SUPERPIXEL-BASED UNSUPERVISED CHANGE DETECTION USING MULTI-DIMENSIONAL CHANGE VECTOR ANALYSIS AND SVM-BASED CLASSIFICATION
In this paper, a novel superpixel-based approach is introduced for unsupervised change detection using remote sensing images. The proposed approach contains three steps: 1) Superpixel segmentation. The simple linear iterative cluster (SLIC) algorithm is applied to obtain lattice-like homogenous superpixels. To avoid discordances of the superpixel boundaries obtained from bi-temporal images, the two images are firstly fused using principle component analysis. And then, the SLIC algorithm is applied on the first three principle components, which contain the main information of the two images. 2) For each superpixel, which is considered as the basic unit of the image space, the multi-dimensional change vector is computed from spectral, textural and structural features. 3) The superpixels are classified into two type: changed and unchanged through two progressive classification processes. The superpixels are firstly cataloged into three types: changed, unchanged and undefined by thresholding the change vectors and a voting process. And then the undefined superpixels are further classified into two classes: changed and unchanged, using a SVM-based classifier, which is trained by the derived changed and unchanged superpixels from the former step. The experiment using Indonesia data set has confirmed that the proposed approach is able to detect the changes automatically, by exploiting multiple change features.