ts or
atrix
ures.
iring
idow
t not
from
ropy
; and
i| for
s the
(ness
om a
ge 1S
nsity
xture
grey
ix of
uring
n an
re, P
vhere
it the
The
bility
rence
LCM
re of
X the
on a
se of.
se of
rence
stical
dY),
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India,2002
An example would help to show a GLCM from a 5x5 pixel
window with Spatial = (0,1) (i.e. Nbr is one pixel below the Ref
pixel).
Image Window Resulting co-occurrence
matrix
Pixel value
21716015 149 114 doopbrog
¥ 9 Den Q
S 47/8 }5 = #10"
S. 510090 712
Sa 333 pz qs = 6 010-040
TT A É 7919] 91
Sq 1 00180
6 17141814
The computation starts from the top left corner and counts the
occurrences of each reference pixel (Ref) to neighbour pixel
(Nbr) relationship. For example, the first relationship is 3 (Ref)
to 5 (Nbr). The second relationship is 7 (Ref) to 4 (Nbr). The
co-occurrence matrix lists the number of occurrences of each
relationship. Hence in the example, the (Ref,Nbr)=(4,6)
relationship occurs twice, and the (3,5) relationship occurs
once. So the (4,6) element of the GLCM is set at 2 while the
(3,5) element is set to 1. Such a GLCM is asymmetric. A
symmetric GLCM, which is the sum of the asymmetric co-
occurrence matrix with its transpose, can also be used. The four
symmetric measures of texture used in this study are:
1) Contrast: It is a measure of the amount of local
variation in the image.
i)Entropy: It is high when the elements of GLCM
have relatively equal values. Low when the
image is uniform in the window.
ill)Angular Second Moment (ASM): This is the
opposite of Entropy. It is high when the
GLCM has few entries of large magnitude,
low when all entries are almost equal. This
is a measure of local homogeneity.
iv)Correlation: Measures the linear dependency of
grey levels of neighbouring pixels.
The computations of these are done by the following equations:
If N is the number of grey levels; P is the normalized
symmetric GLCM of dimension N x N then,
P(1,) is the normalized co-occurrence matrix such that
N-1
> P(i,;))=1
1,j=0
N-1
Contrast: 2 (P(i,))*(i-j)**2)
ij=0
N-1
Entropy: 2 (-P(i,j))*In(P(i,)))
1,3=0
assuming that 0*In(0) = 0
NI
ASM: 2. (1,750, N- 1(P(1,) **2)
1,1=0
N-1
Correlation: X (PCi.j)*G-D)*G-p)Y V (9;*6)
1J=0
3. STUDY AREA AND DATA USED
The study area is Surendranagar district, Gujarat state situated
in western India between 22? 08"N to 23? 3I"N and 70° 58°E to
72? 8'E. This district has a mean annual rainfall of 507 mm but
only 20 percent of the district follows irrigated agriculture. The
soil type is black cotton. During the monsoon season the
predominantly grown crop is cotton, followed by pearl millet
and rice. These crops are sown generally by the second
fortnight of June to early July. Other vegetation is scrub.
3.1 Data Used
Radarsat Standard Beam 5 in descending node was taken for
the study. It is HH polarised, nominal incidence angle is 38
degrees and operates in C band at 5.6 Ghz. The pixel spacing
was 12.5 m. Multi-temporal data (three subsequent dates) was
acquired. The dates of satellite pass were on July 18, Aug 11
and Sep 04, 2001. This data covered the early growth to
vegetative phase of the crops. Raw data, which was only
calibrated and georeferenced, was used for the study.
The ground truth was done synchronous with each pass of the
satellite and for better location accuracy, a hand held Global
Positioning System was used. Plant parameters like age, height,
vigour, density, LAI, row direction and spacing and soil
parameters like moisture status, type, roughness, weed
infestation and field bund utilisation were collected.
The analysis was done using EASI/PACE software on Octane
workstation.
4. METHODOLOGY
The multi-temporal data acquired was calibrated and
georeferenced. The data was calibrated using header
information. The header information was used to compute the
backscatter from the digital number value by computing the
radar brightness and incidence angle for each pixel and to
create the dB image. The digital numbers were converted first
to B° using the calibration coefficients provided in the leader
file of the data using the following:
B° = 10 * logo [(DN; + A3) / A2;] dB
The B° was then converted to 6° using the following :
0, i fi; +10 * log;o(sinl;) dB
Where, A2 is the scaling gain value for the j'" pixel and A3 is
the fixed offset and I; is the incidence angle at the i™ range
pixel.
No filtering was done. The four texture measures were chosen
from available literature. Five window sizes were so chosen so
as to find out the feasibility of feature discrimination and
efficiency of texture based classification based on window
statistics. The window sizes chosen were 3 x 3; 7 x 7; 15 x 15:
25 x 25 and 31 x 31. The vegetation classes chosen were
cotton, rice, pearl millet, scrub. In addition, urban and water
sites were also taken. Texture measures were used on all three-
date data separately and the output images were put on 32-bit
real channel. The statistics obtained from output image were
used as input for computing the spectral distances using