A previsão e emissão de alertas antecipados constituem um dos principais elementos na prevenção dos impactos ocasionados por eventos de cheias. Uma das formas utilizadas para se obter uma ampliação do horizonte de previsão é através do uso da modelagem chuva-vazão associada à previsão de precipitação, tipicamente derivada de modelos meteorológicos. A precipitação, no entanto, é uma das variáveis que impõe maior dificuldade na previsão meteorológica, sendo considerada uma das principais fontes de incerteza nos resultados da previsão de cheias. A previsão por conjunto é uma técnica originalmente desenvolvida nas ciências atmosféricas e procura explorar as incertezas associadas às condições iniciais e/ou deficiências na estrutura dos modelos meteorológicos com intuito de melhorar sua previsibilidade. A partir de diferentes modelos meteorológicos ou de diferentes condições iniciais de um único modelo, são gerados um conjunto de previsões que representam possíveis trajetórias dos processos atmosféricos ao longo do horizonte de previsão. Pesquisas recentes, principalmente na Europa e Estados Unidos, têm mostrado resultados promissores do acoplamento de previsões meteorológicas por conjunto à modelos hidrológicos para realizar previsões de cheia. Essa pesquisa trata da avaliação do benefício da previsão de cheias por conjunto em curto prazo, em uma bacia de médio porte, utilizando dados e de ferramentas para previsão de vazões disponíveis em modo operacional no Brasil. Como estudo de caso foi utilizada a bacia do Rio Paraopeba (12.150km²), de clima tipicamente tropical, localizada na região sudeste do Brasil. A metodologia proposta para geração das previsões hidrológicas utilizou o modelo hidrológico MGB-IPH alimentado por um conjunto previsões de precipitação de diferentes modelos, com diferentes condições iniciais e parametrizações, dando origem a distintos cenários de previsão de vazões. Como parâmetro de referência na avaliação do desempenho das previsões por conjunto foi utilizada uma previsão hidrológica determinística única, baseada em uma previsão de precipitação obtida da combinação ótima de saídas de diversos modelos meteorológicos. As previsões foram realizadas retrospectivamente no período entre ago/2008 e mai/2011, sendo analisadas durante o período chuvoso dos anos hidrológicos (out-abr). Os resultados das previsões de cheia por conjunto foram avaliados através de uma representação determinística, considerando a média dos membros do conjunto, assim como através de uma representação probabilística, considerando todos os membros, através de medidas de desempenho específicas para esse fim. Na avaliação determinística, a média do conjunto hidrológico apresentou resultados similares aos obtido com a previsão determinística de referência, embora tenha apresentado benefício significativo em relação à maior parte dos membros do conjunto. A avaliação das previsões de cheia por conjunto, por sua vez, mostrou a existência de uma superestimativa e de um subespalhamento dos membros em relação às observações, sobretudo nos primeiros intervalos de tempo da previsão. Na comparação dos resultados das previsões de eventos do tipo dicótomos, que consideram a superação ou não de vazões limites de alerta, o 9º decil das previsões por conjunto mostrou superioridade em relação à previsão determinística de referência e mesmo a média do conjunto, sendo possível obter, na maior parte dos casos analisados, um aumento significativo na proporção de eventos corretamente previstos mantendo as taxas de alarmes falsos em níveis reduzidos. Esse benefício foi, de modo geral, maior para maiores antecedências e vazões limites, situações mais importantes num contexto de prevenção de cheias. Os resultados mostraram ainda que, em média, uma diminuição do número de membros do conjunto diminui seu desempenho nas previsões. / The forecasting and issuing of early warnings represent a key element to prevent the impacts of flood events. An alternative to extend forecasting horizon is the use of rainfall-runoff modeling coupled with precipitation forecasts derived from numerical weather prediction (NWP) models. However, NWP models have difficulty to accurately predict precipitation due to the extremely sensitivity of the initial conditions. Therefore, this variable represents one of the major sources of uncertainties in flood forecasting. A probabilistic or ensemble forecasting approach was originally developed in the atmospheric sciences and then applied to other research areas. This procedure explores the uncertainties related to initial conditions and deficiencies in the structure of NWP models intending to improve its predictability. Using different NWP models or different initial conditions of a single model, an ensemble forecast showing possible trajectories of atmospheric processes over the forecast horizon are produced. Recent studies developed in Europe and the United States have shown promising results in flood forecasting using hydrological models fed by NWP ensemble outputs. The present research assess the performance of short term ensemble flood forecasting in a medium size tropical basin, based on data and streamflow forecasting tools available in operational mode in Brazil. The Paraopeba River basin (12,150 km²), located in the upper portion of the São Francisco River basin, in Southeastern Brazil, was selected as a case study. The proposed methodology used the MGB-IPH hydrological coupled to an ensemble of precipitation forecasts generated by several models with different initial conditions and parameterizations. The results are several scenarios of streamflow forecasts. A single deterministic streamflow forecast, based on a quantitative precipitation forecast derived from the optimal combination of several outputs of NWP models, was used as a reference to assess the performance of the streamflow ensemble forecasts. The streamflow forecasts were performed between aug/2008 and may/2011 and were analyzed during the rainy seasons (austral summer). The results from the ensemble flood forecasting were assessed by deterministic and probabilistic performance measures, with the ensemble mean being used by the former, and specific assessment measure by the later. Based on the deterministic assessment, the ensemble mean showed similar results to those obtained by the deterministic reference forecast, although showing better performance over most of the ensemble members. Based on the probabilistic performance measures, however, results showed the existence of an ensemble overforecasting and underspread of the members in regard to observed values, especially during the first lead times. The results for predictions of dichotomous events, which mean exceeding or not flood warning thresholds, showed that the 9th decile of the ensemble over performed the deterministic forecast and even the ensemble mean. In most cases, it was observed an increase in the proportion of correctly forecasted events while keeping false alarm rates at low levels. This benefit was generally higher for higher flow thresholds and for longer lead times, which are the most important situations for flood mitigation. The results show, also, that, in average, a reduction in the number of ensemble members decreases the performance of ensemble flood forecasts.

