Search

Classical calibration methods in hydrology are commonly performed with a single cost function computed on long time series. Even though the hydrological model has acceptable scores in NSE and KGE, unbalancing problems can still arise between overall score and the model performance for flood events, and particularly flash floods. Enhancing multi-criteria calibration methods with multi-scale signatures to improve distributed flood modeling remains a challenge. In this study, the potential of hydrological signatures computed continuously and at the scale of flood events on long time series, is employed within various multi-criteria calibration approaches to attain a more efficient hydrological model. This work presents an improved and original signature-based calibration approach, implemented in the variational data assimilation algorithm of SMASH (Spatially distributed Modelling and ASsimilation for Hydrology) platform, applied over 141 catchments mostly located in the French Mediterranean region. Several signatures, especially flood event signatures are firstly computed, relying on a proposed automatic hydrograph segmentation algorithm. Suitable signatures for constraining the model are selected based on their global sensitivity analysis to model parameters. Several multi-criteria calibration strategies with the selected signatures are eventually performed, including a multi-objective optimization approach, and a single-objective optimization approach, that transforms the multi-criteria problem into a single-objective function. Note that in the first approach, the proposed technique based on a simple additive weighting method is used to select an optimal solution obtained from a set of non-inferior solutions. The suggested methods show that, for a global calibration, the average relative error in simulating the peak flow has been dropped from about 0.27 to 0.01-0.08 and from about 0.30 to 0.18-0.21 with various multi-criteria optimization strategies, respectively in calibration and temporal validation. For a distributed calibration, while the average NSE (resp. KGE) still slightly decreases from 0.78 (resp. 0.86) to 0.75 (resp. 0.81) in calibration, the quality of simulated peak flow has been enhanced about 1.5 times in average. In particular, the NSE (resp. KGE) calculated solely on 111 flood events which are picked from 23 downstream gauges has been improved from 0.80 (resp. 0.71) up to 0.83 (resp. 0.78) in median. These results have demonstrated the robustness and delicacy of the model constrained by the signatures for enhancing flash flood forecasting systems.

International audience; The distributed SMASH platform is based on a gridded mesh and on a modular design. On each cell, the model features different hydrological components. Each component offers different modeling options such as snow modules, surface interception, production, transfer and percolation functions. At the grid scale, different routing models are implemented via a cell-to-cell numerical routing scheme. S MASH comes with its numerical adjoint model which is obtained by automatic differentiation with Tapenade. A variational data assimilation algorithm is implemented and helps to calibrate the distributed parameters or evaluate the model states. This algorithm uses the quasi-Newton lbfgs-b descent algorithm and the gradient of the cost function relative to the model parameters and states. This gradient is computed by a run of the adjoint model. T he numerical SMASH platform is a Fortran code. To gain in modularity and facilitate the use of SMASH in the research and engineering communities, a Python interface has been created with the new F90Wrap software. The original Fortran code has been revamped. The new structure enables to 1) control any inputs and outputs with Python, 2) keep an automatically differentiable and computationally efficient numerical Fortran model, 3) call a binary from the shell to preserve a backward compatibility with old practices. T he key to achieve this Python interface is to use Fortran modules and derived types to store all inputs and outputs variables. These Fortran structures are stored in different modules. F 90Wrap automatically generates the fortran functions and wrappers to give access to every component of each derived type. A Python class is generated to facilitate the use of these wrappers inside a Python code. T he Python object "model" aggregates all inputs/outputs variables required by SMASH. The "model" object comes with built-in methods to allow end users to perform simulations, calibrations, plotting and hdf5 export. The Python binding facilitates all post-processing since it does not requires I/O into text files anymore.

