Classification of multi-sensor remote sensing images using self-organizing feature maps and radial basis function networks

Neural networks have been known for its classification capability and have been used quite extensively in pixel classification for remote sensing in recent years. The authors propose the use of self-organizing feature map (SOFM) neural networks to evaluate the centers for the hidden neurons in RBF neural networks for pixel classification. The combined use of both neural networks leads to a much better classification than the use of only one of the two networks. For the experimental study the image data considered is from both SAR and ATM sensors and for each pixel the combined data form a 15 dimensional feature vector. Five pattern classes are defined for 5 crops. The use of SOFM alone provides only 62.7% correct at best. If only RBF network is used the best reported performance is 89.5% correct. However by combining SOFM and RBF, the best average performance is 95.15% correct, a dramatic improvement over the use of either network. Performance degradation is about 5% if the authors use only the SAR data. This points to the needs for the data level fusion for most effective pixel classification. Performance comparison is also made with the traditional k-nearest neighbor decision rule which provides 86.5 correct. The results clearly demonstrate the advantage of multi-sensor remote sensing using neural networks