A previsão hidrológica possibilita a identificação antecipada de eventos de cheia potencialmente causadores de inundação, o que é de grande importância para a atuação de entidades como a Defesa Civil. Quando se deseja estender a antecedência no tempo em relação a estes eventos, principalmente nos casos onde a bacia de interesse é relativamente rápida, torna-se necessária a incorporação de previsões quantitativas de precipitação (QPF) na modelagem hidrológica, as quais podem ser obtidas a partir de modelos numéricos de previsão do tempo. Entretanto, a falta de acurácia atribuída a estas previsões de chuva, dadas de forma determinística, vem promovendo sua substituição por sistemas de previsão meteorológica por conjunto (EPS - Ensemble Prediction Systems), cuja finalidade é a geração de possíveis estados futuros da atmosfera para considerar as incertezas associadas ao seu estado inicial e às deficiências na representação física dos modelos de previsão do tempo. Neste contexto, o presente estudo teve por objetivo avaliar uma metodologia de previsão de cheias por conjunto na bacia do Taquari-Antas/RS até a cidade de Encantado (19.000 km²), localizada na região Sul do Brasil. Para tanto, foi utilizado o modelo hidrológico MGB-IPH acoplado a diferentes sistemas de previsão, sendo eles: (i) EPS Regional ETA, de curto prazo (até 72 horas) com 5 membros de diferentes parametrizações; (ii) EPS Global ECMWF de médio prazo (até 10 dias) com 50 membros de condições iniciais perturbadas, incluindo perturbação estocástica nos parâmetros de ajuste do modelo e; (iii) Previsão Determinística do Modelo Regional ETA (até 7 dias). A avaliação das previsões consistiu em dois hindcastings distintos, envolvendo uma análise visual de eventos singulares ocorridos em 06/06/2014 e 21/07/2011 além de uma análise estatística no período de Mar/2014 - Nov/2014. Durante a análise visual foi possível identificar, a partir de antecedências de 5 a 6 dias, uma persistência na previsão dada pelo crescente número de membros do conjunto de médio prazo (ECMWF) com superação dos limiares de referência, na medida em que se aproximavam os eventos de cheia. Apesar da grande incerteza na magnitude das previsões hidrológicas para o conjunto de curto prazo, a vazão máxima foi relativamente bem prevista por pelo menos 1 membro em quase todas as antecedências, enquanto que a previsão do timing dos eventos foi considerada de boa confiabilidade. Durante a avaliação estatística foi possível notar uma falta de espalhamento nos conjuntos, com tendência de subestimativa de acordo com o aumento da antecedência. Em uma comparação com previsões determinísticas, as previsões por conjunto demonstraram maior acurácia principalmente até 72 horas de antecedência, com destaque para a maior probabilidade de detecção dos limiares de referência e manutenção de falso alarme a níveis reduzidos. Além disso, verificou-se também que a agregação de previsões efetuadas em tempo anterior naquelas atuais acarreta em ligeira ampliação do espalhamento do conjunto e maiores probabilidades de detecção dos limiares de alerta para os membros mais elevados, apesar da redução no desempenho em termos de acurácia e viés. De forma geral, as previsões por conjunto apresentam potencial para servir como uma informação complementar em sistemas de alerta contra cheias, possibilitando uma melhor preparação dos agentes envolvidos durante a ocorrência destes eventos. / Hydrological forecasting plays an important role for issuing flood warnings, allowing for anticipation and better preparation of authorities at the occurrence of such events. In order to extend lead time in a flood forecast, especially when the catchment response time is relatively fast, it may be useful to couple a hydrological model to quantitative precipitation forecasts (QPF), usually obtained directly from numerical weather prediction (NWP). However, deterministic (i.e. single) QPF are usually referred to many errors and lack of accuracy, mainly caused by uncertainties on initial state of the atmosphere and on physical representation of weather forecasting models. To address these shortcomings, it becomes necessary to take into account the uncertainties associated to rain forecasts, which can be represented by Ensemble Prediction Systems (EPS). The purpose of such systems is to provide different trajectories of the atmosphere by perturbations on its initial condition and on parameterization schemes of the models, generating an ensemble of forecasts that can be used as input to hydrological modelling (HEPS). In this context, the present study aimed to assess a methodology of ensemble flood forecasting on Taquari-Antas basin up to the city of Encantado/RS (19.000 km²), located in southern Brazil. Therefore, the hydrological model MGB-IPH was coupled to different forecasting systems: (i) Short Range EPS ETA (up to 72 hours), a regional model with 5 members of different parameterization schemes; (ii) Medium Range EPS ECMWF (up to 10 days), a global model with 50 members of perturbed initial conditions and stochastic perturbation in the model parameters; (iii) Deterministic ETA Model (up to 7 days). The forecasts were evaluated by two different hindcastings, which includes a visual assessment of singular events occurred on 2011 and 2014 and a statistical analysis for the period between Mar/2014 and Nov/2014. It was possible to identify a forecast persistence on medium-range for the selected events, by