Author(s): Afshin Fouladi Semnan; Mohammad Javad Ostad Mirza Tehrani

Linked Author(s):

Keywords: No Keywords

Abstract: This study presents a cloud-native machine learning framework for 30m flood depth mapping in the Aqqala basin of northern Iran. Open Earth-observation and geospatial data from Google Earth Engine captured terrain, hydrography, land cover, soils, and hydrologic forcing. Three ensemble-tree models, namely Random Forest, XGBoost, and LightGBM, were trained for return periods of 30, 100, 300, and 1000 years. Gradient-boosted trees achieved the highest accuracy, with XGBoost yielding an R² of 0.95 and an RMSE ≈ of approximately 0.11 m, while the Random Forest was slightly less precise but more robust. SHAP analysis identified Height Above Nearest Drainage, elevation, and distance to river as dominant predictors, confirming strong topographic control on inundation depth. The automated, solver-free workflow eliminates the need for gauge calibration while maintaining hydrologic realism, allowing for rapid and reproducible flood mapping using open data. The results demonstrate the potential of cloud-native ensemble learning for accurate, interpretable, and scalable inundation modelling in low-gradient, data-limited floodplains such as Aqqala.

DOI:

Year: 2026