The consequences of flash floods can be dramatic in terms of casualties or economic losses. Jonkman (2005), in a global assessment of flood-related casualties, showed that flash floods lead to the highest mortality (number of fatalities divided by the number of affected people). For example, in the recent flash flood that occurred in the French Riviera around Cannes on 3 October 2015, 20 casualties and 650 billion euros of insured damage (source: http://www.ccr.fr/) were reported. Flash flood forecasting systems are critically needed to better organize crisis management and rescue operations.As mentioned in chapter 1.3.4 (Gaume et al. 2016), flash floods are characterized by a rapid increase of river water levels. They often affect small watersheds, generally ungauged. The spatial and temporal variability of rainfall, landscape characteristics and pre-event catchment wetness are important influential factors in flash flood generation, contributing to the large space-time variability of hydrological responses (Borga et al. 2011).Forecasting flash floods is therefore a complex task. It necessitates the monitoring of large areas, where each small watershed of a few square kilometres can possibly be affected. Real-time observation networks and models must run at small temporal and spatial scales, on the order of a few minutes and kilometres. Furthermore, discharge time series are not available for the majority of the possibly affected watersheds, posing a real challenge for model calibration and evaluation. In this context, radar based precipitation products and/or meteorological forecasts with a high resolution (typically 1 km2 grid size) are crucial (Creutin and Borga, 2003). Slight misplacements of the precipitation may for instance lead towarnings attributed to the wrong river network and to inappropriate flood management decisions.

International audience; The opportunities to study and understand flash floods still remain highly limited by the lack of observations. If not destroyed or damaged during the flood events, the existing river gauging station networks are too sparse to capture in detail the spatio-temporal pattern of these floods, generally affecting small rivers. Moreover. the short duration of flash floods (few hours) limits the possibility for conducting real time field observations. In this context, extensive post-event surveys can be of great interest to collect information that can be useful to better understand the phenomena, and to define mitigation measures that may reduce damages in the future: spatio-temporal pattern of the flood, flows based on discharge estimations and interviews of witnesses, delineation of the flooded areas based on high-water marks, detail and causes of damages, circumstances of deaths... The present communication presents the outcomes of the post-event survey conducted after the extreme flash floods that occurred on the 15th and 16th of June 2010 around the town of Draguignan, in the South of France. The floods were induced by an extraordinary rainfall event: rainfall accumulations exceeding 400mm within 24-hours were recorded. It is by far the record daily accumulation for the region according to the data of the existing raingauge network, its estimated return period exceeding largely 100-years. The induced flash floods caused important damages to several towns and 25 casualties. Technical services of the French State and local authorities as well as research institutions involved in the HYMEX project collaborated to the extensive field survey conducted after the floods. The collected data set includes discharge estimates at the gauged stations by water level reconstitution and rating curve extrapolation (half of the hydrological network crashed down during the event), more than 35 peak discharge estimates at various ungauged locations of the stream network, with often an idea of the time of the peak obtained through witnesses’ interviews, and more than 1200 geo-referenced high water marks. This data set provides interesting information about the spatial variability of the hydrological response of the watersheds and about the dynamics of propagation of the flood. Based on this data set, it was possible to reconstitute the flood genesis and also to estimate the possible return periods of the estimated peak discharge values along the river network. The circumstances of the casualties were also analysed, in order to understand how they could have been avoided. Some recommendations – already partly put into action - were formulated to reduce in the future the impact of such an event, such as the strengthening of the gauging network, the development of a flood warning system based on the experience gained through the post event survey, or the application of more constraining building regulations in the exposed areas. The affected area (South Eastern part of France where no extraordinary flash flood had been so deeply documented in the past) and the collaboration between technical services and a large number of research institutions during a post-event survey make of this survey a particularly interesting case study.

