ification. To
ised on their
ruary, 1993.
issifier. The
D) of 1 and
vith different
e better with
mation alone
ere obtained
were initially
method. The
on to utilise
t al. (1997),
not easy to
nambiguous.
grey level
the spatial
of an image
ches such as
or statistical
| adopted to
ence matrix
rocedure on
this study
is by far the
to compute
the GLCM
h grey levels
x pixels and
the texture
V. However,
easures viz.,
rd deviation,
relation as
n discussed
as Inverse
neity as it
nces in pair
the presence
p(i,j) is the
equency, the
values of a
presents a
GLCM concentrated around the principal diagonal. This means
that high contrast values imply very coarse texture.
CON 2 X," Xo "* (-jy pj)
1.1.3 Dissimilarity (DSM): Dissimilarity feature akin to
contrast, tells about the heterogeneity of the grey levels. Higher .
values of dissimilarity in GLCM indicate coarser textures.
DSM - Xi.) "*! X4 "*' Abs(-j)p(ij)
1.1.4 Mean (MEAN): Mean is an indicator of the distribution
of grey levels with respect to the central position. Interpretation
of this feature in association with variance will provide textural
information.
MEAN 2 X," X9" ip(ijj)
1.1.5 Standard Deviation (SD): Standard deviation or variance
of GLCM denote dispersion of the grey levels as defined by the
sum of squares. Generally, coarse textured features associate
with higher standard deviations.
SD - sqrt(X;. "*' S p NE G**mean - p(i,j))^
1.1.6 Entropy (ENT): Entropy measures the disorder of an
image. When the image is not texturally uniform, many GLCM
elements have very low values implying Entropy is very large.
Conceptually, homogeneity and entropy are inversely
correlated.
ENT « -X9 "* X. .o "* pij) Jog(p(i.j))
1.1.7 Angular Second Moment (ASM): Angular Second
Moment is also called Energy and Uniformity and is a measure
of textural uniformity i.e., pixel pair repetition. High ASM
values occur when the grey level distribution has either a
contrast or a periodic form.
ASM= Yi, Na Yo Ne pij»
1.1.8 Correlation (COR): Correlation is a measure of grey
tone linear dependencies in the image. High correlation values
imply a linear relationship between the grey levels of pixel
Pairs
COR = Lio 51 5 1-0 "5" [(i-0).(i-H)-pGj)V0102
2. MATERIALS AND METHODS
2.1 STUDY AREA:
Study area selected is a part of East Godavari district in Andhra
Pradesh, India. It covers, Rajahmundry town on the banks of
river Godavari and a wide range of land use land covers occur
in the site, which include paddy, pulses, tobacco, sunnhemp,
plantations, permanent and current fallows, water bodies and
settlements. The upper part is characterized by upland areas and
dry crops and the lower part by irrigated wet lands.
2.2 SATELLITE DATA AND ITS PROCESSING:
ERS-1 SAR data acquired at 23° look angle and VV
polarization in C -band forms the data set for the study. The
data were acquired at 35 day interval during October '92 -
February '93, corresponding to two crop seasons, kharif and
rabi. The data were geocoded and resampled to 36m. The
resultant imagery are of the size of 750 x 750. During the
observation within a crop season, considerable changes were
observed with the dynamic features such as crops, water bodies
etc. In addition, on ERS-1 SAR image, wide range of texture
patterns could be associated with different land covers. For
example, water bodies and dry fallow fields manifested with
very fine texture while settlements and coconut plantations
exhibited very coarse texture.
The SAR data sets were co-registered and filtered employing
3x3 Gamma Maximum Apriori Probability (GMAP) filter to
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
suppress the speckle. Based on the preliminary field survey,
training and test area sites for major land use and land cover
classes were identified and their statistics generated. Pair wise
separability was computed for all these classes considering
ERS-1 SAR intensity data alone by Jeffries-Matusita distance.
It is clear from the separability analysis that majority of the
classes have a maximum possible separability of 2.0. However,
discrimination of pulses from tobacco (0.916), shallow water
from early transplanted paddy (1.09), tobacco and pulses from
mixed plantation (1.423 and 0.763, respectively), pulses from
coconut plantation (1.69) are poor. Further the standard
deviation of training set pixels was high for early transplanted
paddy, mixed and coconut plantations. These observations
confirm the need to account for textural variations of different
land covers to improve class discrimination on radar imagery.
2.2.1 Classification by tone and texture: Haralick's texture
features from grey level co-occurrence matrix formed the basis
in the present study as detailed in Section 1.1. However, since
the possible derivable textural features were too numerous, with
correlation existing amongst the features, it was necessary to
identify the effective minimum set of textural features by
optimization based on interclass separability. To optimize the
number of texture features, a mosaic of sample segments of
representative textures identified in the ERS-1 SAR imagery
have been used initially. Subsequently, Haralick's texture
features have been computed for all the sample segments as a
function of inter pixel distance and direction and optimized
based on their performance in terms of number of separable
textures with each of them independently and with their
combinations. Angular second moment and Mean at Inter Pixel
Distance (IPD) of 1 and Dissimilarity at an IPD of 3 were found
to best discriminate various features on the test data set. Hence
these features have been used in generating the texture images
corresponding to the ERS-1 SAR data used in the study.
The tonal and the optimized textural information were used in
classifying the SAR data sets, corresponding to rabi (17 Dec
'92. 21 Jan '93 and 25 Feb '93) and kharif (8 Oct. '92 and 12
Nov '92) crop seasons. The SAR data of rabi season was used
as test data while that of kharif was used as validating data set.
When intensity data was alone used, the classification accuracy
as computed by Kappa coefficient (k) were observed to be
higher with two date data (k=0.612) than either with December
(k=0.330) or January (k=0.413) data sets alone. However,
texture information alone yielded k of 0.795 and 0.612,
respectively for December and January months (Table 1).
Classification accuracy parameters pertaining to kharif season
are presented in Table 2. The raw and classified images of the
study area of kharif and rabi seasons are presented in Figure 1.
Significant improvements in classification accuracy were
obtained when tone and texture were combined, for all the
single date data sets. When two date intensity data alone were
used for classification, kappa accuracy was 0.612, which .
significantly improved to 0.838 with texture alone and 0.867
when tone was combined with texture. Only marginal
improvements in kappa coefficient were observed in using tone
as well as texture with two date intensity data. Similar results
were obtained for the data set acquired during kharif crop
season with October and November data over the same study
area during the overcast conditions.
Addition of textural features in classification scheme clearly
improved the accuracy of the classes that showed distinct
textural pattern as in the case of pulses, tobacco, coconut,
cashew and settlements. Similar results were obtained (Figure
2) for the dataset acquired during another crop season (October