exhibit some dispersion around a mean value (i.e., spectral signatures 
are statistical in character). This should be expected, since it is 
well known that taxonomy based on any characteristic shows dispersion. 
Thus, as we will use the term, a spectral signature is a probability 
density function (or set of such functions) which characterize the 
statistical attributes of a finite set of observations of a material and 
can be used to classify the material or its condition to some degree 
of fineness. 
At the foundation of discrimination theory is the necessity to 
realize that optimum discrimination techniques require not only that 
the procedures be tailored to recognize the item or material of interest, 
but also simultaneously, that they be tailored to reject other items or 
materials that lie in the vicinity of the desired materials but that are 
not of interest, i.e., the backgrounds in which the items of interest 
are embedded. Two types of error are possible: failure to classify 
all of the desired class actually present as that class and misclassi- 
fication of other classes as that class. Photo interpreters commonly 
call these errors of omission and commission, respectively. Errors of 
the first class can be reduced by matching the decision process as well 
as possible to the desired class. This is not very useful, however, 
because the errors of class two will be very large; i.e., many things 
will be misclassified as the desired class, and whatever information is 
to be extracted will be grossly in error. It can be shown that in 
all but trivial cases it will always be found that class-two errors 
are large when the discrimination technique is matched only to the item 
of interest. To do any better requires simultaneous tailoring of the 
process to discriminate for the item of interest and to reject the items 
not of interest. It is this need that gives rise to the central 
importance of signatures of both items of interest and the backgrounds 
in which they may be imbedded. 
Examination of measurements of the spectral reflectivity and 
emissivity of materials can aid in the development of effective discrim- 
ination procedures by providing insight into the basic optical properties 
of the materials of interest. To aid investigators in understanding 
these properties an automated data library of laboratory spectral 
reflectance and emittance measurements gathered primarily from the 
literature has been instituted by the NASA Manned Spacecraft Center 
[9, 10] under the name Earth Resources Spectral Information System. 
This computer-based library also includes programs for performing 
statistical analyses on the data which can give needed insight into 
the variability to be expected in data from the same material class. 
However, quantitative multispectral sensing studies which can lead 
to development of improved automatic techniques require adequate 
theoretical models that relate laboratory spectra to radiance detected 
from a given object by an airborne or spaceborne scanner. Unless some 
insight is achieved in connecting causative factors with detected 
effects, there is no foundation for claiming that a specific cause is 
uniquely coupled with a detected effect. Some detected effects could be 
due to spurious causes which may be transient and be fundamentally 
unconnected with the condition of interest even though the occurrence 
of the detected effect appears to be associated with this condition at 
some time and location. 
9