Machine learning-based rapid emulation of pluvial flood inundation using hydraulically derived flow depths and high-resolution physical-geographical predictors

Abstract

Abstract Pluvial flooding is becoming increasingly frequent and severe due to rapid land use/cover changes and changing rainfall patterns, creating an urgent need for fast and reliable flood inundation mapping. Although physically-based hydraulic models can provide highly accurate simulations, their substantial computational cost often limits their applicability for real-time forecasting and scenario testing. This study aimed at emulating pluvial flood extent and flow depth using machine learning (ML). In particular, we utilized extreme gradient boosting (XGB), random forest (RF), and neural network (NN) algorithms, which were tested on the Gidra River domain (western Slovakia). The training data for ML consisted of hydraulically modeled flood extents and flow depths, using a 2D MIKE + hydraulic model, for four rainfall scenarios with constant rainfall intensities of 20, 40, 60, and 80 mm/h. Seven high-resolution physical-geographical predictors were used to train the ML models under the tenfold cross-validation method. The ML-based modeling was evaluated using relevant metrics. Overall, all ML models resulted in a comparable, acceptable, and computationally efficient flood extent and flow depth classification and regression. The F1-score values ranged between 0.70 and 0.95 across different rainfall scenarios and tested folds. Higher F1-score was achieved for rainfall intensities of 20 and 40 mm/h, while slightly lower performance occurred for 60 and 80 mm/h. The ML models performed better in upstream and central test folds compared to downstream folds. Because all held-out folds were located within one hydrologically connected river corridor, these results represent spatial transferability within a corridor rather than unseen basin generalization. The regression metrics of RMSE (5–23%), MAE (2–15%), and Bias (−6 to 13%) had very similar ranges of values among ML models. Training and inference times confirmed the potential of ML models for rapid pluvial flood hazard assessment, with training times from seconds (XGB and RF) to a few minutes (NN) and testing times within a few seconds for the held-out test folds.