Neural classification with genetic algorithms

Classification of terrain cover using polarimetric radar is an area of considerable interest and research. This paper describes the application of neural network approach to the classification of a fully polarimetric SAR image. The structure and connection weights of the network are decided based on the cascade correlation algorithm. The procedures are performed by interleaving the minimization of the total output error and the maximization of the correlation of the new inserted hidden unit with the residual error. While the training is stuck at a local minimum, it is possible that a useless unit will be permanently installed to the network. This problem can be solved partially by applying other optimization methods such as genetic algorithms. Experimental results showed that the proposed genetic learning method has higher classification rate and can create more compact networks, in terms of number of hidden nodes, than that of the standard cascade correlation algorithm