The Challenge of Flash Flooding in France

Last week was spent in Toulouse, the home of MeteoFrance and SCHAPI, the national centre for flood forecasting (operational forecasts are made in 22 regional offices).   Most of the week was taken up with a project meeting for the EU IMPRINTS project (see, with the last day spent in a workshop joining meteorologists, consultants, emergency planners and some other academics.   The IMPRINTS project has the aim of providing tools for users for flash flood and debris flow prediction.  Led by Daniel Sempere at the Universitat Polytechnica de Barcelona, the EU funded project involves teams from France, UK, Netherlands, Spain, Italy and Switzerland, with Case Studies on test bed basins including the Llobgregat and Guadalhorce in Spain; Linth and Verzasca basins in Switzerland; Gardon d’Anduze in France; and Destra Sele near Salerno in Italy.   At the workshop impressive videos of floods in France and elsewhere were shown, include in the Gard in 2002 and the Var in 2010.

Flash floods and debris flows are particularly difficult to forecast.  One definition of a flash flood is that the natural response time of a catchment is smaller than the lead time required for warning (though we were told that even at the scale of the Garonne as it passes through Toulouse, in 1875 the river rose 8 metres in 3 hours).    This means that for forecasts to be useful for warning purposes they need to make use of rainfall forecasts.   The IMPRINTS project is exploring the utility of both radar rainfall projections, numerical weather predictions and the use of similar historical analogues in forecasting flash floods and debris flows.  Both, of course, will be uncertain even for relatively short times and both are being evaluated in terms of ensemble forecasts.

Another challenge for these types of predictions is that such events tend to be rather localised (especially debris flows).   This partly results from local rainfall intensities, but even in some high volume events the response can be strongly conditioned by the initial state of the soil as dependent on past rainfalls, land cover and evapotranspiration and a host of soil properties that are not well known even if a soil map is available.

The final challenge is whether the characteristics of such events are in the process of changing.  Such events are relatively rare in any basin, even if some events occur somewhere in the basins draining to the north Mediterranean nearly every year.   So the data available for model calibration and testing in any application (including the test bed basins) is very limited.   It is therefore even more difficult to test for the impacts of change on such events.

Results so far from the test bed basin applications suggest that the spatial and temporal characteristics of the rainfall forecasts are critical in getting accurate predictions.   The ensemble rainfall forecasts often have a wide range but do not always span what actually happens.   This is not unexpected but means that it might only be able to predict some relative risk over larger areas than local predictions of flash flooding.    In larger basins, such as the Llobregat, it is possible to use data assimilation to improve the predictions with useful lead times, but then these situations do not really fall in the definition of flash floods.

This is really another example of where the dominant uncertainties (in this case the input rainfalls) are not really statistical in nature but result from lack of knowledge (epistemic uncertainties).   The ensemble forecasts of the inputs are then not really probabilistic but are rather potential scenarios.   That really needs to be conveyed to users of such forecasts.   How to do so, is an interesting issue in the IMPRINTS project, common to the issues being addressed by CCN.

With no limitations on the number of beta testers, it’s a bit of a free-for-all, and with facebook’s mobile prowess having come along in leaps and bounds just recently, this will certainly help the improvements continue

Comments are closed.