S 
ple, 
that 
he 
tion, 
e to 
ry, 
me ly 
for 
S. 
er 
ropy) 
ry, 
tion 
on of 
esents 
y 
annon 
have 
ly 
he 
lly 
alues 
s are 
1) 
obably 
sure 
ation 
nt-in 
2 
in the 
ne 
Visual 
the 
ing of 
2 
It 
theory 
nake a 
avant 
ion 
ne 
che 
sing 
ne 
ts of 
1 
ized 
reural 
structures is a characteristic repeated in 
the visual cortex (Hubel and Wiesel, 
1979). It is believed that each layer of 
specialized neurons performs specific 
spatial information processing functions, 
and that it's output propogates in concert 
with the functioning of adjacent layers. 
Visual information processing is an 
example of what may be called a "cascade" 
or "pipeline" process, with the retinal 
transform representing the earliest 
stages. 
Due to the complex spatial interaction 
among underlying neurons, every 
photoreceptor in the retina forms the 
center of a receptive field where light 
striking the photoreceptor tends to 
increase the activity of its associated 
neurons. The activity of these neurons 
tends to dampen the activity of 
neighboring neurons associated with other 
photoreceptors. The effect is mutual, 
resulting in a phenomenon known as 
"lateral inhibition." The degree of such 
spatial interaction between photoreceptors 
varies inversely with distance, but there 
is no limit at which it may be regarded as 
zero. It has been hypothesized that the 
retina compresses the bandwidth of visual 
information by suppressing the constant 
attributes of the visual field while 
enhancing deviations from the background 
defined by the whole scene (Mahowald and 
Mead, 1991). 
Visualizations of the metrics of 
information theory share several 
interesting similarities with retinal 
functioning. In both, each instantaneous 
field of view is a function of the entire 
visual field as well as stimulation at a 
point. Both functions replace the 
bandwidth occupied by the amplitude of 
stimuli with representations of the 
relevance of stimuli to a statistical 
background defined by the entire visual 
field. Both tend to radically enhance any 
variation relative to the statistical 
background. Finally, neither depicts the 
visual space in an absolute metric. 
Output remains constant despite variations 
integrated over the field of view. 
Conversely, changes in the reference field 
of view, or of specific structures within 
it, will result in changes in the 
representation of everything in the visual 
field, even things which have otherwise 
remained constant. This explains how the 
sudden appearance of subtle events in an 
otherwise monotonous visual experience can 
startle us. The retinal function 
sacrifices the physical objectivity of the 
visual world by emphasizing those aspects 
which reveal the most information about 
the environment. This is the primary 
strength which the information theoric 
visualization attempts to exploit. 
The Utility of Visualizing the Metrics of 
Information 
Frequently, visualizations reveal coherent 
structures which are absent or barely 
perceptible in untransformed imagery. 
These structures are aspects of the 
natural world which are being revealed, 
685 
however nothing more can be inferred about 
the structures beyond their relative 
mathematical "interest" and their spatial 
structure. Since the mathematical 
"interest" of an area is a function of the 
entire image, any change in the spatial 
context of the image, or any change in the 
image contents (such as clouds) will 
potentially change the brightness 
signature of the structures revealed 
within it. This extreme relativity can be 
unnerving for those seeking repeatable, 
quantitative measurements. Since the 
structures are real phenomena, their 
spatial form and location remains 
constant. 
An unfortunate liability with the approach 
as it now stands is frequent encounters 
with "granularity," "waves," and 
"streaking" which arise as the result of 
noise in the scanner-based imaging systems 
which contaminates image statistics. Even 
visually "clean" imagery can contain 
significant levels of noise which becomes 
visualized along with other subtle 
components of the image information field. 
Noisy imagery can be immediately 
recognized due to the above effects and 
inferences derived from the imagery 
subjected to deserved scrutiny. The 
potential of this technique with 
relatively noise-free imagery is an 
exciting prospect. 
Conclusion 
A system has been described which enables 
digital, multispectral imagery to be 
viewed via the metrics of information 
theory. Such images possess utility by 
revealing subtle environmental structures 
in satellite imagery, and their ability to 
reveal distortions of the image 
acquisition process. The images possess a 
theoretical richness extending from the 
realm of visual psychophysics to that of 
the study of natural systems from space. 
Often useless, occasionally startling and 
always intriguing, the images reveal the 
world from the perspective of a seamless 
mathematical model which begins at the 
imaging system in space and ends in the 
brain. They allow us to approach an 
understanding of the visual and natural 
world on terms other than a static, 
cartesian grid of calibrated measurements. 
These techniques hold exciting research 
potential into natural systems. 
References 
Gnedenka, B. and Khinchin, A. 1962. An 
Elementary Introduction to the Theory of 
Probability, Dover, New York, 130 pages. 
Hubel, D.H. and Wiesel, T.N. 1979. "Brain 
Mechanisms of Vision", Scientific 
American, 252(8):150-163. 
Khinchin, A. I. 1957. Mathematical 
Foundations of Information Theory, New 
York University, New York, 120 pages. 
Mahowald, Misha and Mead, Carver. 1991. 
The Silicon Retina, Scientific American, 
264(5):76-82.