Significant challenges in water resources management arise because of the ever-increasing pressure on the world’s heavily exploited and limited water resources. These stressors include demographic growth, intensification of agriculture, climate variability, and climate change. These challenges to water resources are usually tackled using a top-down approach, which suffers from many limitations including the use of a limited set of climate change scenarios, the lack of methodology to rank these scenarios, and the lack of credibility, particularly on extremes. The bottom-up approach, the recently introduced approach, reverses the process by assessing vulnerabilities of water resources systems to variations in future climates and determining the prospects of such wide range of changes. While it solves some issues of the top-down approach, several issues remain unaddressed. The current project seeks to provide end-users and the research community with an improved version of the bottom-up framework for streamlining climate variability into water resources management decisions. The improvement issues that are tackled are a) the generation of a sufficient number of climate projections that provide better coverage of the risk space; b) a methodology to quantitatively estimate the plausibility of a future desired or undesired outcome and c) the optimization of the size of the projections pool to achieve the desired precision with the minimum time and computing resources. The results will hopefully help to cope with the present-day and future challenges induced mainly by climate. In the first part of the study, the adequacy of stochastically generated climate time series for water resources systems risk and performance assessment is investigated. A number of stochastic weather generators (SWGs) are first used to generate a large number of realizations (i.e. an ensemble of climate outputs) of precipitation and temperature time series. Each realization of the generated climate time series is then used individually as an input to a hydrological model to obtain streamflow time series. The usefulness of weather generators is evaluated by assessing how the statistical properties of simulated precipitation, temperatures, and streamflow deviate from those of observations. This is achieved by plotting a large ensemble of (1) synthetic precipitation and temperature time series in a Climate Statistics Space (CSS), and (2) hydrological indices using simulated streamflow data in a Risk and Performance Indicators Space (RPIS). The performance of the weather generator is assessed using visual inspection and the Mahalanobis distance between statistics derived from observations and simulations. A case study was carried out using five different weather generators: two versions of WeaGETS, two versions of MulGETS and the k-nearest neighbor weather generator (knn). In the second part of the thesis, the impacts of climate change, on the other hand, was evaluated by generating a large number of representative climate projections. Large ensembles of future series are created by perturbing downscaled regional climate models’ outputs with a stochastic weather generator, then used as inputs to a hydrological model that was calibrated using observed data. Risk indices calculated with the simulated streamflow data are converted into probability distributions using Kernel Density Estimations. The results are dimensional joint probability distributions of risk-relevant indices that provide estimates of the likelihood of unwanted events under a given watershed configuration and management policy. The proposed approach offers a more complete vision of the impacts of climate change and opens the door to a more objective assessment of adaptation strategies. The third part of the thesis deals with the estimation of the optimal size of SWG realizations needed to calculate risk and performance indices. The number of realizations required to reach is investigated utilizing Relative Root Mean Square Error and Relative Error. While results indicate that a single realization is not enough to adequately represent a given stochastic weather generator, results generally indicate that there is no major benefit of generating more than 100 realizations as they are not notably different from results obtained using 1000 realizations. Adopting a smaller but carefully chosen number of realizations can significantly reduce the computational time and resources and therefore benefit a larger audience particularly where high-performance machines are not easily accessible. The application was done in one pilot watershed, the South Nation Watershed in Eastern Ontario, yet the methodology will be of interest for Canada and beyond. Overall, the results contribute to making the bottom-up more objective and less computationally intensive, hence more attractive to practitioners and researchers.

