401 
'IBIS' Spatial Filtering Approach. The 'IBIS' procedure for spatial 
filtering differs morphologically iron the Modified Davis and Peet 
routine in that the latter is a single algorithm and the former utilizes 
a sequence of modular VICAR programs to perform the desired task. The 
technique is referred to as the 'IBIS' procedure (in the context of 
this paper) only to differentiate it from the modified Davis and Peet 
approach. 
The 'IBIS* spatial filtering approach was designed with the specific in 
tent of breaking all polygons having diagonal pixel connections and re 
assembling them in a logical and desirable manner with only hortizonal 
and vertical connections. Like the modified Davis and Peet approach, 
polygons below a specified threshold size can be removed and a class 
weighting scheme utilized. 
The heart of the approach is the enlarging of the classified image by 
a factor of three (e.g. a 100 x 100 image becomes 300 x 300). Each 
pixel becomes nine, making it possible to perform subsequent operations 
at a 'sub-pixel' level. A spatial filter is applied with a box size 
of three and a set of hierarchical class conversion weights. The 
weights are hierarchical in that a given class will always take prec 
edence over classes with lower weights. Equal weights are not permit 
ted because they would allow diagonal connections to persist. The 
weights are designed so that at every diagonal decision junction there 
is always a specific winning polygon that receives a hortizonal and 
vertical connection (Figure 1). Only two sub-pixels at each diagonal 
connection change class labels, preserving local spatial distributions. 
However, unavoidable changes will occur along other angular frontiers 
of the polygon. 
An example of hierarchical class conversion weights for five classes 
would be: 1.00, 1.27, 1.60, 2.01, 2.53. With this weighting structure, 
there is a strong bias for higher weighted classes to grow in size at 
the expense of lower weighted classes, which is particularly undesirable 
for certain classes such as water. Thus, the class frequency histogram 
of a classified image can be changed rather significantly (Table 1). 
One advantage of this procedure, however, is that the center sub-pixel 
of each enlarged 3x3 pixel is never changed, so the pre-filtered 
spatial distribution is not lost. On the other hand, image size is 
increased by a factor of nine, increasing computer processing costs. 
Several subsequent steps are necessary to complete the IBIS post-pro 
cessing technique: 1) All polygons are individually labeled and their 
areas calculated; 2) Polygons less than a specified size are marked for 
replacement, then set to a class label of zero; and 3) A modal filter 
with a box window size of nine is passed iteratively over the clas 
sified image to remove all zero class pixels and replace them with 
new class labels (Figure 4). 
Occasionally, several polygons of less than threshold size are adjacent 
to each other and result in a rather large block of class zero pixels. 
The modal filter extrapolates from the surrounding classes to fill the 
void, but the result can be unsatisfactory. Alternative handling 
methods for these areas include: 1) Giving them a separate class iden 
tify as 'high frequency' areas; 2) assigning them to the unknown class; 
or 3) individually assigning a label that best fits the combined area. 
The 'IBIS' technique can also be applied to a classified image first 
processed by the Modified Davis and Peet spatial filter. While apply 
ing a simplification routine to an already simplified classification