The effect of some internal neural network parameters on SAR texture classification performance

Artificial neural networks have been successfully applied to image processing, and have shown a great potential in the classification of a wide range of remote sensing data. The major advantages of neural network algorithm over traditional classifiers are its nonparametric nature and its easy adaptation to different types of data format from multiple sources. However, a successful application of neural networks in remote sensing data classification requires a good comprehension of the effect of some internal parameters related to the neural network structure and training process. In this work we report the application of backpropagation neural network in classifying natural wetlands vegetation using SAR data. The effect of some parameters related to the architecture and the training process on classification performance was investigated and new techniques for ameliorating this performance are discussed. The results showed that the variations of the number of hidden layers and the number of nodes by layer have not a substantial effect on classification accuracy but affect only the training time. However, other parameters related to the neural algorithm computation (such as the threshold value) affect significantly the overall classification. It is concluded that, although the neural network method have a great potential in remote sensing data classification, a rigorous choice of the threshold value still necessary to optimize the ratio of the incorrectly and the correctly classified pixels