Multiscale Attention Coupled With Adaptive Guidance for Change Detection With Remote Sensing Images

Multiscale strategy based on deep learning networks has been widely used in land cover change detection (LCCD) with remote sensing images (RSIs). However, ground targets with varying shapes and sizes often exhibit different feature performances at different scales. In this article, we proposed two submodules to extend the classical UNet and then improve the performance of LCCD with RSIs. First, a hierarchical multiscale attention fusion (HMAF) was proposed to cover the changing area with multiscale shapes and sizes. Second, a change weight learning module (CWLM) was proposed to describe the change probability between the learned features from each encoding layer. Finally, the adaptive weight acquired by CWLM was adopted to guide the decoding process of the proposed network. Compared to five state-of-the-art methods using three pairs of real RSIs, the proposed network is feasible to improve the change performance of LCCD with RSIs, such as it achieved an improvement of 0.75% in overall accuracy (OA) and 14.03% in precision in terms of Dataset-A.