraquat
exposure, and their relationship to Parkinson disease. Neurology 79 (20):2061-6.
Lee, P. C., S. L. Rhodes, J. S. Sinsheimer, J. Bronstein, and B. Ritz. 2013. Functional
paraoxonase 1 variants modify the risk of Parkinson's disease due to
organophosphate exposure. Environ Int 56:42-7.
Lerner, K. L., and B. W. Lerner. 2008. Crop Rotation. In The Gale Encyclopedia of
Science, 1187-1190. Detroit.
Lippitt, C. D., L. L. Coulter, M. Freeman, J. Lamantia-Bishop, W. Pang, and D. A. Stow.
2012. The effect of input data transformations on object-based image analysis.
Remote Sensing Letters 3 (1):21-29.
Manthripragada, A. D., S. Costello, M. G. Cockburn, J. M. Bronstein, and B. Ritz. 2010.
Paraoxonase 1, agricultural organophosphate exposure, and Parkinson disease.
Epidemiology 21 (1):87-94.
Marusek, J. C., M. G. Cockburn, P. K. Mills, and B. R. Ritz. 2006. Control selection and
pesticide exposure assessment via GIS in prostate cancer studies. Am J Prev Med
30 (2 Suppl):S109-16.
Mather, P. M., and M. Koch. 2011. Chapter 7: Filtering Techniques. In Computer
Processing of Remotely-Sensed Images: An Introduction, 4th Edition
Oxford, UK: Wiley Blackwell.
Maxwell, S., J. Meliker, and P. Goovaerts. 2010a. Use of land surface remotely sensed
satellite and airborne data for environmental exposure assessment in cancer
research. J Expo Sci Environ Epidemiol 20 (2):176-85.
Maxwell, S. K. 2010. Generating land cover boundaries from remotely sensed data using
object-based image analysis: overview and epidemiological application. Spat
Spatiotemporal Epidemiol 1 (4):231-7.
———. 2011. Downscaling Pesticide Use Data to the Crop Field Level in California
Using Landsat Satellite Imagery: Paraquat Case Study. Remote Sensing 3
(9):1805-1816.
Maxwell, S. K., M. Airola, and J. R. Nuckols. 2010b. Using Landsat satellite data to
support pesticide exposure assessment in California. Int J Health Geogr 9:46.
Ministry of Agriculture. 2013. Pesticide Application Equipment. British Columbia 2013
[cited April 3 2013]. Available from
http://www.agf.gov.bc.ca/pesticides/f_2.htm#3.

140

Montello, D. R. 2001. Scale in Geography. In International Encyclopedia of the Social &
Behavioral Sciences, ed. N. J. Smelser, Baltes, P.B., 13501-13504: Pergamon
Press.
National Atlas. The Public Land Survey System (PLSS) 2013 [cited 2013. Available from
http://www.nationalatlas.gov/articles/boundaries/a_plss.html.
NCI (National Cancer Institute). SEER Stat Fact Sheets: All Sites 2012 [cited 2013.
Available from http://seer.cancer.gov/statfacts/html/all.html#incidence-mortality.
NPR (National Public Radio). California's Central Valley 2002 [cited 2013. Available
from http://www.npr.org/programs/atc/features/2002/nov/central_valley/.
Nuckols, J. R., R. B. Gunier, P. Riggs, R. Miller, P. Reynolds, and M. H. Ward. 2007.
Linkage of the California Pesticide Use Reporting Database with spatial land use
data for exposure assessment. Environ Health Perspect 115 (5):684-9.
Nuckols, J. R., M. H. Ward, and L. Jarup. 2004. Using geographic information systems
for exposure assessment in environmental epidemiology studies. Environ Health
Perspect 112 (9):1007-15.
O'Sullivan, D. O., and D. J. Unwin. 2010. Geographic Information Analysis. 2nd ed:
John Wiley & Sons.
Oates, L., and M. Cohen. 2011. Assessing diet as a modifiable risk factor for pesticide
exposure. Int J Environ Res Public Health 8 (6):1792-804.
Persson, E. C., B. I. Graubard, A. A. Evans, W. T. London, J. P. Weber, A. Leblanc, G.
Chen, W. Lin, and K. A. McGlynn. 2012. Dichlorodiphenyltrichloroethane and
risk of hepatocellular carcinoma. Int J Cancer.
Pickle, L. W., L. A. Waller, and A. B. Lawson. 2005. Current practices in cancer spatial
data analysis: a call for guidance. Int J Health Geogr 4 (1):3.
Reynolds, P., J. Von Behren, R. B. Gunier, D. E. Goldberg, M. Harnly, and A. Hertz.
2005. Agricultural pesticide use and childhood cancer in California. Epidemiology
16 (1):93-100.
Reynolds, P., J. Von Behren, R. B. Gunier, D. E. Goldberg, A. Hertz, and M. E. Harnly.
2002. Childhood cancer and agricultural pesticide use: an ecologic study in
California. Environ Health Perspect 110 (3):319-24.
RHRC (Rural Health Research Center). Rural-Urban Commuting Area Codes (RUCAs)
2000 [cited. Available from http://depts.washington.edu/uwruca/.
Ritz, B., and S. Costello. 2006. Geographic model and biomarker-derived measures of
pesticide exposure and Parkinson's disease. Ann N Y Acad Sci 1076:378-87.

