Remote sensing semantic change detection via hybrid Mamba-Transformer architecture and cross-attention guidance

Semantic change detection (SCD) in remote sensing aims to identify change regions and recognize semantic categories between bi-temporal images simultaneously. In practical applications, remote sensing images are often susceptible to irrelevant environmental factors such as illumination fluctuations and seasonal vegetation succession, which frequently lead traditional models to pseudo-change detections or semantic misclassifications. To address these challenges, we propose a novel hybrid Mamba-Transformer dual-branch architecture designed to balance efficiency of global context modeling with precision of fine-grained feature extraction. Specifically, vision Mamba blocks are incorporated into encoder stage to leverage their ability to model long sequences with linear complexity, effectively capturing spatiotemporal dependencies of ground objects across large scales. Furthermore, to mitigate feature drift caused by illumination variations and vegetation growth, a cross-attention-guided fusion module is introduced. By calculating correlation between bi-temporal features, this module dynamically enhances feature responses associated with genuine changes while suppressing environmental noise. Finally, decoder integrates Transformer self-attention mechanism to further refine delineation of object boundaries in complex scenes. Extensive experiments were conducted on SECOND and Landsat-SCD datasets to evaluate proposed method. On SECOND dataset, HMTNet achieved 73.85% mIoU and 23.12% SeK, while on Landsat-SCD dataset, it yielded 96.15% overall accuracy. Results confirmed significant performance advantages over mainstream baselines and validated core contributions of proposed hybrid architecture in enhancing environmental robustness and semantic consistency. Integration of Mamba’s global modeling and Transformer’s local refinement provides a robust solution for large-scale remote sensing monitoring under varying environmental conditions, effectively distilling change-relevant information from noisy backgrounds while maintaining high-fidelity boundary localization.