· · · Repository hosted by TU Delft Library

The book presents machine learning as an approach to build models that learn from data, and that can be used to complement the existing modelling practice in aquatic and geotechnical environments. It provides concepts of learning from data, and identifies segmentation (clustering), classification, regression and control as the learning tasks. A unified methodology based on the concepts of machine learning, information theory and statistics is presented that can be followed to build models using data as well as expert knowledge. Several machine learning methods (neur mentatlon a new algorithm is developed for segmentation of data series where the constraint of contiguity has to be satisfled. In classification multi-scale data (signal) transformatlon methods (e.g. scale-space transformation) are used to extract features to build data-driven models in geotechnics. A set of regression models are built to predict sediment transport rates and in assessing harbour sedimentation. Controllers that replicate the control strategy of model-based optimal controllers of water systems are built for situations where fast and accurate decisions are needed. The models built demonstrate excellent performance; they can complement or even replace the existing models and can be used In practice. The performance of the models proves the effectiveness of the methodology and machine learning in general.

The complementary nature of physically based and datadriven models in their demand of 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 minimize 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 network and fuzzy rule-based models. Complementary modelling offers the possibility of incorporating processes and data that are not considered by the model without affecting the routine operation of physically based models. The possibility of obtaining information that can help to improve the physically based model is also demonstrated.

Day-to-day water management is challenged by meteorological extremes, causing floods and droughts. Often operational water managers are informed too late about these upcoming events to be able to respond and mitigate their effects, such as by taking flood control measures or even requiring evacuation of local inhabitants. Therefore, the use of weather forecast information with hydrological models can be invaluable for the operational water manager to expand the forecast horizon and to have time to take appropriate action. This is called Anticipatory Water Management. Anticipatory actions may have adverse effects, such as when flood control actions turn out to have been unnecessary, because the actual rainfall was less than predicted. Therefore the uncertainty of the forecasts and the associated risks of applying Anticipatory Water Management have to be assessed. To facilitate this assessment, meteorological institutes are providing ensemble predictions to estimate the dynamic uncertainty of weather forecasts. This dissertation presents ways of improving the end-use of ensemble predictions in Anticipatory Water Management.

The intensive uses of limited water resources, the growing population rates and the various increasing human activities put high and continuous stresses on these resources. Major problems affecting the water quality of rivers, streams and lakes may arise from inadequately treated sewage, poor land use practices, inadequate controls on the discharges of industrial waste waters, uncontrolled poor agricultural practices, excessive use of fertilizers, and a lack of integrated watershed management. This study explores the impact of these pollution problems and the water quality degradation of Irrigated agricultural watersheds When the watersheds have a complex physical basis of interacting water bodies such as canals, drains and coastal lagoons as in the case of irrigated watersheds in coastal river Deltas, and when these environments are ‘data scarce environments’, the problems of managing water quality becomes more obvious and the need for reliable solutions becomes an urgent requirement. This study focused on the management of surface water quality problems in such watersheds and the importance of taking into consideration all the watershed components and the effects of pollution from the upstream canals on the downstream coastal lakes. In this study a generic framework for a (Water Quality Management Information System) is developed depending on the integration of physically based hydrodynamic and water quality models with GIS capabilities and the spatial and temporal capabilities of remote sensing in water quality modeling. The application is developed and tested for the Edko drainage catchment and shallow lake system in the western part of the Nile Delta, Egypt. The developed framework includes a hierarchy of modeling tools: a 1D-2D basic hydrodynamic model for a combined shallow lake-drainage system, a detailed 2D hydrodynamic model of the shallow lake, and a 2D water quality and eutrophication screening models for the lake system. The basic water quality model for the lake system simulates the main water quality parameters including the oxygen compounds, nutrients compounds, temperature, salinity and the total suspended matter (TSM). The complexity of the physical and ecological properties of the lake system implied the use of different methodologies for models calibration using remote sensing. The combination of remote sensing with mathematical modelling, for the calibration and verification of TSM and chlorophyll-concentrations in the shallow lake system showed reliable and successful results.

In the last few decades, the consequences of floods and flash floods in many parts of the world have been devastating. One way of improving flood management practice is to invest in data collection and modelling activities which enable an understanding of the functioning of a system and the selection of optimal mitigation measures. A Digital Terrain Model (DTM) provides the most essential information for flood manager. Light Detection and Ranging (LiDAR) surveys which enable the capture of spot heights at a spacing of 0.5m to 5m with a horizontal accuracy of 0.3m and a vertical accuracy of 0.15m can be used to develop high accuracy DTM but it need careful processing before it can be used for any application. The research presents the augmentation of an existing Progressive Morphological filtering algorithm for processing raw LiDAR data to support a 1D/2D urban flood modelling framework. The key characteristics of this improved algorithm are: (1) the ability to deal with different kinds of buildings; (2) the ability to detect elevated road/rail lines and represent them in accordance to the reality; (3) the ability to deal with bridges and riverbanks; and (4) the ability to recover curbs and the use of appropriated roughness coefficient of Manning’s value to represent close-to-earth vegetation (e.g. grass and small bush).

A hydroinformatics system represents an electronic knowledge encapsulator that models part of the real world and can be used for the simulation and analysis of physical, chemical and biological processes In water systems, for a better management of the aquatic environment. Thus, modelling is at the heart of hydroinformatics. The theory of nonlinear dynamics and chaos and the extent to which recent improvements in the understanding of inherently nonlinear natural processes present challenges to the use of mathematical models in the analysis of water and environmental systems are elaborated in this work. In particular, it demonstrates that the deterministic chaos present in many nonlinear systems can impose fundamental limitations on our ability to predict behaviour even when well-defined mathematical models exist. On the other hand, methodologies and tools from the theory of nonlinear dynamics and chaos can provide means for a better accuracy of short-term predictions as demonstrated through the practical applications in this work.

This thesis presents powerful machine learning (ML) techniques to build predictive models of uncertainty with application to hydrological models. Two different methods are developed and tested. First one focuses on parameter uncertainty analysis by emulating the results of Monte Carlo simulations of hydrological models using efficient ML techniques. Second method aims at modelling uncertainty by building an ensemble of specialised ML models on the basis of past hydrological model’s performance. Methods employed include artificial neural networks, model trees, locally weighted regression and fuzzy logic. The application of the methods to several real-world case studies demonstrates the capacity of machine learning techniques for building accurate and efficient predictive models of uncertainty.

Flooding in urbanized areas has become a very important issue around the world. The level of service (or performance) of urban drainage systems (UDS) degrades in time for a number of reasons. In order to maintain an acceptable performance of UDS, early rehabilitation plans must be developed and implemented. In developing countries the situation is serious, little investment is done and there are smaller funds each year for rehabilitation. The allocation of such funds must be “optimal” in providing value for money. However this task is not easy to achieve due to the multicriteria nature of the rehabilitation process, taking into account technical, environmental and social interests. Most of the time these are conflicting, which make it a highly demanding task. The present book introduce a framework to deal with multicriteria decision making for the rehabilitation of urban drainage systems, and focuses on several aspects such as the improvement of the performance of the multicriteria optimization through the inclusion of new features in the algorithms and the proper selection of performance criteria. The use of Genetic Algorithms, parallelization and application in countries like Brazil, Colombia y Venezuela are treated in this book.