the increasing number of members exceeding the reference thresholds from lead times up to 5 - 6 days. On the short range, although large uncertainties in the magnitude of hydrological forecasts were found, the peak discharge was well forecasted - at least for a single member - in nearly all lead times, whereas the prediction of the peak timing was considered reliable. Regarding to statistical evaluation, an inadequate spread in the ensemble was observed from short- to medium-range, with a tendency of underestimation for longer lead times. In a comparison with deterministic forecasts, the ensemble forecasts showed higher accuracy especially up to 72 hours in advance, including highlights on greater probability of detection (POD) above the reference thresholds even with low false alarm rates. It also was found that the a combination of previous forecasts on the recent ones leads to a slight increase of ensemble spread and POD for higher members, despite the performance reduction in terms of accuracy and bias. In summary, the hydrological ensemble forecasts demonstrated a good potential to serve as an additional information within a Flood Alert System.

A previsão e emissão de alertas antecipados constituem um dos principais elementos na prevenção dos impactos ocasionados por eventos de cheias. Uma das formas utilizadas para se obter uma ampliação do horizonte de previsão é através do uso da modelagem chuva-vazão associada à previsão de precipitação, tipicamente derivada de modelos meteorológicos. A precipitação, no entanto, é uma das variáveis que impõe maior dificuldade na previsão meteorológica, sendo considerada uma das principais fontes de incerteza nos resultados da previsão de cheias. A previsão por conjunto é uma técnica originalmente desenvolvida nas ciências atmosféricas e procura explorar as incertezas associadas às condições iniciais e/ou deficiências na estrutura dos modelos meteorológicos com intuito de melhorar sua previsibilidade. A partir de diferentes modelos meteorológicos ou de diferentes condições iniciais de um único modelo, são gerados um conjunto de previsões que representam possíveis trajetórias dos processos atmosféricos ao longo do horizonte de previsão. Pesquisas recentes, principalmente na Europa e Estados Unidos, têm mostrado resultados promissores do acoplamento de previsões meteorológicas por conjunto à modelos hidrológicos para realizar previsões de cheia. Essa pesquisa trata da avaliação do benefício da previsão de cheias por conjunto em curto prazo, em uma bacia de médio porte, utilizando dados e de ferramentas para previsão de vazões disponíveis em modo operacional no Brasil. Como estudo de caso foi utilizada a bacia do Rio Paraopeba (12.150km²), de clima tipicamente tropical, localizada na região sudeste do Brasil. A metodologia proposta para geração das previsões hidrológicas utilizou o modelo hidrológico MGB-IPH alimentado por um conjunto previsões de precipitação de diferentes modelos, com diferentes condições iniciais e parametrizações, dando origem a distintos cenários de previsão de vazões. Como parâmetro de referência na avaliação do desempenho das previsões por conjunto foi utilizada uma previsão hidrológica determinística única, baseada em uma previsão de precipitação obtida da combinação ótima de saídas de diversos modelos meteorológicos. As previsões foram realizadas retrospectivamente no período entre ago/2008 e mai/2011, sendo analisadas durante o período chuvoso dos anos hidrológicos (out-abr). Os resultados das previsões de cheia por conjunto foram avaliados através de uma representação determinística, considerando a média dos membros do conjunto, assim como através de uma representação probabilística, considerando todos os membros, através de medidas de desempenho específicas para esse fim. Na avaliação determinística, a média do conjunto hidrológico apresentou resultados similares aos obtido com a previsão determinística de referência, embora tenha apresentado benefício significativo em relação à maior parte dos membros do conjunto. A avaliação das previsões de cheia por conjunto, por sua vez, mostrou a existência de uma superestimativa e de um subespalhamento dos membros em relação às observações, sobretudo nos primeiros intervalos de tempo da previsão. Na comparação dos resultados das previsões de eventos do tipo dicótomos, que consideram a superação ou não de vazões limites de alerta, o 9º decil das previsões por conjunto mostrou superioridade em relação à previsão determinística de referência e mesmo a média do conjunto, sendo possível obter, na maior parte dos casos analisados, um aumento significativo na proporção de eventos corretamente previstos mantendo as taxas de alarmes falsos em níveis reduzidos. Esse benefício foi, de modo geral, maior para maiores antecedências e vazões limites, situações mais importantes num contexto de prevenção de cheias. Os resultados mostraram ainda que, em média, uma diminuição do número de membros do conjunto diminui seu desempenho nas previsões. / The forecasting and issuing of early warnings represent a key element to prevent the impacts of flood events. An alternative to extend forecasting horizon is the use of rainfall-runoff modeling coupled with precipitation forecasts derived from numerical weather prediction (NWP) models. However, NWP models have difficulty to accurately predict precipitation due to the extremely sensitivity of the initial conditions. Therefore, this variable