Superpixel-based Multiple Change Detection in Very-High-Resolution Remote Sensing Images

This paper presents a novel unsupervised superpixel-based change detection approach to detect multiple changes in Very-High-Resolution remote sensing images. The proposed approach investigates the spectral-spatial variations at superpixel level which aims to enhance the traditional pixel level change detection performance. In particular, superpixel representation of the spectral change vectors is built by exploiting the homogeneity of local objects associating with the change and no-change classes. A decision-level ensemble strategy is designed to generate a reliable binary change detection result. Then the multi-class changes are identified by automatic clustering. Sensitivity of the relevant parameters are analyzed and discussed. Experimental results obtained on a pair of real VHR images confirm the effectiveness of the proposed approach.