Content uploaded by Michael J. Falkowski
Author content
All content in this area was uploaded by Michael J. Falkowski on Jun 26, 2015
Content may be subject to copyright.
Content uploaded by Michael J. Falkowski
Author content
All content in this area was uploaded by Michael J. Falkowski on Jun 26, 2015
Content may be subject to copyright.
Content uploaded by Michael J. Falkowski
Author content
All content in this area was uploaded by Michael J. Falkowski on Jun 26, 2015
Content may be subject to copyright.
!
1!
Research Paper 1!
2!
Citation: Falkowski, M.J., and J. Manning. 2010. Parcel-based classification of agricultural 3!
crops via multi-temporal Landsat imagery in support of habitat availability monitoring for 4!
western burrowing owls in the Imperial Valley agro-ecosystem. Canadian Journal of Remote 5!
Sensing. Vol. 36, No. 6, 750-762. 6!
7!
Parcel-based classification of agricultural crops via multi-temporal Landsat 8!
imagery in support of habitat availability monitoring for western burrowing owls in 9!
the Imperial Valley agro-ecosystem. 10!
11!
Michael J. Falkowski and Jeffrey A. Manning 12!
13!
14!
M.J. Falkowski.1
15!
School of Forest Resources and Environmental Science, Michigan Technological 16!
University, Houghton, MI 49931-1295, USA
17!
18!
J.A. Manning. 19!
Fish and Wildlife Resources, College of Natural Resources 107, University of Idaho, 20!
Moscow, ID 83844-1136, USA 21!
22!
1Corresponding author (e-mail:mjfalkow@mtu.edu) 23!
24!
!
2!
Abstract 1!
Agricultural production has a large impact upon the sustainability of wildlife 2!
populations. Certain species decline in the presence of intensive agriculture, while others 3!
thrive. Thus, quantifying changes in habitat within agricultural systems is paramount to 4!
understanding the underlying ecological mechanisms influencing species range shifts, 5!
and may facilitate the development of management strategies for sensitive wildlife 6!
species that depend on agricultural systems. Remotely sensed data can provide an 7!
efficient means to assess and monitor agricultural crop dynamics across large spatial 8!
extents. The objective of this study was to classify field-level agricultural crop types via a 9!
time series of Landsat imagery across the 3,620,000 ha agro-ecosystem in the Imperial 10!
Valley of California, USA. The primary impetus was to generate data for characterizing 11!
short-term changes in habitat for the western burrowing owl (Athene cunicularia), a 12!
species for concern that has an affinity for agricultural systems. The parcel based 13!
classification method employed attained overall accuracies of 84% and 69% for Level II 14!
(3 crop groups) and Level III (31 crop types) agricultural crops, respectively. Given the 15!
large number of crop categories classified, these accuracies were quite high, especially 16!
when compared to existing crop type classifications across the region (e.g., The National 17!
Cropland Data Layer). These results suggest that the crop type classifications presented 18!
herein could potentially provide a means to evaluate owl demographic and space use 19!
parameters in light of shifting crop patterns across the study area. 20!
21!
22!
23!
!
3!
Introduction 1!
Increasing awareness of the ecological changes that occur with agricultural 2!
conversion in large-scale agro-ecosystems (Peterjohn 2003) has offered new perspectives 3!
for studying how animals respond to land use and climate change. Although many fish 4!
and wildlife populations have declined in areas of agricultural production (Carlson 1985), 5!
some species nest, forage, and utilize cover in agricultural habitats (Moulton et al. 2006). 6!
Moreover, emerging evidence (Moulton et al. 2006) suggests that agricultural habitats 7!
may attenuate northern range expansions of some species that show an affinity for 8!
agriculture. Thus, quantifying changes in habitats within agro-ecosystems is a critical 9!
step toward understanding the underlying ecological mechanisms that drive species range 10!
shifts, and may facilitate the development of management strategies for sensitive wildlife 11!
species that depend on agricultural systems. 12!
Vegetation cover type can be assessed through a variety of methods, including ground 13!
surveys, organic deposits (Witt et al. 2006), mapping or historical records (Fensham 14!
2008), and remotely sensed imagery (Smith et al. 2010). Ground surveys and mapping of 15!
individual crops across agro-ecosystems is not practical due to the large spatial extent, 16!
high number of property owners, diversity of crops, and frequency of crop rotation. 17!
Satellite remote sensing, however has been successfully used to map, inventory, and 18!
monitor land cover types (e.g., Haack 1987, Vogelman et al. 2001, Falkowski et al. 19!
2005), including agricultural crops (Bauer 1985, Ortiz et al. 1997). The Landsat family of 20!
satellite sensors was designed specifically with such information needs in mind (Beck 21!
and Gessler 2008), and the synoptic view provided via remotely sensed data are ideal for 22!
gathering, mapping, and monitoring information on land surface and vegetation 23!
!
4!
characteristics across large areas such as agro-ecosystems. Landsat imagery can be used 1!
to measure a variety of land cover and crop types classified according to the habitat 2!
characteristics identified for a specific wildlife species in an agro-ecosystem. Compared 3!
to traditional field-based crop inventory techniques, maps produced via remotely sensed 4!
data require a relatively small amount of field data that reduces field sampling while 5!
maintaining accuracy. Lastly, it provides complete coverage across large areas as well as 6!
a long-term data archive ideal for quantifying changes in wildlife habitats through time. 7!
Traditional, pixel-based classifications of remotely sensed data are often ill suited to 8!
modeling real landscapes because they cannot accommodate multiple, hierarchical scales 9!
of biological organization (Townshend et al. 2000, Burnett and Blaschke 2003). Because 10!
Landsat image pixels are 30 x 30 m and agricultural fields are comprised of numerous 11!
pixels, variable crop conditions (e.g., stages of development) and soil types among pixels 12!
within a field leads to a high degree of within field classification variation. This variably 13!
is represented by a ‘salt and pepper’ appearance in pixel-based classification maps. This 14!
within field variably can reduce accuracy of classifying field-level crop types correctly. A 15!
solution is to use existing field polygon boundaries as parcels in an object based 16!
classification (Benz et al. 2004). This approach, which is a form of objected-based image 17!
analysis (OBIA; Hay and Castilla, 2008), aggregates pixels within GIS polygon layer 18!
depicting agricultural field boundaries, and uses within-field frequency distributions of 19!
pixel values to classify entire agricultural fields into desired crop type categories. 20!
