lia 641 003.
information
wever offers
fusion may
> (PAN) and
racy of land
ising simple
s were then
of the parent
n degree of
land cover
Agriculture,
ear Imaging
1atic sensor
> used. The
were used as
rocessing 2)
ata is the
grid. When
tion images,
vith a 5.8 m
ement of the
5 m spatial
1985). For
ies namely,
Dempster —
al networks
nsor fusion
e categories
rmed; pixel-
ased fusion,
y pixel basis
| techniques
level. Image
segmentation is often performed for the different sources, and
then the segmented images are fused together. In decision-level
fusion the outputs of each of the single-source interpretation
modules are integrated to create a consensus interpretation
(Kim and Swain, 1990; Benediktsson and Swain, 1992).
The intensity, hue and saturation (IHS) transformation were
used which refers to the parameters of human color perception
(Lillesand and Kiefer, 1987). “Intensity” (Munsells
value:Munsell, 1950) refers to the total brightness of a colour.
“Hue” generally refers to the dominant or average wavelength
of light contribution to a colour. “Hue” generally refers to the
dominant or average wavelength of light contributing to a
colour; and “saturation” specifies the purity of a colour relative
to a gray. To begin with, the digital LISS III and PAN data
covering a common area were digitally registered using an
image to image registration algorithm taking 20 tie points
(ground control points) and using a third order polynomial
transformation. Subsequently, LISS III data were resampled to
a 6 m pixel dimension using nearest neighborhood algorithms
for further processing. Later three sets of hybrid products were
generated.. In the first set, the three bands 0.52 to 0.59 um, 0.62
to 0.68 um and 0.77 to 0.86 um of LISS III data were
transformed into intensity, hue and saturation (HIS) and
Intensity (I) was replaced by PAN data during back
transformation of IHS data to RGB components. In the second
set, band-2 (0.52 to 0.59 um) of LISS III data were replaced by
PAN data In the third set, three principal components were
generated by using the four bands of LISS III data. The first
principal component includes the largest percentage of the total
variance and the principal component 2 and 3 contains a
decreasing percentage of the variance. The first principal
component was replaced by the high resolution PAN data.
4.2 Image analysis
Ground truths were collected to represent different land use/
land cover classes present in the study area. Care was taken to
represent pure land use/land cover class in each site. USGS
classification system was used to define the classification
themes. During ground truth collection, the land use
composition was collected for the level I, level II and level III
of the land use classification system (Table 1). The land
use/land cover categories include three levels of classification
system. Four major categories viz, settlements, agricultural
land, wasteland, water bodies and others were considered at
level one classification. In agricultural lands, three subclasses
namely crop land, plantation crops and fallow lands were
defined for level two classification. In cropland, three
subclasses namely banana, sugarcane and other crops were
defined at level three class.
After displaying the digital LISS III data and identifying the
training sets where field observations were made, the spectral
response patterns of different land use were generated. A
similar procedure was adopted for classifying LISS-III and
PAN merged data, IHS transformed data and PCA-fused data.
All the data sets were classified using per-pixel Gaussian
maximum likelihood classifier and the color coded digital
outputs were generated
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India,2002
Table 1. Classification scheme adopted for land use
mapping
S.No Level I Level II Level III
1. Built-upland
2. Agricultural 2.1 Cropland 2.1.1 Banana
land 2.1.2 Sugarcane
24.3 Other
crops
2.2 Fallowland
23
Plantationcrops 2.3.1 Coconut
3. Wastelands 3.1 Stony waste
4. Water bodies — 4.1 Tanks
S. Others 5.1 Roads
4.3 Accuracy estimation
For quantitative estimation of the classification accuracy of
color-coded maps generated from LISS IIL PAN data and
combinations there of, sample areas representing different land
use/land cover categories were selected randomly (Congalton et
al., 1983). An adequate number of sample points representing
different land use/land cover classes were identified on the
color-coded maps for accuracy estimation. Because all four
data sets were registered digitally to each other with sub-pixel
accuracy and resampled to a 6m pixel dimension, the sample
areas used for accuracy estimation of one data set were also
valid for the other data sets. A one —to — one comparison of the
categories mapped from all the data sets and ground truth data
was made. An accuracy estimation in terms of overall accuracy,
errors of omission and errors of commission and .kaxpa
^
coefficient (k) was subsequently made after generating
A
confusion matrix. The kappa coefficient ( K ) was computed as
follows (Bishop et al., 1975).
r T
NS i en S x. Xy
= i=1 i=l .*
2
N -Y x, Xu
izl
D>
Where r= number of rows in the error matrix, x;= the number
of observations in row i and column i (on the major diagonal),
x;,= total of observations in row i (shown as marginal total to
right of the matrix), x,= total of observations in column i
(shown as marginal total at bottom of the matrix), N= total
number of observations included in matrix.
5. RESULT AND DISCUSSION
The discussion focuses on the thematic accuracy of different
land use / land cover classes and evaluating the performance of
various combinations thereof.
Land use maps derived from IRS 1C LISS III and PAN merged
LISS III using IHS substitution are included in Figures 1 and 2.
The accuracy estimate of the thematic map serves as a valuable
and objective tool for evaluation of its reliability. The overall
accuracy and the percentage of correctly classified pixels in