141

Ritz, B., and R. P. Rull. 2008. Assessment of environmental exposures from agricultural
pesticides in childhood leukaemia studies: challenges and opportunities. Radiat
Prot Dosimetry 132 (2):148-55.
Ritz, B. R., A. D. Manthripragada, S. Costello, S. J. Lincoln, M. J. Farrer, M. Cockburn,
and J. Bronstein. 2009. Dopamine transporter genetic variants and pesticides in
Parkinson's disease. Environ Health Perspect 117 (6):964-9.
Roberts, E. M., P. B. English, J. K. Grether, G. C. Windham, L. Somberg, and C. Wolff.
2007. Maternal residence near agricultural pesticide applications and autism
spectrum disorders among children in the California Central Valley. Environ
Health Perspect 115 (10):1482-9.
Rothman, K. J., S. Greenland, and T. L. Lash. 2008. Modern epidemiology: Lippincott
Williams & Wilkins.
Rull, R. P., R. Gunier, J. Von Behren, A. Hertz, V. Crouse, P. A. Buffler, and P.
Reynolds. 2009. Residential proximity to agricultural pesticide applications and
childhood acute lymphoblastic leukemia. Environ Res 109 (7):891-9.
Rull, R. P., and B. Ritz. 2003. Historical pesticide exposure in California using pesticide
use reports and land-use surveys: an assessment of misclassification error and
bias. Environ Health Perspect 111 (13):1582-9.
Rull, R. P., B. Ritz, and G. M. Shaw. 2006a. Neural tube defects and maternal residential
proximity to agricultural pesticide applications. Am J Epidemiol 163 (8):743-753.
Rull, R. P., B. Ritz, and G. M. Shaw. 2006b. Validation of self-reported proximity to
agricultural crops in a case-control study of neural tube defects. J Expo Sci
Environ Epidemiol 16 (2):147-55.
Rural Assistance Center. What is Rural? Frequently Asked Questions 2012 [cited 2013.
Available from
http://www.raconline.org/topics/ruraldef/ruraldeffaq.php#principal.
SAS. 2013. SAS. Cary, North Carolina, United States.
Strahler, A. H., C. E. Woodcock, and J. A. Smith. 1986. On the nature of models in
remote sensing. Remote sensing of environment 20 (2):121-139.
Szklo, M., and F. Nieto. 2007. Epidemiology: beyond the basics. Johns and Bartlett
Publishers. Inc. Maryland 2.
U.S. Census Bureau. Urban and Rural Classification 2000 [cited 2013. Available from
http://www.census.gov/geo/reference/urban-rural.html.
———. TIGER Products 2013 [cited 2013. Available from
http://www.census.gov/geo/maps-data/data/tiger.html.

142

USDA (U.S. Department of Agriculture). Farms, Land in Farms, Value of Land and
Buildings, and Land Use: 2007 and 2002 2007a [cited 2013. Available from
http://www.agcensus.usda.gov/Publications/2007/Full_Report/Volume_1,_Chapte
r_2_County_Level/California/st06_2_008_008.pdf.
———. Fertilizers and Chemicals Applied: 2007 and 2002 2007b [cited 2013. Available
from
http://www.agcensus.usda.gov/Publications/2007/Full_Report/Volume_1,_Chapte
r_2_County_Level/California/st06_2_042_042.pdf.
———. Selected Crops Harvested: 2007 2007c [cited 2013. Available from
http://www.agcensus.usda.gov/Publications/2007/Full_Report/Volume_1,_Chapte
r_2_County_Level/California/st06_2_025_025.pdf.
———. Rural-Urban Commuting Area Codes: Overview 2012 [cited 2013. Available
from http://www.ers.usda.gov/data-products/rural-urban-commuting-area-
codes.aspx#.UWSs-kprYRw.
———. Four Band Digital Imagery: Information Sheet 2013a [cited 2013. Available
from www.fsa.usda.gov/Internet/FSA_File/fourband_info_sheet_2013.pdf.
———. State Fact Sheets 2013b [cited 2013. Available from
http://www.ers.usda.gov/data-products/state-fact-sheets/state-
data.aspx?StateFIPS=06&StateName=California#.UWSfVkprYRw.
USGS (U.S. Geological Survey). NDVI, the Foundation for Remote Sensing Phenology
2011 [cited 2013. Available from
http://phenology.cr.usgs.gov/ndvi_foundation.php.
———. Landsat Processing Details 2013a [cited 2013. Available from
http://landsat.usgs.gov/Landsat_Processing_Details.php
———. Landsat: A Global Land-Imaging Mission 2013b [cited 2013. Available from
http://landsat.usgs.gov/.
———. Path/Row Shapefiles 2013c [cited 2013. Available from
http://landsat.usgs.gov/tools_wrs-2_shapefile.php.
Vassiliou, A., M. Boulianne, and J. Blais. 1988. On the application of averaging median
filters in remote sensing. Geoscience and Remote Sensing, IEEE Transactions on
26 (6):832-838.
Waller, L. A., and C. A. Gotway. 2004. Applied spatial statistics for public health data:
Wiley-Interscience.
Wang, A., S. Costello, M. Cockburn, X. Zhang, J. Bronstein, and B. Ritz. 2011.
Parkinson's disease risk from ambient exposure to pesticides. Eur J Epidemiol 26
(7):547-55.

