International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
48 
from the edges. The criterion of a simple form of boundaries 
can easily be put into effect. It is a disadvantage of purely edge- 
based segmentation methods that the homogeneity of the 
segments is not guaranteed automatically. 
Various combinations of region-based and edge-based concepts 
for segmentation exist. E.g. region growing may be performed 
with a boundary criterion in addition to the homogeneity 
condition. 
Watershed segmentation (Vincent and Soille, 1991) also uses 
edge information. An edge image obtained by applying a 
gradient operator is considered as a terrain relief, with high 
pixel values (edges) representing ridges and areas of low pixel 
values representing valleys and basins. A catchment basin is 
defined around each local minimum of this edge image as the 
set of all pixels that can be connected with the minimum pixel 
by a path descending from the pixel to the minimum pixel. The 
catchment basins found in this way represent segments. In order 
to obtain satisfactory results, a smoothing filter has to be 
applied before the gradient filter, and neighbouring resulting 
segments of similar mean pixel values have to be merged. The 
degree of smoothing and the similarity criterion used in the 
merging step control the mean size of the final segments. 
Various methods of segmentation of Landsat TM images for 
landcover mapping purposes have been tested. The results have 
been compared to delineations prepared by visual interpretation. 
As illustrated in Figures 2 and 3, different segmentation 
methods produce different delineations of the regions. In 
general, region growing gives the most satisfactory results. 
The result of the segmentation process may be represented as a 
labelled image band, the labels counting through the segments, 
and/or in form of a list, which contains information on the 
attributes of the segments. Topological information may be 
provided in a graph structure. 
Fig. 2. Watershed segmentation. 
2.2. Subpixel Methods 
Depending on the size of the pixels in relation to the spatial 
details of the scene, mixed pixels occur in the original image in 
a varying proportion. They may cause problems in the 
segmentation process. Mixed pixels may form segments on their 
own, although these segments have no meaning in the scene. 
One way to deal with the mixed pixel problem is to adapt the 
segmentation method in order to avoid mixed pixel segments. 
As an example, a region growing algorithm can be modified in 
the following way: Firstly, one has to avoid mixed pixels as 
seed pixels. This can be achieved by selecting as seed pixels 
low-gradient pixels, or pixels representing local minima or 
maxima. The next step is to grow regions with rather strict 
homogeneity requirements. A search for subpixel candidates 
follows. These are pixels with contiguous segments on opposite 
sides, where the pixel value in each band lies between the 
corresponding values of the adjacent segments. These subpixel 
candidates are assigned to the spectrally nearest segment 
(although they do not comply with the similarity criterion). This 
algorithm is applied in the method described in chapter 4. 
The problems with mixed pixels can also be alleviated by 
applying subpixel analysis as a preprocessing step, before 
segmentation. Spectral and spatial subpixel analysis may be 
distinguished. 
Spectral subpixel analysis tries to decompose pixel vectors into 
prototype pixel vectors (Settle and Drake, 1993). Spectral 
subpixel analysis yields information on the fractions of 
individual categories (corresponding to the prototypes) within 
every pixel, but no information on the spatial arrangement 
within the pixels. The number of prototypes has to be smaller 
than the number of (linearly independent) spectral bands. It is 
difficult to establish reliable prototypes. 
Fig. 3. Region growing segmentation.