Spectral-spatial hyperspectral image classification via SVM and superpixel segmentation

Integration of spatial information has recently emerged as a powerful tool in improving the classification accuracy of hyperspectral image (HSI). However, partitioning homogeneous regions of the HSI remains a challenging task. This paper proposes a novel spectral-spatial classification method inspired by the support vector machine (SVM) and superpixel segmentation. Core ideas of the proposed method are twofold: 1) the HSI is first classified by the pixel-wise classifier (i.e. SVM); 2) a fast superpixel segmentation-based spatial processing is, for the first time, introduced in this study to refine the homogeneity and consistency of the classification maps. Experiments are conducted on two benchmark HSIs (i.e. the Indian Pines data and the Washington, D.C. Mall data) with different spectral and spatial resolutions. It is found that the proposed method yields more accurate classification results compared to the state-of-the-art techniques.