143

Ward, M. H., J. Lubin, J. Giglierano, J. S. Colt, C. Wolter, N. Bekiroglu, D. Camann, P.
Hartge, and J. R. Nuckols. 2006. Proximity to crops and residential exposure to
agricultural herbicides in iowa. Environ Health Perspect 114 (6):893-7.
Ward, M. H., J. R. Nuckols, S. J. Weigel, S. K. Maxwell, K. P. Cantor, and R. S. Miller.
2000. Identifying populations potentially exposed to agricultural pesticides using
remote sensing and a Geographic Information System. Environ Health Perspect
108 (1):5-12.
Wardlow, B. D., and S. L. Egbert. 2008. Large-area crop mapping using time-series
MODIS 250 m NDVI data: An assessment for the US Central Great Plains.
Remote sensing of environment 112 (3):1096-1116.
Wood, A. 2013. Compendium of Pesticide Common Names 2010 [cited 2013]. Available
from http://www.alanwood.net/pesticides/class_pesticides.html
YCEO (Yale Center for Earth Observatoin). Yale Guide to Landsat 8 Image Processing
2013 [cited 2013. Available from
http://www.yale.edu/ceo/Documentation/Landsat%208%20image%20processing.
pdf

144

APPENDIX A: PESTICIDE DATABASE
Table A1 Organochlorine pesticides
1

Pesticide name CDPR chemical code
Aldrin 9
Allidochlor 114
Chlordane 130
Chlorobenzilate 132
Chloroneb 135
DCPA 179
Dalapon 180
TDE 184
DDT 186
Dichlone 202
Dieldrin 210
Endosulfan 259
Endrin 262
Heptachlor 317
Dicofol 346
Chlordecone 347
Gamma-HCH 359
Methoxychlor 384
Mirex 402
Quintozene 464
Pentachlorophenol 465
Dienochlor 468
Ethyl-DDD 472
Tetradifon 581
Toxaphene 594
Trichlorobenzoic Acid 602
Azacosterol dihydrochloride 2026
Acetochlor 2349
HCH 5835

1
Data from Dich et al. (1997); Gunier et al. (2001); Alavanja,
Hoppin, and Kamel (2004); Greene and Pohanish (2005);
Rull et al. (2006a, 2009); Wood (2010); and AgroPages (2013)

Table A2 Organophosphate pesticides
1

Pesticide name CDPR chemical code
Temephos 1
Monocrotophos 52

145

Table A2 continued

Pesticide name CDPR chemical code
Fenthion 63
Bensulide 70
Dicrotophos 72
Trichlorfon 88
Carbophenothion 110
Crotoxyphos 140
Coumaphos 165
Fensulfothion 181
Dichlorvos 187
Tribufos 190
Dioxathion 192
Diazinon 198
Dimethoate 216
Disulfoton 230
Chlorpyrifos 253
Fonofos 254
Butonate 255
EPN 263
Ethion 268
Famphur 282
Tetrachlorvinphos 305
2,4-DEP 306
Azinphos methyl 314
Phosmet 335
Malathion 367
Oxydemeton-methyl 382
Methyl parathion 394
Ethoprophos 404
Naled 418
Schradan 446
Parathion 459
Phorate 478
Phosalone 479
Mevinphos 480
Phosphamidon 482
Fenchlorphos 517
Crufomate 519
Sulfotep 558
Demeton 566
TEPP 577
Dichlofenthion 614

146

Table A2 continued

Pesticide name CDPR chemical code
Phosacetim 1523
Ethephon 1626
Leptophos 1676
Acephate 1685
Methidathion 1689
Methamidophos 1697
Dialifos 1799
Glyphosate-isopropylammonium 1855
Fospirate 1856
Fenamiphos 1857
Fosamine ammonium 1921
Edifenphos 1964
Sulprofos 2006
Profenofos 2042
Propetamphos 2122
Isofenphos 2194
Fosetyl-al 2210
Pirimiphos-methyl 2217
Glyphosate-sesquisodium 2275
Isazofos 2282
Omethoate 2285
Glyphosate-trimesium 2327
Isocarbophos 2414
Butathiofos 2433
Chlorpyrifos-methyl 2468
Chlorthiophos 2469
Fenitrothion 2520
Pirimiphos-ethyl 2781
Terbufos 2925
Thionazin 2939
Glyphosate 2997
Triazophos 3543
Vamidothion 3544
Glufosinate-ammonium 3946
Azinphos-ethyl 4053
Demeton-methyl 4063
Paraoxon 4082
Prothiofos 4094
Trichloronate 5001
Chlorethoxyfos 5106

147

Table A2 continued

Pesticide name CDPR chemical code
Tebupirimfos 5122
Glyphosate-diammonium 5810

1
Data from Dich et al. (1997); Gunier et al. (2001); Alavanja,
Hoppin, and Kamel (2004); Greene and Pohanish (2005);
Rull et al. (2006a, 2009); Wood (2010); and AgroPages (2013)

Table A3 Carbamate pesticides
1

Pesticide name CDPR chemical code
Terbucarb 51
Barban 55
Propoxur 62
Bufencarb 91
Carbaryl 105
Carbofuran 106
Formetanate hydrochloride 111
Bcpc 141
Propham 339
Methiocarb 375
Methomyl 383
Aldicarb 575
Pebulate 590
Mexacarbate 623
Phenmedipham 675
Dichlormate 690
Karbutilate 691
Benomyl 1552
Thiophanate 1684
Thiophanate-methyl 1696
Desmedipham 1748
Pirimicarb 1875
Oxamyl 1910
Bendiocarb 1924
Propamocarb 2147
Carbendazim 2176
Carbosulfan 2182
Butoxycarboxim 2201
Thiodicarb 2202
Aldoxycarb 2265

148

Table A3 continued

Pesticide name CDPR chemical code
Fenoxycarb 2283
Cimectacarb 2345
Aminocarb 2435
Thiofanox 2938
Trimethacarb 2962
Ammonium carbamate 3041
Propamocarb hydrochloride 4022
Dioxacarb 4067
Promecarb 4092
Swep 4098
Asulam 5076
Pyraclostrobin 5759
Iprovalicarb 5938

