Classification of multispectral image using neural network

The classical land cover classification methods for the remotely sensed data have often led to unsatisfactory results. The parametric procedures such as the maximum likelihood classifier are statistically stable and robust. However they lack flexibility and and are incapable of making correct area estimates. On the other hand, nonparametric classifiers are generally very sensitive to distribution of anomalies and critically dependent on training sample size. These problems can be improved by providing prior probabilities derived from a non-parametric process in a conventional parametric classifier. In this paper, a method to merge the advantages of parametric and non-parametric strategies is presented by using Kohonen neural network and multilayer neural network with a backpropagation algorithm.