Leveraging geospatial foundation model embeddings for local-scale landslide susceptibility mapping via positive-unlabeled learning

Geospatial foundation models (GeoFMs) provide transferable embeddings that may reduce manual feature engineering in landslide susceptibility mapping (LSM). However, their application to local-scale LSM remains constrained by unreliable negative samples and redundancy in high-dimensional representations. This study proposes a deployment-oriented workflow that integrates AlphaEarth embeddings with positive-unlabeled (PU) learning, thereby aligning transferable geospatial representations with the intrinsic positive-unlabeled structure of landslide inventories for practical local-scale LSM, using the Hong Kong landslide inventory as a testbed. Methodologically, we first diagnosed and mitigated multicollinearity within embeddings and traditional physical features through targeted optimization, including dimensionality reduction and regularization. We then trained PU classification models using the Positive and Background Learning with Constraints algorithm under three feature configurations: Embedding-only, Traditional-only, and Fused. Furthermore, we evaluated model interpretability and decision stability using SHapley Additive exPlanations and rank-correlation metrics. The results show that the Embedding-only configuration provides a competitive, low-preprocessing baseline. When multicollinearity is mitigated, the Fused model achieves the best overall performance and the most spatially selective delineation of high-susceptibility areas, with the majority of landslides captured within only 6.3% of the study area and a landslide density approximately 24 times the background level. The results further suggest that GeoFM embeddings complement traditional physical features by providing fine-grained surficial semantics, helping reduce false positives in complex built environments, and partially substituting for some satellite-derived indices. Overall, this workflow supports rapid engineering-geological screening and more detailed susceptibility assessment under realistic label constraints.