The interaction between climate and land-surface hydrology is extremely important in relation to long term water resource planning. This is especially so in the presence of global warming and massive land use change, issues which seem likely to have a disproportionate impact on developing countries. This thesis develops tools aimed at the study and prediction of climate effects on land-surface hydrology (in particular streamflow), which require a minimum amount of site specific data. This minimum data requirement allows studies to be performed in areas that are data sparse, such as the developing world. ¶ A simple lumped dynamics-encapsulating conceptual rainfall-runoff model, which explicitly calculates the evaporative feedback to the atmosphere, was developed. It uses the linear streamflow routing module of the rainfall-runoff model IHACRES, with a new non-linear loss module based on the Catchment Moisture Deficit accounting scheme, and is referred to as CMD-IHACRES. In this model, evaporation can be calculated using a number of techniques depending on the data available, as a minimum, one to two years of precipitation, temperature and streamflow data are required. The model was tested on catchments covering a large range of hydroclimatologies and shown to estimate streamflow well. When tested against evaporation data the simplest technique was found to capture the medium to long term average well but had difficulty reproducing the short-term variations. ¶ A comparison of the performance of three limited area climate models (MM5/BATS, MM5/SHEELS and RegCM2) was conducted in order to quantify their ability to reproduce near surface variables. Components of the energy and water balance over the land surface display considerable variation among the models, with no model performing consistently better than the other two. However, several conclusions can be made. The MM5 longwave radiation scheme performed worse than the scheme implemented in RegCM2. Estimates of runoff displayed the largest variations and differed from observations by as much as 100%. The climate models exhibited greater variance than the observations for almost all the energy and water related fluxes investigated. ¶ An investigation into improving these streamflow predictions by utilizing CMD-IHACRES was conducted. Using CMD-IHACRES in an 'offline' mode greatly improved the streamflow estimates while the simplest evaporation technique reproduced the evaporative time series to an accuracy comparable to that obtained from the limited area models alone. The ability to conduct a climate change impact study using CMD-IHACRES and a stochastic weather generator is also demonstrated. These results warrant further investigation into incorporating the rainfall-runoff model CMD-IHACRES in a fully coupled 'online' approach.

The first part of this thesis aims at defining the conditions triggering muddy floods in the Belgian loess belt. On average, each municipality is confronted with 3.6 muddy floods each year. Annual costs associated with their off-site impacts are estimated at € 16-172 millions for the entire Belgian loess belt. A topographic threshold is derived to predict the source areas of muddy floods. Furthermore, the storms required to produce a flood are, on average, smaller in May and June (25 mm) than between July and September (46 mm). This difference is explained by the variability of soil surface characteristics that determine the runoff potential of cultivated soils (soil cover by crops and residues, soil surface crusting and roughness). Steady state infiltration rates of cropland and grassed areas were characterised in the field using a 0.5 m2-portable rainfall simulator. Overall, grassed areas have a lower infiltration rate (16-23 mm h-1) than croplands (25-52 mm h-1). Muddy floods are mostly observed between May-September because of the coincidence of critical soil surface conditions for runoff generation with the most erosive storms. After an adaptation of its decision rules to the local conditions, the STREAM expert-based model provides satisfactory runoff/erosion predictions at the catchment scale. The second part of the thesis aims at evaluating the effectiveness of measures to control muddy floods. A modelling case-study showed that peak discharge was reduced by more than 40% by installing a grassed waterway and a dam at the outlet of a 300 ha-catchment. Monitoring the same catchment (2002-2007) demonstrated that the grassed waterway as well as three dams prevented any muddy flood in the downstream village despite the occurrence of several extreme storms (with a maximum return period of 150 years). Peak discharge was reduced by 69%. Specific sediment yield dropped from 3.5 t ha-1 yr-1 to a mean of 0.5 t ha-1 yr-1 after the installation of the control measures, thereby reducing drastically sediment transfer to the alluvial plain. Finally, a methodology is provided to implement grassed waterways and earthen dams in other dry valleys in the Belgian loess belt and comparable environments.

