John Bosco Kyalo Kiema 
  
is possible to implicitly incorporate topological information through the use of triangulation segmentation procedures 
e.g., Delaunay triangulation. 
Different classifiers may be employed in the supervised classification of remotely sensed data, namely: maximum 
likelihood, minimum distance and parallelepiped classifiers (Lillesand and Kiefer, 1994). For the study presented here, 
a maximum likelihood classification approach is adopted. The selection of the training data is basically done using 
manual digitising. Five basic urban object classes are identified: Buildings, Pavement, Trees, Grass-covered areas and 
Special. The class Special is introduced in order to take care of the many miscellaneous urban objects of limited 
dimension that are imaged (e.g., vehicles, tents, water fountains, sculptures etc.). Because of diverse spectral diffusion 
in the multispectral data depending on different sunlight conditions, the digitisation of the training areas is often a very 
time consuming exercise. Consequently, it would be preferable to procure the training data automatically using existing 
GIS databases as proposed in (Walter, 1998). 
  
* 
(a) Results based on Daedalus scanner imagery (b) Results based on Daedalus scanner and ALS data 
Trees Pavement Special Grass Buildings 
  
  
Figure 1: Classification results 
For the purpose of this study, the elimination of shadow effects in the classification is carried out using an approach 
similar to that described in (Haala and Walter, 1999). Firstly, the shadow areas on the imagery are automatically 
identified. A separate class Shadow is then introduced for each of the apriori defined object classes. This means that 
each object class is divided into one separate subclass for shadow areas and another for non-shadow areas. Separate 
training data is digitised for each subclass. The classification is then carried out after which the shadow and non-shadow 
subclasses for each object class are combined. This results in one unique class for each apriori defined object class. 
Figs. la and b compare the final classification results in the absence and presence of the ALS data respectively. A 
comparison of these figures clearly demonstrates the improvement realised in the classification upon fusing the 
multispectral and geometric datasets. In particular, the improved ability to discriminate between the low lying urban 
object classes (e.g., pavements, grass-covered areas etc.) from features that are significantly above the terrain (e.g., 
buildings, trees etc.) is noted. 
Triangulation-based segmentation methods are usually employed in mid-level image processing procedures in order to 
combine structured image regions into semantically homogeneous clusters. In general, these methods use the delaunay 
triangle as the basic image segmentation primitive. Through this, the topological relationship between the image 
segments is implicitly exploited. This is in contrast to traditional segmentation methods which use the spectral 
information of the pixels. Details on spatial tessellations in general, and delaunay triangulation in particular, are 
discussed in (Okabe er. al, 1992). In order to smoothen the contours of the identified segments as well as minimise the 
effect of noise, use is also made of the connected components technique and morphological operations. A more 
comprehensive treatment of mathematical morphology and connected components is given in (Serra, 1986) and 
(Haralick and Shapiro, 1992) respectively. Segments are extracted for the different urban object classes. A delaunay 
  
490 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.