Genetic neural networks for image classification

This paper introduces the application of genetic neural networks for spectral classification of remotely sensed images. Genetic neural networks have combined the features of neural networks and genetic algorithms in the way that the coded-instructions of evolvable genes specify the architecture of neural networks. This enables consistent reductions of mean square errors of spectral classification with respect to sample training pixels. While supervised classification is usually confined to the data with which the training was done, genetic neural networks have a strong flexibility to cope with various attributes of the data, such as sensor types, stretching, solar angles and so on. Additionally, for the problems of mixed and ambiguous pixels, the algorithm of simulated annealing was examined to test if it helps genetic algorithms climb up from semi optima of fitness landscape.