on 
es 
nt 
es 
se 
sis 
ble 
ıte 
  
. The map polygons represent meaningful units in 
terms of the purpose of the database. For 
example, the forester is primarily interested in 
which forest stands have been logged, not which 
pixels. In a paper explaining how GIS can be 
applied in animal ecology, Haslett (1990) 
recommends using vector- rather than raster- 
systems, since analysis based on vector 
polygons, defined by parameters of relevance to 
the animals, are of much greater ecologial 
significance than analysis based on individual 
pixel values. 
e When two data sets with different spatial 
resolution are combined, the data set with the 
coarser resolution must determine the resolution 
of the analysis. 
° Map attributes are generalizations which are valid 
on the level of the map polygons, but which are 
not necessarily valid in every single pixel. 
° The influence from natural spectral heterogeneity 
present in the image data is reduced. 
° Standard methods for data analysis, such as one- 
dimensional statistics, can be applied, e.g. to 
change detection. 
These arguments support the approach of concentrating 
the image analysis efforts to extraction of one or several 
relevant parameters, which are valid at a map polygon 
level. New polygon attribute(s) are created from the 
image data, and change detection is performed in the 
attribute database environment, rather than in the spatial 
(vector/raster) environment. This approach has 
previously been adopted to extract model parameters 
from remotely sensed data for ecological modelling (Band 
et al., 1991). 
3. IMAGE GENERALIZATION FUNCTIONS 
As of yet, only a handfull GIS functions exist to 
generalize image data to GIS-polygons. These are found 
in raster-based GIS. The need for development in this 
area has previously been pointed out e.g. by Trotter 
(1991). 
Image generalization functions in raster GIS operate on 
regions of connected pixels, rather than on individual 
pixels. Pixel values are generalized from one data layer 
to regions defined in another data layer (figure 1). 
Conceptually, two principally different categories of 
image generalization functions can be distingushed, 
based on image data type. Raw image data is 
quantitative in nature, i.e. the pixel values correspond to 
measurements on an interval scale. Image data which 
has been categorized is qualitative in nature, i.e. the 
pixel values correspond to classes on a nominal scale. 
Image categorization can take place e.g. by statistical 
classification, clustering or image segmentation. 
3.1 Quantitative image generalization 
functions 
Generalization of raw image data to GIS-polygons may 
Include as well condensation of multiple image bands into 
  
  
171 | 62 | Image | 
  
1/72 | 61 73} 651; 
67 60 68 63 
  
  
  
  
  
  
  
Figure 1. The shaded pixels represent a region in the map 
data layer, defined by 4-connectivity. Image pixel values 
are generalized into one mean value (66), valid for the 
map region. 
a single band, as extraction of representative values 
from a set of pixels in this single band. None of these 
problems are new to the remote sensing community, but 
they need to be expanded and adopted to work on sets of 
image pixels, defined by regions or polygons in other GIS 
layers. 
The normalized vegetation index (NDVI), Prinicipal 
Component Analysis, Tasseled cap transformation, as 
well as other image transformations are commonly 
accepted methods to reduce the dimensionality of the 
data (Lillesand and Kiefer, 1987). In some applications, 
image transformations are required, in others it may be 
sufficient to use individual image bands. By using 
existing methods for reduction of data dimensionality, it 
is possible to concentrate development efforts to image 
generalization functions which operate on singel image 
bands. 
Primarily, a full set of statistical functions to compute 
measures of central tendency and dispersion for image 
pixels within GIS-polygons is required. These statistical 
functions should include descriptive measures such as 
mean, median, trimmed mean, variance, and standard 
deviation (table 1). 
Tablet. Statistical measures and methods to generalize 
quantitative image data to GIS polygons 
  
  
  
  
  
  
  
  
  
  
Measure 
Central mean, 
tendency trimmed mean, 
median 
Dispersion variance, 
standard deviation 
Shape of skewness, 
distribution kurtosis 
Uni-modal histogram peak detection, 
distribution? clustering 
  
  
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996