GeoPhysAdapter: Scale-Matched Adaptation for Cross-Domain Landslide Mapping

A novel approach for mapping landslides utilizing vision foundation models has been introduced in a paper available on arXiv (2608.09325). This technique, referred to as GeoPhysAdapter, tackles the issue of cross-domain transferability in landslide identification, particularly when new landslides lack immediate annotations. Although vision foundation models excel in representational transfer, they tend to produce high-confidence false alarms in unfamiliar areas, events, and data sources. The study highlights that factors like terrain, material, and rainfall triggering can help mitigate these inaccuracies, but their effectiveness is limited to local, regional, and event-specific contexts. GeoPhysAdapter leverages a frozen vision foundation model and applies constrained adaptation at both the pixel and candidate landslide body levels, ensuring reversion to the original model when adaptation is unnecessary. The authors aim to enhance the accuracy of landslide mapping for emergency response and regional risk evaluation.