window over the image and checking for the situation where none of the 
eight neighboring pixels is the same class as the center pixel. When 
this situation is encountered, the class label of the center pixel is 
changed. The criterion used to determine the new class of the center 
pixel has been changed from the original algorithm. In the original 
algorithm, the classes are assumed to be nominal and a simple majority 
rule (mode of the distribution) is used for class conversion. In the 
version implemented for FOCIS, the classes are assumed to be at least 
weakly ordinal and apriori class conversion weights are used. In this 
approach, the counts of pixels in each class are multipled by the class 
conversion weights and the largest value of the resulting products 
determines the new class of the center pixel. 
Creation of a polygon table is the second step in the spatial filtering 
process. An entry is made in the polygon table for each polygon in the 
image that stores among other items, the number of pixels in the poly 
gon, their class label, and the frequency of occurrence of each class 
in the pixels bordering on the polygon. Diagonal pixels are considered 
to be bordering on polygons. The entries in the polygon table are 
sorted by size and row number before use in polygon conversion. 
For polygon conversion, each polygon is evaluated in the order of its 
occurrence in the polygon table. If the polygon is larger than the 
minimum size specified by the user for that class then the polygon is 
not changed. If the polygon is below the minimum size then the entire 
polygon is converted to another class depending on the class composi 
tion of the surrounding pixels. The pixel counts for each class are 
multiplied by the class conversion weights and the largest product 
determines the new class. Following class conversion, a polygon table 
entry is created for the new polygon that includes the current polygon. 
At this time, all polygon entries that are subsets of the new polygon 
are removed. If the new polygon does not meet tire minimum size require 
ment then it is converted to a new class in the same manner as before. 
In this way each polygon is evaluated until all polygons in the clas 
sified image meet the minimum size requirement for their respective 
classes. 
One result of the spatial filtering of the classified image is that 
pixels change classes. Due to the algorithm that is used, large 
classes tend to grow and small classes shrink. However, class conver 
sion weighting reduces the magnitude of this effect by increasing the 
chances of small classes having polygons of similar classes converted 
to their class label. 
Research in spatial filtering of classified images has often stemmed 
from two different motivations and their associated assumptions con 
cerning the accuracy of the classification. In one case, the clas 
sification is assumed to have errors and the use of spatial filtering 
is motivated by the intention to improve classification accuracy 
(Guptill, 1978; Goldberg and Goodenough, 1978). The contrasting sit 
uation assumes that the classification is accurate prior to spatial 
filtering and improved spatial coherence is the desired result. Based 
on the assumption that the original classification is accurate, ap 
plication of the spatial filtering algorithm degrades accuracies by 
adding pixels to classes they do not represent (Table 1). This added 
variance is an undesirable by-product of producing a stand map with 
improved spatial coherence. Thus, there is a direct trade-off between 
spatial coherence and classification accuracy. The result of applying 
the modified Davis and Peet spatial filter to a test classification map 
can be seen in Figure 3 and compared against the original (Figure 2).