represents one of the major sources of uncertainties in flood forecasting. A probabilistic or ensemble forecasting approach was originally developed in the atmospheric sciences and then applied to other research areas. This procedure explores the uncertainties related to initial conditions and deficiencies in the structure of NWP models intending to improve its predictability. Using different NWP models or different initial conditions of a single model, an ensemble forecast showing possible trajectories of atmospheric processes over the forecast horizon are produced. Recent studies developed in Europe and the United States have shown promising results in flood forecasting using hydrological models fed by NWP ensemble outputs. The present research assess the performance of short term ensemble flood forecasting in a medium size tropical basin, based on data and streamflow forecasting tools available in operational mode in Brazil. The Paraopeba River basin (12,150 km²), located in the upper portion of the São Francisco River basin, in Southeastern Brazil, was selected as a case study. The proposed methodology used the MGB-IPH hydrological coupled to an ensemble of precipitation forecasts generated by several models with different initial conditions and parameterizations. The results are several scenarios of streamflow forecasts. A single deterministic streamflow forecast, based on a quantitative precipitation forecast derived from the optimal combination of several outputs of NWP models, was used as a reference to assess the performance of the streamflow ensemble forecasts. The streamflow forecasts were performed between aug/2008 and may/2011 and were analyzed during the rainy seasons (austral summer). The results from the ensemble flood forecasting were assessed by deterministic and probabilistic performance measures, with the ensemble mean being used by the former, and specific assessment measure by the later. Based on the deterministic assessment, the ensemble mean showed similar results to those obtained by the deterministic reference forecast, although showing better performance over most of the ensemble members. Based on the probabilistic performance measures, however, results showed the existence of an ensemble overforecasting and underspread of the members in regard to observed values, especially during the first lead times. The results for predictions of dichotomous events, which mean exceeding or not flood warning thresholds, showed that the 9th decile of the ensemble over performed the deterministic forecast and even the ensemble mean. In most cases, it was observed an increase in the proportion of correctly forecasted events while keeping false alarm rates at low levels. This benefit was generally higher for higher flow thresholds and for longer lead times, which are the most important situations for flood mitigation. The results show, also, that, in average, a reduction in the number of ensemble members decreases the performance of ensemble flood forecasts.

abstract: This doctoral thesis investigates the predictability characteristics of floods and flash floods by coupling high resolution precipitation products to a distributed hydrologic model. The research hypotheses are tested at multiple watersheds in the Colorado Front Range (CFR) undergoing warm-season precipitation. Rainfall error structures are expected to propagate into hydrologic simulations with added uncertainties by model parameters and initial conditions. Specifically, the following science questions are addressed: (1) What is the utility of Quantitative Precipitation Estimates (QPE) for high resolution hydrologic forecasts in mountain watersheds of the CFR?, (2) How does the rainfall-reflectivity relation determine the magnitude of errors when radar observations are used for flood forecasts?, and (3) What are the spatiotemporal limits of flood forecasting in mountain basins when radar nowcasts are used into a distributed hydrological model?. The methodology consists of QPE evaluations at the site (i.e., rain gauge location), basin-average and regional scales, and Quantitative Precipitation Forecasts (QPF) assessment through regional grid-to-grid verification techniques and ensemble basin-averaged time series. The corresponding hydrologic responses that include outlet discharges, distributed runoff maps, and streamflow time series at internal channel locations, are used in light of observed and/or reference data to diagnose the suitability of fusing precipitation forecasts into a distributed model operating at multiple catchments. Results reveal that radar and multisensor QPEs lead to an improved hydrologic performance compared to simulations driven with rain gauge data only. In addition, hydrologic performances attained by satellite products preserve the fundamental properties of basin responses, including a simple scaling relation between the relative spatial variability of runoff and its magnitude. Overall, the spatial variations contained in gridded QPEs add value for warm-season flood forecasting in mountain basins, with sparse data even if those products contain some biases. These results are encouraging and open new avenues for forecasting in regions with limited access and sparse observations. Regional comparisons of different reflectivity -rainfall (Z-R) relations during three summer seasons, illustrated significant rainfall variability across the region. Consistently, hydrologic errors