Because this approach removes within field classification variability it should improve 21!
the accuracy of classifying crops, and may prove effective for classifying field level crop 22!
types and quantifying field level habitat change. 23!
!
5!
Although individual crop types could be classified from single date Landsat imagery, 1!
many studies have found increased accuracy through the incorporation of multi-date 2!
imagery (Oetter et al. 2001; Wardlow and Egbert, 2008). This is because different plant 3!
species have unique phenologic trajectories, or stages of development and growth 4!
patterns, that can be detected via a time series of remotely sensed data (Oetter et al. 5!
2000). Thus, multi-date imagery series embedded in an OBIA may further improve 6!
classification of crop types in agro-ecosystems where a diversity of crop types typically 7!
exist in different stages of development and growth. 8!
The primary objective of this study is to explore the use of OBIA with time series of 9!
Landsat data embedded in each object (i.e., agricultural fields) to classify field-level crop 10!
types for characterizing short-term changes in availability of wildlife habitat in agro-11!
ecosystems. The crop type classifications developed in this study will eventually be 12!
employed in the future to characterize habitat availability dynamics for the western 13!
burrowing owls (Athene cunicularia) in the Imperial Valley agro-ecosystem in the 14!
southwest United States, an important region for burrowing owls (Desante et al. 2004, 15!
Sauer et al. 2008). This species is a conspicuous inhabitant of grasslands, deserts, 16!
agricultural systems, and other arid areas throughout western North America and Central 17!
and South America (Haug et al. 1993). This species has special conservation status (Klute 18!
et al. 2003) due to range-wide declines, although some populations in agro-ecosystems 19!
within their southern range have increased, suggesting an increasing dependence on 20!
agriculture (Wellicome and Holroyd 2001, Desante et al. 2004, Moulton et al. 2006, 21!
Sauer et al. 2008). Owls in agro-ecosystems commonly nest along borders of agricultural 22!
fields and forage in the adjacent fields (Rosenberg and Haley 2004). Emerging evidence 23!
!
6!
(Rosenberg and Haley 2004) suggests that the home range and space use of burrowing 1!
owls in agro-ecosystems are linked to specific types of agricultural crops (i.e., alfalfa, 2!
bare ground, and fallow fields). A large-scale study of annual abundance and distribution 3!
of burrowing owls in the Imperial Valley agro-ecosystem was completed in 2006 -2007 4!
(Imperial Irrigation District 2009). However, these studies have not conducted a detailed 5!
analysis of the influence that cropping patterns have upon burrowing owl dynamics. 6!
Thus, the GIS-based distribution maps of crop types generated in the current study (crop 7!
types across the Imperial Valley agro-ecosystem each spring from 2003 – 2007) will be 8!
leveraged in the future to determine the influence of crop type and dynamics on owl 9!
abundance. 10!
Although, the US Department of Agriculture‘s National Agriculture Statistics Service 11!
employs remotely sensed imagery to periodically generate a large area agricultural 12!
classification product across the entire US (The National Cropland Data Layer; Craig 13!
2001, Mueller and Ozga 2002), an independent assessment of this product demonstrates 14!
that it lacks the accuracy and precision required for a detailed analysis of the influence of 15!
cropping patterns on owl abundance. The results of the CDL product assessment are also 16!
presented herein. 17!
Materials and methods 18!
Study site 19!
The study area was the 3,620,000-ha agro-ecosystem in the Imperial Valley of 20!
California, USA (32o 58’ N, 115o 31’ W). This site supports one of highest concentrations 21!
of burrowing owls in North America (Desante et al. 2004, Sauer et al. 2008). Extensive 22!
landscape change occurred in this area during the early 20th century, when it was 23!
!
7!
converted from a desert ecosystem to an agro-ecosystem via irrigation water supplied by 1!
the Colorado River (Bailey 1994). The primary land use across this region is irrigated 2!
crop agriculture. However, there are also additional cover types and land uses present 3!
including urban, suburban, industrial, pasture, wetland, abandoned fields, desert dry 4!
wash, riparian woodland, irrigation drains, and roadside embankments. 5!
6!
Ground reference data 7!
Ground reference data were collected at 420 randomly located agricultural crop 8!
fields across the study site from March 31-April 22, 2007. We chose to survey 420 9!
fields because it was the maximum number possible give time and funding constraints. 10!
In each random field, a point location was randomly selected, and the type of crop and 11!
stage of development (bare ground, sprout, mid-stage, mature, abandoned, grazed, 12!
stubble) present in the field were visually identified by a crop specialist and recorded at 13!
that point. Field sample locations were recorded using a high precision GPS unit. Crops 14!
were hierarchically identified in the field as crop groups (Level II crops) and individual 15!
crop species (Level III crops, Table 1). Sample sizes for each crop type ranged from 3 16!
to 52 (Table 1). The level of precision in terms of level III crop types is very high. 17!
Having these detailed crop categories will enable, in the future, a detailed analysis of 18!
potential crops influencing burrowing owl abundance and distribution. 19!
20!
Evaluation of the National Cropland Data Layer 21!
Prior to developing our independent crop classifications we employed the ground 22!
reference data to evaluate the accuracy of the 2007 National Cropland Data Layer (CDL) 23!
!
8!
from California. Specifically, we assessed the accuracy of the 30 m CDL layer, which is 1!
generated from Landsat data via a classification and regression tree approach (USDA 2!
NASS, 2009). Although, according to the metadata (USDA NASS, 2007), the CDL layer 3!
should contain most of the desired level III classification categories, this was not the case 4!
across the Imperial Valley (i.e., according to the CDL data set many of the crop Level III 5!
types observed in the field were not present across the Imperial Valley study area). This 6!
makes it problematic to assess CDL accuracy for the Level III crop categories. Thus, the 7!
CDL layer was re-categorized into one of the Level II crop categories (Table 2). A 8!
traditional accuracy assessment was then conducted on this re-categorized CLD layer. 9!
10!
Landsat image acquisition and processing 11!
We acquired 3 Landsat images (30 x 30 m resolution) for each spring from 2003 to 12!
