High-resolution impact-based early warning system for riverine flooding

Flooding affects more people worldwide than any other natural hazard and is one of the four key climate change hazards. An estimated 1.81 billion people (23% of global population) are directly exposed to 100-year floods. With anthropogenic climate change, inadequate investments, and cognitive biases, flood impacts are expected to worsen. Record-breaking rainfall events have increased since the 1990s and the rarest events are projected to intensify most strongly. As preparedness and defenses are overwhelmed by extremes, forecasting and early warning systems are crucial to safeguard life and reduce losses. Advances in numerical weather prediction (NWP), hydrological forecasting, data assimilation, and computational efficiency have improved FEWS capabilities, yet catastrophic events such as the July 2021 floods in Western Europe show persistent gaps between forecasting skill and societal outcomes. A technologically advanced FEWS requires observational initial conditions, NWP, hydrological forecasting, and ideally hydrodynamic and impact modelling. Major challenges include uncertainty in precipitation forecasts (especially for rare events), computational demands of fine-resolution inundation modelling for ensembles, lack of high-resolution river cross-section data, and limited integration of probabilistic information into decision-making. Pre-calculated hazard maps used operationally (e.g., in EFAS and GloFAS) provide only rough, sometimes inconsistent inundation estimates and lack dynamics. This study addresses these gaps by extending FEWS with high-resolution (1–10 m) quasi-real-time hydrodynamic and impact forecasting, generating lead-time maps and object-based impact indicators. Using the 2021 Ahr River flood as a case, the research question is whether an end-to-end, impact-based, probabilistic FEWS can provide actionable, spatially explicit information (inundation, velocities, impacts) with useful lead time and uncertainty quantification to improve preparedness and response.

The paper reviews advances in hydro-meteorological monitoring and forecasting, noting improvements in NWP resolution to convection-permitting scales (1–4 km), ensemble methods, and better process representation. Hydrological forecasting has progressed toward hyper-resolution (0.1–1 km) but remains challenging due to data and parameter field requirements. Ensemble flood forecasting is still maturing operationally and often not tailored to local managers’ needs for high spatio-temporal resolution. Current FEWS (e.g., EFAS, GloFAS, national systems in Germany, Australia, USA) are summarized (Table 1), including their atmospheric models, resolutions, and update frequencies. Operational approaches to inundation and impact often interpolate pre-calculated hazard maps for limited return periods, which are static and may be inconsistent, especially for unusual events. The literature highlights computational barriers to real-time high-fidelity inundation modelling and the development of simplified or emulated models to balance speed and accuracy, though these may struggle with dynamics, out-of-sample extremes, and diverse landscapes. The need to propagate and communicate uncertainties along the forecast chain is emphasized, as is the shift to impact-based warnings advocated by WMO and demonstrated to improve emergency management and behavioral responses.

An extended, end-to-end warning chain (ICON_D2_EPS–mHM–RIM2D) produces high-resolution, impact-based forecasts. Meteorology: The DWD ICON-D2-EPS limited area ensemble (2.2 km resolution, 20 members) provides 48-h forecasts, initialised every 3 hours. Hydrological initial conditions use near-real-time radar-adjusted precipitation (DWD) and temperature fields generated via External Drift Kriging using station variograms. Hydrology: The mesoscale Hydrologic Model (mHM) runs at 1.1 km resolution, using multiscale parameter regionalization and forced by ICON-D2-EPS to generate streamflow and water level predictions at Altenahr gauge. For hindcast, 16 initialisations (every 3 h from 13 July 02:00 to 14 July 23:00 CEST) × 20 members = 320 ensemble forecasts drive mHM. mHM was calibrated using DDS (500 iterations) against hourly discharge at Altenahr for 2011–2020 with a 5-year warm-up; 2021 event excluded. Hydrodynamics: RIM2D (validated for the Ahr 2021 event) simulates 2D inundation on a 10 m grid. First, HQ100 flood depths are mapped; then for each forecast, the lead time to exceed HQ100 is computed per 10 m cell downstream of Altenahr. Hydrodynamic forecasts are triggered automatically upon reaching user-defined warning thresholds for selected ensemble percentiles (min, 25th, median, 75th, max) to ensure timely 3-hourly updates. RIM2D runs on GPUs (NVIDIA Tesla P100), achieving ~22 minutes per 48-h simulation for a 30 km river reach at 10 m resolution; ensemble members are run in parallel (one per GPU). Impacts: Object-based impacts are computed by intersecting inundation outputs with OpenStreetMap features (buildings, roads, railways) to estimate affected building footprints and infrastructure lengths. Validation and benchmarking: Water level forecasts are evaluated against LfU reconstructed levels at Altenahr. Inundation extents and impacts are compared against Copernicus EMS Rapid Mapping (CEMS) and the inundation extent mapped by LfU. Forecast persistency: To communicate robust lead times, the study defines persistency as three consecutive initialisations exceeding HQ100 at a grid cell; the lead time is taken from the third forecast. Uncertainty propagation: Atmospheric ensemble uncertainty (n=20) is propagated through mHM to RIM2D; products include ensemble statistics (min, 25th, median, 75th, max) for water levels, inundation extent, and impacts. Data requirements: High-resolution DEM (10 m), land use, near-real-time precipitation and temperature, and gauge data. Codes and datasets are openly available (with some datasets under restricted but free access).

