Research On Back-propagation Neural Network-based Remote Sensing Image Classification

With the resolution of remote sensing satellite continuously improved, people can get more and more useful data and information from remote sensing images. So it is an important means to obtain the spatial information from remote sensing images by classifying the remote sensing images. However, the spectrum values of remote sensing images are mixtures of various natural features, so there often exist that different natural features have the same spectrum value or different spectrum values are represented as the same natural feature. Because of these, it is very difficult to only rely on the pixel spectral similarity between each other to raise the whole classification accuracy. On the other hand, artificial neural network have nonlinear characteristics and more strong fault-tolerant capabilities, so to solve problems above can be possible by using artificial neural network. Spot5 remote sensing data 2006, Boao, Hainan have been used as unspecified data in this paper. Based on the standard back propagation (BP) neural network classification method, we have improved this algorithm through normalizing an input vector, increasing testing set, accelerating learning speed, validating the number of neural nerve cell of hidden layer and so on. Then we classify the remote sensing image using the improved BP neural network. In the end, we have compared the classified results with non-supervised and maximum likelihood classifying results. The compared results show that the improved BP neural network classification algorithm is better than those of non-supervised classification and maximum likelihood method.