Nagendra, Harini
evergreen forest, may be present in two different forms, one with a more open canopy. These might be identified as two
classes in an unsupervised classification, but merged in a supervised classification. The desired accuracy therefore may
not be achieved with unsupervised classification.
Unsupervised classification has hitherto been mainly used for landscape mapping in the relatively homogeneous
temperate forest regions (for example Belward et al., 1990; Meiger et al., 1991; Vanoverstraeten and Trefois, 1993).
Nevertheless, Rey-Benayas and Pope (1995) suggest that satellite imagery can be used to provide information on
landscape pattern in the tropics, obviating the need for extensive field input. It is necessary to evaluate the relative
effectiveness of unsupervised classification techniques for deriving parameters of landscape pattern for more
heterogeneous tropical regions, often difficult to classify.
The Western Ghats of India is considered one of the world's biodiversity “hot-spots” (Myers, 1991). Detailed
classification is difficult to achieve for this hill chain due to its high topographic variability, small patch sizes, relatively
large number of land cover types and complex landscape patterns - characteristics rather unfavorable for the application
of remote sensing techniques (Meiger et al., 1991). Previous exercises have demonstrated a decrease in accuracy due to
unsupervised classification, for this region (Nagendra and Gadgil, 1999 a, b). However, misclassification errors have
been believed not to significantly bias metrics of landscape pattern (Wickham et al., 1997). This paper investigates
these issues, comparing metrics of landscape pattern derived from supervised and unsupervised classification
techniques for this tropical landscape.
2 METHODOLOGY
2.1 Study Area
The hill chain of the Western Ghats of India (8°—21°N, 73°-77°E) runs parallel to the western coast of India for over
1600 km. Vegetation cover in this area, mostly moist evergreen forest to begin with, has been influenced to varying
extents by humans, over centuries (Pascal, 1988). The range of environmental regimes across the Western Ghats,
varying in topography, soil, rainfall and temperature, make for a highly heterogeneous landscape. Over time, the extent
and severity of human-affected landscape change has increased. As in several parts of the tropics, the Western Ghats
landscape is now a heterogeneous, highly variegated mosaic of both natural and managed ecosystems. The region is in
need of informed strategies for conservation and management (Gadgil, 1996).
Within this bio-region, thirteen landscapes ranging in area from 9-59 km? were identified. Selection of landscapes was
limited by logistic constraints, and intended for incorporation into a program of biodiversity assessment. The
distribution of landscapes does not therefore reflect the complete variability in environmental regimes that exists in the
Western Ghats. These thirteen landscapes constitute an adequate sample group for the comparison of the utility of
supervised and unsupervised classification techniques for landscape pattern assessment. Details of the geographic
locations (in terms of latitude and longitude) and the area covered by each landscape can be found in Nagendra and
Gadgil (1999, b).
2.2 Landscape Mapping
For these landscapes, one time imagery from the Indian Remote Sensing satellite IRS 1B LISS 2 sensor, taken between
1991-1993 depending on availability of cloud free data, was obtained. Images were selected from the dry season, mid-
February to mid-June, when deciduous trees are leafless, and the contrast between evergreen and deciduous foliage is
maximum (Roy, 1993). The satellite data consists of intensities at four wavelengths - 450 to 520 nm, 520 to 600 nm,
630 to 690 nm and 730 to 900 nm, with a pixel size of 36.25 m (Kasturirangan et al, 1991). This imagery was rectified
to remove geometric distortions with the help of Survey of India 1:50,000 scale topographical sheets.
In conjunction with ground training information collected during the months of February-August 1995, this imagery
was used to carry out supervised classification of the landscapes into land cover types. For all major land cover types, at
least one or more areas of minimum four hectares in extent were used as training sites in the supervised classification.
Land cover types that present in a few small patches in the landscape could not provide good training sites for the
classifier, and were therefore omitted from consideration. Supervised classification was carried out by the maximum
likelihood algorithm (Jensen, 1986). The Grass 4.1 image processing software was used for image processing (United
States Army Corps of Engineers, 1993).
An unsupervised classification was also prepared for each landscape. The classification algorithm, specified in the
Grass 4.1.5 software, uses input parameters set by the user on the initial number of clusters, the minimum allowed
distance between clusters, and the correspondence between iterations which is required, in order to identify class
signatures. Initial cluster means for each band are defined by giving the first cluster a value equal to the band mean
minus its standard deviation, and the last cluster a value equal to the band mean plus its standard deviation. All other
cluster means are equally spaced in between. The first pass through the clustering algorithm then assigns each pixel to
the class to which it is closest in terms of Euclidean distance. All clusters less than the user-specified minimum distance
956 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
—
are tl
pixel
corre
The
com]
clust
Once
Thes
class
Clas:
Gadg
and 1
uniqi
2.3 €
Supe
throu
struc
their
diffe
the r
neare
Defi
Patcl
a] Pa
b] P:
patch
c] Nt
Land
a] M
b] M
c] M
the s;
d] S
and/c
e] It
inter:
f] C
inter:
For e
the t
Wilc
signi
Infor
lands
supei
thirte
to ve
band
great
3 RE
Betw
from
with
Table
estim
