Neural networks classifiers based on geocoded data and multispectral images for satellite image interpretation

In previous papers, we presented an image interpretation system for automatic cartography using remote sensing and photointerpreter knowledge for classification purpose. Two different approaches were proposed: expert system (ICARE) [1] and connexionist approach [5]. Both used a preclassification made with the maximum likelihood method and included texture features derived from grey levels cooccurrences matrices. Then expert knowledge was added to improve classification results. Even if this approach has shown good results, one of the problem we had was expert knowledge acquisition and their expression in a symbolic way (production rules with certainty factors). Up to now, this stage was made in a natural language form using a textual interface. Getting pertinent rules took a long time because of the many trial- and-error needed by such a process. We want to overcome textual acquisition limitations by providing a way of reducing knowledge acquisition time thanks to neural networks generalization capabilities.

Authors and Affiliations

1. I.R.I.T (Institut de Recherche en Informatique de Toulouse), Université Paul Sabatier, 118 Route de Narbonne, 31062, Toulouse Cedex, France

   L. Mascarilla, E. H. Zahzah & J. Desachy

Reprints and permissions

© 1993 Springer-Verlag Berlin Heidelberg

Policies and ethics