Classification of SAR images and retrieval of remotely sensed parameters using neural networks

Neural networks are applied for retrieval and classification of remotely sensed data. Study of various characteristics of the parameters to be retrieved is essential prior to the retrieval process. Fast learning neural networks have been found to successfully retrieve relevant parameters if the input to network is properly selected. Forest scattering parameters from simulated data have been successfully retrieved with excellent results. The use of neural network is extended for supervised classification of AIRSAR images. Two types of networks are used for the classification problem, the multi-layer perceptron which uses the output weight optimization technique and the functional link network which is a polynomial classifier. Both of these networks perform excellently with low training error percentage if the input to the network is properly chosen. The resulting classified images are compared to conventional Bayes-Gaussian Classifier outputs.