Structured neural networks for the classification of multisensor remote-sensing images

This paper proposes the application of Structured Neural Networks to the supervised classification of multisensor remote-sensing images. Purpose of the proposed approach is to exploit neural networks advantages while solving, in the context of the considered application, the problems of `architecture definition' and of `opacity'. The architecture of the proposed neural networks reflects the provenance of data from different sensors. This allows one to easily define a network architecture by exploiting the characteristics of a given multisensor classification problem. In addition, the `structuring' of the architecture notably helps to understand the classification criteria implemented by the neural network classifier. To make possible such an interpretation, a transformation of the representation of original networks into a `simplified representation' has also been defined. In this paper the advantages provided by such networks are pointed out from the viewpoint of the remote-sensing application. Experimental results on multisensor data and comparisons with the Bayesian classifier are reported.