ABSTRACT:
coefficient of 0.76.
1. INTRODUCTION
Agriculture has been a key driver of development in Southeast
Asia countries (Evenson and Rosegrant., 2003; Timmer, 2009).
More than 60% of the labor force in Cambodia, Laos, Myanmar,
and Vietnam depends on agriculture. Rice agriculture in these
countries is still the most dominant agricultural activity taking up
a large part of the harvested area (50-80%) and value of
production (FAOSTAT, 2009). Thailand and Vietnam by far are
the two leading rice suppliers in the world (FAO, 2010). Rice is
the most important staple food for more than 50% of the world's
population with more than 20% of their daily calories (Maclean
et al., 2002). As the world population continues to grow steadily,
while land and water resources are declining. Moreover, climatic
change through global warming has also been a key factor
causing declined rice production (Furuya and Kobayashi, 2009;
Matthews and Wassmann, 2003). Efforts to balance rice
production to meet the food demands of a growing population are
vitally important. Thus, accurate estimates of rice growing areas
are needed to estimate rice production.
Conventional methods of acquiring these data at a regional
scale reveal problems due to the costs of field surveys for
complex farming systems throughout the year. Low-resolution
remote sensing has been proven as an indispensable tool for
providing data for this monitoring purpose at regional and global
scales because the data has advantages of high temporal
resolution and wide coverage. However, low-resolution
remotely-sensed data are often obscured by cloud cover.
Filtering such noise from the data is usually done prior to the
classification. A number of noise reduction algorithms (e.g.,
Fourier transform, wavelet transform, empirical mode
decomposition, and local maximum fitting) have been developed
and commonly used for filtering noise from time series of
satellite vegetation indices. This study used wavelet transform
for noise filtering of time-series NDVI data. As NDVI data are
nonlinear and traditional parametric classification algorithms
based on spectral bands, such as maximum likelihoods, are
insufficient to delineate seasonal farming activities,
non-parametric mapping methods (e.g., artificial neural networks
* Corresponding author.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
MAPPING MAJOR CROPPING PATTERNS IN SOUTHEAST ASIA FROM MODIS DATA
USING WAVELET TRANSFORM AND ARTIFICIAL NEURAL NETWORKS
N.T. Son **, C.F. Chen ?, C.R. Cru?
“Center for Space and Remote Sensing Research — (ntson, cfchen)@csrsr.ncu.edu.tw
d Department of Civil Engineering — ccruncu@gmail.com
National Central University, Jhongli, Taiwan
Commission III, WG III/5
KEY Words: MODIS data, Croplands, Wavelet transform, Artificial neural networks (ANNSs), Southeast Asia.
Agriculture is one of the most important sectors in the economy of Southeast Asia countries, especially Thailand and Vietnam. These
two countries have been the largest rice suppliers in the world and played a critical role in global food security. Yearly rice crop
monitoring to provide policymakers with information on rice growing areas is thus important to timely devise plans to ensure food
security. This study aimed to develop an approach for regional mapping of cropping patterns from time-series MODIS data. Data were
processed through three steps: (1) noise filtering of time-series MODIS NDVI data with wavelet transform, (2) image classification of
cropping patterns using artificial neural networks (ANNs), and (3) classification accuracy assessment using ground reference data. The
results by a comparison between classification map and ground reference data indicated the overall accuracy of 80.3% and Kappa
— ANNE, support vector machines — SVMs) have been proven to
be sufficient to handle complex classification tasks. In this work,
we applied ANNs for mapping cropping patterns in the study
area. This method is developed based on statistical learning
theory (Foody and Mathur, 2004; Haykin, 1994).
The main objective of this study is to develop an approach
for mapping major cropping patterns in Southeast Asia from
MODIS time-series NDVI data using wavelet transform and
ANNE.
2. STUDY AREA
The study area includes four Southeast Asian countries: Vietnam,
Thailand, Laos and Cambodia, lying between 5.62-23.45 N and
97.34-109.51 E (Figure 1). The total area is approximately
1,081,130 km”, in which agricultural land occupies
approximately 23% (Stibig et al., 2004). Rice was a main crop
commonly practiced in plain areas. Field crops such as sugarcane,
cassava, and maize occupied the uplands. Monitoring rice
growing areas becomes an important activity due to the official
initiatives to ensure food supply and security.
Rice crops in the study area are classified according to their
periods of cultivation (cultivating seasons). Basically, there are
three types of rice cropping systems: single-cropped rice,
double-cropped rice, triple-cropped rice. Single cropping system
used long-term rice varieties (160-180 days) was often planted
under predominantly rain-fed conditions, whilst double and triple
rice cropping systems used short-term varieties (90-100 days) are
commonly practiced in the Vietnamese Mekong River Delta
(MRD), Red River Delta (RRD), and Chao Phraya River Delta
(CRD) of Thailand.
As rice area was practiced in the lowlands and our study
focused on investigating rice agriculture, mountainous areas
where the elevation was higher than 500 m were masked out
using the shuttle radar topography mission (SRTM) 90 m digital
elevation model.
24
—L
20
L
Figur
(grou
MOI
grid
were
band
quali
for s
(Ver
Basi
(Sub
2000
prep:
The
ever
8-da
ND
com
miti;
For:
blue
et al
fron
WaVi
time
