MAWNet: Adaptive Spatio-Frequency Representation Learning for Semantic Change Detection in Remote Sensing Images

Semantic change detection (SCD) using multi-temporal remote sensing imagery is critical for land-cover dynamics analysis and wetland ecosystem monitoring. However, most existing SCD methods rely primarily on spatial domain features, which limits their ability to capture fine-grained semantic transitions, scale variations, and ambiguous boundaries in complex wetland scenarios. To address these challenges, this paper proposes a multi-scale adaptive network for wetland semantic change detection (MAW Net). Under a bi-temporal framework, MAW Net integrates spatial-domain and frequency domain representations to enhance the discrimination of subtle semantic changes under complex natural conditions. A multi-scale atrous spatial pyramid pooling module (MASPP) is adopted to effectively model irregular objects and land-cover patterns with diverse scales, while an adaptive frequency aware spatial-temporal modeling (AFSM) mechanism explicitly captures the temporal evolution of semantic features. Furthermore, a semantic change guided feature fusion strategy with boundary-aware constraints is designed to improve semantic consistency and boundary delineation accuracy in changed regions. The proposed method is comprehensively evaluated on three benchmark datasets, namely SECOND, HRSCD, and the Dongting Lake wetland dataset. Experimental results demonstrate that MAW Net consistently outperforms state-of-the-art SCD methods, achieving mloU scores of 75.25% and 71.55% on the SECOND and HRSCD datasets, respectively. On the Dongting Lake wetland dataset, MAW Net further attains an mIoU of 88.86% and an Fscd of 91.01%. These results validate the robustness and generalizability of MAWNetfor semantic change detection in complex wetland environments.