Quantifying Uncertainty in Flash Flood Forecasting using Ensemble Methods and Sensitivity Analysis

Short-term flood and inundation forecasts are challenging due to the short lead time of convective heavy rainfall events and the associated uncertainties of input data or model initial conditions. These uncertainties propagate along the forecast chain to uncertainties of the prediction of flooding extents, flow regimes, and, eventually, potential damages. Within the research project AVOSS (funded by the Federal Ministry of Education and Research) the aim is to quantify the contribution of the accompanying uncertainties of the individual forecast components. Hence, we focus on the uncertainties of the input variables particularly precipitation variability, soil moisture and soil properties, and urban drainage system effects as well as associated model and parameter uncertainties.

The applied forecast model chain consists of three parts. The first part is an ensemble based radar forecast of the temporal and spatial distribution of rainfall intensity. In a second part hydrological models are used to predict surface runoff formation based on the rainfall forecasts and pre-event soil moisture estimates. To capture the variety of different model approaches, two different hydrological models (RoGeR [1] and LARSIM [2]) were used. In a third step, the ensemble of surface runoff estimates from the hydrological models were then used to calculate inundation depths, flow velocities and local discharge applying a hydraulic surrogate model based on neural networks. The surrogate model was trained using a large ensemble of hydrodynamically simulated runoff scenarios generated by the 2D-hydraulic model HydroAS [3]. Uncertainties underlying the 2D-hydraulic model were considered by repeating a subset of hydraulic simulations with two additional hydraulic models.

We applied the forecast approach to an urbanized catchment at the foothills of the Black Forest, Germany, with a catchment extend of about 20 km². Based on the short computation time of the neural network model, which has been found to provide good reproductions of maximum water depths, maximum flow velocities, and maximum discharges, the setup enables the production of large forecast ensemble, suitable for a profound uncertainty estimate. In order to systematically evaluate and rank the influence of the input, parameter and model uncertainties along the forecast chain, a sensitivity analysis using Sobol Indices was carried out with the SAFE toolbox [4].

The results demonstrate which uncertainties plays the dominant role in short-term flash flood forecasting. Our study also enhances knowledge about the overall uncertainties for real events and their specific quantitative effects in pluvial flash floods. Furthermore, we identified the most relevant factors to be considered for the design of real-time flood hazard maps and subsequent damage forecasts. This ultimately has the potential to create more reliable predictions for pluvial flash floods and provide insights for decision-making under uncertainty.

 

[1] Steinbrich et al. (2016): Model-based quantification of runoff generation processes at high spatial and temporal resolution. Environmental Earth Sciences (2016) 75:1423.

[2] Bremicker (2000). Das Wasserhaushaltsmodell LARSIM: Modellgrundlagen und Anwendungsbeispiele. Institut für Hydrologie der Universität Freiburg.

[3] Hydrotec mbH (2021): 2D-Strömungsmodell für die wasserwirtschaftliche Praxis.

[4] Pianosi et al. (2015), A Matlab toolbox for Global Sensitivity Analysis, Environmental Modelling & Software, 70, 80-85.