introduced by the distinct Z-R relations, are significant and proportional (in the log-log space) to errors in precipitation estimations and stream flow magnitude. The use of operational Z-R relations without prior calibration may lead to wrong estimation of precipitation, runoff magnitude and increased flood forecasting errors. This suggests that site-specific Z-R relations, prior to forecasting procedures, are desirable in complex terrain regions. Nowcasting experiments show the limits of flood forecasting and its dependence functions of lead time and basin scale. Across the majority of the basins, flood forecasting skill decays with lead time, but the functional relation depends on the interactions between watershed properties and rainfall characteristics. Both precipitation and flood forecasting skills are noticeably reduced for lead times greater than 30 minutes. Scale dependence of hydrologic forecasting errors demonstrates reduced predictability at intermediate-size basins, the typical scale of convective storm systems. Overall, the fusion of high resolution radar nowcasts and the convenient parallel capabilities of the distributed hydrologic model provide an efficient framework for generating accurate real-time flood forecasts suitable for operational environments. / Dissertation/Thesis / Ph.D. Civil and Environmental Engineering 2012

A previsão e emissão de alertas antecipados constituem um dos principais elementos na prevenção dos impactos ocasionados por eventos de cheias. Uma das formas utilizadas para se obter uma ampliação do horizonte de previsão é através do uso da modelagem chuva-vazão associada à previsão de precipitação, tipicamente derivada de modelos meteorológicos. A precipitação, no entanto, é uma das variáveis que impõe maior dificuldade na previsão meteorológica, sendo considerada uma das principais fontes de incerteza nos resultados da previsão de cheias. A previsão por conjunto é uma técnica originalmente desenvolvida nas ciências atmosféricas e procura explorar as incertezas associadas às condições iniciais e/ou deficiências na estrutura dos modelos meteorológicos com intuito de melhorar sua previsibilidade. A partir de diferentes modelos meteorológicos ou de diferentes condições iniciais de um único modelo, são gerados um conjunto de previsões que representam possíveis trajetórias dos processos atmosféricos ao longo do horizonte de previsão. Pesquisas recentes, principalmente na Europa e Estados Unidos, têm mostrado resultados promissores do acoplamento de previsões meteorológicas por conjunto à modelos hidrológicos para realizar previsões de cheia. Essa pesquisa trata da avaliação do benefício da previsão de cheias por conjunto em curto prazo, em uma bacia de médio porte, utilizando dados e de ferramentas para previsão de vazões disponíveis em modo operacional no Brasil. Como estudo de caso foi utilizada a bacia do Rio Paraopeba (12.150km²), de clima tipicamente tropical, localizada na região sudeste do Brasil. A metodologia proposta para geração das previsões hidrológicas utilizou o modelo hidrológico MGB-IPH alimentado por um conjunto previsões de precipitação de diferentes modelos, com diferentes condições iniciais e parametrizações, dando origem a distintos cenários de previsão de vazões. Como parâmetro de referência na avaliação do desempenho das previsões por conjunto foi utilizada uma previsão hidrológica determinística única, baseada em uma previsão de precipitação obtida da combinação ótima de saídas de diversos modelos meteorológicos. As previsões foram realizadas retrospectivamente no período entre ago/2008 e mai/2011, sendo analisadas durante o período chuvoso dos anos hidrológicos (out-abr). Os resultados das previsões de cheia por conjunto foram avaliados através de uma representação determinística, considerando a média dos membros do conjunto, assim como através de uma representação probabilística, considerando todos os membros, através de medidas de desempenho específicas para esse fim. Na avaliação determinística, a média do conjunto hidrológico apresentou resultados similares aos obtido com a previsão determinística de referência, embora tenha apresentado benefício significativo em relação à maior parte dos membros do conjunto. A avaliação das previsões de cheia por conjunto, por sua vez, mostrou a existência de uma superestimativa e de um subespalhamento dos membros em relação às observações, sobretudo nos primeiros intervalos de tempo da previsão. Na comparação dos resultados das previsões de eventos do tipo dicótomos, que consideram a superação ou não de vazões limites de alerta, o 9º decil das previsões por conjunto mostrou superioridade em relação à previsão determinística de referência e mesmo a média do conjunto, sendo possível obter, na maior parte dos casos analisados, um aumento significativo na proporção de eventos corretamente previstos mantendo as taxas de alarmes falsos em níveis reduzidos. Esse benefício foi, de modo geral, maior para maiores antecedências e vazões limites, situações mais importantes num contexto de prevenção de cheias. Os resultados mostraram ainda que, em média, uma diminuição do número de membros do conjunto diminui seu desempenho nas previsões. / The forecasting and issuing of early warnings represent a key element to prevent the impacts of flood events. An alternative to extend forecasting horizon is the use of rainfall-runoff modeling coupled with precipitation forecasts derived from numerical