Reusable landslide detection from remote-sensing imagery is limited by cross-region domain
shifts, scarce local labels, and weak links between model research and operational mapping.
This paper presents LSDetector, an open-source local workbench that packages LSDFormer and
LSDSAM with model-weight, dataset, adaptation, inference, evaluation, and GIS-export modules.
The workflow supports task-adaptive fine-tuning, domain-adversarial fine-tuning, and targetspecific
fine-tuning, then iteratively converts AI predictions and expert corrections into improved
regional inventories. In the Wuping rainfall-triggered landslide area, LSDSAM-H achieved the
best benchmark performance, whereas LSDFormer provided the fastest deployment option. A
large-area deployment over 17,295 km2 of PlanetScope imagery produced 40,400 candidate
landslide polygons, demonstrating the potential of LSDetector for reviewable, semi-automatic
landslide inventory construction.

Landslide detection, remote sensing, deep learning, foundation model