Due to the complexity of hydrological systems a single model may be unable to capture the full range of a catchment response and accurately predict the streamflows. A solution could be the in use of several specialized models organized in the so-called committees. Refining the committee approach is one of the important topics of this study, and it is demonstrated that it allows for increased predictive capability of models. Another topic addressed is the prediction of hydrologic models' uncertainty. The traditionally used Monte Carlo method is based on the past data and cannot be directly used for estimation of model uncertainty for the future model runs during its operation. In this thesis the so-called MLUE (Machine Learning for Uncertainty Estimation) approach is further explored and extended; in it the machine learning techniques (e.g. neural networks) are used to encapsulate the results of Monte Carlo experiments in a predictive model that is able to estimate uncertainty for the future states of the modelled system. Furthermore, it is demonstrated that a committee of several predictive uncertainty models allows for an increase in prediction accuracy. Catchments in Nepal, UK and USA are used as case studies. In flood modelling hydrological models are typically used in combination with hydraulic models forming a cascade, often supported by geospatial processing. For uncertainty analysis of flood inundation modelling of the Nzoia catchment (Kenya) SWAT hydrological and SOBEK hydrodynamic models are integrated, and the parametric uncertainty of the hydrological model is allowed to propagate through the model cascade using Monte Carlo simulations, leading to the generation of the probabilistic flood maps. Due to the high computational complexity of these experiments, the high performance (cluster) computing framework is designed and used. This study refined a number of hydroinformatics techniques, thus enhancing uncertainty-based hydrological and integrated modelling.
Computers are increasingly used in the simulation of natural phenomena such as floods. However, these simulations are based on numerical approximations of equations formalizing our conceptual understanding of flood flows. Thus, model results are intrinsically subject to uncertainty and the use of probabilistic approaches seems more appropriate. Uncertain, probabilistic floodplain maps are widely used in the scientific domain, but still not sufficiently exploited to support the development of flood mitigation strategies. In this thesis the major sources of uncertainty in flood inundation models are analyzed, resulting in the generation of probabilistic floodplain maps. The utility of probabilistic model output is assessed using value of information and the prospect theory. The use of these maps to support decision making in terms of floodplain development under flood hazard threat is demonstrated.
This book comprehensively accounts the advances in data-based approaches for hydrologic modeling and forecasting. Eight major and most popular approaches are selected, with a chapter for each — stochastic methods, parameter estimation techniques, scaling and fractal methods, remote sensing, artificial neural networks, evolutionary computing, wavelets, and nonlinear dynamics and chaos methods. These approaches are chosen to address a wide range of hydrologic system characteristics, processes, and the associated problems. Each of these eight approaches includes a comprehensive review of the fundamental concepts, their applications in hydrology, and a discussion on potential future directions.
The importance of monitoring networks for resource-management decisions is becoming more recognized, in both theory and application. Quantitative computer models provide a science-based framework to evaluate the efficacy and efficiency of existing and possible future monitoring networks. In the study described herein, two suites of tools were used to evaluate the worth of new data for specific predictions, which in turn can support efficient use of resources needed to construct a monitoring network. The approach evaluates the uncertainty of a model prediction and, by using linear propagation of uncertainty, estimates how much uncertainty could be reduced if the model were calibrated with addition information (increased a priori knowledge of parameter values or new observations). The theoretical underpinnings of the two suites of tools addressing this technique are compared, and their application to a hypothetical model based on a local model inset into the Great Lakes Water Availability Pilot model are described. Results show that meaningful guidance for monitoring network design can be obtained by using the methods explored. The validity of this guidance depends substantially on the parameterization as well; hence, parameterization must be considered not only when designing the parameter-estimation paradigm but also-importantly-when designing the prediction-uncertainty paradigm.
Numerical models of flow and transport processes are heavily employed in the fields of surface, soil, and groundwater hydrology. They are used to interpret field observations, analyze complex and coupled processes, or to support decision making related to large societal issues such as the water-energy nexus or sustainable water management and food production. Parameter estimation and uncertainty quantification are two key features of modern science-based predictions. When applied to water resources, these tasks must cope with many degrees of freedom and large datasets. Both are challenging and require novel theoretical and computational approaches to handle complex models with large number of unknown parameters.
Like all natural hazards, flooding is a complex and inherently uncertain phenomenon. Despite advances in developing flood forecasting models and techniques, the uncertainty in forecasts remains unavoidable. This uncertainty needs to be acknowledged, and uncertainty estimation in flood forecasting provides a rational basis for risk-based criteria. This book presents the development and applications of various methods based on probablity and fuzzy set theories for modelling uncertainty in flood forecasting systems. In particular, it presents a methodology for uncertainty assessment using disaggregation of time series inputs in the framework of both the Monte Carlo method and the Fuzzy Extention Principle. It reports an improvement in the First Order Second Moment method, using second degree reconstruction, and derives qualitative scales for the interpretation of qualitative uncertainty. Application is to flood forecasting models for the Klodzko catchment in POland and the Loire River in France. Prospects for the hybrid techniques of uncertainty modelling and probability-possibility transformations are also explored and reported.
