Abstract: Aiming at support vector machine（SVM） using single kernel learning not handling the classification problem of hyperspectral imagery data that the sample distribution is irregular and complex, a hyperspectral imagery data classification method based on sampling strategy and multiple kernel support vector machine is proposed in this paper. Firstly the method does sampling referring to the minority support vectors (SVs) rather than the training data to provide a balanced distribution during multiple kernel support vector machine mode, and then uses the weighted sum approach to multiple kernel learning(MKL) and optimizes parameters by gradient descent algorithm. Finally, a series of two-class classifiers are used to achieve the multi-class classification. Experimental results show that overall classification accuracy increased by 4.07%, average classification accuracy increased by 9.62% compared with the traditional SVM.

Key words: hyperspectral image, imbalanced classification, multiple kernel SVM, over-sampling, gradient descent algorithm