images in, for example, the biological and high energy physics field. This 
can be accomplished because analysis methods have assumed either that all 
objects in the field of view are disjoint in the image plane and can be 
treated individually or, if overlapping, that there are no missing or 
extra features. 
This is not the case, of course, with even the simplest of aerial 
photographs. Some of the problems associated with automatic aerial scene 
(15) 
analysis of natural environments are as follows: 
l. The image-ground separation problem is aggravated by a 
variety of backgrounds, especially those having similar 
surface and shape properties. 
Geometric shape generalization is complicated and ambiguous 
because the projective mapping varies with the point of view, 
and only the projection can be sensed. 
The illumination found in natural environments is not 
controlled ànd, for any scene, its intensity and/or spectral 
ranges may be larger than any available sensor (or technique) 
can accurately accommodate in reasonable time at required 
spatial resolution. 
Any single description (property) is often ambiguous; for 
example, an intensity gradient can be due to a change in 
surface reflectivity, a change of surface angle, or a change 
in illumination resulting in shadows. 
(15) "Processing Visual Data with an Automation Eye", G. E. Forsen, 
Stanford Research Institute, Menlo Park, Calif., May 1967.