1
Data from Dich et al. (1997); Gunier et al. (2001); Alavanja,
Hoppin, and Kamel (2004); Greene and Pohanish (2005);
Rull et al. (2006a, 2009); Wood (2010); and AgroPages (2013)

149
APPENDIX B: PESTICIDE USE REPORT PROCESSING
Table B1 Pesticide Use Report (PUR) logic checks
1,2

Logic check Definition (1974-1989) Definition (1990) Action taken
1. Duplicates
Using county code, acres treated,
product number, AI code, pounds of
applied AI, application date, and
commodity code
Using county code, use number, grower ID, site
location ID, acres planted, acres treated, product
number, AI code, pounds of applied AI,
application date, and commodity code
Kept first
record
2. Spatially
Inconsistent
County
CO-MTRS outside of county
boundary CO-MTRS outside of county boundary Excluded
3. Inconsistent
County Code N/A
First two digits of grower ID does not match
county code Excluded
4. Missing
agricultural field
location IDs N/A Missing grower ID, site location ID, or CO-MTRS Excluded
5. Inconsistent CO-
MTRS for
location N/A
Given a grower ID and site location ID, different
CO-MTRS for different PUR records Excluded
6. Inconsistent acres
planted N/A
Given a grower ID, site location ID, CO-MTRS,
and commodity code, different planted acres Excluded
7. Treated acres
greater than
planted acres N/A Treated acres greater than planted acres Excluded

1
Data adapted from CDPR (2000a)
2
Logic checks 3 through 7 were not applied to PUR data extracted between 1974 and 1989 due to missing variables. These logic checks were
adapted from CDPR (2000a), which were designed to check PUR data beginning in 1990.

150

Table B2 Pesticide Use Reports between 1974 and 1989: logic checks
1,2

Year
PUR
records (N)
Agricultural use
[N (%)]
Logic check [N (%)]
1 2
3

1974 552,244 433,291 (78.5%) 416,238 (96.1%) 398,377 (91.9%)
1975 583,457 447,837 (76.8%) 394,528 (88.1%) 379,419 (84.7%)
1976 569,142 434,885 (76.4%) 416,626 (95.8%) 397,220 (91.3%)
1977 611,351 472,164 (77.2%) 431,371 (91.4%) 409,957 (86.8%)
1978 476,981 363,844 (76.3%) 347,916 (95.6%) 326,429 (89.7%)
1979 689,568 531,559 (77.1%) 462,887 (87.1%) 436,380 (82.1%)
1980 619,809 454,306 (73.3%) 436,565 (96.1%) 413,197 (91.0%)
1981 691,734 503,078 (72.7%) 481,823 (95.8%) 453,613 (90.2%)
1982 662,702 465,068 (70.2%) 444,310 (95.5%) 420,345 (90.5%)
1983 724,774 464,274 (64.1%) 445,556 (96.0%) 422,128 (90.9%)
1984 832,385 542,628 (65.2%) 521,387 (96.1%) 498,076 (91.8%)
1985 929,918 522,256 (56.2%) 501,877 (96.1%) 482,042 (92.3%)
1986 1,021,166 554,470 (54.3%) 531,502 (95.9%) 504,596 (91.0%)
1987 1,072,329 591,611 (55.2%) 566,131 (95.7%) 535,457 (90.5%)
1988 1,092,688 615,188 (56.3%) 587,519 (95.5%) 552,262 (89.8%)
1989 1,305,573 599,535 (45.9%) 568,657 (94.8%) 539,582 (90.0%)

1
Data from CDPR (2013)
2
Logic checks were sequentially applied to data. Row percentages for logic check 2 use the
number of agricultural PURs as the denominator.
3
By design, PURs with an invalid CO-MTRS are excluded at logic check 2.

Table B3 Pesticide Use Reports in 1990: logic checks
1

PUR
records
[N]
Agricultural
use
[N (%)]
Logic check [N (%)]
2

2,657,840
2,157,190
(81.2%)
1 2 3 4
2,092,940
(97.0%)
1,937,646
(89.8%)
1,860,144
(86.2%)
1,859,988
(86.2%)
5 6 7
1,751,727
(81.2%)
1,574,012
(73.0%)
1,540,315 (71.4%)

1
Data from CDPR (2013)
2
Logic checks were sequentially applied to data. Row percentages for the logic checks use the
number of agricultural PURs as the denominator.

151

Table B4: PUR Outliers
1

Year
Agricultural
use (N)
Chemical class
[N (%)]
Outlier 1
2
[N (%)]
Outlier 2
3
[N (%)]
Outlier 3
4
[N (%)]
1974 398,377 194,488 (48.8%) 3 (0.002%) 126 (0.1%) …
1975 379,419 182,453 (48.1%) 8 (0.004%) 172 (0.1%) …
1976 397,220 182,132 (45.9%) 14 (0.008%) 164 (0.1%) …
1977 409,957 194,199 (47.4%) 17 (0.01%) 189 (0.1%) …
1978 326,298 149,823 (45.9%) 3 (0.002%) 89 (0.1%) …
1979 436,378 188,438 (43.2%) 61 (0.04%) 150 (0.1%) …
1980 413,195 174,153 (42.1%) 326 (0.19%) 419 (0.2%) …
1981 453,613 193,440 (42.6%) 37 (0.01%) 168 (0.1%) …
1982 420,344 166,253 (39.6%) 26 (0.02%) 119 (0.1%) …
1983 422,127 163,010 (38.6%) 22 (0.02%) 108 (0.1%) …
1984 498,075 198,488 (39.9%) 17 (0.01%) 199 (0.1%) …
1985 482,041 189,742 (39.4%) 25 (0.01%) 219 (0.1%) …
1986 504,594 198,915 (39.4%) 43 (0.02%) 265 (0.1%) …
1987 535,456 215,810 (40.3%) 21 (0.01%) 218 (0.1%) …
1988 552,261 234,047 (42.4%) 36 (0.02%) 158 (0.7%) …
1989 539,582 226,483 (42.0%) 35 (0.02%) 168 (0.1%) …
1990 1,540,315 357,930 (23.2%) 89 (0.02%) 699 (0.2%) 1,765 (0.5%)

1
Data from CDPR (2013)
2
Outlier 1 refers to application rates >200 lb/ac (>1,000 lb/ac if fumigation).
3
Outlier 2 refers to application rates >50 times the median rate for all uses of that pesticide
product, commodity code, unit type, and record type.
4
Outlier 3 refers to identification via a neural network.

