582 
effects). In addition, it is highly desirable for the system to provide a rigorous estimate of the accuracy of 
inference. Systems must be flexible enough to handle many decisions at the expert level. 
A knowledge based system called VEG was developed to overcome many of these problems. VEG infers 
vegetation characteristics using nadir and/or directional reflectance data as input VEG is also a tool for aid 
ing a researcher in developing new and more accurate techniques for inferring vegetation characteristics. 
VEG has many tools to test techniques for inferring vegetation characteristics. Finally, VEG has options to 
process files of remotely sensed data collected over large regions of the earth. 
Part of VEG is designed to have an array of extraction techniques for inferring vegetation characteristic 
using nadir and/or directional reflectance data as input Currently, the system has many techniques for infer 
ring spectral hemispherical reflectance, percent ground cover, and view angle extension. It is designed to be 
easily expanded to handle other inferences such as total hemispherical reflectance, leaf area index, biomass, 
photosynthetic capacity, etc. The system intelligently and efficiently integrates traditional spectral data with 
diverse knowledge bases that occur in the literature, from field data sets of directional scattering behavior, 
and from human experts. VEG accepts any combination of nadir and/or off-nadir spectral data of an unk 
nown target as input, determines the best strategy(s) for inferring the desired vegetation characteristic, applies 
the strategy(s) to the target data, and provides a rigorous estimate of the accuracy of the inference. 
VEG is also a research tool with provisions (1) for testing and developing new extraction techniques on 
a directional, spectral database, (2) for browsing and analysis (multiple plotting schemes) of data in the 
system’s spectral database and (3) for a Learning System that learns spectral and directional reflectance rela 
tionships that discriminate between user defined vegetation classes. 
The Learning System in VEG addresses an important problem. For years, researchers have studied the 
physical characteristics of directional reflectance distributions from vegetation canopies as a function of solar 
zenith angle, vegetation structure, vegetation density, and optical properties of vegetation and the soil (e.g. 
Kimes 1990). Many researchers over the years have recognized that directional reflectance contains much 
information that could be used to extract vegetation characteristics, however, little success has been realized 
in this area. Direct inversion techniques hold promise but are limited in that a large number of view angles 
or specific view directions (which may be difficult to obtain even from pointable satellites) are required 
before stable and accurate solutions can be obtained. These problems were the driving force in developing a 
learning system (Kimes et al ., 1992) that is part of the expert system VEG. 
The learning system "learns" class descriptions from samples (both positive and negative) of spectral, 
directional reflectance data of natural surfaces (bare soils, natural vegetation, and agricultural vegetation). In 
theory the user can define any class(es) they wish. Classes can include broad categories such as (soil, vegeta 
tion, and other), more specific classes (such as com, soybeans, wheat, etc.), subclasses of continuous parame 
ters (such as sparse--0-30% ground cover, intermediate-30-60% ground cover, and dense--60-100% ground 
cover of vegetation), or even structurely related classes (such as homogeneous structured vegetation canopies, 
and non-homogeneous vegetation canopies). Other classes could be defined that are based on species, 
biomass, leaf area index, physical location, vegetation structure, or portion of green leaf material. The learn 
ing system is designed to be able to handle any combination of directional view angles or spectral bands. 
The system finds class descriptions which contain the most important features that distinguish each class 
from the others. The explicit relationships used in the class descriptions include greater-than relationships 
between all combinations of view angles or spectral bands, and maximum and minimum value relationships 
for all view angles or spectral bands, forward and backscatter relationships for all view angles, and concave 
and convex relationships. The class descriptions are used to classify an unknown target using the available 
directional views or spectral bands. 
The system VEG is described and the results from studies using VEG are summarized. 
2-METHODS 
The current version of VEG provides the scientist with three different capabilities: estimation of vege 
tation parameters, a Learning System, and estimation of atmospheric effects. Vegetation parameter tech 
niques enable the scientist to apply various techniques to calculate the spectral hemispherical reflectance, 
view angle extension, and proportion ground cover. The Learning System enables VEG to learn class 
descriptions of different vegetation classes and then use the learned classes to classify unknown targets).