Hyperspectral image classification with SVM-based domain adaption classifiers

A common assumption in hyperspectral image classification is that the distribution of the classes is stable for all the areas of hyperspectral image. However, this assumption is often incorrect due to the inner-class variety over even short distance on the ground. In this paper, we present a semi-supervised support vector machine (SVM) framework to learn the cross-domain kernels from both the source and target domain in hyperspectral data. The proposed method simultaneously learns the cross-domain kernel mapping and a robust SVM classifier, which is done by minimizing both the Maximum Mean Discrepancy and structural risk functional of SVM. Experiments are carried out on two real data sets and results show that the proposed model can achieve high classification accuracy and provide robust solutions.