2007 (totaling 15 images) from the United States Geological Survey (USGS). The 13!
image acquisition dates of the 2007 images (Table 2) coincided with the collection of 14!
ground reference data. Due to large amounts of cloud cover during March and April it 15!
was necessary to use a combination of Landsat 5 and Landsat 7 data to avoid 16!
contamination of the data by clouds (i.e., if a particular Landsat 5 image was 17!
contaminated by clouds, we used a cloud free Landsat 7 image acquired within a 18!
similar time frame). This strategy provided imagery coverage during the desired time 19!
frame for each of the 5 years between 2003 and 2007, inclusive. 20!
The Landsat 7 sensor scan line corrector (SLC) failed on May 31st , 2003, resulting 21!
in imagery that contains systematic data gaps, which are more pronounced toward the 22!
edges of a Landsat scene. Indeed, the SLC errors present a potential disadvantage of 23!
!
9!
using Landsat 7 data for this study. However, the OBIA approach implemented in this 1!
project (described below) should be relatively unaffected by any systematic gaps 2!
within a Landsat 7 scene since the classification rules rely upon distributional statistics 3!
of spectral reflectance within each individual agricultural field. Furthermore, since the 4!
Imperial Valley Region is approximately at the center of the Landsat scene (WRS-2 5!
Path 39, Row 27), the systematic data gaps are either non-existent or relatively small 6!
across the entire study site. 7!
Each Landsat image was systematically processed to correct for radiometric and 8!
geometric errors via a streamlined image processing methodology developed by Beck 9!
and Gessler (2008). This was accomplished by geo-reregistering the April 2007 image 10!
to high-resolution digital orthophotos, acquired by the National Agriculture Imagery 11!
Program (NAIP), spanning the study area. The remaining 14 Landsat images were then 12!
co-registered to the April 2007 image. This geo-registration process ensured that each 13!
Landsat scene had a geo-location accuracy of < 15 m on the ground. 14!
Following geometric correction, each image was radiometrically calibrated and 15!
corrected to remove atmospheric effects via an integrated image correction procedure. 16!
This image correction procedure converts raw digital numbers to radiance and 17!
eventually percent reflectance via equations and coefficients described by Markhan and 18!
Barker (1986), Chander and Markham (2003), and Williams (2004)). Atmospheric 19!
effects are removed (e.g., haze and dust) via the COST atmospheric correction model 20!
(Chavez 1996). The COST model employs a modified dark body subtraction 21!
atmospheric correction that corrects for variation in atmospheric transmittance effects 22!
by including the cosine of the solar zenith angle (Chavez 1996). This process 23!
!
10!
radiometrically calibrates each image and removes radiometric errors, ensuring that 1!
each image was directly comparable to another image within the time series, regardless 2!
of acquisition date and/or radiometric differences between the Landsat 5 and Landsat 7 3!
sensors (i.e., this process allows the direct comparison of Landsat 5 and Landsat 7 4!
images). 5!
6!
Creation of image objects for classification 7!
Object based image analysis (OBIA) is becoming a more common approach for 8!
image classification. A recent review papers by Blaschke (2010) as well as Lu and Weng 9!
(2007) present detailed descriptions of the state-of-the-art in OBIA; thus, a detailed 10!
review is beyond the scope of this paper. However, in general there are two distinct 11!
approaches to OBIA; one in which objects are directly derived from remotely based on 12!
photomorphic units (often termed image segmentation) and another where predefined 13!
objects in a GIS layer are intersected with remotely sensed data prior to classification 14!
(often termed parcel-based image classification) (Lu and Weng, 2007). In the current 15!
study, we employed a parcel based image classification where image objects were based 16!
on a GIS layer depicting Common Land Units (CLUs; U.S. Department of Agriculture, 17!
Farm Service Agency, Aerial Photography Field Office, Salt Lake City, Utah; 18!
www.apfo.usda.gov). A CLU is a GIS polygon layer outlining individual agricultural 19!
fields; buildings, residential areas, impervious surfaces, and irrigation ditches are not 20!
included within the polygon boundary. The CLU layer was intersected with each Landsat 21!
image to create groups of pixels representing image objects (hereafter referred to as 22!
agricultural fields). 23!
!
11!
Distributional statistics (e.g., mean, standard deviation, range, minimum values and 1!
maximum value of pixels) were calculated for each Landsat band within each agricultural 2!
field. This process created a GIS layer containing the aforementioned distributional 3!
statistics for each agricultural field across the Imperial Valley Study area for each of the 5 4!
years in the Landsat time series (Figure 1). The resulting GIS layer contained 5!
approximately 7,000 individual polygons (representing agricultural fields across the study 6!
area), each containing 35 different data attributes quantifying the distributional statistics 7!
of Landsat pixels within each agricultural field. Following this process, fields sampled 8!
during data collection were separated out of the 2007 GIS layer and used as training data 9!
in the classification, as described below. 10!
Distributional statistics were also calculated for the Normalized Difference 11!
Vegetation Index (NDVI), which is sensitive to living plant biomass (Tucker 1979). 12!
These metrics provided an opportunity to improve our classification of individual crops 13!
in the Imperial Valley Study area because individual crops display distinct variation in 14!
their spectral responses over time due to phenologic development. For example, Sudan 15!
grass fields in our study area showed low mean NDVI values in late March, and these 16!
rise through April, corresponding to plant establishment and growth during this time 17!
period (Figure 2). Sugar beet fields displayed a decreasing trend in the mean NDVI, 18!
which is mostly likely related to a gradual senesce of this crop (Figure 2), and the NDVI 19!
for alfalfa fields remained relatively constant through the time series (Figure 2). 20!
21!
Image classification 22!
The field data collected during 2007 were employed to classify crop types from the 23!
!
12!
