XXIX-B3, 2012
MODIS DATA
VORKS
and Vietnam. These
ty. Yearly rice crop
plans to ensure food
DIS data. Data were
iage classification of
] reference data. The
f 80.3% and Kappa
have been proven to
n tasks. In this work,
atterns in the study
. Statistical learning
294).
levelop an approach
outheast Asia from
velet transform and
| countries: Vietnam,
n 5.62-23.45 N and
a is approximately
] land occupies
ce was a main crop
)s such as sugarcane,
s. Monitoring rice
y due to the official
d according to their
Basically, there are
ingle-cropped rice,
ole cropping system
) was often planted
Ist double and triple
ss (90-100 days) are
ekong River Delta
Phraya River Delta
ands and our study
mountainous areas
1 were masked out
J RTM) 90 m digital
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
98 402 406 410
L 1 1
24
ZA
T
2
12
4 Chao Phraya
River Delta y
d ROAST. Pe-^ Mekohg Riv
vo) M amos g River
a T : Delta
Sa
z
wo} 3 1 $ I- o
d^; s
: X 0 100 200 400 Kitometers
o: ——— A
$n
a 8" RN ^ :
E Areas higher 500 m
P^ Mekong River
KE Sampling areas
T T T T
98 102 106 110
Figure 1. Map of the study area showing the sampling areas
(ground reference data) used for extracting training data used in
ANNs and evaluating accuracy classification results.
3. DATA
MODIS/Terra surface reflectance 8-day L3 global 500 m SIN
grid v005 (MODO9A1 product) acquired from NASA for 2010
were used in this study. This data product has seven spectral
bands. The spatial resolution is 500 m x 500 m. The data include
quality control lags for image artifacts. It has also been validated
for stage 2 and is ready for use in scientific publications
(Vermote et al., 2008). Other data including the Lower Mekong
Basin land-use map (MRC, 2010), the 2000 VMD land-use map
(Sub-NIAPP, 2002), and the 2000 Thailand land-use map (LDD,
2000) were also collected and used for field investigation and
preparation of ground reference data.
4. METHODS
4.1 Time-series NDVI data noise filtering
The time-series NDVI data were created by first calculating
every 8-day MODIS scene. They were then stacked into one
8-day composite scene with 46 bands. Time-series MODIS
NDVI data are often obscured by noise due to the cloud cover
commonly seen in the tropical climate. Thus, it is essential to
mitigate such noise before the data can be used for classification.
For this reason, we first masked out thick cloud cover using the
blue band, where its reflectance value was greater than 0.2 (Xiao
et al., 2006). The missing values were replaced with new values
from the time-series profile using linear interpolation. The
wavelet transform was then applied to filter noise from the
time-series NDVI data. The wavelet transform W(s,t) of a signal
x(t) is defined as follows:
W(s, 1) = S x(t)w (=) dt, (1)
where s > 0 and x € R, x(t) is the analyzed input signal; V(t) is the
mother wavelet; and s and c are scaling and translation
parameters. In this study, we used Coiflet wavelet (order 4)
(Torrence and Compo, 1998). This wavelet function has been
demonstrated to give the best results among Daubechies and
Symlet wavelet functions for determining rice crop phenology
(Sakamoto et al., 2005).
4.2 Image classification
The ANNs back-propagation algorithm was utilized for
classification. This algorithm uses the delta rule of a steepest
descent to adjust weights based on the backward propagation of
errors in the network (Paola and Schowengerdt, 1995). The
ANNs has a learning process where training signatures are
randomly selected and fed to adjust the internal weight matrix.
This is made by a repetition of back-propagations of the answer
into the weight matrix. When the learning process is complete,
the weight matrix is ready to process any new signature from the
imagery dataset. In this study, we designed ANN architecture:
one input layer, one output layer and one hidden layer. The
number of 93 neurons were calculated based on the existing
literature recommendations: 2m + 1 (Atkinson and Curran, 1995),
where n is the number of bands. The commonly-used
back-propagation algorithm using tansig function was utilized to
train the network.
The training samples used to train ANNS for classification
were extracted from the ground reference data prepared after
field investigation. A total of samples of 14,602 (single-cropped
rice: 2,333, double-cropped rice: 3,638, triple-cropped rice:
1,336, field crops: 1,020, forests/orchards/perennial trees: 5,741,
and built-up areas: 534) seems to most of variability of the study
was used. Water bodies were excluded from the analysis. We
identified water bodies and masked them out by analysis of
smooth NDVI and land surface water index (LSWI) profiles
(Xiao et al., 2006). The training samples were divided into three
parts: training samples (50% of the total samples), validation
samples (25% of the total samples), and testing samples (25% of
the total samples).
The network’s performance is measured using a mean
squared error (MSE) and a confusion matrix. The training
process was carried out until the mean squared error (MSE)
reached 0.00044 at the epoch 366. The confusion matrix shows
the percentages of correct classification of 98.8%. There was
significant correlation between testing samples and outputs R=
0.99). The classification results values range from 0 to 1. The
winner-take-all algorithm was used to get the class indices as the
position of the highest element in each output vector of ANNs.
4.3 Accuracy assessment
The classification results were compared with the field data. A
total of 2,000 pixels for each class were randomly extracted from
the ground reference data. The classification accuracy
assessment was performed such that these ground reference
pixels were compared with that of the classified map using the
confusion matrix. Kappa coefficient and other parameters (i.e.,
overall, producer and user accuracies) were used to measure the
classification accuracy.
