LS-FPSAM: a SAM-based frequency-prompt model for coseismic landslide detection

Coseismic landslides pose severe threats to human life and infrastructure, necessitating rapid and accurate detection for effective emergency response. Traditional methods, such as visual interpretation and conventional deep learning models, are often time-consuming and labor-intensive, limiting their utility in time-critical scenarios. With the emergence of large vision models (LVMs) and the increasing availability of remote sensing data, there is significant potential to improve the efficiency and accuracy of coseismic landslide detection. In this study, we propose LS-FPSAM, a novel frequency-prompt fine-tuning framework based on the Segment Anything Model (SAM) for coseismic landslide detection. Key innovations include: (1) a lightweight frequency-prior guided prompt generator that accentuates high-frequency landslide features; (2) inter-block residual prompt injection for fine-grained feature modulation without altering SAM's original encoder-decoder architecture; and (3) a parameter-efficient fine-tuning strategy that reduces computational costs while preserving the generalization capability of the foundation model. Experimental results on the Luding coseismic landslide dataset demonstrate that LS-FPSAM achieves a mean Intersection over Union of 0.737 and an F1-score of 0.915, outperforming several mainstream segmentation models. The approach enables rapid, high-accuracy landslide mapping with limited labeled data, offering a scalable and practical solution for disaster management and emergency response in seismically active regions.