Additionally, it is important to describe the shape of the 
distribution of pixel values within a polygon. This is 
crucial to indicate whether or not the polygon is 
spectrally homogeneous. Basically, three situations can 
occur. The values are uni-modally distributed without 
tails (figure 2a), the values are uni-modally distributed 
with a tail (figure 2b), the values are not uni-modally 
distributed (figure 2c). Such tests should be performed 
prior to the computation of image means. Skewness and 
kurtosis can be applied to find tails in the distribution, 
e.g. to detect local within-unit variation, such as that 
introduced by the road in the forest stand (figure 2b). 
There is no standard statistical method to determine if a 
set of values are uni-modally distributed. Several 
alternatives exist, such as histogram peak detection 
algorithms, or clustering algorithms. One concern here is 
processing time, another is that the method should be 
applicable to small data sets, since a map polygon may 
include few pixels. Research and tests are ongoing. 
  
  
  
  
  
  
  
  
  
  
Figure 2. a) A spectrally homogeneous forest stand, b) a 
forest stand with uni-modally distributed pixel values, 
with a tail caused by logging roads, c) a forest stand that 
has been partially logged, resulting in a bi-modal 
histogram. 
3.2 Qualitative image generalization 
functions 
Multi-spectral image classification reduces the 
dimensionality of the data, at the same time converting 
raw measurements into information. Mode (highest 
frequency of occurence) is a measure of central 
tendency in the data (table 2). Percentage coverage of a 
certain class, and composition (i.e. the relative 
proportion) of all classes are two measures which 
describe the "dispersion" or heterogeneity (figure 3). 
Table 2. Measures to generalize qualitative image data to 
  
  
  
  
  
  
GIS polygons 
Measure 
Central mode 
tendency 
"Dispersion" percentage coverage of 
or each class 
homogeneity relative proportion of classes 
Spatial fragmentation index, 
distribution of other spatial pattern 
classes indeces. 
  
  
  
  
   
  
1) open land 
2) forest 
NS 
  
  
N 
Figure 3. Measures for qualitative image data: mode = 1; 
percent coverage of water is 30, and composition of 
categories is {40,30,30} 
The spatial distribution of classes can be just as 
important as the proportional coverage classes, when 
extracting image information for GIS polygons. Consider 
for example two forest stands composed of 
approximately the same proportion of the image 
categories forest and bare ground (figure 4). One has 
been partially logged, while the other contains numerous 
rock outcrops. The major difference between the two 
stands is the spatial distribution of categories. 
  
B) Forest stand with 
rock outcrops 
FI = 0.042 
A) Logged forest stand 
Fl = 0.0042 
Bl rores 
Bare ground 
Figure 4. The partially logged forest stand has a lower 
pattern complexity than the forest stand with rock 
outcrops. Pattern complexity has been measured using à 
fragmentation index , Fl (Monmonier, 1974) 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
For 
inte 
cho 
and 
199 
whe 
ima 
ima 
bas 
bas 
Figu 
com) 
The 
reco 
ecol 
capti 
Mari 
cate 
Sour 
4.1 
The « 
was 
Colui 
(Syst 
proje 
SEIL 
Reset 
Nature 
Minist 
Parks,