Source based feature extraction for support vector machines in hyperspectral classification

Classification of hyperspectral remote sensing data with support vector machines (SVMs) is investigated. SVMs have shown to perform well in terms of classification accuracies for hyperspectral data sets. On the other hand, the computational burden of SVMs in hyperdimensional space can be quite intense. Therefore, it is important to explore approaches, which lighten the computational burden without sacrificing the overall classification accuracies. Two different feature extraction methods, decision boundary feature extraction and nonparametric weighted feature extraction are tested. The hyperspectral data are split into several "independent data sources". The data from each data source are transformed using feature extraction, then two approaches are investigated. In the first approach the data from all sources are classified together with a multisource SVM kernel. In the second approach, the data are classified separately using classical SVM RBF kernel. The results from the SVMs are then fused for final classification. Results are compared and discussed.