Transferable machine learning prediction of urban pluvial flood water depth using multi-resolution digital elevation models
Abstract
Machine learning methods, particularly Random Forest (RF), present an effective solution for predicting urban pluvial flood water depth, especially in data-scarce regions where traditional hydrological and hydraulic observations are lacking. This study proposes a DEM-centric RF framework for predicting spatially continuous urban flood water depth across various precipitation return periods, evaluating its sensitivity to terrain resolution degradation and potential spatial transferability. The model is trained and tested in Milan using a 1 m LiDAR-derived DEM and evaluated with a coarser 5 m EU-DEM to assess sensitivity to reduced resolution. Transferability is tested by applying the trained model to Cologne using independent 1 m LiDAR data. Model performance is assessed across five precipitation scenarios (5–100-year return periods) using absolute error metrics, tolerance thresholds, spatial diagnostics, and confidence-stratified analyses. Results show that the framework retains useful predictive capability under reduced DEM resolution and cross-city transfer, although performance decreases compared with the original Milan 1 m configuration. Mean absolute errors remain below 0.30 m, with more than 58% of grid cells within ± 0.25 m of the reference flood depth. Prediction errors are mainly found in hydraulically complex areas, while the majority of the urban domain shows well-constrained predictions. Confidence diagnostics indicate that high- and very-high-confidence predictions comply with tolerance thresholds under both DEM resolution degradation and spatial transfer. These findings suggest that relative terrain structure governs urban pluvial flooding patterns, enabling screening-level flood-depth estimation using medium-resolution DEMs. This framework provides a practical approach for large-scale screening for urban flood screening, hazard assessment, and climate adaptation planning in areas with limited data.