Change detection in heterogeneous images based on multiple pseudo-homogeneous image pairs

Abstract

Due to the significant disparities in feature spaces of multi-source images, change detection (CD) of heterogeneous remote sensing images (HRSIs) remains a highly challenging problem. Currently, CD methods based on domain transfer networks (DTNs) have garnered significant attention. However, the computer scientists underutilize knowledge in the field of CD during DTNs design, and the existing CD methods do not fully utilize the heterogeneous complementary features contained in HRSIs. Therefore, this study proposes a novel CD method based on multiple pseudo-homogeneous image pairs. First, a cycle-consistent generative adversarial network with knowledge constraints (named as KCGAN) was designed for obtaining good pseudo-homogeneous images. In detail, both the domain knowledge that there are land cover changes in multi-temporal images and that the objects in an image can be described from different scales were well integrated into the design of KCGAN. Then, a multi-modal difference Siamese fusion network (named as MDSiamF) was proposed to extract change information from the multiple pseudo-homogeneous image pairs generated with KCGAN. Experiments conducted on three datasets showed that: 1) compared to existing domain transfer methods, the unchanged areas in the pseudo-homogeneous images obtained by KCGAN exhibit better feature consistency (with a peak signal-to-noise ratio higher than 20.85 and a PHash value higher than 0.9); 2) compared to state-of-the-art methods for CD of HRSIs, the proposed method shows stable and good CD performance (with an overall accuracy higher than 0.98 and a F1 Score higher than 0.78).