IASFF-CTIT: a remote sensing image change detection method based on multiscale cross-temporal feature fusion

Abstract

Change detection is a fundamental task in remote sensing image processing, with the goal of accurately identifying areas of land cover change in pairwise images. To address the limitations of existing methods—such as insufficient multiscale information fusion, susceptibility to loss of boundary details, and inadequate interaction between pairwise features—we propose a lightweight dual-branch change detection network. This network uses MobileNetV2 as a shared backbone to extract multi-scale features from the two-temporal images. It introduces an IASFF module to perform channel alignment, followed by adaptive fusion of features at adjacent scales, thereby enhancing the consistency and complementarity of semantic information across different scales. Building on this, a CTIT module is introduced to fuse the two-temporal features and model global interactions, thereby improving the ability to localize change areas. The decoding stage employs a deep supervision strategy, using multi-scale supervised outputs to constrain the model to learn more stable change representations and further strengthen boundary responses. To validate the feasibility of the proposed method, we conducted comparisons with state-of-the-art models on the public datasets LEVIR-CD and SYSU-CD. Experimental results show that the proposed method achieved F1 scores of 91.42% and 82.60%, respectively, outperforming several state-of-the-art methods.