LSDetector: An Open-Source Tool Bridging Landslide Detection Models and Practical Deployment through Three-Stage Transfer Learning

Zijin Fu, Fawu Wang, Sansar Meena, Senlin Luo, Zihan Tu, Filippo Catani 

Rapid and reusable landslide detection from remote-sensing imagery remains challenging because practical deployment often requires cross-region transfer learning, limited local labels, and reproducible model-to-product workflows. This paper presents LSDetector, an open-source local workbench that bridges advanced landslide detection models and real-world deployment through three-stage transfer learning. The platform integrates LSDFormer and LSDSAM model families, registry-based management of model weights and datasets, and a workflow consisting of task-adaptive fine-tuning, domain-adversarial fine-tuning, and target-specific fine- tuning. Using the Wuping rainfall-triggered landslide area as a demonstration case, we evaluate model adaptation behavior, computational efficiency, and large-area deployment performance. Benchmark experiments show that LSDSAM-H achieves the best overall performance among the tested models, while LSDFormer provides the lightest and fastest deployment option. In practical mapping, LSDSAM-H was used to support two rounds of AI-assisted refinement and full-area inference over PlanetScope imagery covering 17,295 km2, producing 40,400 candidate landslide polygons. These results demonstrate that LSDetector can reduce manual annotation effort and support semi-automatic construction of reviewable regional landslide inventories.