IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
AN EMPIRICAL INVESTIGATION ON THE THEMATIC ACCURACY OF
LAND COVER CLASSIFICATION USING FUSED IMAGES
V. Arunkumar*“, S. Natarajan * and R. Sivasamy *
* Remote Sensing Unit, Dept. of Soil Science and Agrl. Chemistry, Tamil Nadu Agricultural University, Coimbatore, India 641 003.
email::varrunl @yahoo.co.in
KEYWORDS: Remote sensing, image fusion, land cover classification, Thematic accuracy
ABSTRACT:
Over the years remote sensing has been a key data source to provide the land use/land cover information. However, the information
utility of multispectral image is limited by spectral and spatial data of the imaging system. Current imaging system however offers
trade off between high spatial and spectral resolution. In order to achieve both high spatial and spectral resolution, image fusion may
be employed. In this research, three image fusion methods were tested using high spatial resolution (5.8 m) Panchromatic (PAN) and
Linear Imaging Self Scanning imagery from Indian Remote Sensing Satellite IRS — 1D for evaluating the thematic accuracy of land
cover classification through an example using IRS 1C PAN and LISS III images. Three fused images were generated using simple
band substitution, Intensity-Hue-Saturation (IHS) and Principal Component analysis (PCA) methods. All the images were then
classified under supervised classification approaches of Maximum likelihood classification. Using the classified result of the parent
(original multispectral) image as the benchmark, the integrative analysis of the overall accuracy indicated a certain degree of
improvement in the classification from using the fused images. The validity and limitations of image fusion for land cover
classification are finally drawn.
1. INTRODUCTION
Over the years, remote sensing has been a key data source to
provide the land use/land cover information particularly at
regional scale. This information is typically derived from the
remote sensing data products using digital image classification
techniques. With the launch of the Indian Remote sensing
satellite (IRS-1A) in 1988 space borne multispectral data with a
spatial resolution comparable to the Landsat (MSS/TM) and a
push broom scanning mode of data acquisition similar to SPOT
became available. The potential of Linear Imaging Self
Scanning Sensors (LISS I and IT) on board the IRS series of
satellites i.e., IRS 1A and 1B for providing the desired
information on land use/land cover was exploited in several
studies (Sudhakar et al. (1999).
Subsequently, space borne spectral measurements from three
payloads, i.e., the panchromatic (PAN) camera, the Linear
Imaging Self Scanning Sensor (LISS-III) and the Wide Field
Sensor (WiFS) representing the state of the art sensors with
respect to spatial, spectral and temporal resolution, became
available with the launch of the Indian Remote sensing satellite.
IRS-1C in late 1995 and IRS -1D in 1997. In this context, an
attempt was made to evaluate the potential of IRS 1D LISS III,
PAN merged LISS III data for land use/land cover mapping in
parts of Coimbatore district, Tamil Nadu, Southern India.
2. STUDY AREA
Study area covered 1706.48 hectares was located between 11?
08’ to 11° 11’ and 76° 58’ to 77° O1’E of Coimbatore district,
Tamil Nadu. The climate of the area is subtropical. The study
area receives rainfall both from south-west and North-East
monsoons. The mean monthly maximum and minimum
temperature of the region are 32° and 21.5°C, respectively. The
mean annual precipitation in the area is 612 mm. soil moisture
and soil temperature regimes of the area qualify as ustic and
isohyperthermic, respectively (U.S. Department of Agriculture,
1975).
3. THE DATA SET
The Indian Remote Sensing Satellite (IRS 1C) Linear Imaging
Self Scanning Sensor (LISS III) and panchromatic sensor
(PAN) data acquired on 24^ March, 2000 were used. The
survey of India topographic maps at 1:25000 scale were used as
collateral information.
4. METHODOLOGY
The methodology was comprised of i) image processing 2)
image analysis and 3) accuracy estimation.
4.1 Image processing
The first step in generating multisensor data is the
georeferencing of the image to a common map grid. When
merging higher resolution data with lower resolution images,
usually a high resolution image (here PAN data with a 5.8 m
spatial resolution) is used as a reference for enhancement of the
lower resolution data (LISS III data with a 23.5 m spatial
resolution) (Hay dn et al, 1982, Cliche et al, 1985). For
multisensor data fusion, three main approaches namely,
statistical methods (Welch and Ehlers, 1987), the Dempster —
Shafer theory (lehrer et al, 1987) and Neural networks
(Benediktsson and Swain, 1989) are used. Sensor fusion
techniques, in general, could be divided into three categories
according to the stage at which the fusion is performed; pixel-
feature, and decision-level-based fusion. In pixel-based fusion,
the sensor measurements are merged on a pixel by pixel basis
(Ringot and Kwok, 1990). Feature based fusion techniques
merge the different data sources at intermediate level. Image
segn
then
fusk
mod
(Kir
The
usec
(Lill
valu
“Hu
of li
dom
colo
to a
COVé
ima;
(gro
tran
ab
for |
gene
to (
tran
Inte
tran
set,
PA?
gen
prin
vari
decı
com
4.2
Gro
land
repr
clas
then
com
of 1
use/
Syst
lanc
leve
nair
defi
sub
defi
Aft:
trail
resp
sim
PA]
All
ma»
out]