2007 Landsat imagery. Crops were hierarchically classified as groups (Level II crops) 1!
and individual species (Level III crops, Table 1) using the Random Forest (RF) 2!
classification algorithm (Breiman 2001). The RF algorithm is a classification and 3!
regression tree (CART) technique that has achieved excellent results in classifying 4!
remotely sensed imagery (Lawrence et al. 2006). The RF algorithm develops 5!
classification rules by growing numerous (> 1,000) classification trees from random 6!
subsets of training data, and randomly selects which predictor variables (e.g., percent 7!
reflectance) to use for each decision rule. The correct classifications, or predictions, are 8!
then determined by selecting the most common decision rule at each node within the 9!
group of multiple trees (Breiman 2001, Prasad et al. 2006, Lawrence et al. 2006). As the 10!
algorithm runs, errors are error estimates are calculated with the training data not used in 11!
the random selection process (approximately 37% of the training data). After the entire 12!
forest of classification trees is grown, error rates and accuracies from every tree in the 13!
forest are averaged to calculate overall errors and accuracies; a process that is somewhat 14!
analogous to conducting a cross-validation to estimate error and accuracy (Cutler et al. 15!
2007). The algorithm produces precise, unbiased decision rules, and does not overfit the 16!
training data (Breiman 2001). Because RF’s bootstrapped error estimates are robust and 17!
unbiased, researchers have recommended that it is unnecessary to estimate error from an 18!
independent dataset; thus, all ground reference data can be used to develop classification 19!
rules (Breiman 2001, Lawrence et al. 2006, Cutler et al. 2007, Evans and Cushman, 2009, 20!
Falkowski et al., 2009; Murphy et al., 2010). 21!
Following an accuracy assessment (described below), the RF classification tree 22!
ensemble was applied to retrospectively classify crop types in agricultural fields present 23!
!
13!
across the study site from late March to early April in 2003, 2004, 2005, 2006, and 2007 1!
(Figure 3). The results from the image classification generated 5 separate GIS polygon 2!
layers for each year in the time series (Figure 3). Retrospectively classifying agricultural 3!
crops prior to 2007 assumes that the types of crops cultivated in the Imperial Valley from 4!
2003-2006 were not drastically different than the crops present in 2007. In addition any 5!
future classifications will only be valid if the types of crops in cultivation remain 6!
consistent. Given state-level agricultural statistics for California this seems to be a 7!
reasonable assumption. Although the total area dedicated to the cultivation of specific 8!
crops varies slightly, according to the most recent agriculture crop statistics, the types of 9!
crops in cultivation do not vary considerably (CDFA, 2010). 10!
11!
Accuracy assessment 12!
Two separate methods were employed to assess accuracy of the level III crop 13!
classification: 1) with all of the ground reference data and 2) with a random selection of 14!
25% of the ground reference data held back for an independent accuracy assessment. 15!
Standard error matrices were used to calculate the overall accuracy as well as users and 16!
producers accuracies (and errors of omission and commission) for each individual crop 17!
type in the level II and level III classifications. We also used kappa statistics (KHAT; 18!
Cohen, 1960) to determine if our classifications were significantly better than that 19!
expected by chance (Congalton and Green 1999). Although we acknowledge that Kappa 20!
statistics are often poor indicators of overall accuracy (Pontius, 2000; Foody, 2002), since 21!
they are still a commonly used statistics for assessing the classification of subjects into 22!
categorical groupings, we present the kappa statistics along with the full errors matrices 23!
!
14!
for the Level II and Level III classifications. 1!
Results 2!
National Cropland Data Layer Accuracy 3!
As previously mentioned, the 2007 CDL layer did not contain many of the crop 4!
categories that were observed in the field during the 2007 data collection campaign. 5!
Specifically, many of the grasses (e.g., Sudan grass, Bermuda grass, Kline grass) that 6!
may be of importance to the burrowing owl (Rosenberg and Haley 2004) are absent from 7!
the 2007 CDL data layer. Furthermore, many of the detailed crop categories (e.g., 8!
artichoke, bamboo, bell pepper, cantaloupe, carrot, sugar beat, among others) are also 9!
absent from the CDL layer. The re-categorized level II CDL crop classification had an 10!
overall accuracy of 58.82%. Users accuracies were 27.85% for bare ground/fallow 11!
category, 80.5% for the broadleaf crop category, and 54.65% for the grass category, 12!
while producers accuracies were 73.33% for bare ground/fallow category, 51.28% for the 13!
broadleaf crop category, and 67.31% for the grass category. The kappa statistic is 0.335, 14!
which suggests only a fair agreement between the re-categorized CDL layer and the 15!
independent field observations. The full error matrix and associated accuracy statistics 16!
are presented in Table 4. 17!
Crop Classification Accuracy 18!
Preliminary results from the 2 methods of assessing error (RF estimates based upon 19!
the full training dataset and 25% of the ground reference data held back for an 20!
independent assessment) were similar (Level III species-level classification error was 21!
68.6% and 69.05% for reduced and full datasets, respectively). Based on these 22!
comparable results and the conclusions of Breiman (2001), Lawrence et al. (2006), and 23!
!
15!
Cutler et al. (2007) we only present the results form the RF classifications the utilized the 1!
entire ground reference dataset. 2!
The level II crop classification had an overall accuracy of 84%, with users 3!
accuracies of bare ground = 66.76%, grass crop = 81.41%, and broad leaf/other crop = 4!
88% (Table 1). The level III classification, classified 31 individual crops, and had an 5!
overall accuracy of 69.05%, with individual class accuracies ranging from 0-100% 6!
(Table 1). There was a good degree of agreement beyond chance alone (kappa statistics = 7!
0.71 and 0.67) for level II and level III classifications, respectively. A full error matrix 8!
for the level II crop classification is presented in Table 5. 9!
In the level III classification, five crops (artichoke, asparagus, cauliflower, 10!
cilantro, and parsley) were classified with 100% users accuracies, while four (alfalfa, 11!
sugar beets, watermelon, and triticale) were classified with 80-99% user accuracy, and 12!
six others (corn, potato, radish, Bermuda grass, Sudan grass, and wheat) with 70-80%. 13!
Bell pepper, broccoli, carrot, kline grass, lettuce, mixed flowers all had user accuracies 14!
>50%, while the cabbage, citrus, sugar cane, and wild oat classes each had user 15!
accuracies of 0%. A full error matrix for the level III crop classification is presented in 16!
Table 6. 17!
According to the level III classification, the total land area each crop occupied in 18!
the study area varied annually (Table 7). Alfalfa was the most common crop in 19!
cultivation, occupying about 45,000 – 70,000 ha of the HCP between March and April 20!
from 2003 to 2007 (Table 7; Figure 4). The amount of bare ground, alfalfa, Sudan grass, 21!