In order to find effective measures that meet the requirements for proper groundwater quality and quantity management, there is a need to develop a Decision Support System (DSS) and a suitable modelling tool. Central components of a DSS for groundwater management are thought to be models for surface- and groundwater flow and solute transport. The most feasible approach seems to be integration of available mathematical models, and development of a strategy for evaluation of the uncertainty propagation through these models. The physically distributed hydrological model ECOMAG has been integrated with the groundwater model MODFLOW to form a new integrated watershed modelling system - ECOFLOW. The modelling system ECOFLOW has been developed and embedded in Arc View. The multiple-scale modelling principle, combines a more detailed representation of the groundwater flow conditions with lumped watershed modelling, characterised by simplicity in model use, and a minimised number of model parameters. A Bayesian statistical downscaling procedure has also been developed and implemented in the model. This algorithm implies downscaling of the parameters used in the model, and leads to decreasing of the uncertainty level in the modelling results. The integrated model ECOFLOW has been applied to the Vemmenhög catchment, in Southern Sweden, and the Örsundaån catchment, in central Sweden. The applications demonstrated that the model is capable of simulating, with reasonable accuracy, the hydrological processes within both the agriculturally dominated watershed (Vemmenhög) and the forest dominated catchment area (Örsundaån). The results show that the ECOFLOW model adequately predicts the stream and groundwater flow distribution in these watersheds, and that the model can be used as a possible tool for simulation of surface– and groundwater processes on both local and regional scales. A chemical module ECOMAG-N has been created and tested on the Vemmenhög watershed with a highly dense drainage system and intensive fertilisation practises. The chemical module appeared to provide reliable estimates of spatial nitrate loads in the watershed. The observed and simulated nitrogen concentration values were found to be in close agreement at most of the reference points. The proposed future research includes further development of this model for contaminant transport in the surface- and ground water for point and non-point source contamination modelling. Further development of the model will be oriented towards integration of the ECOFLOW model system into a planned Decision Support System. / QC 20101007

Bien que les modèles hydrologiques soient aujourd'hui utilisés pour la prévision des crues, leur mise en place sur un territoire particulier peut encore poser question. Classiquement, elle nécessite des chronologies observées de pluies et de débits. Cependant, ces données peuvent s'avérer incertaines, insuffisantes voire indisponibles. Cela impact alors l'identification des modèles ainsi que leurs prévisions en les rendant elles-mêmes incertaines et potentiellement inopérantes.Dans cette thèse, nous avons cherché à revisiter ces problématiques liées à la mise en place des modèles, en proposant des approches innovantes étalonnant les modèles sans certaines données pouvant s'avérer trop incertaines ou indisponibles. Ces dernières, habituellement exigées pour l'identification des modèles, sont de plus estimées au cours de l'étalonnage. Deux méthodes applicables quelque-soit la nature du modèle considéré sont proposées. La première est une méthode d'inversion des pluies et des paramètres. Elle permet d'étalonner un modèle hydrologique à partir de chronologies observées de débits et de cumuls de pluies seulement. Cette méthode identifie de plus simultanément les chronologies des pluies au pas de temps horaire. Une application particulière de cette méthode à des modèles déjà établis généralise par ailleurs à des modèles non-analytiques l'approche d'inversion des pluies proposée par Kirchner (2009). La seconde est une méthode de calibration pluie-hauteur d'eau. À partir de chronologies observées de pluies et de hauteurs d'eau seulement, elle permet de mettre en place un modèle hydrologique et d'identifier une courbe de tarage. Outre les analyses originales pouvant être menées sur les chronologies de pluies ou les courbes de tarage estimées par ces méthodes, ces dernières étendent à des contextes plus restreints en données, la possibilité d'identification de modèles pluie-débits et offrent des applications en opérationnel.En parallèle du développement de ces méthodes, cette thèse nous a aussi amené à construire une représentation conceptuelle appelée espace des connaissances. Celle-ci permet d'appréhender de manière commune les questions nouvelles ayant été étudiées ainsi que des approches plus classiques comme la calibration ou la simulation. / Rainfall-runoff models are used for flood forecasting by warning authorities. However their implementation on a particular territory is still a challenge. Indeed, they generally need to be calibrated by using observed rainfall and discharge time series. These data may be subject to errors and uncertainties. They are not always available. Then, the model calibration relevancy may be affected and the forecasts may also be subject to significant uncertainties. This research would like to address such issues related to the rainfall-runoff models calibration, by proposing original methods which may set up a model by using less data than the ``classical'' calibration. The unused data might be either subject to uncertainties or not available. Moreover, these unused data may be estimated by the methods. Two model independant approach were suggested. Both are an heuristic inversion algorithm of rainfall-runoff models. The first method estimates simultaneously hourly rainfall time series and models parameters, by using only observed hourly discharge time series and total areal rainfall of flood events. A specific application of this method to set up models (with fixed parameters), generalises to models which are not invertible analyticaly, the ``hydrology backward'' approach proposed by Kirchner (2009).The second method estimates simultaneously models parameters and a rating-curve, by using only observed hourly rainfall and stage time series. Original analysis may be performed on the rainfall time series and the rating-curve estimated by the methods. Also, they extend the applicability rainfall-runoff models to hydrological context with restricted available data and offer promising operational applications. Yet, this research lead us to build a conceptual framework, denoted knowledge space. This framework unifies not only the original approaches which were proposed, but also some more ``classical'' approaches to hydrology as the calibration and the simulation.