152

APPENDIX C: LANDSAT MOSAICS, 1990
Landsat band 3 (red) and band 4 (near infrared) images for Paths 41-42 and Rows 35-36
were mosaicked for the months between January and October 1990. The following two
mosaics are examples of the radiometrically corrected (to at-sensor reflectance) images
contributing to the crop signature library.

Figure C1 Landsat mosaic (band 3), Paths 41-42 and Rows 35-36, from October 1990
(Data from U.S. Census Bureau 2013; and USGS 2013b)

153

Figure C2 Landsat mosaic (band 4), Paths 41-42 and Rows 35-36, from October 1990
(Data from U.S. Census Bureau 2013; and USGS 2013b)

154

APPENDIX D: CROP SIGNATURE LIBRARY
Boxplots are presented for agricultural land use classes included in the crop signature
library (CSL) via stratified random sampling (SRS) (N=55). Agricultural land use classes
with few SRS samples will not show distinct boxplot features (e.g. avocado). Broad land
use classes, such as without subclass designations in the 1990 Kern County land use
survey (e.g. class=F, field crop) are included. Refer to Table D1 for land use class strata
sample sizes.

Table D1 Sampled land use class polygons in crop signature library
1,2

Land use class
SRS
samples (n) Land use class
SRS
samples (n)
Alfalfa 30 Vacant 18
Almond 30 Olive 17
Apple 30 Storage 17
Bean (dry) 30 Pepper 16
Carrot 30 Sweet potato 16
Corn 30 Apricot 14
Cotton 30 Miscellaneous truck 14
Farmstead 30
Idle-new lands
prepared for crops 12
Feed lot 30 Unpaved area 12
Field crop 30 Grain sorghum 11
Flowers and nursery 30 Industrial 10
Freeway 30
Fruit and vegetable
cannery 9
Grain and hay crop 30 Airport runway 7
Idle-cropped in past
year 30 Asparagus 7
Lawn area: irrigated 30 Cemetery: irrigated 7
Lettuce 30 Fig 7
Melon, squash,
cucumber 30 Turf farm 7
Mixed pasture 30 Grapefruit 6
Native vegetation 30 Jojoba 6
Onion and garlic 30 Cole crop 4
Orange 30 Prune 4

155

Table D1 continued
Land use class
SRS
samples (n) Land use class
SRS
samples (n)
Peach and nectarine 30 Cabbage 2
Pistachio 30 Cherry 2
Plum 30 Municipal auditorium 2
Potato 30 Native pasture 2
Residence: 3-4
houses/ac 30 Oil refinery 2
Safflower 30 School 2
Sugar beet 30 Avocado 1
Tomato 30 Bushberry 1
Urban 30 Commercial 1
Vineyard 30 Idle 1
Water surface 30 Manufacturing 1
Lemon 29
Miscellaneous
establishment 1
Kiwi 27
Miscellaneous
subtropical fruit 1
Walnut 26 Motel 1
Miscellaneous
deciduous 24 Paved area 1
Truck, nursery, and
berry crop 24 Pea 1
Bean (green) 23 Pear 1
Extractive industry 22 Rice 1
Dairy 21
Urban: residential 1
Sudan 21

1
Data from CDWR (2013)
2
A total of 1,423 samples across all 81 land use classes included in the crop signature library
were included.

Table D2 Monthly NDVI values by land use strata
1,2

Land use class
NDVI
values

(N) Land use class
NDVI
values

(N)
Native vegetation 149,648 Grain sorghum 4,728
Pistachio 26,878 Dairy 4,118
Urban 26,540 Sweet potato 3,846
Safflower 20,960 Fig 3,757
Almond 19,746 Vacant 3,405
Cotton 17,531 Jojoba 3,271

156

Table D2 continued
Land use class
NDVI
values

(N) Land use class
NDVI
values

(N)
Grain and hay crop 16,031 Asparagus 3,232
Apple 14,798 Kiwi 3,050
Idle-cropped in
past year 14,316
Miscellaneous
truck 3,022
Vineyard 13,556 Apricot 2,992
Lettuce 13,154 Airport runway 2,464
Olive 12,744 Rice 2,427
Sugar beet 11,620 Farmstead 2,199
Potato 11,191 Storage 1,816
Tomato 11,070 Grapefruit 1,711
Lemon 10,706 Industrial 1,648
Freeway 10,702 Unpaved area 1,519
Alfalfa 10,394
Municipal
auditorium 1,232
Bean (dry) 10,359 Turf farm 1,153
Water surface 10,347 Prune 966
Field crop 10,324
Fruit and vegetable
cannery 940
Carrot 9,954
Cemetery:
irrigated 888
Orange 9,360 Cole crop 865
Plum 9,293
Miscellaneous
subtropical fruit 356
Melon, squash,
cucumber 9,252 Cherry 347
Onion and garlic 9,217 Native pasture 315
Idle-new lands
prepared for crops 9,131 Oil refinery 296
Truck, nursery,
and berry crop 9,052 Cabbage 232
Walnut 8,404 Pear 219
Extractive industry 7,628 Commercial 135
Sudan 7,015
Miscellaneous
establishment 122
Corn 6,633 Idle 119
Feed lot 6,559 Manufacturing 100
Residence: 3-4
houses/ac 6,096 Bushberry 88
Miscellaneous
deciduous 5,970 School 88
Bean (green) 5,576 Urban: residential 88

