Yihua Ni, Shengyan Liu, Tengyue Guo, Min Xia
Remote sensing image change detection serves as a core technology in environmental monitoring. While the widespread availability of high-resolution remote sensing data provides essential support for detailed detection, it also presents technical challenges such as complex terrain interference, subtle change recognition, and large-scale scene processing. Current mainstream deep learning methods, despite their global modeling advantages, demonstrate limitations in cross-temporal fine-grained correlation mining and are prone to ambiguous edge localization in changing areas due to spatial detail loss. This paper proposes a high-resolution change detection network (TiBT-Net) that integrates bi-temporal space enhancement with token interaction. The model achieves precise change detection through dynamic token interaction and adaptive enhancement (TDIAE), utilizing deformable attention to capture semantic correlations. It constructs a Bi-Temporal Information Interaction Module (BTII) that enhances spatial details via multi-scale convolutions and channel attention, while introducing a delayed fusion mechanism (DLF) to dynamically balance dual-branch feature contributions. Experimental validations on LEVIR-CD, WHU-CD, and DSIFN-CD datasets achieved F1 scores of 90.38%, 86.74% and 96.28%, respectively, with Intersection-Union Ratios (IoU) of 82.46%, 76.59% and 92.82%. The overall accuracy (OA) reached up to 99.04%. This model effectively resolves the integration conflict between semantic information and spatial details, providing a reliable technical solution for high-precision change detection in complex scenarios.