A novel unsupervised binary change detection method for VHR optical remote sensing imagery over urban areas

Abstract

Change detection (CD) is a hot topic and has been applied in many fields. Very high resolution (VHR) images contain the rich spatial information, and are widely used in CD applications. Compared with supervised CD methods, unsupervised methods are more popular, since they can identify changes automatically. In this paper, a novel unsupervised binary CD method for VHR optical images using an advanced automatic sample selection approach with a lightweight convolutional neural network (CNN) is proposed. First, a pre-trained CNN is employed to produce deep features of two images, and a pseudo change map is yielded based on change vector analysis. Second, classification map of each image is automatically obtained by decision tree, and the other pseudo change map is yielded using post-classification comparison. Third, reliable samples are generated by fusing two pseudo change maps. Finally, image patches, i.e., neighborhood areas of the selected samples, are fed into a lightweight CNN embedded with Ghost Module and Channel Attention Module to perform training, and the CD result is obtained using the trained network. Experimental results confirm the superiority of the proposed method.