157

Table D2 continued
Land use class
NDVI
values

(N) Land use class
NDVI
values

(N)
Flowers and
nursery 5,493 Pea 83
Mixed pasture 5,164 Avocado 61
Peach and
nectarine 5,152 Paved area 53
Lawn area:
irrigated 4,801 Motel 52
Pepper 4,759

1
Data from CDWR (2013)
2
There are 645,127 NDVI values for each month between January and October 1990.
There is a total of 6,451,270 NDVI values across all months contributing to the crop
signature library.

Figure D1 Boxplot characteristics

158

Figure D2 Alfalfa: NDVI in Kern County, 1990

Figure D3 Almond: NDVI in Kern County, 1990

159

Figure D4 Apple: NDVI in Kern County, 1990

Figure D5 Apricot: NDVI in Kern County, 1990

160

Figure D6 Asparagus: NDVI in Kern County, 1990

Figure D7 Avocado: NDVI in Kern County, 1990

161

Figure D8 Bean (dry): NDVI in Kern County, 1990

Figure D9 Bean (green): NDVI in Kern County, 1990

162

Figure D10 Bushberry: NDVI in Kern County, 1990

Figure D11 Cabbage: NDVI in Kern County, 1990

163

Figure D12 Carrot: NDVI in Kern County, 1990

Figure D13 Cherry: NDVI in Kern County, 1990

164

Figure D14 Cole crop: NDVI in Kern County, 1990

Figure D15 Corn: NDVI in Kern County, 1990

165

Figure D16 Cotton: NDVI in Kern County, 1990

Figure D17 Field crop: NDVI in Kern County, 1990

166

Figure D18 Fig: NDVI in Kern County, 1990

Figure D19 Flowers and nursery: NDVI in Kern County, 1990

167

Figure D20 Grain and hay crop: NDVI in Kern County, 1990

Figure D21 Grain sorghum: NDVI in Kern County, 1990

168

Figure D22 Grapefruit: NDVI in Kern County, 1990

Figure D23 Idle: NDVI in Kern County, 1990

169

Figure D24 Idle-cropped in past year: NDVI in Kern County, 1990

Figure D25 Idle-new lands prepared for crops: NDVI in Kern County, 1990

170

Figure D26 Jojoba: NDVI in Kern County, 1990

Figure D27 Kiwi: NDVI in Kern County, 1990

171

Figure D28 Lemon: NDVI in Kern County, 1990

Figure D29 Lettuce: NDVI in Kern County, 1990

172

Figure D30 Melon, squash, cucumber: NDVI in Kern County, 1990

Figure D31 Miscellaneous deciduous: NDVI in Kern County, 1990

173

Figure D32 Miscellaneous subtropical fruit: NDVI in Kern County, 1990

Figure D33 Miscellaneous truck: NDVI in Kern County, 1990

174

Figure D34 Mixed pasture: NDVI in Kern County, 1990

Figure D35 Native pasture: NDVI in Kern County, 1990

175

Figure D36 Olive: NDVI in Kern County, 1990

Figure D37 Onion and garlic: NDVI in Kern County, 1990

176

Figure D38 Orange: NDVI in Kern County, 1990

Figure D39 Pea: NDVI in Kern County, 1990

177

Figure D40 Peach and nectarine: NDVI in Kern County, 1990

Figure D41 Pear: NDVI in Kern County, 1990

178

Figure D42 Pepper: NDVI in Kern County, 1990

Figure D43 Pistachio: NDVI in Kern County, 1990

179

Figure D44 Plum: NDVI in Kern County, 1990

Figure D45 Potato: NDVI in Kern County, 1990

180

Figure D46 Prune: NDVI in Kern County, 1990

Figure D47 Rice: NDVI in Kern County, 1990

181

Figure D48 Safflower: NDVI in Kern County, 1990

Figure D49 Sudan: NDVI in Kern County, 1990

182

Figure D50 Sugar beet: NDVI in Kern County, 1990

Figure D51 Sweet potato: NDVI in Kern County, 1990

183

Figure D52 Tomato: NDVI in Kern County, 1990

Figure D53 Truck, nursery, and berry crop: NDVI in Kern County, 1990

184

Figure D54 Turf farm: NDVI in Kern County, 1990

Figure D55 Vineyard: NDVI in Kern County, 1990

185

Figure D56 Walnut: NDVI in Kern County, 1990

186
APPENDIX E: SEGMENTATION AND CLASSIFICATION

Figure E1 First three principal components derived from 1985 Landsat NDVI images.
These PCA images were used for segmentation.