Bermuda grass, and wheat, classes of of potential interest to burrowing owls (Rosenberg 22!
and Haley 2004), also occupied large tracts of land between 2003 and 2007. These 23!
!
16!
categories occupy approximately 71-76% of the total land area between 2003 and 2007. 1!
The distribution of these crops also shifted through the 2003-2007 time frame, with some 2!
parcels having crops of potential interest to the burrowing owl cultivated in numerous 3!
years, while others had little or no cultivation of these crops (Figure 5). Some of the least 4!
common crops in cultivation included potato, radish, bell pepper, and watermelon, among 5!
others, occupying approximately 12-16 % of the total land area between 2003 and 2007. 6!
7!
Discussion 8!
The overall goal of this project was to use remotely sensed data to classify crop 9!
assemblages and individual crop types to support future assessments of their importance 10!
to burrowing owls across the Imperial Valley of California. We employed an parcel based 11!
RF classification algorithm to classify 31 different crop types with a 3-date time series of 12!
Landsat data (March – April), and achieved reasonably high accuracy, as compared to 13!
similar studies, especially given the large number of crop categories classified. 14!
Specifically, many of the level III individual crop categories had user accuracies >75%; 15!
and overall, the classification approach achieved accuracies equal to or higher than 16!
similar studies. For example, Akbari et al. (2006) classified Landsat imagery into 20 17!
different crop categories and attained an accuracy of 62%, while Belward and Hoyos 18!
(1987) classified Landsat imagery into 8 different crop categories and attained an overall 19!
accuracy of 64.8%. We also note that the crop Level II classification developed herein 20!
had a much higher accuracy than the CDL crop data layer generated by the National 21!
Agriculture Statistics Service. Furthermore, the independent assessment of the CDL layer 22!
demonstrated that the product had an overall accuracy of only 58.82% and for the Level 23!
!
17!
II crop categories. This number is significantly lower than the 97.22% (kappa = 0.966) 1!
state-level accuracy reported for the 2007 California CDL layer (USDA NASS 2007). 2!
These results suggest that users of the CDL data layer should conduct an independent 3!
accuracy assessment for their specific study areas prior to using such products. 4!
Lower numbers of crop categories can yield higher accuracies, as seen in our 5!
classification level II, but such groupings may not yield enough detail for relating crops 6!
to burrowing owl abundance in localized areas within the HCP. In fact, other studies have 7!
reported slightly higher accuracies (e.g., Janssen 1992 and Turker and Arikan 2005 8!
attained 81% accuracy), but this is likely due to classifying only 7 to 11 crop types, 9!
respectively. Given the relatively high number of crops classified in this study, which is 10!
important in order to assess their relative importance to burrowing owls, the accuracy is 11!
reasonably high. 12!
The approximate 20% decline in overall accuracy between level II and III 13!
classifications was expected given the increased number of classes between the 2 14!
categories. The GIS polygon layers produced by the level III classification predicted 15!
individual crop types with reasonable accuracy, given the 69.05% overall accuracy, and 16!
should be suitable for relating individual crops to burrowing owl abundance. 17!
Of particular interest are the accuracies of bare ground (level II = 66.67%, level III = 18!
41.17%) and grass crops (level II = 81.41%, level III: alfalfa=88.46%, Bermuda 19!
grass=70%, Sudan grass=76.09%, triticale=80%, and wheat=77.78%) because the results 20!
from one study suggest that owls may use bare ground more than other cover types near 21!
nests and hay crops more than other types at distances away from nests (Rosenberg and 22!
Haley 2004). The high levels of accuracy for those crops from these analyses could 23!
!
18!
provide reliable predictors of owl demographic and space use parameters in agro-1!
ecosystems. 2!
Acknowledgments 3!
This study was funded in part by the Imperial Irrigation District, with oversight provided 4!
by the Imperial Irrigation District’s Habitat Conservation Plan Implementation Team. We 5!
thank W. Leimgruber for contributing expertise in crop identification and S. Borrego and 6!
L. Macfarland in collection of ground reference data. 7!
8!
9!
10!
11!
12!
13!
14!
15!
16!
17!
18!
19!
20!
21!
22!
23!
!
19!
Literature Cited 1!
Akbari, M., Mamanpoush, A., Gieske, A., Miranzadeh, M., Torabi, M., and Salemi. H.R. 2!
2006. Crop and landcover classification in Iran using Landsat 7 imagery. International 3!
Journal of remote Sensing, Vol. 27, pp. 4117-4135. 4!
5!
Bailey, R.G. 1994. Ecological classification for the United States. USDA Forest Service, 6!
Washington DC. 7!
8!
Bauer, M.E. 1985. Spectral inputs to crop identification and condition assessment. 9!
Proceedings of the IEEE, Vol. 73, pp. 1071-1085 10!
11!
Beck, R., and Gessler, P.E. 2008. Development of a Landsat Time-Series for the Inland 12!
Northwest and an Application in Forest Status Assessment. Western Journal of Applied 13!
Forestry. Western Journal of Applied Forestry, Vol.23, pp. 1:53-60. 14!
15!
Belward, A.S., and de Hoyos, A. 1987. A comparison of supervised maximum likelihood 16!
and decision tree classification for crop cover estimation from multitemporal LANDSAT 17!
MSS data. International Journal of Remote Sensing, Vol. 8, pp. 229-235. 18!
19!
Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen, M. 2004. Multi-20!
resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready 21!
information. ISPRS Journal of Photogrammetry and Remote Sensing, , Vol. 58, pp. 239-22!
!
20!
258. 1!
2!
Blaschke, T. 2010. Object based image analysis for remote sensing. ISPRS International 3!
Journal of Photogrammetry and Remote Sensing, Vol. 65, ppp. 2-16. 4!
5!
Breiman, L., 2001. Random Forests, Machine Learning, Vol. 45, pp. 5–32. 6!
7!
Burnett, C., and Blaschke, T. 2003. A multi-scale segmentation/object relationship 8!
modelling methodology for landscape analysis. Ecological Modelling, Vol. 168, pp. 233–9!
249. 10!
11!
