Extraction of land cover themes from aerial ortho-images in mountainous areas using external information

In mountainous areas, land cover extraction from ortho-images through semi-automatic classification techniques is limited by several factors including large shadow areas, radiometric similarities between distinct themes and inhomogeneous radiometry among regions of the same class. The information in the image is thus not sufficient to separate the different classes. Nevertheless, good results can be obtained by dividing each land cover class into two subclasses, shadow and non-shadow, and introducing external information into the classification process. This information can be derived from an older or more generalised database, geographical knowledge concerning the links between relief and land cover, or prior information concerning shadows. This external knowledge is then interpreted in terms of prior probabilities and merged with radiometric information from the image in a maximum a posteriori (MAP) per region classification process. Moreover, the results can also be improved by the use of combinations of derived channels calculated from the initial spectral bands of the image.