Event-Conditional Flood Susceptibility from Land-Surface Predictors: A Globally Trained Random Forest Framework

We present a globally trained random forest framework for estimating conditional flood susceptibility at 250 m resolution, given that a flood event occurs. Using binary inundation labels from the Global Flood Database, the model integrates terrain, vegetation, imperviousness, and soil predictors without precipitation or drainage infrastructure inputs. Balanced sampling and geographic cross-validation yielded a ROC-AUC of 0.92, with NDVI, HAND, and GFI contributing most strongly. Monte Carlo perturbations propagate assumed input uncertainties into susceptibility scores, producing spatial summaries and uncertainty intervals. The framework supports relative susceptibility assessment and provides a foundation for future flood simulation applications, specifically in data-scarce regions.