LAND-SURFACE CLASSIFICATION BY NEURAL NETWORKS

Spectral data from blue to near-infrared (IR) were sampled at three different dates in 1992 from a fire damaged forest region near Berlin (Germany) and have been analysed by a principal component analysis, by the Normalized Difference Vegetation Index (NDVI) and by a self-organizing feature map (SOM) algorithm. The properties of SOMs are summarized and it is shown that the introduction of lateral network connections allows an easy clustering of the resulting topological feature space. The SOMs reveal interesting land surface features and suggest, with the clustering scheme applied, further work with this new type of classification algorithm.