187

Figure E2 Classification 1 (standard): common land use classes

Figure E3 Classification 2 (subclass-required): common land use classes

188

Figure E4 Classification 3 (strict): common land use classes

Table E1 Agricultural land use classes: 1990 Kern County land use survey
1

Land use class N Percent Land use class N Percent
Cotton 3,036 27.5
Truck, nursery,
and berry crop 24 0.22
Alfalfa 1,420 12.86 Turf farm 23 0.21
Field crop 1,226 11.1 Olive 21 0.19
Almond 776 7.03 Pepper 18 0.16
Vineyard 751 6.8
Miscellaneous
truck 17 0.15
Grain and hay crop 625 5.66 Sweet potato 16 0.14
Orange 490 4.44
Idle-new lands
prepared for crops 13 0.12
Idle-cropped in
past year 486 4.4 Grain sorghum 12 0.11
Potato 224 2.03 Jojoba 9 0.08
Onion and garlic 178 1.61 Fig 8 0.07
Pistachio 139 1.26 Asparagus 7 0.06
Mixed pasture 135 1.22 Grapefruit 6 0.05

189

Table E1 continued

Land use class N Percent Land use class N Percent
Sugar beet 132 1.2 Cole crop 5 0.05
Carrot 130 1.18 Native pasture 5 0.05
Corn 130 1.18 Avocado 4 0.04
Flowers and
nursery 128 1.16 Prune 4 0.04
Peach and
nectarine 118 1.07
Miscellaneous
subtropical fruit 3 0.03
Apple 104 0.94 Bushberry 2 0.02
Bean (dry) 88 0.8 Cabbage 2 0.02
Melon, squash,
cucumber 86 0.78 Cherry 2 0.02
Tomato 72 0.65 Strawberry 2 0.02
Plum 60 0.54 Artichoke 1 0.01
Kiwi 45 0.41 Castor bean 1 0.01
Miscellaneous
deciduous 39 0.35 Celery 1 0.01
Lettuce 37 0.34
Deciduous fruit
and nut 1 0.01
Walnut 37 0.34 Idle 1 0.01
Lemon 32 0.29 Pasture 1 0.01
Safflower 32 0.29 Pea 1 0.01
Bean (green) 25 0.23 Pear 1 0.01
Apricot 24 0.22
Rice 1 0.01
Sudan 24 0.22

1
Data from CDWR (2013)

Table E2 Final CSL-classified land use classes

Land use class N Percent Land use class N Percent
Cotton 1,878 18.76 Mixed pasture 116 1.16
Miscellaneous
subtropical fruit 717 7.16
Peach and
nectarine 100 1.00
Lemon 599 5.99 Apple 99 0.99
Cole crop 525 5.25 Bean (green) 93 0.93
Miscellaneous
truck 387 3.87 Orange 80 0.80
Alfalfa 383 3.83 Turf farm 76 0.76
Prune 381 3.81 Sugar beet 75 0.75

190

Native pasture 329 3.29 Apricot 73 0.73
Potato 311 3.11
Idle-cropped in
past year 67 0.67
Flowers and
nursery 292 2.92 Onion and garlic 54 0.54
Kiwi 290 2.90 Asparagus 50 0.50
Pear 287 2.87 Safflower 44 0.44
Grapefruit 259 2.59 Corn 40 0.40
Walnut 257 2.57 Grain sorghum 38 0.38
Avocado 201 2.01 Lettuce 35 0.35
Bean (dry) 193 1.93 Sweet potato 32 0.32
Melon, squash,
cucumber 192 1.92 Tomato 31 0.31
Sudan 188 1.88
Miscellaneous
deciduous 30 0.30
Plum 183 1.83 Fig 28 0.28
Pea 168 1.68 Pepper 15 0.15
Pistachio 168 1.68 Bushberry 14 0.14
Almond 165 1.65 Rice 10 0.10
Carrot 161 1.61 Cabbage 3 0.03
Olive 148 1.48
Idle-new lands
prepared for crops 2 0.02 Cherry 141 1.41

191

APPENDIX F: APPLIED PESTICIDES AND RURALITY

Figure F1 Kern County Urbanized Areas and Urban Clusters (2000)
(Data from U.S. Census Bureau 2013)

192

Figure F2 Applied pesticides and ZCTA rurality: organochlorines
(Data from CDPR 2013; and U.S. Census Bureau 2013)

193

Figure F3 Applied pesticides and ZCTA rurality: organophosphates
(Data from CDPR 2013; and U.S. Census Bureau 2013)

194

Figure F4 Applied pesticides and ZCTA rurality: carbamates
(Data from CDPR 2013; and U.S. Census Bureau 2013)

195

Figure F5 Applied pesticides and census tract rurality: organochlorines
(Data from CDPR 2013; and U.S. Census Bureau 2013)

196

Figure F6 Applied pesticides and census tract rurality: organophosphates
(Data from CDPR 2013; and U.S. Census Bureau 2013)

197

Figure F7 Applied pesticides and census tract rurality: carbamates
(Data from CDPR 2013; and U.S. Census Bureau 2013)

198

Table F1 Pesticide-treated crop fields and sections intersecting ZCTAs by rurality

Organochlorines
N Mean ± SD Median (IQR) Min. Max.
RUCA
Rural 20 54.25  101.69 2 (34) 0 307
Urban 27 68  110.37 7 (144) 0 457
U.S. Census
Bureau
Rural 18 30.06  76.02 1.5 (7) 0 266
Urban 29 82.07  117.57 14 (155) 0 457
Organophosphates
RUCA
Rural 20 94.75  163.47 10 (75) 0 481
Urban 27 104  156.34 21 (206) 0 658
U.S. Census
Bureau
Rural 18 54.78  121.67 10.5 (24) 0 481
Urban 29 128.17  172.45 32 (220) 1 658
Carbamates
RUCA
Rural 20 76.15  138.84 5.5 (41) 0 415
Urban 27 87  134.49 12 (187) 0 577
U.S. Census
Bureau
Rural 18 42  97.19 9 (15) 0 377
Urban 29 107.45  149.97 23 (191) 1 577

Table F2 Pesticide-treated crop fields and sections intersecting census tracts by rurality