Carlson, C. A. 1985. Wildlife and agriculture: can they coexist? Journal of Soil and 12!
Water Conservation, Vol. 40, pp. 263–266. 13!
14!
CDFA 2010. California Agricultural Production Statistics 2009-2010. California 15!
Department of Food and Agriculture. Available from <!http://www.cdfa.ca.gov/statistics/ 16!
> (accessed 23 November 2010). 17!
18!
Chander, G., and Markham, B. 2003. Revised Landsat-5 TM radiometric calibration 19!
procedure and postcalibration dynamic ranges. IEEE Transaction on Geosciences and 20!
Remote Sensing, Vol. 41, pp. 2647-2677. 21!
22!
Chavez, P.S., jr. 1996. Image-based atmospheric corrections -Revisited and Improved. 23!
Photogrammetric Engineering and Remote Sensing, Vol. 62, pp. 1025-1036. 24!
!
21!
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educational and 1!
Psychological Measures, Vol.20, pp. 37–40. 2!
3!
Congalton, R.G., and Green, K. 1999. Assessing the Accuracy of Remotely Sensed Data. 4!
Lewis Publishers, New York, pp. 1–137. 5!
6!
Craig, M.E. 2001. The NASS cropland data layer program. In: Proceedings, third 7!
international conference on geospatial information in agriculture and forestry, Denver, 8!
Colorado, 5-7 November 2001. 9!
10!
Cutler, R. D., Edwards, T. C., Jr., Beard, K. H., Cutler, A., Hess, K. T., Gibson, J., and 11!
Lawler, J. 2007. Random Forests for classification in ecology. Ecology, Vol. 88, pp. 12!
2783−2792. 13!
14!
Desante, D.F., E.D. Ruhlen, E.D. 2004. Density and abundance of burrowing owls in the 15!
agricultural matrix of the Imperial Valley, California. Studies in Avian Biology, Vol. 27, 16!
pp. 116-119. 17!
18!
Desclee, B., Bogaert, P., and Defourny, P. 2006. Forest change detection by statistical 19!
object-based method. Remote Sensing of Environment, Vol. 102, pp. 1-11. 20!
21!
Evans, J.S., and Cushman, S.A., 2009. Gradient Modeling of Conifer Species Using 22!
Random Forest. Landscape Ecology, Vol. 5, pp. 673-683. 23!
!
22!
1!
Falkowski, M.J., Evans, J.S., Martinuzzi, S., Gessler, P.G., and Hudak, A.T., 2009. 2!
Characterizing forest succession with lidar data: An evaluation for the inland northwest, 3!
USA. Remote Sensing of Environment. Vol. 113, 946-956. 4!
5!
Falkowski, M.J., Gessler, P.E., Morgan, P., Hudak, A.T., and, Smith, A.M.S. 2005. 6!
Characterizing and mapping forest fire fuels using ASTER imagery and gradient 7!
modeling. Forest Ecology and Management, Vol. 217, pp. 129–146. 8!
9!
Fensham, R.J. 2008. Leichhardt’s maps: 100 years of change in vegetation structure in 10!
inland Queensland. Journal of Biogeography, Vol. 35, pp. 141–156. 11!
12!
Foody, G.M. 2002. Status of land cover classification accuracy assessment. Remote 13!
Sensing of Environment, Vol. 80, pp. 185-201. 14!
15!
Haack, B. 1987. An assessment of Landsat MSS and TM data for urban and near-urban 16!
land-cover digital classification Source. Remote Sensing of Environment, Vol. 21, pp. 17!
201–213. 18!
19!
Haug, E.A., Millsap, B.A., and, Martell, M.S. 1993. Burrowing owl (Speotyto 20!
cunicularia). In The Birds of North America, No. 611. Edited by A. Poole and F. Gill. The 21!
Academy of Natural Sciences, Philadelphia, Pennsylvania, USA, and The American 22!
Ornithologists’ Union, Washington, D.C., USA. 23!
!
23!
1!
Hay, G.J., and Castilla, G. 2008. Geographic Object-Based Image Analysis (GEOBIA): 2!
A new name for a new discipline. Object-Based Image Analysis Lecture Notes in 3!
Geoinformation and Cartography, 2008, Section 1, 75-89, DOI: 10.1007/978-3-540-4!
77058-9_4. 5!
6!
Imperial Irrigation District. 2009. 2009 annual report of Imperial Irrigation District 7!
Pursuant to SWRCB Revised Order WRO 2002-013. Imperial, California, USA. 8!
Available from <http://www.iid.com/Media/2009-Report.pdf> (accessed 23 August 9!
2010). 10!
11!
Janssen, L.L.F., and Middelkoop, H. 1992. Knowledge based crop classification of a 12!
Landsat Thematic Mapper image. International Journal of Remote Sensing, Vol. 13, pp. 13!
2872-2837. 14!
15!
Klute, D.S., Ayers, L.W., Green, M.T., Howe, W.H., Jones, S.L., Shaffer, J.A., Sheffield, 16!
S.R., and Zimmerman, T.S. 2003. Status assessment and conservation plan for the 17!
western burrowing owl in the United States. U.S. Department of Interior, Fish and 18!
Wildlife Service, Washington, D.C., USA, Biological Technical Publication FWS/BTP-19!
R6001-2003. 20!
21!
Lawrence, R.L., Wood, S.D., and Sheley, R.L. 2006. Mapping invasive plants using 22!
hyperspectral imagery and Breiman Cutler classifications (RandomForest). Remote 23!
!
24!
Sensing of Environment, Vol. 100, pp. 356-362. 1!
2!
Lu. D., and Weng, Q. 2007. A survey of image classification methods and techniques for 3!
improving classification performance. International Journal of Remote Sensing 4!
Vol. 28, pp. 823–870. 5!
6!
Markham, B.L., and Barker, J.L. 1986. Landsat MSS and TM post-calibration dynamic 7!
ranges, exoatmospheric reflectances and at-satellite temperatures. EOSAT Technical 8!
Notes, Vol. 1, pp. 3-8. 9!
10!
Moulton, C.E., Brady, R.S., and Belthoff, J.R. 2006. Association between wildlife and 11!
agriculture: underlying mechanisms and implications for burrowing owls. Journal of 12!
