classes in a two dimensional band space using the extracted
class statistics. If a sample point satisfying Eq.(7), we
can expect that a sample point belongs to a class number i,
with 10Ox(1 - a ) % confidence.
D 2 (i) < X 2 ( 2a ) (7).
In our study a = 0.05 is used. When the overlap domains
appears as are shown in Figure 13, they are sequentially
numbered as 1001,1002,.. and so on. Such numbers are called
index number. The linked list structure is useful to
determine which classes are involved in the overlap domain.
We use two tables here, i.e., a index table INDEX and a link
list LINK shown in Figure 14. Column 1 in of INDEX contains
nonnegative integers, called pointers, the value of which is
the row of the array LINK. LINK has the class numbers in
column 1 and the pointers in column 2. When the value of
the pointer in column 2 of the array LINK becomes zero, it
indicates the end of the list. For example, we can find the
classes with class number 210 and 110 by which the overlap
domain 1001 is made from the tables in Figure 14.
The confidence regions for pattern classes in a two
dimensional band space are tabulated and stored in a look-up
table LUT. More specifically, row number m and column number
n of the table LUT corresponds to the integer part of 100x
albedo value in band I and in band J, respectively. We have
that 0 ¿ml 100, and 0 £n 1100: m,n are positive integer.
Values of I and J could be chosen from 4,5,6 and 7 , but I ^
J. In a discrete two dimensional I-J space a point (m,n)
fallen in the confidence region of a certain pattern class
is assigned to have a correcponding class number and each
element of LUT(m,n) is, thus, determined to have a
corresponding class number. When a point (m,n) can not
belong to any confidence region, then class number 0 is
assigned. The sample array of a look-up table LUT is shown
in Figure 15.
Consultation of the LUT is done in the table look-up
step. If an unknown albedo pattern having p = ( 15,10) in a
I-J space, the LUT assigns immediately to a class with class
number 440 for such a pattern in the case of Figure 15.
Suppose an assigned class number k > 1000, then a pattern
falls in an overlap domain. In this case tables of INDEX and
LINK are referred to identify classes involved in the
overlap. Then, the Mahalanobis distances from a pattern to
each class center of involved are computed and 2 a pattern is
assigned to a class which gives the minimum D(i). We can
obtain a classification result by applying the foregoing
table look-up approach to every point in a new albedo data
set. As for computer times, the time required to classify
256 x 256 pixels into 11 classes by our table lool-up
approach was about 50 seconds of IBM 4341 CPU time, whereas
8 minutes were needed to do the same classification by a
conventional method based on a maximum likelihood dicision
rule.
DISCUSSIONS