Land-cover classification in remote-sensing images using structured neural networks

This paper proposes the application of structured neural networks to land-cover classification in remote-sensing images (in particular, multisensor images are considered). Purpose of our approach is to give a criterion for network architecture definition and to allow the interpretation of the "network behaviour'. The first result aims to avoid a cumbersome trial-and-error process; the latter one can be used to obtain information about the relevance of sensors and related bands to land-cover classification. First of all, the architecture of structured networks is tailored to a multisensor classification problem. Then, they are trained and transformed into "simplified networks' which allow one to evaluate the relevance of sensors and related bands. Experimental results on a multisensor data set related to an agricultural area are reported. -from Authors