10
Ecosystem types could not be distinguished with a
sufficient accuracy by unsupervised classification, and
supervised classification was therefore essential at this
stage of the exercise.
EXTENDING THIS METHODOLOGY -
CHALLENGES
Such a methodology can in principle be extended
to cover all of India and a variety of natural resources.
However, while conceptualizing such an activity,
several scientific challenges come to mind, which need
to be addressed before beginning this exercise. To begin
with, the extent to which a remote sensing based
classification of LSE types can provide information on
the distribution of other natural resources like biomass,
land erosion or soil quality needs to be assessed. The
first step is to decide on an appropriate set of land and
water cover types for the country. This set should fulfil
the following requirements :
1) An observer in the field should be in a position to
assign any particular locale to a particular land/
water cover type on the basis of a pre-defined set
of criteria such that inter-observer variation in such
assignment is within acceptable levels.
2) It should be possible to.assign a pixel characterized
by certain spectral signature to a particular cover
type on the basis of either visual interpretation,
supervised or unsupervised classification so that
the resultant map is within some acceptable degree
of error in relation to the ground truth.
3) The different land / water cover types should differ
from each other significantly enough in the
distribution of the natural resource of interest that
they harbor, so that assignment of a cover type to a
given locale carries useful information on the
likely occurrence of this resource.
These raise questions of interest from a statistical
perspective. The number of land/water cover types
needs to be defined optimally as increasing or
decreasing this number would have advantages and
disadvantages. As the number of cover types is
increased, keeping the resolution of the remotely sensed
data constant, inter-observer variation in assigning
locales to specific cover types will increase
correspondingly (McGwire, 1992; Fisher, 1994). Inter
observer variability will of course depend upon the
fineness with which one wishes to discriminate between
LSE types on the ground. To quote two extreme cases -
classifying a landscape into just two categories, forest
and non-forest, will result in practically no inter
observer variability. A detailed classification into
evergreen, disturbed evergreen and highly disturbed
secondary evergreen will definitely have higher inter
observer variability during field assignment, as well as
increase the classification accuracy of a remote-sensing
based map.
However, this does not necessarily mean that as
broad as possible a classification scheme be adopted. As
the list of cover types for a defined area increases, the
level of detail in the classification will increase, and the
correspondence between field determined levels of
resource distribution and specific cover types will
correspondingly increase. Some compromise has to be
made considering all of these factors, to decide the level
of detail to which LSE type classification can be taken.
The challenge is to determine that set of land/water
cover types for which this tradeoff results in tightest
connections between cover type and level of resource.
Of course, all of this depends on the resolution
(spatial, spectral and temporal) of remote sensors being
used. Increase in sensor resolution lead to decrease in
classification errors (Moody and Woodcock, 1994), and
hence allow for the discrimination of a larger number of
land/water cover types. In our studies, for example, we
have used IRS LISS-2 sensors with a spatial resolution
of 36.25 by 36.25 m, to carry out detailed mapping in
twelve landscapes distributed across the Ghats, 10-50 sq.
km in area, using supervised classification. In initial
mapping exercises, we initially attempted to map these
twelve landscapes into a total of thirty-five cover types
(between fifteen and twenty-five land and water cover
types per landscape), for Angiosperm diversity
assessment. High classification errors resulted in our
reducing the set of cover types to twenty-four (between
five and nine per landscape). From our experience with
these landscapes, and discussions with a large number of
experienced field investigators working in the Western
Ghats, we prepared a list of about hundred land and
water cover types for mapping the entire Western Ghats
and west coast of India, an area of over 170,000 km".
Based on this list we have carried out mapping exercises
in additional landscapes. However, with the advent of
the LISS-3 sensors in 1996 with a higher data resolution
(Kasturirangan et al, 1996), we can distinguish more
cover types with greater accuracy. This will then
strengthen the correspondence between cover type and
species diversity.
Several other issues remain to be analyzed, before
the wider application of a two-step methodology of the
kind discussed above, for natural resource monitoring.
To begin with, it is apparent that the correspondence
between a cover type and level of the natural resource
under consideration will depend on the following -
