Integrated land-cover mapping from satellite imagery using artificial neural networks

The automatic mapping of land cover from satellite imagery requires optimal classification and spatial generalization procedures. Here we describe the use of functional ink neural networks, based on a flat perceptron net with an augmented feature vector, to generate high accuracy classification products. These can then be trained more rapidly than multi-layer perceptrons. The network output is then used to fix land cover class area statistics which control a low-level generalization procedure based on a combined iterative majority filtering and reduced class growing procedure.