A NEW SPARSE MULTIPLE-KERNEL LEARNING METHOD FOR CLASSIFICATION OF HYPERSPECTRAL IMAGERY

In this paper, we propose a new sparse multiple-kernel learning (MKL) method for classification of hyperspectral images. The proposed method adopts two-step strategy to carry out model learning from multiple basis kernels rather than simultaneously optimizing the kernel combination and learning performance. In the first step, we firstly reformulate the multiple-kernel learning so as to making the combined kernel having statistical significance as much as possible instead of time-consuming search for optimal kernel combination. Then, we impose a constraint controlling the sparsity of the learning to improve the interpretability of the model learned from the basis kernels. In the second step, with the optimal combined kernel, we solve the optimization under routine of support vector machine (SVM). The experiments are conducted on real hyperspectral data. The experimental results indicate that the proposed sparse multiple-kernel learning method outperforms the state-of the-arts SVM algorithms in teens of classification of hyperspectral images.