weather prediction (NWP) models. However, NWP models have difficulty to accurately predict precipitation due to the extremely sensitivity of the initial conditions. Therefore, this variable represents one of the major sources of uncertainties in flood forecasting. A probabilistic or ensemble forecasting approach was originally developed in the atmospheric sciences and then applied to other research areas. This procedure explores the uncertainties related to initial conditions and deficiencies in the structure of NWP models intending to improve its predictability. Using different NWP models or different initial conditions of a single model, an ensemble forecast showing possible trajectories of atmospheric processes over the forecast horizon are produced. Recent studies developed in Europe and the United States have shown promising results in flood forecasting using hydrological models fed by NWP ensemble outputs. The present research assess the performance of short term ensemble flood forecasting in a medium size tropical basin, based on data and streamflow forecasting tools available in operational mode in Brazil. The Paraopeba River basin (12,150 km²), located in the upper portion of the São Francisco River basin, in Southeastern Brazil, was selected as a case study. The proposed methodology used the MGB-IPH hydrological coupled to an ensemble of precipitation forecasts generated by several models with different initial conditions and parameterizations. The results are several scenarios of streamflow forecasts. A single deterministic streamflow forecast, based on a quantitative precipitation forecast derived from the optimal combination of several outputs of NWP models, was used as a reference to assess the performance of the streamflow ensemble forecasts. The streamflow forecasts were performed between aug/2008 and may/2011 and were analyzed during the rainy seasons (austral summer). The results from the ensemble flood forecasting were assessed by deterministic and probabilistic performance measures, with the ensemble mean being used by the former, and specific assessment measure by the later. Based on the deterministic assessment, the ensemble mean showed similar results to those obtained by the deterministic reference forecast, although showing better performance over most of the ensemble members. Based on the probabilistic performance measures, however, results showed the existence of an ensemble overforecasting and underspread of the members in regard to observed values, especially during the first lead times. The results for predictions of dichotomous events, which mean exceeding or not flood warning thresholds, showed that the 9th decile of the ensemble over performed the deterministic forecast and even the ensemble mean. In most cases, it was observed an increase in the proportion of correctly forecasted events while keeping false alarm rates at low levels. This benefit was generally higher for higher flow thresholds and for longer lead times, which are the most important situations for flood mitigation. The results show, also, that, in average, a reduction in the number of ensemble members decreases the performance of ensemble flood forecasts.

Les inondations sont un risque naturel majeur pour les biens et les personnes. Prévoir celles-ci, informer le grand public et les autorités sont de la responsabilité des services de prévision des crues. Pour ce faire ils disposent d'observations in situ et de modèles numériques. Néanmoins les modèles numériques sont une représentation simplifiée et donc entachée d'erreur de la réalité. Les observations quant à elle fournissent une information localisée et peuvent être également entachées d'erreur. Les méthodes d'assimilation de données consistent à combiner ces deux sources d'information et sont utilisées pour réduire l'incertitude sur la description de l'état hydraulique des cours d'eau et améliorer les prévisisons. Ces dernières décennies l'assimilation de données a été appliquée avec succès à l'hydraulique fluviale pour l'amélioration des modèles et pour la prévision des crues. Cependant le développement de méthodes d'assimilation pour la prévision en temps réel est contraint par le temps de calcul disponible et par la conception de la chaîne opérationnelle. Les méthodes en question doivent donc être performantes, simples à implémenter et peu coûteuses. Un autre défi réside dans la combinaison des modèles hydrauliques de dimensions différentes développés pour décrire les réseaux hydrauliques. Un modèle 1D est peu coûteux mais ne permet pas de décrire des écoulement complexes, contrairement à un modèle 2D. Le simple chainage des modèles 1D et 2D avec échange des conditions aux limites n'assure pas la continuité de l'état hydraulique. Il convient alors de coupler les modèles, tout en limitant le coût de calcul. Cette thèse a été financée par la région Midi-Pyrénées et le SCHAPI (Service Central d'Hydrométéorolgie et d'Appui à la Prévisions des Inondations) et a pour objectif d'étudier l'apport de l'assimilation de données et du couplage de modèles pour la prévision des crues. Elle se décompose en deux axes : Un axe sur l'assimilation de données. On s'intéresse à l'émulation du filtre de Kalman d'Ensemble (EnKF) sur le modèle d'onde de crue. On montre, sous certaines hypothèses, qu'on peut émuler l'EnKF avec un filtre de Kalman invariant pour un coût de calcul réduit. Dans un second temps nous nous intéressons à