Lim Gn = H.
N-> oo
Once the entropy has been computed, the
redundancy is defined as one minus the
ratio of the measured entropy to its
maximum (Shannon, 1949c),
Redundancy - (1 - (Gn / MAX)).
Computation of the conditional entropy or
"equivocation" across spectral channels is
defined by Theorem 6 (Shannon, 19494).
The conditional entropy assumes that
probabilities representing uncertainties
are conditional upon co-occurring events.
THEOREM 6: Let p(Bi,Sj) be the
probability of sequence Bi followed by
symbol Sj and pBi(Sj) = p(Bi, Sj) / p(Bi)
be the conditional probability of Sj after
Bi. Let
Fn - -). p(Bi,Sj) log pBi(sj)
1,7
where the sum is over all blocks Bi of N-1
symbols and over all symbols Sj. Then Fn
is a monotonic decreasing function of N,
Fn = NGn - (N - 1) G(n-1),
N
NZ Fn,
1
Fn <= Cn,
Gn
Il
and Lim Fn = H.
N->00
What Are We Seeing?
Perhaps the most remarkable aspect of
visualizations of the metrics of image
information is the extent to which the
visualizations are easily recognizable
images of the natural world. This is in
spite of the fact that the metrics of
information are mathematical abstractions
calculated from tables of uncertainties.
Concepts such as "information,"
"redundant" or "equivocal" do not
effectively describe what we are seeing
since these concepts imply complex
unconscious value judgements which have
little meaning relative to the
mathematical abstractions visualized.
Paradox in the use of familiar language
and concepts is inevitable. Probably the
best approach on casual inspection is to
just view the visualizations as "pictures"
which reveal the world in a different
light.
In visualizations of image entropy, bright
areas represent the "mathematically most
interesting" parts of an image, while dark
areas represent the "mathematically least
interesting” parts. Another
interpretation is that bright areas
delineate regions conveying the most
information about the image field, and
dark areas convey the least.
In visualizations of image redundancy,
more subtlety of interpretation is
involved. Redundancy provides a measure
684
of the amount of information which is
pre-determined by context. For example,
the redundancy of written English is
approximately fifty percent, meaning that
almost half of what we write is
pre-determined by the structure of the
language. In a redundancy visualization,
bright areas represent regions which
convey the least information relative to
the image field. In satellite imagery,
the redundancy is dominated by extremely
low probability events and is ideal for
locating point events which may be
spectrally subtle yet statistically
prominent, such as fires or hot spots.
The redundancy is a logarithmic rather
than a linear inverse of the entropy.
The equivocation (or conditional entropy)
requires the most subtlety in
interpretation. In information theory,
equivocation is a measure of information
lost during transmission as a function of
channel capacity and noise. It represents
the ambiguity or residual uncertainty
associated with our measurement. Shannon
calls it "...the uncertainty when we have
received a signal of what was actually
sent." In an image visualization, the
equivocation seems to represent the
coherence of statistical structures.
Isolated clouds, for example, generally
have very low to zero equivocation values
since their multi-spectral statistics are
generally very specific (unequivocal!)
relative to those of other image
components. The equivocation can probably
best be thought of as an inverse measure
of "clustering" relative to the
statistical background. The equivocation
occasionally reveals structures absent in
both the entropy and the redundancy.
Conversely, structures visible in the
entropy or redundancy may be absent in the
equivocation. Noise in the imaging
process is a definite component of the
equivocation.
The Entropy Function and Biological Visual
Systems
An interesting analog to visualizing the
metrics of information in digital
multispectral imagery is the processing of
the inverted image projected upon the
retina of biological visual systems. It
must be emphasized that information theory
is not a model of retinal processing.
Visualizations of the metrics of
information theory do share enough
characteristics with known retinal
operations at the cellular level to make a
discussion of retinal processing relevant
to their visual interpretation.
In the human visual system, information
which reaches the visual cortex in the
brain has already been subjected to
extensive information processing by the
time it leaves the eye. This processing
occurrs in the retina, a thin membrane
lining the back of the eyeball. The
retina is an extension of the neural
architecture of the brain and consists of
two types of photoreceptors (rods and
cones) and several layers of specialized
neurons. This layering of distinct neural
Mc ODO <0 NON = ctr Wn
cn No onm A 0 nm
thoH Un o'Odn.t' 56 r£ F.-- 2000 HRNIRZN