A NEURAL SYSTEM FOR DEFORESTATION MONITORING ON LANDSAT IMAGES OF THE AMAZON REGION

We deal with the problem of automating the interpretation of satellite images of the Amazon region for deforestation monitoring. Our approach is based on a combination of image segmentation and classification techniques, the latter employing a neural-network architecture that works on a fuzzy model of classification. The architecture implements a relaxation mechanism on top of a feedforward neural network, in order to take advantage of the interrelations among neighboring image segments. Our fuzzy, segment-based approach has numerous advantages over more traditional, pixel-based approaches employing statistical techniques. These advantages range from the possibility of treating transition and interference phenomena in the images to the ease with which complex information related to a region's geometry, texture, and contextual setting can be used. We report on a great variety of experiments on representative portions of the Amazon region, employing neural networks trained by the back-propagation algorithm. The results indicate very good overall performance, and allow us to draw some conclusions regarding the effectiveness of the various sources of information available as input to the system. In particular, it appears that simple spectral information, together with textural information on a region's entropy and correlation and simple contextual information, are effective in the classification for deforestation monitoring. It also appears that the effective incorporation of geometric information would require further investigation on possible enhancements to the system.