National audience; The 1000 municipalities in the region Provence-Alpes-Cote d'Azur are exposed to at least one of the five following hazards: flash floods, mudslides, landslides, avalanches and forest fires. In the three departments of this mountain region (Alpes de Haute Provence, Hautes Alpes and Alpes Maritimes), a hundred of them is exposed to five. The occurrence and intensity of these fluctuations depend more or less directly on the quantity and nature of precipitation. However, in the French Alps, the visibility of the radar national network ARAMIS is considerably reduced by the masking effects of terrain, clutter and altitude of the beam. Accordingly, there is a poor quality for the estimated rainfall. In this context, the project RHYTMME jointly run by Météo-France and Cemagref, provides: - To supplement the existing radar network by new technology radar deployment (radar Doppler polarimetric X-band) in the Southern Alps; - To develop downstream of this network, a service platform to manage natural hydrometeorological risks, especially in the mountains. This article presents the strategy adopted to improve radar coverage in the Southern Alps and relates an example obtained by a first X-band radar information (radar Hydrix) already in service in the Maritime Alps.; Les quelques 1000 communes de la région Provence-Alpes-Côte d'Azur sont exposées à au moins l'un des cinq aléas suivants : crues éclairs, laves torrentielles, mouvements de terrain, avalanches et feux de forêts. Dans les trois départements de montagne de cette région (Alpes de Haute Provence, Hautes Alpes et Alpes Maritimes), une centaine d'entre elles est exposée aux cinq. L'occurrence et l'intensité de ces aléas dépendent plus ou moins directement de la quantité et de la nature des précipitations. Or, sur l'arc alpin français, la visibilité des radars du réseau national ARAMIS est considérablement réduite par les effets de masque du relief, d'échos fixes et d'altitude du faisceau. En conséquence, l'estimation des cumuls de pluie y est de qualité médiocre. Dans ce contexte, le projet RHYTMME, piloté conjointement par Météo-France et le Cemagref, prévoit : - de compléter le réseau de radars existant par le déploiement dans les Alpes du Sud de radars de nouvelle technologie (radars Doppler polarimétriques en bande X) ; - de développer, en aval de ce réseau, une plateforme de services d'aide à la gestion des risques naturels hydrométéorologiques, notamment en montagne. Cet article présente la stratégie adoptée pour améliorer la couverture radar dans les Alpes du sud et rapporte un exemple d'observation de phénomène torrentiel permise par un premier radar en bande X (radar Hydrix) déjà en service dans les Alpes Maritimes.

International audience; The real-time vision of a large portion of French rivers is not a coloured one, not even a black and white, but a blind one, du to the large number of basins which are not monitored. The event of June 2010 in Southern France, where 25 people died during the flood of the Nartuby and Argens rivers gives a dramatic example of this blind vision. In such a context a simple insight, aiming at providing flood alerts rather than pretending to represent exactly the physical process at stake in the catchment, seems welcomed. The aim of this study is to assess the ability of a simple conceptual and parsimonious rainfall-runoff model to deliver flood alerts in ungauged catchments. We use a downward approach: the model is calibrated on gauged catchments and its performances are checked on nested catchments which are not used for the calibration. The model is calibrated over 20 catchments of southern France with areas ranging from 150 km² to 2170 km². The rainfall information is provided by a radar network at an hourly time-step for 15 events that have occurred between 2005 and 2008. The "ungauged" catchments used for the model evaluation are 47 nested basins located within the 20 calibration catchments. We derive the model parameters for these ungauged catchments, from the gauged ones. The model used belongs to the GR models family developed at the Cemagref institute. It works on en event basis thank to a production parameter S and two routing parameters: a shape parameter B and a lag parameter C. Our knowledge of the production parameter benefits from a previous regionalisation work based on several characteristics of the catchments such as the description of their drainage network, the ground-water typology and a climatic variable. The model calibration rests upon the production parameter only. We implement two versions of the model: a lumped one and a distributed one. The results of this downward approach show that the model tested here is able to deliver successful flood alerts in ungauged areas. The distributed version performs slightly better than the lumped version. This study also show that fully distributing the routing parameters results in better alerts than using uniform routing parameters in the distributed version. To fully assess the performance of the model, further evaluation should be carried out in an operational mode, i.e. using an initialisation procedure to determine the value of the S parameter at the beginning of each event instead of using a calibrated value. We anticipate this study by checking how well the calibrated S correlate with the parameter of a continuous soil moisture accounting model run at a daily time-step. The correlation being better in the case of the distributed version of the model, this version seems hence quite promising for an operational use.