a 
ng 
old 
te 
ity 
1a] 
)n. 
ps 
n 
his 
Nagendra, Harini 
  
ESTIMATING LANDSCAPE PATTERN FROM SUPERVISED AND UNSUPERVISED 
CLASSIFICATION: STUDIES IN THE WESTERN GHATS, INDIA 
Harini NAGENDRA 
Centre for Ecological Sciences 
Indian Institute of Science 
Bangalore 560012, India 
Harini@ces.iisc.ernet.in 
KEY WORDS: Accuracy, Classification, Ecosystems, Interpretation, Land use/Land cover, Pattern recognition, Spatial 
data, Sustainability 
ABSTRACT 
Current research suggests that metrics of landscape pattern can act as indicators of ecological processes and biodiversity 
maintenance. Land cover maps, created through remote sensing, enable the evaluation of such pattern. This paper 
compares supervised and unsupervised classification techniques for the estimation of landscape pattern, in tropical 
landscapes that are often inaccessible and difficult to classify. In the mountains of the Western Ghats of India, thirteen 
landscapes ranging from 9-54 km^ were delineated for mapping. Supervised classification accuracy ranged from 70- 
92%. Unsupervised classification accuracy was uniformly worse, ranging from 31-75%. Misclassification errors have 
been previously believed not to bias metrics of landscape pattern. This paper reports a significant bias in patch and 
landscape metrics estimates concomitant with misclassification resulting from unsupervised classification. For all 
landscapes, patch size, shape and nearest neighbor distance metrics derived from unsupervised classification were 
significantly greater than those derived from supervised classification. Landscape metrics of mean patch size, mean 
patch shape, mean nearest neighbor distance and the Shannon index of landscape diversity determined from 
unsupervised classification were also significantly greater than those from supervised. Metrics of interspersion- 
juxtaposition and contagion do not however demonstrate significant differences. Possible explanations for the observed 
bias are discussed. Whether the bias noticed extends to other methods of unsupervised classification, requires 
examination. This exercise has strong implications for the development of country level methodology for monitoring 
biodiversity in India. 
1 INTRODUCTION 
The management of tropical landscapes is an increasingly pressing imperative. While the tropics are increasingly under 
threat due to human extraction of resources, these landscapes are also highly complex and difficult to evaluate. Due in 
large part to their inaccessibility, relatively little is known about these regions (Roy, 1993). Remote sensing provides an 
effective means of studying these parts of the world, though it is only recently that its potential for ecological studies is 
being utilized (Roughgarden et al., 1991). Under the auspices of “ecosystem management”, land cover maps, created 
using remote sensors, are becoming increasingly critical inputs for biodiversity management (Naveh, 1993; Noss, 
1996). Aspects of landscape structure such as shape, size and inter-patch distance for patches of various ecotope types, 
or land cover types, and of organization, such as the diversity and connectivity of the larger landscape in which these 
patches are embedded, are believed to have significant bearing on species distributions. Current research also suggests 
that landscape metrics, quantifying the amount and arrangement of land cover types, can act as indicators of ecological 
processes operating within these landscapes (Naveh, 1994; Palang, 1998). This has increased interest in quantifying 
landscape pattern with such metrics (Forman, 1995). 
Land cover maps are commonly created from remotely sensed data through supervised or unsupervised classification 
techniques (Jensen, 1986). In a supervised classification, the identity and location of certain representative patches of 
the land cover types present in a landscape need to be identified prior to classification. Initial field input is normally 
required for adequate map accuracy (Lark, 1995). This can prove a limitation in relatively inaccessible tropical areas 
(Rey-Benayas and Pope, 1995). Unsupervised classification of satellite imagery does not require such field input. The 
spectral data is organized into spectral classes, typically through the use of a clustering algorithm. Post-classification, 
these clusters are then associated with land cover types by the analyst. This can save a great deal of effort and time. In 
cases where the training data is unavailable or of poor quality, unsupervised classification is preferable to supervised 
(Jensen, 1986). 
However, unsupervised classification assumes a simple model where each spectral class directly corresponds to a single 
land cover type. In practice, this does not always apply (Lark, 1995). For example, a single cover type, such as a semi- 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 955