This master thesis deals with the influence of temporal distribution of precipitation on hydrological characteristics of the rainfall-runoff process. The practical part of the thesis was carried out on the drainage basin Volšovka. In the theoretical part, the origin, division, and temporal distribution of rainfall is described, as well as surface runoff and some chosen hydrological characteristics of flow-off. Moreover the work deals with hydrological simulation, there is also described in further detail the rainfall-runoff model HEC-HMS, which is furthermore used in the practical part of this thesis. In the practical part, the characteristics of the area of interest is introduced. Next is described the preparation of the rainfall data with the use of DES_RAIN programme, also creation of the digital model of the terrain in the ArcGIS programme and the preparation of hydrological model in HEC-HMS model. In the conclusion of the thesis, the flow capacities of the closure profile of Volšovka were simulated, as well as other particular closure profiles of the catchment area. There were 16 simulations carried out, which were evaluated on the basis of the peak flow capacities and the time of culmination.

Lagoons are coastal waterbodies which are sensitive to meteorological and hydrological changes. This study focused on the spatial distribution of salinity in Gialova lagoon, Greece. The area in which the lagoon is located is under pressure from agriculture and truism in the area. Besides that, the lagoon is an important stop for migratory birds as well as the home for rare species. To investigate how salinity is distributed in the lagoon a salinity gradient was produced. Also, an attempt was made to model the salt and fresh water fluxes in the lagoon using a mass balance approach. The water fluxes are either from fresh water sources or influx of saline water from the sea. Manual electric conductivity measurements were taken during a field campaign to the lagoon. Time series data used in the model was obtained from stations in the lagoon and the surroundings. An investigation in how the lagoon measurement station differ in salinity was also performed in this study. Results from the gradient map and manual measurements show that the water in the lagoon is the freshest in the North-Eastern parts and the most saline to the South-West. The modelled water fluxes show an inverse relationship from each other. High fresh water fluxes correspond to precipitation events, lower salinity concentration. Whereas high salt water fluxes correspond to high salinity concentration and lowered precipitation. The stations concentrated to the middle lagoon show corresponding values in salinity whereas the station to the South-West differs. The canal surrounding the lagoon shows interactions since measurements show that the water is brackish. A longer time series could provide patterns in water fluxes over time. Trying to find the portions of terrestrial groundwater and surface water, and further investigation of the regional aquifer could provide new information to develop this model. The region is expected to experience water stress which makes further studies and monitoring important.

