Neural networks have been proposed in recent years as appropriate techniques for satellite image classification. Being non-parametric models, neural networks have the advantage over traditional methods of making no assumption about the probability density function of the pixel values in an image. Another additional advantage over traditional statistical methods is that neural networks can more naturally incorporate ancillary information, such as topographical height and terrain slope, to the spectral information. This ancillary information is useful to eliminate spectral ambiguities when analyzing two or more different terrain types having similar reflectance. We have found, however, that even though the use of topographical information improves unsupervised neural network classification it creates a problem known as class proliferation, that is, too many clusters are created or many pixels are left unclassified. To address this problem we propose a two-step process to satellite image classification. The first step uses a fuzzy neural network to perform a preliminary unsupervised clustering using the spectral information. The resulting preliminary classification together with the topographic information are then fed to a discrimination tree generated by the ID3 learning algorithm. We present the approach proposed and discuss its results when compared to methods based on neural networks only.