Computational learning and data-driven modeling for water resources management and hydrology

Water scarcity, changing climate, and hydrologic uncertainty present serious challenges for water resources management and hydrologic modeling. Development of surface and groundwater resources, success in harnessing the power of flowing water, mitigation of the effects of floods and droughts, and provision for clean water require models with high predictive capability. Computational learning theory and data-driven modeling techniques are new and rapidly expanding areas of research that examine formal models of induction with the goal of developing efficient learning algorithms. This dissertation introduces new and improved modeling frameworks and systematic guidelines to integrate various forms of available data to provide reliable forecasts for the behavior of hydrologic systems that are important in water resources management. The objective is to advance the concepts of support vector machines, relevance vector machines, and locally weighted projection regression learning algorithms to capture the convoluted physical processes, provide decision-relevant information, model chaotic dynamic systems, and detect drift and novelty in the systems. These learning machines are applied to evaluate their plausibility and utility in diverse water resources-related settings. The models are designed in this dissertation to be parsimonious and robust, and to have the ability to quantitatively describe various aspects of uncertainty in model forecasts. Promising simulation results using real-life case studies show the ability of learning machines to build accurate models with competitive predictive capabilities and hence constitute a valuable means for extracting knowledge and improving modeling techniques.