Multiscale Superpixel Kernel Learning for Spatial-Spectral Hyperspectral Image Classification

In this paper, to further explore the spatial-spectral information, we propose a novel multiple kernel learning (MKL) algorithm embedded with multiscale superpixels, whose shapes and sizes are changed adaptly according to the local structural features. Specifically, we first introduce the superpixel to generate multiscale homogeneous regions. Then, we cluster neighboring superpixels and measure their similarity in the high-dimensional kernel space. Subsequently, we adopt the Gaussian radial basis kernel function with different bandwidths to compute the spectral kernels, intra-superpixel spatial kernels, and inter-superpixel spatial kernels on different scales based on the extracted spatial-spectral feature. Finally, we use the MKL algorithm to adaptively combine the multiscale superpixel kernels according to the data and the problem at hand to exploit the multiscale spatial-spectral information within and among superpixels effectively. The experimental results on two well-known HSI datasets demonstrate that the proposed algorithm outperforms several competing algorithms.