sensor spectral response 
sensor signal (voltage) 
A simple but useful physical model for the reflective region of 
the spectrum derived from the equation above and referred to later is 
L = 0^, 0)E, (A) +L tL 
Ai pA nA 
where 
LA = the noise equivalent spectral radiance. 
In this model the atmosphere contributes both a multiplicative term T(X3, 
the transmittance, and an additive term L the path radiance to the 
reflectance, the basic attribute of the object selected for recognition. 
Detection and discrimination of an object by multispectral sensing 
requires differences in the radiation received from the object and its 
surroundings. This radiation "contrast" is due to differences in 
reflectance p^, emittance €, or temperature T between the object and 
its surroundings. 
Discrimination means devising a decision rule, based on measure- 
ments from a sample from each of two or more given classes, which will 
enable us to assign new samples to the correct classes when we do not 
know to which they belong [5]. Classification means to assign samples 
into groups which shall be as distinct as possible. In discrimination 
the existance of the classes is given; in classification it is a matter 
to be determined. Classification has been termed "learning without 
teacher" [6] and discrimination is "learning with teacher" or supervised 
learning. Supervised learning should achieve greater efficiency because 
it takes advantage of available human knowledge and intelligence [7]. 
In some cases, such human assistance is not possible, and an unsupervised 
learning approach may be required [8]. The user defines which process is 
employed. Having made the distinction I will now use the terms almost 
interchangeably but limiting my discussion to supervised learning 
approacnes. The terms identification and recognition will also be 
employed for convenience. 
Basic to this process of recognition, classification, or discrimina- 
tion, is the concept of a signature. In general, a signature is any 
collection of observable features of a material or its condition that 
can be used for precise classification. The features that make up a 
signature may all be observed simultaneously or in a sequence of observa- 
tions spread over a considerable time period. The research reported here 
is directed primarily at instantaneous signatures because it is more rapid 
and economical to use them; but there is a need to use time-distributed 
signatures in some applications. 
A basic element of spectral discrimination theory is the realization 
that spectral signatures cannot be completely deterministic. That is 
spectral reflectivity and emissivity measurements of natural objects 
4 
8