- Predictability and thresholds: Despite ensemble precipitation underestimation relative to the most realistic observed estimate (119 mm between 07/14 07:00–21:00 CEST), the key decision variable—probability of exceeding warning thresholds—was informative. For all forecasts issued within 17 hours of the flood peak, the probability of exceeding HQ100 at Altenahr was ≥50%. At 11 hours before peak (14 July, 14:00 CEST), the exceedance probability surged to 90%. - Uncertainty evolution: The probability of exceeding HQ100 increased by 30% from 20 h to 17 h before peak, then decreased by 20% in the next forecast, reflecting atmospheric forecast uncertainty. - Lead-time maps: Raster-based lead-time maps (10 m resolution) show for the ensemble median a window of 6–30 hours to exceed HQ100 downstream of Altenahr; for the maximum ensemble member, 24–48 hours. The ensemble median inundation extent underestimates LfU’s mapped extent (11.33 km²), while the maximum member aligns closely, consistent with observed precipitation lying near the ensemble maximum range. - Impact benchmarking: Comparison with Copernicus Rapid Mapping shows the maximum ensemble member issued 47 h before peak overestimated inundated building footprint by ~10%. Across several initialisations, maximum members often matched benchmarks closely; median and lower percentiles tended to underestimate, consistent with lower water levels. - Uncertainty representation: Whisker plots of inundated area across 16 initialisations (47 h to 2 h before peak) illustrate propagated uncertainty from NWP to hydrology to inundation. The ensemble median inundation area exceeded HQ100 and HQextreme hazard map extents by 20 and 17 lead hours, respectively, offering actionable preparation time despite underestimation of absolute extent. - Computational feasibility: GPU-accelerated RIM2D delivered near-real-time 10 m-resolution ensemble inundation and impact forecasts for a 30 km river reach in about 22 minutes per 48-h simulation (per ensemble member), demonstrating operational feasibility even for small, fast-reacting rivers.

The study demonstrates that integrating high-resolution hydrologic and hydrodynamic modelling with ensemble NWP within an operationally oriented FEWS enables probabilistic, spatially continuous inundation and impact forecasts. This addresses a key gap in traditional gauge-based warnings by providing dynamic lead-time information, depths, and velocities over floodplains. Although ensemble precipitation underestimation limited absolute flood peak prediction, the probabilistic exceedance of critical thresholds (e.g., HQ100) and persistent lead times provided meaningful, actionable insights for emergency management. Lead-time maps that consider forecast persistency help balance false alarms and missed events, providing robust windows for response. Comparisons with satellite-based rapid mapping illustrate both the value of impact-based forecasts and the uncertainties in remote-sensing benchmarks (e.g., SAR misclassification, timing). The approach supports decision-making frameworks (e.g., PADM, prospect theory) by offering probabilistic information that can be aligned with agency-specific thresholds. Overall, the findings support a shift toward impact-based services that transparently communicate uncertainty and better align forecasts with the needs of local authorities for targeted, timely actions.

This proof-of-concept shows that an end-to-end, impact-based FEWS can deliver near-real-time, high-resolution inundation and impact forecasts with quantified uncertainties, even for small, rapidly responding catchments. The system provided actionable lead-time maps (median 6–30 h; maximum 24–48 h) and probabilistic exceedance information (HQ100 exceedance probability ≥50% within 17 h, peaking at 90% 11 h before the flood peak). The maximum ensemble member closely matched observed inundation extents, highlighting the importance of using full ensembles for risk-aware decisions. Future work should: (1) operationalize at national/regional scales, optimizing computational scheduling and data flows; (2) improve precipitation estimation and post-processing to better capture extremes; (3) systematically evaluate performance across more events to calibrate decision thresholds; (4) co-develop communication protocols, tailored warnings, and training with emergency managers and social scientists; and (5) expand to data-scarce regions leveraging remote sensing and open datasets while ensuring robust uncertainty propagation and communication.

- Scaling and operations: National-scale, real-time impact-based FEWS face trade-offs among computational power, operational scheduling, and data storage. - Data availability: Many regions lack high-quality, high-resolution DEMs, soil and land data, real-time observations, and high-resolution NWP, as well as long-term discharge records for calibration and thresholds. - NWP uncertainty: Ensemble size, model structure, and representation of convection lead to precipitation uncertainty; real-time precipitation may be underestimated and needs post-processing. - Integration complexity: Complex data integration, validation, and workflow management are needed; operational tools (e.g., ecFlow) and user-friendly interfaces are important but add complexity. - Evaluation challenges: Limited hindcast data and short operational histories hinder robust verification; reliability is critical to avoid the cry-wolf effect. - Rarity of extremes: Limited observations of rare, extreme floods (>100-year return periods) constrain calibration and validation; manager training is essential. - Communication and actionability: Real-time impact-based warnings must be coupled with tailored messages, action protocols, and decision support, requiring interdisciplinary collaboration. - Cognitive and systemic biases: Communicating unprecedented events requires impartial, transparent uncertainty presentation; continuous calibration and consideration of megafloods are needed to avoid surprise. - Event-specific performance: For the 2021 event, ensemble precipitation underestimation led to underpredicted water levels and inundation in median forecasts; while maximum members aligned better, reliance on single extremes can increase false alarms.