This book explores a new realm in data-based modeling with applications to hydrology. Pursuing a case study approach, it presents a rigorous evaluation of state-of-the-art input selection methods on the basis of detailed and comprehensive experimentation and comparative studies that employ emerging hybrid techniques for modeling and analysis. Advanced computing offers a range of new options for hydrologic modeling with the help of mathematical and data-based approaches like wavelets, neural networks, fuzzy logic, and support vector machines. Recently machine learning/artificial intelligence techniques have come to be used for time series modeling. However, though initial studies have shown this approach to be effective, there are still concerns about their accuracy and ability to make predictions on a selected input space.
Hydrologic modelling has benefited from significant developments over the past two decades, which has led to the development of distributed hydrologic models. Parameter adjustment, or model calibration, is extremely important in the application of these hydrologic models. Multi-criteria calibration schemes and several formal and informal predictive uncertainty estimation methodologies are among the approaches to improve the results of model calibration. Moreover, literature indicates a general agreement between formal and informal approaches with respect to the predictive uncertainty estimation in single-criterion calibration cases. This research extends the comparison between these techniques to multi-criteria calibration cases, and furthermore, proposes new ideas to improve informal multi-criteria calibration and uncertainty estimation in hydrological modelling. GLUE is selected as a candidate informal methodology due to its extreme popularity among hydrological modellers, i.e., based on the number of applications in the past two decades. However, it is hypothesized that improvements can be applied to other certain types of informal uncertainty estimation as well. The first contribution of this research is an in-depth comparison between GLUE and Bayesian inference in the multi-criteria context. Such a comparison is novel because past literature has focused on comparisons for only single criterion calibration studies. Unlike the previous research, the results show that there can be considerable differences in hydrograph prediction intervals generated by traditional GLUE and Bayesian inference in multi-criteria cases. Bayesian inference performs more satisfactorily than GLUE along most of the comparative measures. However, results also reveal that a standard Bayesian formulation (i.e., aggregating all uncertainties into a single additive error term) may not demonstrate perfect reliability in the prediction mode. Furthermore, in cases with a limited computational budget, non-converged MCMC sampling proves to be an appropriate alternative to GLUE since it is reasonably consistent with a fully-converged Bayesian approach, even though the fully-converged MCMC requires a substantially larger number of model evaluations. Another contribution of this research is to improve the uncertainty bounds of the traditional GLUE approach by the exploration of alternative behavioural solution identification strategies. Multiple behavioural solution identification strategies from the literature are evaluated, new objective strategies are developed, and multi-criteria decision-making concepts are utilized to select the best strategy. The results indicate that the subjectivity involved in behavioural solution identification strategies impacts the uncertainty of model outcome. More importantly, a robust implementation of GLUE proves to require comparing multiple behavioural solution identification strategies and choosing the best one based on the modeller's priorities. Moreover, it appears that the proposed objective strategies are among the best options in most of the case studies investigated in this research. Thus, it is recommended that these new strategies be considered among the set of behavioural solution identification strategies in future GLUE applications. Lastly, this research also develops a full optimization-based calibration framework that is capable of utilizing both standard goodness-of-fit measures and many hydrological signatures simultaneously. These signatures can improve the calibration results by constraining the model outcome hydrologically. However, the literature shows that to simultaneously apply a large number of hydrological signatures in model calibration is challenging. Therefore, the proposed research adopts optimization concepts to accommodate many criteria (including 13 hydrologic signature-based objectives and two standard statistical goodness-of-fit measures). In the proposed framework, hydrological consistency is quantified (based on a set of signature-based measures and their desired level of acceptability) and utilized as a criterion in multiple calibration formulations. The results show that these formulations perform better than the traditional approaches to locate hydrologically consistent parameter sets in the search space. Different hydrologic models, most of which are conceptual rainfall-runoff models, are used throughout the thesis to evaluate the performance of the developed strategies. However, the developments explored in this research are typically simulation-model-independent and can be applied to calibration and uncertainty estimation of any environmental model. However, further testing of these methods is warranted for more computationally intensive simulation models, such as fully distributed hydrologic models.
The complementary nature of physically-based and data-driven models in their demand for physical insight and historical data, leads to the notion that the predictions of a physically-based model can be improved and the associated uncertainty can be systematically reduced through the conjunctive use of a data-driven model of the residuals. The objective of this thesis is to minimise the inevitable mismatch between physically-based models and the actual processes as described by the mismatch between predictions and observations. Principles based on information theory are used to detect the presence and nature of residual information in model errors that might help to develop a data-driven model of the residuals by treating the gap between the process and its (physically-based) model as a separate process. The complementary modelling approach is applied to various hydrodynamic and hydrological models to forecast the expected errors and accuracy, using neural
