Mask-Guided Local–Global Attentive Network for Change Detection in Remote Sensing Images

Change detection in remote sensing images is a challenging task due to object appearance diversity and the interference of complex backgrounds. Self-attention and spatial attention-based solutions face limitations, such as high memory consumption and an inadequate ability to capture long-range relations, leading to imprecise contextual information and restricted performance. To address these challenges, this paper introduces a novel mask-guided local-global attentive network (MLA-Net). MLA-Net incorporates a memory-efficient local-global attention (LGA) module that leverages the benefits of both self-attention and spatial attention to accurately capture the local-global context. Through simultaneous exploitation of context within inter and intra patches and information refinement, the feature representation capability is significantly enhanced. Additionally, we introduce a change mask to refine feature differences and eliminate interference from irrelevant changes caused by complex backgrounds. Accordingly, a mask loss is defined to guide the generation of the mask. Extensive experiments on the LEVIR-CD, WHU-CD, and CLCD datasets show that our MLA-Net performs better than state-of-the-art methods. The code and models will be publicly available athttps://github.com/bearshng/mla-netfor reproducible research.