Delineating Direct Seeded Rice (DSR) Cultivation Using SAR Data
The
RISAT-1
data was processed using ENVI software. The lev-
el-2,
HH
polarization data was imported in ENVI raster format
(.hdr). To reduce noise, Enhanced Lee Adaptive Speckle Filter
with kernel size of 5 × 5 was applied (Lopes
et al
., 1990).
Radiometric calibration for
HH
polarization amplitude image
was carried out by using the metadata provided along with the
imagery. The scaled values of digital numbers were converted
into backscattering coefficient using the equation given in
(Chakraborty
et al
., 2013; Laur
et al
., 2002). The temporal im-
ages were co-registered by fitting a second order polynomial
model to the manual
GCP
s with 50 well distributed points
throughout the scene. The temporal
HH
polarization
RISAT-1
data was stacked. The rice crop delineated from Landsat-8 was
used to subset the
RISAT-1
data stack for further processing
.
In Raichur, direct seeded rice cropping is practiced along with
transplanted rice which is most prevalent. Transplanted rice is
of two types depending on sowing date. The early transplanted
rice is sown in August and the late transplanted rice in Septem-
ber.
SAR
data is unique since it can distinguish between different
rice sowing dates. The rice crop generates a unique temporal
backscatter profile. It is clear from Figure 3 that during the estab-
lishment stage of transplanted rice, the backscatter is higher than
direct seeded rice due to tilling and low moisture. During trans-
plantation, backscatter is quite low due to standing water which
reflects minimal energy in the backscatter direction. Further,
the backscatter increases as the plant grows because of multiple
reflections from the crop canopy (Choudhury
et al
., 2012).
In the establishment phase of direct seeded rice, low
backscatter was observed due to moisture conditions. This
is the key to distinguishing 80 percent of the direct seeded
rice from transplanted fields. Also, some of the transplanted
rice fields cannot be distinguished from direct seeded fields
due to high standard deviation in transplanted fields because
of roughness. Additionally, early transplanted rice could be
distinguished at the transplantation stage due to very low
backscatter. However, it is difficult to discriminate between
direct seeded rice and late transplantation due to their similar
growth stages. The deviation of backscatter was found to be
low. The decision tree hybrid classification algorithm was
used here in order to separate the above mixed classes. June
is the crucial month to separate transplanted rice form direct
seeded rice (80 percent). The early transplanted rice was eas-
ily separated from directed seeded rice by applying decision
rules on the August imagery. This was possible because the
difference in backscatter was more than 2
dB
. The late trans-
plantation was not separable from direct seeded rice because
of the low difference in backscatter. In order to resolve these
types of fields, unsupervised k-means algorithm with a con-
vergence threshold of 0.99 was applied on the remaining area
Figure 3. Temporal changes in backscatter coefficient in different
rice growing practices.
Plate 1. Land-use/land-cover during the 2014 monsoon season (
kharif
) in Raichur District, with rice classes temporal NDVI profiles.
876
November 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
