Landslide Extraction Based on a Dual-Stream Cross-Modal Network with Hierarchical Fusion from GF-2 Optical Imagery and LiDAR Data

Abstract

Accurate landslide extraction plays an important role in landslide inventory mapping, geological hazard assessment, and risk management in mountainous areas. However, optical imagery alone often has difficulty distinguishing landslides from spectrally similar background objects, especially under vegetation cover and in complex terrain conditions. To address this problem, this study proposes a dual-stream cross-modal network for landslide extraction by integrating Gaofen-2 optical imagery and LiDAR-derived topographic data. This study designed a dualstream encoder to separately extract multiscale features from optical and topographic data, thereby preserving the representation capability of each modality. To improve crossmodal interaction, this study further introduces a hierarchical fusion module, which progressively integrates heterogeneous information through gated fusion, channel attention enhancement, and spatial attention refinement. In addition, this study incorporated the PKI module and Triplet Attention mechanism to strengthen multi-scale feature representation, improve boundary perception, and suppress background interference. Experimental results confirm the effectiveness of the proposed network. Compared with simple concatenation-based fusion, the hierarchical fusion module improves the IoU from 0.8316 to 0.8548. The proposed model achieves the best overall performance, with a precision of 0.9182, an F1 score of 0.9263, an IoU of 0.8548, and an overall accuracy of 0.9891. The extraction results further show that the proposed model reduces omission and commission errors, produces clearer boundaries, and provides more complete landslide delineation. It has a strong potential for practical applications in landslide identification and geological hazard risk assessment.