Integrating Geospatial Foundation Model Embeddings with Environmental Predictors for Spatially Robust Flood Susceptibility Mapping

Flood susceptibility mapping commonly relies on manually selected environmental and terrain variables, which may not fully capture the complex spatial information available from satellite observations. Recent geospatial foundation models provide a new way to represent the Earth's surface, but their usefulness for flood susceptibility mapping has not been well tested, particularly under spatially independent validation. In this study, we evaluate whether AlphaEarth Foundation (AEF) embeddings can improve flood susceptibility mapping in the Koshi Terai floodplain of Nepal. A multi-year Sentinel-1 SAR flood inventory was developed for five flood years (2017, 2019, 2021, 2022, and 2024) using VV/VH change detection, Refined Lee filtering, and Otsu thresholding. Random Forest and XGBoost models were trained using three feature configurations: conventional environmental predictors, AEF embeddings, and their combination. Model performance was evaluated using spatially blocked nested cross-validation with a 750 m buffer between training and test samples to reduce spatial dependence. A coordinate-only model was also used as a baseline to assess whether predictions were mainly driven by geographic location. During model development, an unusually high AEF-only performance led to a detailed check of the feature extraction process. The analysis showed that inconsistent temporal compositing between flood and non-flood samples was responsible for the anomalous result, and the pipeline was corrected before final evaluation. After correction, the conventional-feature model achieved a mean ROC-AUC of 0.81 ± 0.05, while the AEF-only model achieved 0.84 ± 0.06. The combined XGBoost model performed best, with a ROC-AUC of 0.88 ± 0.04 and a PR-AUC of 0.83. SHAP analysis showed that AEF embeddings and conventional environmental variables contributed complementary information to the predictions. The best-performing model was then used to produce a flood susceptibility map, which was combined with population exposure and social vulnerability to identify areas with higher flood risk. The results indicate that AEF embeddings can provide useful information beyond conventional environmental predictors and may improve the spatial performance of machine-learning-based flood susceptibility models when evaluated using rigorous spatial validation.