Les eaux résiduaires urbaines du bassin versant du lac du Bourget sont largement collectées et traitées. Les gestionnaires du lac s'interrogent, dans le contexte de l'évolution climatique et de l'évolution des usages des sols, sur l'impact et les modalités de maîtrise des pertes en phosphore et en particulier de la contribution des sources diffuses d'origine agricole. La thèse vise à modéliser les régimes d'exportation du phosphore vers le lac, dans la perspective de tester des scénarios de maîtrise des flux. Elle s'inscrit dans un dispositif de recherche opérationnelle CIFRE entre le bureau d'étude SAFEGE et l'UMR CARRTEL, en partenariat avec le CISALB (Comité Intersyndical pour l'Assainissement du Lac du Bourget). Dans ce cadre, deux sous bassins versants expérimentaux emboîtés représentatifs de la géographie du bassin du Bourget, (l'un rural dominé par l'élevage laitier de 250 ha, l'autre périurbain et agricole de 6800 ha) ont été équipés et suivis à leur exutoire durant une période de 18 mois. Les suivis hydrochimiques ont permis une caractérisation qualitative et quantitative des régimes d'exportations à l'exutoire de chaque bassin versant (flux hydrologique, de matières en suspension, du phosphore total, particulaire, et dissous, et de certaines formes d'azote – nitrate, ammonium). Ces flux ont été modélisés à l'aide du modèle SWAT. Pour rendre compte de certains processus spécifiques de mobilisation et de transfert de la charge en phosphore, des développements ont été mis en œuvre (extension de l'aire de saturation des sols en relation avec la mise en charge des nappes, stockages transitoires, remobilisations au sein du réseau hydrographique) qui constituent une des principales avancées de la thèse. Les résultats obtenus apportent une amélioration des connaissances scientifiques sur les mécanismes de transfert du phosphore. Les développements en termes de modélisation constituent un outil adapté pour accompagner la mise en place d'un plan d'actions visant la réduction des flux de phosphore parvenant au lac du Bourget, ceci dans la perspective des évolutions climatiques. / Urban wastewater are now widely collected and treated. Many questions still remain about the impact of climatic or anthropic evolution on the mobilisation and the transfer of phosphorus to Lake Bourget. This work deals with hydrological and nutrient modelling to understand phosphorus mobilization and transfers processes and to test mitigation options. Thesis is supported by a CIFRE agreement between SAFEGE office and UMR CARRTEL, in partnership with the CISALB. In this context, 2 experimental watersheds nested and variable in land uses (one agricultural of 250 ha, the other suburban and agricultural of 6,800 ha) were instrumented and followed at their outlet over a period of 18 months (continuous measurement of flow and some key parameters of physico-chemical water quality). Hydro-chemical measurements led to a qualitative and quantitative characterization of the exports regime at the outlet of each watershed (water flow, suspended solids, some forms of phosphorus – total, particulate, dissolved - and nitrogen). The exports were then modeled. To account for some specifics processes of mobilization and transfer of phosphorus (variable source areas, "in-stream" processes), some developments of the SWAT model were implemented. IIn this context, this works provides an improved scientific understanding of the mechanisms of transfer of phosphorus. Modeling tool can support the development of mitigation options for best management practices included in the framework of the "Lake Bourget watershed contract".

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior / Green roofs (GRs) have as one of its principal advantages the stormwater retention. To evaluate this quality at different spatial scales, and face to not observed events, hydrological modelling is required. In Brazil, studies related to the hydrological modelling of GRs are incipient. Thus, this research had as main objective the analysis of the efficiency of GRs in the stormwater runoff control at different spatial scales, simulating an urban watershed, where conventional roofs were replaced by GRs. To reach this objective were accomplished the following processes: monitoring of rainfall-runoff events of an experimental GR; calibration of the GR effective rainfall volume by the modified SCS-CN method; development of a rainfall-runoff hydrological model for the GR; design of a micro drainage networks system simulating the watershed without GRs and using design storms; simulation of the watershed with the implementation of GRs to evaluate the impact on the reduction of peak flows and volumes propagated into the drainage networks face to observed rainfalls. Calibrated average CN resulted in 83, staying within the expectations when compared to values found by other authors for GR with similar characteristics. In this calibration was