els
on
Textural Patterns
There is one main obstacle which limits the aptitude of the multi-
spectral classification. This limitation is caused by the classi-
fication performance itself, which assigns each pixel to its class
independent of the spectral information of the surrounding ele-
ments. However, in real picture data extensive objects have an
intensity variation which is often typical to those objects. It
is, therefore, preferable to evaluate the grey level information
within subregions - 'evaluation units' - of the image and to
extract patterns which describe the whole image region. These
evaluation units can be of different size and shape. Depending
on the kind of objects to be classified, the resolution and the
scale of the image data, the size of the units can vary e.g. bet-
ween 10 and 10 000 pixels. The shape of the evaluation units is
normally square or rectangular especially in case they are syste-
matically distributed over the whole image (overlapping, joint,
or gaping). If the evaluation unit corresponds to the extension
of the objects an irregular shape is usual.
Histogrom analysis
The easiest method to extract patterns from the grey level distri-
bution within evaluation units is the analysis of histograms
of grey level frequencies h(i). Fig. 2 gives examples of grey
level histograms for a) a forest and b) a housing area. Patterns
to discriminate these objects are the mean, the variance and mo-
ments of higher order (1) of the histograms, as well as the
number of local maxima, the distance between local maxima or
the ratio of the height and width of the local maxima. The histo-
£rams and, therefore, the derived patterns have however, only a
weak connection to the texture within the evaluation unit due to
the neglect of the spatial relationship of the pixels.