PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2014
81
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2014
81
Abstract
Increasing pressure to feed the growing population with
scarce water resources requires accurate and routine cropland
mapping. This paper develops and implements a rule-based
automated cropland classification algorithm (
ACCA
) using
multi-sensor remote sensing data. Pixel-by-pixel accuracy
assessments showed that
ACCA
produced an overall accuracy
of
≥
96 percent (K
hat
= 0.8) when tested using independent
data layers. Furthermore,
ACCA
-generated county cropland
areas showed high agreement (R-square values
≥
0.94) when
compared with three independent data sources: (a) US
Department of Agriculture (
USDA
) cropland data layer derived
cropland areas, (b) county specific crop acreage data from the
Farm Service Agency, and (c) the Census of Agriculture data
for the 58 counties in California. Our results demonstrate the
ability of
ACCA
to generate cropland extent and areas over
space and time, in an automated fashion with high degree of
accuracies year after year, greatly contributing to food and
water security analysis and decision making.
Introduction
The world population tripled in the Twentieth century, even-
tually reaching seven billion in 2011. The world population
is projected to increase by another 50 percent within the next
fifty years as per United Nations estimates. Rapid popula-
tion growth, coupled with urbanization and industrializa-
tion, will inevitably result in increasing pressure on food and
water supply. Meanwhile, recent climate change models have
projected that more extreme events such as severe droughts
(IPCC, 2007) will occur more frequently in the future, putting
higher risks on water scarcity, and food production. We are
facing a greater challenge of food security than ever before,
and the first step to ensure a secure food supply is to routinely
map cropland, i.e., a land cover type subjected to most rapid
changes over continents such as North America and Asia
(Lepers
et al.
, 2005), from which crop growth status and water
consumption can be monitored.
Previously, cropland mapping has been conducted across
multiple spatial scales using various methods (Ozdogan and
Woodcock, 2006; Wardlow
et al.
, 2007; Wardlow and Egbert,
Zhuoting Wu is with the Western Geographic Science Center,
US Geological Survey, Flagstaff, AZ 86001, and the Merriam-
Powell Center for Environmental Research, Northern Arizona
University, Flagstaff, AZ 86001 (
).
Prasad S. Thenkabail is with Western Geographic Science
Center, US Geological Survey, Flagstaff, AZ 86001.
James P. Verdin is with the USGS EROS Data Center, Sioux
Falls, SD 57198.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 1, January 2014, pp. 81–90.
0099-1112/14/8001–81/$3.00/0
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.1.81
2008; Thenkabail
et al.
, 2009a and 2009b; Dheeravath, 2010;
Thenkabail
et al.
, 2010; Gumma
et al.
, 2011; Thenkabail
et al.
,
2011; Thenkabail and Wu, 2012). Remote sensing data serve
as vital sources in delivering accurate and timely informa-
tion on the area, condition, and major crop types across
the globe, given its implication in food security, water use
assessment, land management, trade decisions, policy, and
economics. Many classification methods have been used in
cropland mapping, including maximum likelihood classifica-
tion (EL-Magd and Tanton, 2003), decision tree classification
(Morton
et al.
, 2006; Wardlow and Egbert, 2008; Biradar
et
al.
, 2009; Pittman
et al.
, 2010), neural network methods (Liu
et al.
, 2005; Atzberger and Rembold, 2010), support vector
machine (Mathur and Foody, 2008), spectral unmixing tech-
niques (Lobell and Asner, 2004; Yang
et al.
, 2007; Chen
et al.
,
2008), spectral matching techniques (Thenkabail
et al.
, 2007),
and spectral angle mapper (Rembold and Maselli, 2006). Most
of the existing cropland mapping methods rely extensively
on the human interpretation of spectral signatures, making
the process labor-intensive and difficult to repeat over time
and space.
In addition, due to the low temporal availability of high-
quality and high-resolution imagery, accuracies of time series-
based cropland classification is limited by coarse spatial
resolution imagery (Wardlow and Egbert, 2008). Nevertheless,
dense time series of remote sensing data have been proved
to be critical in determining crop types and intensity (Lobell
and Asner, 2004; Biggs
et al.
, 2006; Wardlow
et al.
, 2007;
Thenkabail
et al.
, 2009a and 2009b; Velpuri
et al.
, 2009;
Dheeravath
et al.
, 2010; Lv and Liu, 2010; Shao
et al.
, 2010;
Biradar and Xiao, 2011; Thenkabail
et al.
, 2011). The combi-
nation of high temporal resolution of
MODIS
and high spatial
resolution of Landsat imagery allows for producing 30 m
resolution cropland maps. Multiple-resolution data fusion has
been used successfully in cropland mapping to improve clas-
sification accuracy (Watts
et al.
, 2011).
Some progress has been made recently in automated
land-cover classification techniques such as an unsupervised
algorithm called independent component analysis (Ozdogan,
2010) and a modified subspace classification method (Bagan
and Yamagata, 2010). An automatic rule-based decision tree
classifier such as
ACCA
, as proposed here, can take advantage
of multiple data sources with various resolutions and per-
form supervised classification of cropland extent/area and
Automated Cropland Classification Algorithm
(ACCA) for California Using Multi-sensor
Remote Sensing
Zhuoting Wu, Prasad S. Thenkabail, and James P. Verdin
