More Sparsity in Hyperspectral SVM Classification Using Unsupervised Pre-Segmentation in the Training Phase

Support vector machines (SVM) have been shown to outperform classical supervised classification algorithms, and have therefore been recently used for classification of hyperspectral images. This paper present hyperspetral image classification based on support vector machines with two different unsupervised pre-segmentation methods applied to hyperspectral training data before the training phase of SVM classification. The pre-segmentation step, in a way compresses the training data by combining similar hyperspectral data, as a result of which the number of training samples is reduced, resulting in an overall smaller support vector amount after training. In this paper, compression is achieved using kmeans and phase correlation based unsupervised segmentation methods before the SVM training phase. It is shown that with the proposed approach it is possible to trade of accuracy against sparsity and also provide faster training time. Sparsity is important, particularly considering the high data amount encountered in hyperspectral imaging, because sparsity determines the model complexity and therefore the computational burden of the classification phase.