This paper presents a study of neural networks and version spaces for classification of remote sensing data. In the first network, precomputed textures based on the Spatial Grey Level Dependence (SGLD) method are fed to the net in conjunction with the spectral data. The second system is the sliding window network which uses all pixels in a small neighbourhood for classification of the central pixel. The third system is based on the candidate elimination implementation of the version space method for concept acquisition and is shown to achieve a performance similar to that of the neural systems but with faster training and symbolic rule generation.