points, but collected at a much finer spatial scale. 
Linking these two different sorts of data to come up with 
an understanding of natural resource distribution is not 
trivial. 
One possible solution is using remote sensors to 
map areas of interest into land and water cover types. 
Each type can also be correlated with a particular level 
or distribution of the resource of interest - say water 
quality, or biomass, based on field measurements. If 
such information is available, a map of a particular 
locality in terms of cover types can then be directly 
linked to levels of resource distributed in those types. 
This assumes the following - that remote sensing based 
classification corresponds to the distribution of different 
land/water cover types (in spite of problems relating to 
lack of spectral separability) and that these cover types 
are tightly correlated with resource distributions. Both of 
these are open questions, for the following reasons. 
Land cover types such as forests are complex 
mixtures of different percentages and kinds of soil, with 
different levels of moisture and are composed of 
different species, each with its own characteristic set of 
coloured pigments, leaf size ranges and orientations 
(Verbyla, 1995). These factors combine in theory to 
make each cover type distinct from others in terms of 
spectral reflectance values. In actuality, boundaries 
between these types may often be arbitrary - while some 
boundaries such as those between water bodies and land 
may be clear-cut, others between different vegetation 
types may be fuzzy, and hard to determine from a 
remote sensor (Forman, 1995). 
Even if boundaries are clear-cut, cover types 
distinct on the ground are not always separable using 
remote sensors, mainly due to mixed pixels. When cover 
types vary at a scale finer than the resolution of the 
remote sensor, each signal received by the sensor is 
actually a composite of signals coming from various 
types, which cannot always be separated. This depends 
on the fineness with which one wishes to delineate these 
ecosystem types - at a coarser scale of mapping, each 
pixel would correspond to a single type. This problem, 
therefore, fixes the spatial scale at which different types 
can be discriminated from each other (Moody and 
Woodcock, 1994). 
Even if distinct, land and water cover types are 
spectrally separable due to various characteristics - of 
which difference in resource levels may be only one 
(Kroner and Running, 1993). For example, land cover 
types can be discriminated and defined in terms of 
several factors, such as vegetation composition, 
structure, biomass levels, percentage leaf cover, 
phenology and soil composition. The extent to which a 
type defines the distribution of a single specific resource 
such as biomass is not very clear. 
LINKING REMOTE SENSING WITH FIELD 
DATA COLLECTION 
We have investigated these questions through the 
application of a two-level methodology, combining 
relatively coarse though extensive remotely sensed data 
with locally sampled, intensive and detailed field data on 
species distributions. Our study was specifically aimed 
towards assessment of a specific natural "resource", 
namely biodiversity, using a combination of remote 
sensors and field measurements. In theory, remote 
sensors can be used to map ecosystems, as their spatial 
and spectral resolution is adequate for this purpose 
(Running et at., 1995). Such broad ecosystem-scale 
mapping is in fact recommended for country wide 
exercises in biodiversity assessment in the Global 
Biodiversity Assessment (UNEP, 1995). 
Species are constituents of ecosystems, along with 
abiotic components. In principle, ecosystems delineate 
species boundaries (Noss, 1996). Maps of ecosystem 
distributions may therefore be used, in conjunction with 
information on species frequencies in these ecosystem 
types, to generate information on species diversity levels 
(Condit, 1996). The Global Biodiversity Assessment in 
fact recommends such a method, stating “broad scale 
sampling generally requires the use of remote sensing 
methods and measures at the levels of the biotope and 
the landscape, while point-sampling involves measuring 
a representative selection of localized sampling points. 
Data from a series of points, when coupled with remote 
sensing, may provide information that can be 
extrapolated for global coverage”. However, the 
correlation between species and ecosystems is not 
absolute and ecosystem boundaries are often hard to 
define (Hansen and di Castri, 1992). 
Earlier studies in this area have not been 
conclusive, (Treitz et at., 1992; Franklin et at., 1994; 
Ravan et at., 1995), but our studies in the Western Ghats 
of India (Nagendra and Gadgil, 1999) clearly 
demonstrate the feasibility of such a methodology for 
assessing Angiosperm (flowering plant) species 
diversity. We observe that supervised classification of a 
small (36.5 km") landscape into ecosystem types of a 
hectare or more, can distinguish between ecosystem 
types accurately enough that the classified types differ 
significantly in the diversity of their flowering plants. 
Unsupervised classification, with a specification of the 
number of types to be classified, was unable to do so.