Pattern classification for remote sensing using neural network

The authors propose a pattern classification method for remote sensing data based on neural network theory. From geographical knowledge and Kohonen´s self-organization feature maps, training areas for each pattern are selected. Using the backpropagation algorithm, a layered neural network is trained such that the training patterns can be classified within a level. After training the network, some pixels are omitted from the training areas if they are incorrectly classified and new training ones are determined. Once training is complete, remote sensing data are applied to the trained neural network. Experiments on Landsat TM (Thematic Mapper) data show that this approach produces excellent classification results which are more realistic and noiseless compared with the conventional Bayesian approach