International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
Spatial subpixel analysis, on the other hand, tries to derive 
subpixel information from the spatial pattern of pixels in a 
certain neighbourhood (e.g. 3x3 pixels) of a given pixel 
(Schneider, 1993, Steinwendner and Schneider, 1998). 
Different models about the scene pattern within the 
neighbourhood can be assumed. A very simple model is 
illustrated in Figure 4. The scene is composed of 2 
homogeneous areas, separated by a straight boundary. 4 
parameters are necessary to define this model (2 parameters for 
the position of the boundary line, d and fr, and 2 parameters for 
the pixel values within the homogeneous regions, p! and p 2 . 
Further models can be defined. The maximum number of 
parameters of any model must not exceed 8, as it must be 
smaller than the number of independent input pixel values, 
which is 9 in the case of a 3x3 neighbourhood. For every model, 
the parameters are computed from the 9 given pixel values by 
least squares adjustment. The sum of squares of the residuals 
characterises the appropriateness of the model. The model most 
appropriate in this sense is selected and provides subpixel 
information on the central pixel and on its 8 neighbours. 
The result of subpixel analysis may be used in two different 
ways for segmentation: One may use the subpixel parameters 
for resampling with a smaller pixel size, with the advantage that 
the percentage of mixed pixels will be smaller in the 
oversampled image (Figures 5 and 6). Or, secondly, the 
subpixel parameters may be used directly for segmentation. In a 
method implemented at IVFL, the result of region growing 
segmentation employing subpixel parameters is directly output 
in vector format (Steinwendner et al., 1998). One big advantage 
of this approach is the fact that shape parameters are obtained in 
a much more precise manner. 3 
3. CLASSIFICATION 
Classification in general denotes the process of assigning class 
labels to objects on the basis of a set of features. The objects 
may be individual pixels, or segments. According to the 
definition given above for segmentation, classification in 
general can be regarded as a special case of segmentation. 
Fig. 5. Landsat TM 3 (original). 
Fig. 6. Landsat TM 3 (in subpixel resolution). 
Parametric and nonparametric classification methods may be 
distinguished. In parametric classification, the probability 
density distribution of the feature values within each class is 
assumed to be known. Normal distributions are usually 
assumed. This may be appropriate for pure spectral features, but 
it is problematic for texture classification and, in particular, for 
classification based on shape parameters. For classification 
based on context information, parametric classification cannot 
be applied. Therefore, for knowledge-based image analysis as 
discussed here, the use of parametric classification methods is 
limited.