Assessing the riverine flood forecast skill of GloFAS with streamflow observations and impact data: a case study for Mali

Riverine floods are one of Mali's most devastating and frequently occurring disasters. However, so far, actions linked to it are mainly post-disaster ones. For this reason, the Mali Red Cross has recently established with partners a Forecast-based Financing mechanism that triggers early actions to reduce the impacts of floods once a predefined trigger is reached. Given the lack of forecasts at the national scale, the current trigger model is based only on real-time observations from the hydrological monitoring network of the National Directorate of Hydraulics (DNH): if the observed upstream water level exceeds the 5-year return period, an action is triggered to prepare for floods downstream, four days ahead, taking into account the delays in the propagation of the flood. Global flood forecasting systems can possibly complement this local flood monitoring model, especially in large transboundary river basins. This research aims to investigate the riverine flood forecast skill of the Global Flood Awareness System (GloFAS version 3.1, part of the Copernicus Emergency Management Service) in the Niger river basin by evaluating reforecast data against two reference datasets: river flow observations and impact data. The False Alarm Ratio (FAR) and the Probability of Detection (POD) have been calculated for all available extended-range reforecasts (lead times up to 46 days) over a 20-year period and for 15 river gauge station locations. For the skill assessment of GloFAS against river flow observations, most river gauge stations with enough observed data (8 out of 15) show good and robust skill scores for all lead times up to 10 days. For the skill assessment based on impact data, even though at some stations the POD is good, the FAR is too high. A preliminary conclusion is that setting trigger levels for longer lead times (up to 10 days) - to complement the existing monitoring system with a four-day lead time - can be done only for those locations where enough historical observed data is available. Using impact data to set triggers is currently hampered by limitations of the impact dataset, such as no precise event dates and locations.