also possible to notice the GR retained in average 45% of each rainfall event volume. In the development of the GR model, initially was calibrated the runoff propagation using models known in the hydrological field: kinematic wave (KW) and the synthetic unit hydrograph of the SCS (SCS-UH). None of these models produced satisfactory results, then was developed a propagation model based in the SCS-UH method, which was called synthetic unit hydrograph of modular GR (MGR-SUH), proposing changes in some of its parameters. MGR-SUH produced satisfactory results, with the Nash-Sutcliffe coefficient resulting in 0,86, which was substantially bigger than the values obtained using the KW and SCS-UH models (0,67). Related to the reduction of peak flow in the micro drainage networks, the implementation of GR in the plots allowed to reduce one commercial diameter in 33 and 52% of the drainage networks stretches, for systems designed to 5 and 10-year return period design rainfall, respectively. Reductions were possible at different scales, both in stretches located at beginning of network, with few blocks contributing, and stretches located at end of network, with several blocks contributing. Besides, it was evaluated the GR effect at the different analyzed scales, with bigger efficiency at small scale (in average, the volume and peak flow reductions at the plot scale were 48 and 57%, respectively), than at block(s) scale, where these reductions were 32 and 38%. Also was found that the biggest reductions happened in events of smallest rainfall volume. Thus, this research allowed to conclude that the effectiveness of the GR in the stormwater runoff control decreases as the scale used increases, and also as increases the volume and intensity of the rainfall. / Os telhados verdes (TVs) têm como uma das suas principais vantagens a retenção de águas pluviais. Para avaliar essa característica em diferentes escalas espaciais, e frente a eventos ainda não monitorados, é imprescindível o uso de modelagem hidrológica. No Brasil, os estudos relacionados à modelagem hidrológica de TVs são incipientes. Portanto, esta pesquisa teve como principal objetivo analisar a eficiência de TVs no controle do escoamento pluvial em diferentes escalas espaciais, a partir de simulações de uma bacia hidrográfica urbana, onde telhados convencionais foram substituídos por TVs. Para isso, foram realizados os seguintes procedimentos: monitoramento de eventos de chuva-vazão de um TV experimental; calibração do volume de chuva efetiva do TV pelo método do CN-SCS modificado; desenvolvimento de um modelo hidrológico do tipo chuva-vazão para o TV; dimensionamento de um sistema de redes de microdrenagem para a bacia simulada sem os TVs utilizando chuvas de projeto; simulação da bacia com uso de TVs para verificar o impacto na redução das vazões de pico e volumes propagados nas redes de drenagem frente a chuvas observadas. O CN médio calibrado para o TV resultou em 83, ficando dentro das expectativas quando comparado com valores encontrados por outros autores para TV com características semelhantes. Nessa calibração também foi possível observar que, em média, o TV reteve 45% do volume de cada evento de chuva. No desenvolvimento do modelo de TV, inicialmente foi calibrada a propagação do escoamento superficial com o uso de modelos conhecidos no meio hidrológico: de onda cinemática (OC) e o hidrograma unitário sintético do SCS (HU-SCS). Como nenhum desses modelos produziu resultados satisfatórios, foi desenvolvido um modelo de propagação baseado na metodologia do HU-SCS, chamado de hidrograma unitário sintético de TV modular (HUS-TVM), sugerindo-se alterações em alguns dos seus parâmetros. O HUS-TVM produziu resultados satisfatórios, com o coeficiente Nash-Sutcliffe resultando em 0,86, substancialmente maior que o obtido com os modelos de OC e HU-SCS (0,67). Com relação à redução das vazões de pico nas redes de microdrenagem, a introdução de TVs nos lotes possibilitou reduzir um diâmetro comercial em 33 e 52% dos trechos das redes de drenagem, para sistemas dimensionados para chuvas de projeto com 5 e 10 anos de período de recorrência, respectivamente. As reduções foram possíveis em diferentes escalas, tanto em trechos de início de rede, com poucos quarteirões contribuintes, como os de final de rede, com vários quarteirões contribuintes. Ainda, verificou-se o efeito do TV nas diferentes escalas analisadas, sendo que sua eficiência foi mais pronunciada em menor escala (em média, as reduções de volume e vazão de pico no lote foram de 48 e 57%, respectivamente), do que em escala de quarteirão(ões), onde essas reduções foram de 32 e 38%. Constatou-se, também, que as maiores reduções ocorreram nos eventos de menor volume de chuva. Portanto, o trabalho permitiu concluir que a efetividade do TV na redução do escoamento pluvial diminui à medida que se aumenta a escala utilizada, e quanto maior é o volume e intensidade da chuva.

