Hydraulic hotspots vs. erosion reality - a GAM approach to quantifying bank erosion using HEC-RAS and LiDAR difference images

Hydrological models such as HEC-RAS are widely used for flood hazard mapping, but translating these hydraulic outputs into spatially explicit bank erosion risk maps remains challenging. This difficulty arises primarily from the lack of sedimentological data and the highly localized nature of bank erosion during flash floods. Machine learning (ML) offers a powerful tool to uncover complex, non-linear relationships between hydraulic parameters and observed erosion patterns. This study aims to develop and validate a logistic Generalized Additive Model (GAM) that predicts the probability of bank erosion using hydraulic parameters derived from HEC-RAS simulations (e.g., flow velocity, bed shear stress). The model is calibrated and validated using erosion maps derived from high-resolution LiDAR-based elevation change detection. To assess real-world applicability, we validate model predictions against field survey data from the 2021 Ahr flood, providing ground-truthed erosion locations for independent evaluation beyond LiDAR-based mapping.

To quantify the spatial relationship between hydraulic conditions and observed bank erosion, we implemented a two-step workflow combining hydraulic modeling and remote sensing-based erosion mapping. First, we performed a 2D hydrodynamic simulation of the Ahr catchment using HEC-RAS v6.6, driven by a high-resolution 5-m digital elevation model (DEM) and precipitation input derived from the RADOLAN dataset for the extreme July 2021 flood event. The simulation yielded spatially distributed hydraulic parameters: including flow velocity, shear stress and flow depth. Second, we derived a high-resolution erosion map using LiDAR-based change detection. Pre-event (2019) and post-event (2021) 1-m DEMs were differenced to compute elevation changes along the river corridor. Areas exhibiting a depth loss exceeding 0.5 m were classified as “significant erosion” and used as the binary response variable (erosion/no erosion) in subsequent modeling. Finally, we developed a GAM to predict erosion probability as a function of the HEC-RAS-derived hydraulic variables. Our logistic GAM achieved AUC-ROC of 0.8 through non-linear s-terms and physically meaningful te-interactions (shear×depth, velocity×depth). Comparison with field survey results confirmed that the model reliably identifies zone prone to bank erosion. This approach successfully bridges the gap between hydraulic modeling and observed erosion patterns, revealing non-linear, spatially variable relationships that simple thresholds miss. The proposed methodology provides a robust, data-driven framework for translating HEC-RAS outputs into high-resolution erosion risk maps. Future research should integrate spatially explicit sediment characteristics to quantifiy the local mobilisation potential and test the model across a range of geomorphological and lithological settings to further improve its transferability and predictive accuracy.

This research forms part of the MABEIS III project ("Mass Movement Information System for Rhineland-Palatinate"), funded by the Rhineland-Palatinate´s State Office for Mobility (LBM) and the State Authority for Geology and Mining (LGB).