Organochlorines
N Mean ± SD Median (IQR) Min. Max.
RUCA
Rural 28 24.86  93.71 1 (5) 0 486
Urban 112 21.58  62.88 2 (8) 0 424
U.S. Census
Bureau
Rural 7 34.71  72.70 3 (27) 0 198
Urban 133 21.58  69.81 1 (6) 0 486
Organophosphates
RUCA
Rural 28 46.93  163.41 4 (6) 0 828
Urban 112 33.44  89.33 4 (12) 0 631
U.S. Census
Bureau
Rural 7 78  101.80 38 (205) 0 239
Urban 133 33.93  107.86 4 (9) 0 828
Carbamates
RUCA
Rural 28 35.71  128.7 2 (6) 0 657
Urban 112 28.29  78.42 3 (11) 0 567
U.S. Census
Bureau
Rural 7 57  79.97 33 (118) 0 212
Urban 133 28.34  90.76 3 (8) 0 657

Abstract (if available)

Abstract Pesticide exposure estimation in epidemiologic studies can be constrained to analysis scales commonly available for cancer data—census tracts and ZIP codes. Research goals included (1) demonstrating the feasibility of modifying an existing geographic information system (GIS) pesticide exposure method using California Pesticide Use Reports (PURs) and land use surveys to incorporate Landsat remote sensing and to accommodate aggregated analysis scales, and (2) assessing the accuracy of two rurality metrics (quality of geographic area being rural), Rural-Urban Commuting Area (RUCA) codes and the U.S. Census Bureau urban-rural system, as surrogates for pesticide exposure when compared to the GIS gold standard. Segments, derived from 1985 Landsat NDVI images, were classified using a crop signature library (CSL) created from 1990 Landsat NDVI images via a sum of squared differences (SSD) measure. Organochlorine, organophosphate, and carbamate Kern County PUR applications (1974‐1990) were matched to crop fields using a modified three‐tier approach. Annual pesticide application rates (lb/ac), and sensitivity and specificity of each rurality metric were calculated. The CSL (75 land use classes) classified 19,752 segments [median SSD 0.06 NDVI]. Of the 148,671 PUR records included in the analysis, Landsat contributed 3,750 (2.5%) additional tier matches. ZIP Code Tabulation Area (ZCTA) rates ranged between 0 and 1.36 lb/ac and census tract rates between 0 and 1.57 lb/ac. Rurality was a mediocre pesticide exposure surrogate

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Comparing Landsat7 ETM+ and NAIP imagery for precision agriculture application in small scale farming: a case study in the south eastern part of Pittsylvania County, VA

PDF

Using Landsat and a Bayesian hard classifier to study forest change in the Salmon Creek Watershed area from 1972–2013

PDF

Use of remotely sensed imagery to map sudden oak death (Phytophthora ramorum) in the Santa Cruz Mountains

PDF

Remote analysis of avalanche terrain features: identifying routes, avoiding hazards

PDF

Using census data, urban-land cover classification, and dasymetric mapping to measure urban growth of the lower Rio Grande Valley, Texas

PDF

Detection and accuracy assessment of mountain pine beetle infestations using Landsat 8 OLI and WorldView02 satellite imagery: Lake Tahoe Basin-Nevada and California

PDF

A comparative study of ground and satellite evapotranspiration models for southern California

PDF

Quantifying changes in monarch butterfly habitat in California

PDF

Use of GIS for analysis of community health worker patient registries from Chongwe district, a rural low-resource setting, in Lusaka Province, Zambia

PDF

A comparison of urban land cover change: a study of Pasadena and Inglewood, California, 1992‐2011

PDF

Geographic information systems and marketing: a transdisciplinary approach to curriculum development

PDF

Exploring remote sensing and geographic information systems technologies to understand vegetation changes in response to land management practices at Finke Gorge National Park, Australia Between ...

PDF

Preparing for immigration reform: a spatial analysis of unauthorized immigrants

PDF

An evaluation of Esri’s tapestry segmentation product in three Southern California communities: Manhattan Beach, Santa Monica, and Venice Beach

PDF

Distribution and correlates of feral cat trapping permits in Los Angeles, California

PDF

Spatiotemporal analysis of the SLOSH and ADCIRC storm surge models: a case study of hurricane Ida

PDF

Geospatial analysis of the Round Fire: a replication of burn severity analyses in the Sierra Nevada

PDF

Developing a replicable approach for the creation of urban climatic maps for urban heat island analysis: a case study for the city of Los Angeles, California

PDF

Radio frequency identification queuing & geo-location (RAQGEO): a spatial solution to inventory management at XYZ Logistics, Inc.

PDF

Increase in surface temperature and deep layer nitrate in the California Current: a spatiotemporal analysis of four-dimensional hydrographic data

Asset Metadata

Creator Vopham, Trang Minh (author)

Core Title Integrating Landsat and California pesticide exposure estimation at aggregated analysis scales: accuracy assessment of rurality

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 01/23/2014

Defense Date 10/31/2013

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

Tag agricultural pesticides,landsat,OAI-PMH Harvest,pesticide exposure,remote sensing,rurality

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Wilson, John P. (committee chair), Rashed, Tarek (committee member), Ruddell, Darren M. (committee member)

Creator Email tmv7@pitt.edu,vopham@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-360988

Unique identifier UC11296577

Identifier etd-VophamTran-2231.pdf (filename),usctheses-c3-360988 (legacy record id)

Legacy Identifier etd-VophamTran-2231.pdf

Dmrecord 360988

Document Type Thesis

Format application/pdf (imt)

Rights Vopham, Trang Minh

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