Hyperspectral image classification using Support Vector Neural Network algorithm

With the developing technology, Hyperspectral images can be obtained with the satellites, aircraft and even unmanned aerial vehicles. Therefore, the classification applications made on the HSI are becoming increasingly important. In particular, fast and reliable classification algorithms are needed. The basic principle in classification algorithms is using characteristics of the data to find classification function that separate the data from each other. Neural Networks are among the non-linear classification method that can perform with high success. But, syntactic classifier has some problems that occur during training. One of this problems is called over-fitting. In many cases, especially in hyperspectral images, regularization is required for preventing the learning algorithm from over fitting the training data. In this study, a regularization scheme that named eigenvalue decay is used to make to this regularization in the training phase of networks. A training method that uses such a regularization scheme provides a margin maximization as in SVM for NNs. The two well-known data sets that are AVIRIS image of the Salinas Valley in California and image of Okavango Delta in Botswana acquired by The Hyperion sensor on NASA EO-1 satellite are used to test this classifier. The effectiveness of this algorithm on the HSI is evaluated using a series of experiments.