As part of efforts to make weather predictions useful for emergency planning, ECMWF has developed a prototype flash flood forecast system based on the total precipitation Extreme Forecast Index (EFI) product. Results from tests in Europe show that areas susceptible to flash flooding may be identifiable up to five days ahead. The system has the potential to provide global flash flood awareness products for the medium range.
A widespread hazard
Flash flooding is a subset of flooding hazards not currently well captured in flood forecasting systems. It poses an especially significant risk in urban areas and in rapidly responding river catchments. Many countries regard flash flooding as one of the most important natural hazards in their territory. There is no universally agreed definition of what constitutes a flash flood. Meteorologically, flash floods are driven by extreme rainfall intensities which are confined both spatially (between tens and hundreds of square kilometres) and temporally (less than 24 hours). Such situations are often associated with convective activity or orographic enhancement. Hydrologically, the land surface converts most of the precipitation into surface runoff. Flooding then arises either through the rapid accumulation of surface runoff on urbanised or poorly drained surfaces, or the rapid rise of a river within a small, steep catchment.
ECMWF produces global forecasts of weather phenomena that contribute to flash flooding. The aim is to use those forecasts to support existing flood forecast systems, such as the European and the Global Flood Awareness Systems (EFAS and GloFAS) or the World Meteorological Organization Flash Flood Guidance System, which currently does not have continuous global coverage.
A relevant forecast product is ECMWF’s Extreme Forecast Index (EFI) for 24‑hour total precipitation. The EFI indicates how extreme the value of a predicted variable is by integrating the difference between the cumulative distribution functions of the forecast and the model reforecast-derived climatology. High EFI values highlight areas liable to receive more extreme precipitation than would normally be expected at that time of year. Since the EFI is calculated relative to a model-derived climatology, it is less susceptible than precipitation forecasts to the underestimation of extreme rainfall totals in ECMWF’s ensemble forecasts.
Using the EFI product, a threshold could be applied to highlight flash flood susceptible areas which exceed this value. For example, on 9 August 2018 flash flooding occurred in the Gard, Ardèche and Drôme departments of southern France. Reported rainfall totals include 105 mm falling in 1 hour in Saint-Martin-d’Ardèche and 167 mm in 24 hours at Méjannes-le-Clap. Plotting the 24-hour total precipitation EFI from 00 UTC on 7 August 2018 for 9 August shows that locations where flooding was reported were in an area of high EFI values. Applying a threshold, such as EFI ≥ 0.8, could highlight broader-scale ‘flash flood susceptible areas’, which could be very useful for emergency planners.
To explore the range at which such forecasts can be considered skilful, we extracted all 00 UTC forecasts of 24-hour total precipitation EFI for Europe from March 2016 to March 2017, for lead times up to 5 days. For each forecast at each lead time, different EFI thresholds were applied. Areas exceeding the threshold were compared against 2,663 heavy rain reports from the European Severe Weather Database (ESWD) which mentioned flood impacts. The number of hits, misses, false alarms and correct negatives were then collated over the entire one-year period. Forecast skill was computed as the area underneath the Relative Operating Characteristics curve (aROC). aROC values ranged from 0.78 at 0–24 hours lead time to 0.75 at 96–120 hours lead time. Values greater than 0.5 are considered to be skilful, so these are promising results warranting further analysis.
A next step could be to apply the flash flood forecasts globally with a view to integrating them into GloFAS. This will require a repeat of the verification procedure at the global scale using flash flood reports from FloodList.com. The results may make it possible to identify regionally and seasonally specific EFI thresholds that would work best to identify areas at risk. Combining the warnings with exposure information, such as population density and critical infrastructure, could refine the identification of areas where floods would have the greatest impacts.
Further work could consider the added value of using more forecast variables, such as surface runoff and convective activity. The ecPoint rainfall product (see Newsletter No. 153) could be used alongside the EFI to estimate point-rainfall totals within a flash flood susceptible area.