LAND-SURFACE CLASSIFICATION BY NEURAL NETWORKS

Spectral data from blue to near-infrared (IR) were sampled at three di fferent dates in 1992 from a fire damaged forest region near Berlin (G ermany) and have been analysed by a principal component analysis, by t he Normalized Difference Vegetation Index (NDVI) and by a self-organiz ing 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 th e clustering scheme applied, further work with this new type of classi fication algorithm.