l'application de l'EnKF sur l'Adour maritime avec un modèle Saint-Venant. Nous en montrons les limitations dans sa version classique et montrons les avantages apportés par des méthodes complémentaires d'inflation et d'estimation des covariances d'erreur d'observation. L'apport de l'assimilation des données in situ de hauteurs d'eau sur des cas synthétiques et sur des crues réelles a été démontré et permet une correction spatialisée des hauteurs d'eau et des débits. En conséquence, on constate que les prévisions à court terme sont améliorées. Nous montrons enfin qu'un système de prévisions probabilistes sur l'Adour dépend de la connaissance que l'on a des forçages amonts ; un axe sur le couplage de modèles hydrauliques. Sur l'Adour 2 modèles co-existent : un modèle 1D et un modèle 2D au niveau de Bayonne. Deux méthodes de couplage ont été implémentées. Une première méthode, dite de "couplage à interfaces", combine le 1D décomposé en sous-modèles couplés au 2D au niveau frontières liquides de ce dernier. Une deuxième méthode superpose le 1D avec le 2D sur la zone de recouvrement ; le 1D force le 2D qui, quand il est en crue, calcule les termes d'apports latéraux pour le 1D, modélisant les échanges entre lit mineur et lit majeur. Le coût de calcul de la méthode par interfaces est significativement plus élevé que celui associé à la méthode de couplage par superposition, mais assure une meilleure continuité des variables. En revanche, la méthode de superposition est immédiatement compatible avec l'approche d'assimilation de données sur la zone 1D. / Floods represent a major threat for people and society. Flood forecasting agencies are in charge of floods forecasting, risk assessment and alert to governmental authorities and population. To do so, flood forecasting agencies rely on observations and numerical models. However numerical models and observations provide an incomplete and inexact description of reality as they suffer from various sources of uncertianties. Data assimilation methods consists in optimally combining observations with models in order to reduce both uncertainties in the models and in the observations, thus improving simulation and forecast. Over the last decades, the merits of data assimilation has been greatly demonstrated in the field of hydraulics and hydrology, partly in the context of model calibration or flood forecasting. Yet, the implementation of such methods for real application, under computational cost constraints as well as technical constraints remains a challenge. An other challenge arises when the combining multidimensional models developed over partial domains of catchment. For instance, 1D models describe the mono-dimensional flow in a river while 2D model locally describe more complex flows. Simply chaining 1D and 2D with boundary conditions exchange does not suffice to guarantee the coherence and the continuity of both water level and discharge variables between 1D and 2D domains. The solution lies in dynamical coupling of 1D and 2D models, yet an other challenge when computational cost must be limited. This PhD thesis was funded by Midi-Pyrénées region and the french national agency for flood forecasting SCHAPI. It aims at demonstrating the merits of data assimilation and coupling methods for floof forecasting in the framework of operational application. This thesis is composed of two parts : A first part dealing with data assimilation. It was shown that, under some simplifying assumptions, the Ensemble Kalman filter algorithm (EnKF) can be emulated with a cheaper algorithm : the invariant Kalman filter. The EnKF was then implemented ovr the "Adour maritime" hydraulic network on top of the MASCARET model describing the shallow water equations. It was found that a variance inflation algorithm can further improve data assimlation results with the EnKF. It was shown on synthetical and real cases experiments that data assimilation provides an hydraulic state that is in great agreement with water level observations. As a consequence of the sequential correction of the hydraulic state over time, the forecasts were also greatly improved by data assimilation over the entire hydraulic network for both assimilated and nonassimilated variables, especially for short term forecasts. It was also shown that a probabilistic prediction system relies on the knowledge on the upstream forcings ; A second part focusses on hydraulic models coupling. While the 1D model has a great spatial extension and describes the mono-dimensional flow, the 2D model gives a focus on the Adour-Nive confluence in the Bayonne area. Two coupling methods have been implemented in this study : a first one based on the exchange of the state variables at the liquid boundaries of the models and a second one where the models are superposed. While simple 1D or chained 1D-2D solutions provide an incomplete or discontinuous description of the hydraulic state, both coupling methods provide a full and dynamically coherent description of water level and discharge over the entire 1D-2D domain. On the one hand, the interface coupling method presents a much higher computational cost than the superposition methods but the continuity is better preserved. On the other hand, the superposition methods allows to combine data assimilation of the 1D model and 1D-2D coupling. The positive impact of water level in-situ observations in the 1D domain was illustrated over the 2D domain for a flood event in 2014.

