A framework for automated event-based monitoring of flood forecast performance

The Destination Earth (DestinE) programme of the European Commission is developing high-resolution digital twins of the Earth system to improve the simulation of extreme weather events and their impacts. Within this initiative, the global Extremes Digital Twin (G-EDT) provides meteorological simulations at 4.4 km resolution that are coupled to ECMWF’s Land Surface Modelling System (ecLand) and the CaMa-Flood routing model to produce global river discharge forecasts. As digital twins move towards operational use, there is a growing need for automated approaches to assess forecast performance as events unfold, rather than relying solely on manual, delayed, aggregated evaluations. In this context, continuous verification becomes an integral part of operational monitoring, supporting both scientific development and system oversight.

This contribution presents the methodology behind an automated framework for weekly post-event flood analysis, designed to support near-real-time monitoring of flood activity and forecast skill. The system runs on a fixed weekly cycle and analyses recent hydrological conditions using river discharge reanalysis as a proxy for observations. Flood events are identified based on exceedance of return-period thresholds.

Individual station exceedances are grouped into spatially coherent flood events using a density-based clustering approach. For each detected event, river discharge forecasts are evaluated using event-based metrics that target key flood characteristics, including peak timing error, peak magnitude error and flood duration error, assessed across lead times and affected locations. In addition, contingency-table-based verification is applied to threshold exceedances, enabling assessment of forecast event detection and discrimination using metrics derived from hits, misses, false alarms and correct negatives, such as the equitable threat score (ETS) and critical success index (CSI).

Beyond weekly reporting, the framework supports temporal analysis of forecast performance, allowing changes in skill to be tracked over time and across evolving model configurations. While the current implementation focuses on DestinE flood predictions, the methodology is generic and extensible, with planned integration of additional systems and observational datasets.