International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
147 
Fig. 9. Result of the knowledge based segmentation for the object class ‘settlement’. 
Assuming that the appearance of landuse classes in satellite 
images can be characterized by a significant accumulation of 
pixels with typical reflectance values compared to the 
surrounding areas, the spatial distance of pixels with similar 
reflectance values can be a criterion to find object boundaries. 
As an example, Fig. 3 shows an original detail of a Landsat TM 
image and Fig. 4 the significant vegetation-free pixels (in black) 
as candidates for the object class 'settlement'. A successful 
method to determine the object contours is the use of 
triangulation networks, e.g. Delaunay triangulation, of the 
marked pixels (see Fig. 5). Based on all DLM200 settlement 
objects in the satellite image, statistics of the triangle perimeters 
(mean values, standard deviations) can be determined. This is 
used as a structural parameter related to the 'compactness' of 
the object. On the other hand, valid triangles can be extracted 
by means of this perimeter statistic (Fig. 6). After a fusion of all 
adjacent valid triangles, the object contours can be obtained 
(Fig. 7). The same procedure can be applied to all other - also 
the homogeneous ones - object classes. Fig. 8 shows the 
enclosing object contours for the accumulation of the marked 
pixels having a specific vegetation index and the corresponding 
DLM objects in this area. The result of the segmentation 
process is a symbolic scene description with polygonal objects 
of the three land use classes. 
In the last step of feature extraction, texture parameters inside 
the object contours are calculated. A very simple but efficient 
parameter is the standard deviation of the grey values inside the 
objects. Because of some disturbances, for instance caused by 
errors in the topographic database or errors in the geocoding of 
the satellite image, a histogram analysis of the grey value 
distribution is needed. In this histogram analysis, the peak of 
the grey value distribution is extracted and the minor peaks are 
eliminated. The standard deviation is estimated for this changed 
distribution. The result is a Modified Standard Deviation, 
which can be used as robust texture parameter (Fig. 10). Also 
test series with the well-known texture parameters proposed by 
Haralick and Shapiro (1992) have been carried out. The tests 
have shown that homogeneity is one of the best parameters to 
distinguish between the above defined object classes (Fig. 11). 
With both texture parameters the separability of homogeneous 
from inhomogeneous classes is very good. This is important for 
the spectrally similar classes ‘settlement’ and ‘agriculture’. 
n m 
homogeneity = xi 
j=l i=l 
w(i,j) 
1 + (g(i) — g(j)) 2 
w(i, j) ...element of the co - occurrence matrix 
g(i), g(j) ...grey values in a row / column 
Neigbourhood relations for an image object are determined by 
storing the identification code numbers of adjacent objects.