Le pourtour du bassin méditerranéen subit fréquemment des épisodes de pluie diluvienne à l’origine de crues rapides pouvant provoquer de nombreuses victimes et des dégâts considérables. Afin de faire face à ce phénomène, la prévision hydrologiques, permettant au Service Central d’Hydrométéorologie et d’Appui à la Prévision des Inondations de produire des vigilances crues, tient une place centrale. Durant les dernières décennies l’efficacité des réseaux de neurones formels pour la prévision des crues rapides a été montrée sur différents bassins versant. Les travaux menés au cours de cette thèse visent à développer une méthodologie générique de mise en œuvre de réseaux de neurones, testée sur les bassins versants du Gardon d’Anduze et du Lez à Lavalette, dont le comportement hydrodynamique est particulièrement non-linéaire. Afin de limiter l’incertitude des performances en prévision en fonction de l’initialisation du modèle, nous avons, dans un premier temps, proposé un modèle d’ensemble, basé sur la médiane à chaque pas de temps des sorties d’un nombre adéquat de modèles variant uniquement par leur initialisation. D’autre part, sur le bassin du Gardon d’Anduze, afin d’améliorer les performances des réseaux de neurones récurrents par l’introduction d’informations sur l’état du bassin versant avant et pendant l’épisode de crue, différentes variables susceptibles de représenter l’état du système ont été introduites successivement afin de sélectionner celles fournissant les modèles les plus performants. Sur le bassin karstique du Lez, dont la structure est très hétérogène, nous avons appliqué la méthode KnoX permettant d’estimer les contributions au débit à l’exutoire de quatre zones aux comportements hydrologique et hydrogéologique considérés comme homogènes. Ainsi les zones les plus contributives ont été identifiées ; ceci permettra dans un second temps de rechercher les variables les mieux à même de représenter l’humidité dans ces zones. Les performances des modèles développés montrent que la méthodologie générale de conception d’un modèle pluie-débit par réseaux de neurones s’adapte de manière satisfaisante aux deux bassins cibles dont les fonctionnements hydrologiques et hydrogéologiques sont pourtant très différents. Certaines pistes de progrès restent à investiguer parmi lesquelles l’amélioration de l’information d’état est prépondérante. / The Mediterranean region is frequently subjected to intense rainfalls leading to flash floods. This phenomenon can cause casualties and huge material damages. Facing to this phenomenon, hydrologic forecasting is a major tool used by Service Central d’Hydrométéorologie et d’Appui à la Prévision des Inondations to produce flood warning.During past decades, artificial neural networks showed their efficiency for flash flood forecasting on different type of watershed. The present thesis aims thus to contribute to the development of a generic methodology to design artificial neural networks, that is tested on Gardon d’Anduze and Lez at Lavalette watersheds, both displaying non-linear hydrodynamic behavior. To reduce uncertainties on forecasts, ensemble models, based on the median of forecasts calculated at each time step for an adequate number of models varying only by their initialization, have been proposed. In addition, in order to improve forecasting performances on Gardon d’Anduze, with artificial neural networks, we tried to introduce knowledge about the state of the watersheds before and during the flood. Several variables have thus been tested each one its turn, to select the one given the best performances. On the Lez karst system, that has a strongly heterogeneous structure, the KnoX method have been applicated in order to estimate the contribution to outflow from four geographical zones displaying hydrologic and hydrogeologic behavior considered as homogeneous. Thus, the most contributive zones to the discharge zones have been identified. This will help the investigation of representing humidity variables in these zones.The performances of models underlined that the general methodology of rainfall-runoff model conception could be applied on both basins, even though their hydrological and hydrogeological behavior are very different.The contribution of each zone, estimated from the KnoX methodology, improved comprehension of Lez karst system during flash floods. Selection of relevant variables representing the state of the Lez hydrosystem will be possible thanks to this new knowledge. Performances of models developed in this study underlined the difficulty to find satisfactory models, and showed the interest of the generic methodology used to design neural network adapted to the two targeted basins.