Decrypting Neural Network Data: A Gis Case Study

Abstract

The problem of data encoding for training back-propagation neural networks is well known. The basic principle is to avoid encrypting the underlying structure of the data. This is not easy in the real world, where we often receive data which has been processed by at least one previous user. The data may contain too many instances of some class, and too few instances of other classes. Then topology, and parameters settings designed. Finally, the network produces some results which need to be explained, or decrypted.

We present our experience and results on some satellite data augmented by a terrain model. The task was to predict the forest supra-type based on the available information. In this process we were forced to invent some methods to deal with very large amounts of erratically reliable data, and to produce meaningful predictions at the end.