MHT-MambaNet: a multi-source heterogeneous terrain-aware Mamba network for landslide detection in remote sensing imagery

Abstract

Landslide detection from high-resolution remote sensing imagery is critical for geological disaster monitoring and risk assessment. However, existing approaches often struggle to effectively model heterogeneous multimodal data and capture complex spatial dependencies or boundary structures in rugged mountainous environments. To address these challenges, this study proposes MHT-MambaNet, a multi-source heterogeneous terrain-aware Mamba framework. The proposed method integrates a multi-source feature alignment mechanism to mitigate modality discrepancies among optical, SAR, and terrain data, alongside a direction-adaptive Mamba interaction module designed to model long-range spatial dependencies while preserving local structural details. Furthermore, a change-guided object-level decoding module is introduced to enhance boundary consistency and structural integrity. By explicitly modeling geomorphic characteristics, the framework demonstrates an inherent capability to perceive and interpret complex topographic structures, such as steep slopes and deep valleys. Extensive experiments verify that MHT-MambaNet consistently outperforms state-of-the-art approaches in both segmentation accuracy and boundary delineation. Beyond its performance on benchmark datasets, the proposed framework provides a robust technical foundation for domain-adaptive application in high-risk mountainous corridors, such as the southeastern margin of the Qinghai-Tibet Plateau, offering a scalable solution for geological disaster mitigation in complex engineering-geological environments.