Annual geospatial foundation model embeddings for landslide susceptibility assessment in Taiwan

Abstract

Landslide susceptibility models (LSMs) must identify unstable terrain before failure is visible, yet it remains unclear whether compact geospatial foundation-model embeddings encode such pre-event information. We evaluate the efficacy of annual 64-dimensional AlphaEarth Foundations embeddings for LSM using Taiwan landslide inventories from 2017–2024. A lightweight classifier trained with hard negatives is used to predict newly mapped landslides from the previous year’s embeddings under temporal and spatial holdout validation. One-year-ahead susceptibility modeling reaches a mean ROC-AUC over 0.95, and wall-to-wall island-scale susceptibility ranking captures about 90–95% of future landslide area within the top 5% of susceptible terrain during 2019-2023. Performance remains strong across regional spatial holdouts but declines in 2024, indicating sensitivity to trigger-regime shifts. Comparisons with remote-sensing covariates show that embeddings encode useful landscape-state information while benefiting from fusion with explicit geophysical variables. These results support annual geospatial embeddings as scalable baselines for regional landslide susceptibility assessment.