Wildlife Management, Vol. 70, pp. 708–716. 13!
14!
Mueller, R., and Ozga, M. 2002. Creating a cropland data layer for an entire state. 15!
Proceedings, 2002 American congress on surveying and mapping/American Society for 16!
Photogrammetry and Remote Sensing conference and technology exhibition. Bethesda, 17!
MD: American Society for Photogrammetry and Remote Sensing. Available from 18!
<http://www.nass.usda.gov/Education_and_Outreach/Reseach_Reports/02_01_CreateCD19!
L_State_RWM.pdf> (accessed 23 November 2010). 20!
21!
Murphy M, J.S. Evans, and Storfer, A. 2010. Quantifying Bufo boreas connectivity in 22!
Yellowstone National Park with landscape genetics. Ecology, Vol., pp. 252-261. 23!
!
25!
1!
Oetter, D.R., Cohen, W.B., Berterretche, M., Maiersperger, T.K., and Kennedy, R.E. 2!
2001. Land cover mapping in an agricultural setting using multiseasonal Thematic 3!
Mapper data. Remote Sensing of Environment, Vol. 76, pp. 139-155. 4!
5!
Ortiz, M.J., Formaggio, A.R., and Epiphanio, J.C.N. 1997. Classification of croplands 6!
through integration of remote sensing, GIS, and historical database. International Journal 7!
of Remote Sensing, Vol. 18, pp. 95-105. 8!
9!
Peterjohn, B. G. 2003. Agricultural landscapes: can they support healthy bird populations 10!
as well as farm products? Auk, Vol. 120, pp. 14–19. 11!
12!
Pontius Jr., R.G. 2000. Quantification error versus location error in the comparison of 13!
categorical maps. Photogrammetric Engineering and Remote Sensing Vol. 66, p.p. 1011-14!
1016. 15!
16!
Prasad, A. M., L. R. Iverson, and A. Liaw. 2006. Newer classification and regression tree 17!
techniques: bagging and random forests for ecological prediction. Ecosystems, Vol. 9, pp. 18!
181–199. 19!
20!
Rosenberg, D.K., and Haley, K.L. 2004. The ecology of burrowing owls in the 21!
agroecosystem of the Imperial Valley, California. Studies in Avian Biology, Vol. 27, pp. 22!
120-135. 23!
!
26!
1!
Sauer, J.R., Hines, J.E., and Fallon, J. 2008. The North American breeding bird survey, 2!
results and analysis 1966–2007. Version 5.15.2008. U.S. Geological Survey Patuxent 3!
Wildlife Research Center, Laurel, Maryland, USA., Available from <http://www.mbr-4!
pwrc.usgs.gov/bbs/> (accessed 26 April 2010). 5!
6!
Smith, A.M.S., Falkowski, M.J., Hudak, A.T., Evans, J.S. and Robinson, A.P. 2010. 7!
Comparing field and remote estimates of forest canopy cover. Canadian Journal of 8!
Remote Sensing, In Press 9!
10!
Townshend, J.R.G., Huang, C., Kalluri, S.N.V., Defries, R.S., Liang, S., and Yang, K. 11!
2000. Beware of per-pixel characterization of land cover. International Journal of Remote 12!
Sensing, Vol. 21, pp. 839–843. 13!
14!
Tucker, C.J. 1979. Red and photographic infrared linear combinations for monitoring 15!
vegetation. Remote Sensing of the Environment, Vol. 8, pp. 127-150. 16!
17!
Turker, M., and Arikan, M. 2005. Sequential masking classification of multi-temporal 18!
Landsat7 ETM+ images for field-based crop mapping in Karacabey, Turkey. 19!
International Journal of Remote Sensing, Vol. 26, pp. 3813 – 3830. 20!
21!
USDA NASS, 2007. USDA, National Agricultural Statistics Service, 2007 California 22!
Cropland Data Layer. Available from 23!
!
27!
<http://www.nass.usda.gov/research/Cropland/metadata/metadata_ca07.htm > (accessed 1!
23 November 2010). 2!
3!
USDA NASS, 2009. The USDA/NASS Cropland Data Layer. Available from 4!
<http://www.nass.usda.gov/research/Cropland/Method/cropland.pdf> (accessed 23 5!
November 2010). 6!
7!
Vogelman, J.E., Howard, S.M., Yang, L., Larson, C.R., Wylie, B.K., and Van Driel, N. 8!
2001. Completion of the 1990s National Land Cover Data set for the conterminous 9!
United States for Landsat Thematic Mapper data and ancillary data sources. 10!
Photogrammetric Engineering and Remote Sensing, Vol. 67, pp. 650-655. 11!
12!
Wardlow, B.D., and Egbert, S.L. 2008. Large-area crop mapping using time-series 13!
MODIS 250 m NDVI data: An assessment for the U.S. Central Great Plains. Remote 14!
Sensing of Environment. Vol. 112, pp. 1096-1116. 15!
16!
Wellicome, T.I., and Holroyd, G.L. 2001. The second international burrowing owl 17!
symposium: background and context. Journal of Raptor Research, Vol. 35, pp. 269-273. 18!
19!
Williams, D., 2009. Landsat-7 Science Data User's Handbook [online] Available from < 20!
http://landsathandbook.gsfc.nasa.gov/handbook.html > [cited Aug. 8th 2010]. 21!
22!
!
28!
Witt, G.B., Luly, J. and Fairfax, R.J. 2006. How the west was once: vegetation change in 1!
south-west Queensland from 1930 to 1995. Journal of Biogeography, Vol.33, pp. 1585–2!
1596. 3!
4!
!
29!
Table 1. Crop classes and associated class accuracies. Accuracies for the level II and Level III classifications are 84.05%and 68.81%, 1!
respectively. 2!
