Fusing Large Datasets and Models to Improve Understanding of Hydrologic and Hydraulic Processes

Abstract

Global water systems are being stressed by aging infrastructure, climate change, and resource withdrawals. The ability to model large water systems has attempted to keep pace with these challenges, with modern water models now operating effectively across massive spatial and temporal scales. Despite great advances in numerical modeling, many decision makers require high-resolution information that large-scale models do not yet provide. Simultaneously, the affordability and ease of use of sensing platforms has improved dramatically, enabling even small communities and research groups to deploy observation networks. Unfortunately, these real-time measurements are often not attached to a physical or numerical model, which prevents their use in predictive applications. To that end, this dissertation poses the question: how can the domain knowledge embedded in large-scale models be fused with new forms of sensor data to improve understanding of hydrologic and hydraulic processes? Three primary issues currently prevent this question from being answered. First, many datasets are irregular or noisy, making integration with models difficult. This dissertation addresses this issue by providing a methodology for integrating non-standard and distributed measurements with large numerical models. The approach is applied to an unprecedented data set of over one million ship observations across the Great Lakes to generate new insights about distributed hydrometeorological processes. Second, the scales across which water models operate do not often match the scales at which we measure. This dissertation addresses this issue by providing a methodology for dynamically mapping large-scale model outputs to site-scale forecasts. The approach is applied to flood forecasting across the entire state of Iowa, where nearly two hundred sensors are fused with the US National Water Model. Third, since many numerical models of water systems are often heavily parameterized, it is difficult to determine how to update these models when novel sources of sensor data emerge. This dissertation addresses this issue by providing a methodology for abstracting simple models from complex water networks to enable efficient detection and localization of change. The approach will underpin a real-time asset management methodology for stormwater systems. Ultimately, this dissertation seeks to contribute to the emergence of Big Data Hydrology by discovering opportunities in data-driven water modeling that will be enabled by systems engineering and data science.