Classification of multispectral remote sensing image using an improved backpropagation neural network

Recent research has demonstrated that a backpropagation neural network classifier is a useful tool for multispectral remote sensing image classification. However, its training time is too long and the network's generalization ability is not good enough. Here, a new method is developed not only to accelerate the training speed but also to increase the accuracy of the classification. The method is composed of two steps. First, a simple penal term is added to the conventional squared error to increase the network's generalization ability. Secondly, the fixed factor method is used to find the optimal learning rate. We have applied it to the classification of landsat MSS data. The results show that the training time is much shorter and the accuracy of classification is increased as well. The results are also compared to the maximum likelihood method which demonstrate that the back-propagation neural network classifier is more efficient.