abstract: The National Oceanic and Atmospheric Administration (NOAA)’s National Water Model (NWM) will provide the next generation of operational streamflow forecasts at different lead times across United States using the Weather Research and Forecasting (WRF)-Hydro hydrologic system. These forecasts are crucial for flood protection agencies and water utilities, including the Salt River Project (SRP). The main goal of this study is to calibrate WRF-Hydro in the Oak Creek Basin (OCB; ~820 km2), an unregulated mountain sub-watershed of the Salt and Verde River basins in Central Arizona, whose water resources are managed by SRP and crucial for the Phoenix Metropolitan area. As in the NWM, WRF-Hydro was set up at 1-km (250-m) resolution for the computation of the rainfall-runoff (routing) processes. Model forcings were obtained by bias correcting meteorological data from the North American Land Data Assimilation System-2 (NLDAS-2). A manual calibration approach was designed that targets, in sequence, the sets of model parameters controlling four main processes responsible for streamflow and flood generation in the OCB. After a first calibration effort, it was found that WRF-Hydro is able to simulate runoff generated after snowmelt and baseflow, as well as magnitude and timing of flood peaks due to winter storms. However, the model underestimates the magnitude of flood peaks caused by summer thunderstorms, likely because these storms are not captured by NLDAS-2. To circumvent this, a seasonal modification of soil parameters was adopted. When doing so, acceptable model performances were obtained during calibration (2008-2011) and validation (2012-2017) periods (NSE > 0.62 and RMSE = ~2.5 m3/s at the daily time scale). The process-based calibration strategy utilized in this work provides a new approach to identify areas of structural improvement for WRF-Hydro and the NWM. / Dissertation/Thesis / Masters Thesis Civil, Environmental and Sustainable Engineering 2020

Key theoretical and empirical results from the past two decades have established that peak discharges exhibit power-law, or scaling, relation with drainage area across multiple scales of time and space. This relationship takes the form Q(A)= $#945;AΘ where Q is peak discharge, A is the drainage area, Θ is the flood scaling exponent, and α is the intercept. Motivated by seminal empirical studies that show that the flood scaling parameters α and Θ change from one rainfall-runoff event to another, this dissertation explores how certain rainfall and catchment physical properties control the flood scaling exponent and intercept at the rainfall-runoff event scale using a combination of extensive numerical simulation experiments and analysis of observational data from the Iowa River basin, Iowa. Results show that Θ generally decreases with increasing values of rainfall intensity, runoff coefficient, and hillslope overland flow velocity, whereas its value generally increases with increasing rainfall duration. Moreover, while the flood scaling intercept is primarily controlled by the excess rainfall intensity, it increases with increasing runoff coefficient and hillslope overland flow velocity. Results also show that the temporal intermittency structure of rainfall has a significant effect on the scaling structure of peak discharges. These results highlight the fact that the flood scaling parameters are able to be estimated from the aforementioned catchment rainfall and physical variables, which can be measured either directly or indirectly using in situ or remote sensing techniques. The dissertation also proposes and demonstrates a new flood forecasting framework that is based on the scaling theory of floods. The results of the study mark a step forward to provide a physically meaningful framework for regionalization of flood frequencies and hence to solve the long standing hydrologic problem of flood prediction in ungauged basins.

Flash flood is among the most hazardous natural disasters, and it can cause severe damages to the environment and human life. Flash floods are mainly caused by intense rainfall and due to their rapid onset (within six hours of rainfall), very limited opportunity can be left for effective response. Understanding the socio-economic characteristics involving natural hazards potential, vulnerability, and resilience is necessary to address the damages to economy and casualties from extreme natural hazards. The vulnerability to flash floods is dependent on both biophysical and socio-economic factors. This study provides a comprehensive assessment of socio-economic vulnerability to flash flood alongside a novel framework for flash flood early warning system. A socio-economic vulnerability index was developed for each state and county in the Contiguous United States (CONUS). For this purpose, extensive ensembles of social and economic variables from US Census and the Bureau of Economic Analysis were assessed. The coincidence of socio-economic vulnerability and flash flood events were investigated to diagnose the critical and non-critical regions. In addition, a data-analytic approach is developed to assess the interaction between flash flood characteristics and the hydroclimatic variables, which is then applied as the foundation of the flash flood warning system. A novel framework based on the D-vine copula quantile regression algorithm is developed to detect the most significant hydroclimatic variables that describe the flash flood magnitude and duration as response variables and estimate the conditional quantiles of the flash flood characteristics. This study can help mitigate flash flood risks and improve recovery planning, and it can be useful for reducing flash flood impacts on vulnerable regions and population.

A geographical information system (GIS) was used to develop a regression model designed to predict flood magnitudes in the Sandy and Clackamas river basins in Oregon. Manual methods of data assembly, input, storage, manipulation and analysis traditionally used to estimate basin characteristics were replaced with automated techniques using GIS-based computer hardware and software components. Separate GIS data layers representing (1) stream gage locations, (2) drainage basin boundaries, (3) hydrography, (4) water bodies, (5) precipitation, (6) landuse/land cover, (7) elevation and (8) soils were created and stored in a GIS data base. Several GIS computer programs were written to automate the spatial analysis process needed in the estimation of basin characteristic values using the various GIS data layers. Twelve basin characteristic data parameters were computed and used as independent variables in the regression model. Streamflow data from 19 gaged sites in the Sandy and Clackamas basins were used in a log Pearson Type III analysis to define flood magnitudes at 2-, 5-, 10-, 25-, 50- and 100-year recurrence intervals. Flood magnitudes were used as dependent variables and regressed against different sets of basin characteristics (independent variables) to determine the most significant independent variables used to explain peak discharge. Drainage area, average annual precipitation and percent area above 5000 feet proved to be the most significant explanatory variables for defining peak discharge characteristics in the Sandy and Clackamas river basins. The study demonstrated that a GIS can be successfully applied in the development of basin characteristics for a flood frequency analysis and can achieve the same level of accuracy as manual methods. Use of GIS technology reduced the time and cost associated with manual methods and allowed for more in-depth development and calibration of the regression model. With the development of GIS data layers and the use of GIS-based computer programs to automate the calculation of explanatory variables, regression equations can be developed and applied more quickly and easily. GIS proved to be ideally suited for flood frequency modeling applications by providing advanced computerized techniques for spatial analysis and data base management.