An improved spectral-spatial classification framework for hyperspectral remote sensing images

It needs both spectral and spatial information to refine classification of hyperspectral images. There is a general spectral-spatial framework to address the issue. It consists of three major steps: classification, segmentation and combination, to which we have made two improvements. First, superpixels generated by over-segmentation are clustered according to superpixel-wise distances as to balance homogeneity and heterogeneity. Second, a fuzzy-logic-based combination rule is proposed. It harnesses the fuzzy state of the contribution of segmented and classified maps. Experiments have demonstrated that our adaptation is able to largely increase overall accuracy for simple segmentation method and different classifiers; while conventional methods based on majority voting heavily relies on classifiers. Moreover, recommendations are provided on the best superpixel-wise distance to be selected.