Crucial need of accurate flood forecasting systems in context of global change

Affecting 2.3 billion people over the 1995-2015 period, floods are the first weather-related disaster in the world [24]. In France, 1 inhabitant over 4 is exposed to this hazard [71] and over the 1982-2019 period, inundations represented in term of insured costs 21 billion of euros, i.e. 55% of the total cost of all natural disasters [16]. With global warming and its consequences [55], insurance companies even estimate that flood-related impact will increase by more than 300% by the end of the century [75]. In this context, as noted in the Sendai Framework for Disaster Risk Reduction (DRR) 2015-2030 [108], effective early flood warning systems are crucially needed, both to facilitate rescue operations and crisis management, and also to strengthen society resilience. Nevertheless, performing accurate flood forecasts in terms of location, magnitude and timing of runoff and flooding, and identifying areas "at risk", i.e. prone to trigger fatalities and economic losses, remains a key challenge.

Challenges faced

Challenge 1. The processes generating floods and their associated damages are multiple and complex. They involve different "coupled" disciplines (meteorology, hydrology, hydraulics and socio-economic sciences) and spatio-temporal scales. Consequently, the available timing to manage the crisis can range from minutes to days, depending if assets at risk are located downstream a small watershed (<1km²) or a large catchment (thousands of km²).

Challenge 2. Furthermore, forecasting models sometimes critically suffer from a lack of data and robustness: they propagate strong uncertainties particularly when they are "extrapolating". This is the case when warnings are needed at high spatial resolutions, on "un-gauged" locations (i.e. location with no observation data), or when forecasts exceed what has been observed in the past. Meanwhile, those "locally-never-observed" events are often the most devastating.

Challenge 3. Last but not least, warning systems are often developed following a "top-down" approach, driven by "hazardbased" modellers, not sufficiently involving end-users. Therefore, exposure, vulnerability and  risk assessments are not always well integrated and operational chains do not enough meet the real operational needs.

In France, even if all regions are affected, the Mediterranean area concentrates all those issues, sometimes during one single event. For example, in 2020, 500mm of rainfall in 24h at Saint-Martin de Vésubie (Alpes-Maritimes) caused both a very fast and destructive local reaction of the un-gauged Boréon small tributary (65 km²), and a large flood of the Var downstream river (2800 km²). In Dec. 2003, up to 300 mm of rainfall resulted in flash-floods of the main downstream tributaries of the Rhône river, and finally caused a large historical flood on the Rhône (10 days of inundation in some settlements of the city of Arles).

In those cases, spatio-temporal variability of rainfall, catchment characteristics, initial soil moisture and finally of hydraulic characteristics, are recognized as important factors in flood generation. However, better quantifying and accounting for this multi-scale variability, from hillslopes to floodplains, in integrated approaches, while taking advantage of the information available at the different spatial and temporal scales to reduce modeling uncertainties, but also reaching high accuracy with affordable computational costs over large domains, remain open challenges. 

Objectives of the project

Therefore the overarching objective of the MUFFINS project is to develop new accurate and computationally efficient flood forecasting approaches, enabling transferring information between modelings (meteo-hydrology-hydraulic-damage) and scales (from local runoff generation over areas lesser than 1-km² to flood propagation on catchments of thousands km²), and taking advantage of innovative data (in situ, remote observation, opportunistic) to reduce forecasts uncertainties.

More precisely, the MUFFINS project is expected to provide significant advances on the following aspects:

• Understanding and specifying the different requirements and expectations of actors having to take decisions at different times: hours before the crisis (i.e.: internal pre-alerts in a rescue service for preparedness actions), during the crisis (i.e.: immediate rescue of people at risk) and hours after (i.e: first loss assessments by insurance companies);
• Design of new methodologies and tools, in adequacy with user’s needs, for next generation of spatially distributed flood warning methods (improved handling of hydro-meteorological inputs in high uncertainty & low predictibility contexts, seamless regionalizable and multi-scale hydrological-hydraulic modeling chains, fast & accurate computations, flood impact mapping, enhanced synergies with multi-source data);
• Integration of information from multi-source data for improving the accuracy of integrated chains and their adequacy with flood impacts mapping;
• Demonstrators of the methods on Mediterranean catchments with multi-scale issues and rich datasets.Developing effective flood forecasting systems is faced with serious challenges

Originality of the MUFFINS project

The originality of the MUFFINS project lies in:

• Developing seamless tools and products at a large range of spatio-temporal scales, meanwhile end-users still have to ’juggle’ with different products, from different sources, at different resolutions in space and time.
• Exploring the potential of newly available data and methods such as machine learning (ML) to improve forecasts, meanwhile actual flood forecasts still suffer from strong uncertainties, especially in the case of "un-gauged" basins.

Indeed, new technologies offer today a huge panel of "under-explorated" information, representing new opportunities for calibrating, validating and improving modeling approaches. Among those new data, we can cite remote sensing data from satellites that are still underused in flood forecasting (rainfall, lightning, soil moisture, snowcover, landcover, temperature/evaporation, ...), flow video from optical sensors (in-situ or embedded on a drone), but also other data issued from existing databases (i.e. provided by insurance companies or rescue services), or issued from webmining (i.e. volunteered geographic information from social networks or crowdsourcing platforms).

Those improvements will be reachable thanks to a multidisciplinary consortium, gathering experts in hydro-meteorology, hydraulics, applied mathematics and hydro-informatics. Those experts will contribute to elaborating robust mathematical and computational frameworks that enable to address the following issues.

Problems investigated

• Seamless probabilistic precipitation inputs for hydrological models. New approaches will be studied in view to for merging QPE/QPFs data in seamless and probabilistic products. This will be achieved by merging current operational products but available at different spatio-temporal scales (raingauges, radar observation/extrapolation, NWP), and adding other potential already existing data (i.e. lightning, satellite...). This will enable to reduce and propagate QPE/QPF uncertainties into hydrological models, by the means of ensemble simulation, in order to better support decisions.
• Distributed hydrological prediction in ungauged basins.  Improving spatially distributed state-flux representation in scalable hydrological models, particularly on small flash flood prone ungauged streams, remains a tough challenge that the MUFFINS project will help to move ahead, by taking advantage of multi-source data integration in distributed modeling with hypothesis testing and in coherence with downstream hydraulic models as explained later.
• Hydrological-hydraulic coupling. Using a rich panel of modeling approaches, the MUFFINS project will leverage advanced couplings of cartographic and dynamic hydrological (SMASH and MARINE for distributed hydrological modeling) and hydraulic models (DassFlow2D, Telemac2D, FLoodos for 2D hydraulic modeling) to study multi-scale hydrological signatures including local runoff generation, flash flood hydrographs in upstream watercourses and potentially fast concurrency effects down to accurate floodplain flows representation.
• River network-floodplains hydraulic modeling. Based on a panel of hydraulic approaches of various complexities/resolution, the MUFFINS project will propose a modern and efficient panel of integrated flood forecasting methods, capable of both high resolution accuracy and fast computation times adapted to crisis management on large domains. The project will enable to benchmark these approaches based on complex multi-scale flood-inundation cases from local to full catchment scale.
• Inverse problems and data fusion in hydrology-hydraulics. The MUFFINS project will enable: to enhance synergies between constraining multi-source data and hydrological-hydraulic models; to improve the models accuracy by solving the encountered inverse problems by developing original mix of ML / DA methods (hybrid AI).
• Modeling exposure and impacts. The involvement of such impacts data and modeling components in the MUFFINS project will help to meet the needs of decision makers, and to provide an end-users oriented evaluation of the new hydrometeorological modeling chains developed in the project.