Level I
Level II
Users Accuracy
(%)
Producers Accuracy
(%)
Level III
Sample
Size
Users Accuracy
(%)
Producers Accuracy
(%)
Crops
Broad Leaf / Other
88.00
0.86
Alfalfa
52
88.46
74.19
Artichoke
4
100.00
80.00
Asparagus
4
100.00
100.00
Bamboo
4
50.00
100.00
Bell Pepper
5
40.00
100.00
Broccoli
9
44.44
44.44
Cabbage
4
0.00
0.00
Cantaloupe
13
61.54
61.54
Carrot
7
42.85
100.00
Cauliflower
4
100.00
80.00
Cilantro
4
100.00
100.00
Citrus
3
0.00
0.00
Corn
27
74.07
46.50
Cotton
6
66.67
66.67
Lettuce
7
42.85
75.00
Mixed
Flowers
4
25.00
50.00
Onion
23
69.56
94.12
Parsley
4
100.00
80.00
Potato
4
75.00
100.00
Radish
4
75.00
100.00
Sugar Beets
33
93.94
75.60
Sugar Cane
4
0.00
0.00
Watermelon
5
80.00
100.00
Grass Crops
81.41
0.85
Bermuda
Grass
50
70.00
70.00
!
30!
Fox Tail Grass
3
0.00
0.00
Kline Grass
11
36.36
80.00
Sudan Grass
46
76.09
63.60
Triticale
4
80.00
100.00
Wheat
36
77.78
80.00
Wild Oat
6
0.00
0.00
Bare Ground /
Fallow
66.67
0.65
Bare Soil
17
41.17
41.17
Fallow Field
13
53.85
50.00
!
31!
Table 2. Original and Re-categorized CDL classes 1!
Original CDL Class
Re-categorized CDL
Class
Alfalfa
Broadleaf
Corn
Broadleaf
Fallow
Bare/Fallow
Grass
Grass
Grass/Pasture/Non-Ag
Grass
Herbs
Broadleaf
NLCD - Shrubland
Bare/Fallow
Onion
Broadleaf
Other Hays
Grass
Sugarbeets
Grass
Wheat
Grass
Wild oats
Grass
2!
!
32!
1!
2!
3!
Table 3. Acquisition dates and sensor of Landsat data used. 4!
Year
Acquisition
Date
Sensor
2007
28-Mar
Landsat 5
13-Apr
Landsat 5
29-Apr
Landsat 5
2006
1-Mar
Landsat 7
10-Apr
Landsat 5
26-Apr
Landsat 5
2005
30-Mar
Landsat 7
7-Apr
Landsat 5
15-Apr
Landsat 7
2004
27-Mar
Landsat 7
12-Apr
Landsat 7
20-Apr
Landsat 5
2003
25-Mar
Landsat 7
10-Apr
Landsat 7
26-Apr
Landsat 7
5!
!
33!
Table 4. Full error matrix and associated accuracy statistics for the CDL classification 1!
Crop Type
Bare/Fallow
Broadleaf
Grass
Users Accuracy
Bare/Fallow
22
32
25
27.85%
Broadleaf
3
120
26
80.54%
Grass
5
82
105
54.69%
Producers Accuracy
73.33%
51.28%
67.31%
2!
!
34!
1!
Table 5. Full error matrix and associated accuracy statistics for the level II crop 2!
Crop Type
Broadleaf / Other
Grass
Bare Ground / Fallow
Users Accuracy
Broadleaf / Other
206
20
8
88.03%
Grass
26
127
3
81.41%
Bare Ground / Fallow
7
3
20
66.67%
Producers Accuracy
86.19%
84.67%
64.52%
!
35!
Table 6. Full error matrix and associated accuracy statistics for the level III crop 1!
2!
3!
!
36!
1!
!
37!
1!
Table 7. Total area (hectares) in Level III crops on a yearly basis. 2!
Crop
2003
2004
2005
2006
2007
Alfalfa
70,046
61,656
58,968
45,084
54,019
Artichoke
1,446
956
411
3,861
836
Asparagus
19
32
19
4
81
Bamboo
20
101
172
60
174
Bell Pepper
3
2
0
0
33
Broccoli
505
668
577
1,040
589
Cabbage
112
8
63
106
113
Cantaloupe
2,754
1,677
4,333
1,423
3,132
Carrot
1,745
2,147
1,395
2,398
1,305
Cauliflower
15
8
15
6
105
Cilantro
258
168
170
19
53
Citrus
161
0
108
108
242
Corn
7,831
7,733
10,736
12,033
7,794
Cotton
76
31
767
24
602
Lettuce
295
559
403
1,635
1,092
Mixed
Flowers
18
4
23
0
25
Onion
6,228
5,777
4,998
4,549
4,642
Parsley
5
177
100
88
133
Potato
0
0
0
14
23
Radish
0
3
1
0
43
Sugar Beets
13,345
16,771
10,273
9,061
9,450
Sugar Cane
70
226
238
201
117
Watermelon
109
109
0
0
80
Bermuda
Grass
17,776
24,178
29,767
22,319
30,154
Sudan Grass
13,929
18,127
11,647
41,136
20,608
Kline Grass
1,188
1,663
270
848
2,941
Triticale
0
9
0
31
226
Wheat
25,132
18,142
16,878
12,096
18,509
Wild Oat
572
52
168
647
1600
Bare Soil
5,744
8,089
19,234
6,470
9,282
Fallow Field
10,166
10,492
7,854
14,306
11,506
3!
4!
5!
!
38!
1!
2!
Figure 1. Distributional statistic GIS layer color coded by the mean (A), maximum (B) 3!
and standard deviation (C) of a single Landsat band for a portion of the Imperial Valley 4!
Study area. 5!
6!
Figure 2. Phenologic trajectories for three selected crops derived from the OBIA based 7!
on common land units in the Imperial Valley agro-ecosystem, USA, 2007. 8!
9!
Figure 3. Agricultural crop classification in the Imperial Valley agro-ecosystem, USA in 10!
(a) 2003, (b) 2004, (c) 2005, (d) 2006, and (e) 2007. White areas represent portions of the 11!
study are that are not in cultivation (i.e., waterways, urban areas, impervious surfaces, 12!
among others). 13!
14!
Figure 4. Annual change in area of hay crops and bare ground available to burrowing 15!
owls in the Imperial Valley agro-ecosystem, USA, 2003-2007. 16!
17!
Figure 5. Spatial change assessment of crop classes of potential interest to the burrowing 18!
owl (grass and bare ground classes) in the Imperial Valley agro-ecosystem, USA from 19!
2003-2007. The legend depicts the number of years corps of potential interest were 20!
cultivated in a particular field. 21!
