Machine Learning-enhanced Hydrologic Modeling under Changing Climate

Hydrological modeling has been widely used to predict the response of a hydrologic system to changing drivers in short to long terms, thus providing quantitative decision-making support for strategic planning and adaptation policies development. Despite advancements in hydrological modeling, challenges remain for improving the quality and realism of hydrologic predictions. Machine Learning (ML), known for its proficiency in extracting patterns from large datasets, has attracted interest within the hydrology community and has been applied to tackle various challenges in predicting the hydrologic cycle over the past few years. This dissertation focuses on the enhancement of hydrologic models through ML and evolving datasets, with the goal of improving their accuracy and utility. The first part addresses the issue of missing or inadequately represented components in the existing large-scale hydrologic models. Taking the simulation of regulated flow conditions as an example, the study presents a hierarchical temporal scale framework for all data-driven reservoir release models, enabling more effective use of limited data sources and ensuring that practical significance aligns with model configuration. The second part addresses the computational challenges associated with model parameterization. The development of an ML-based surrogate model expedites the parameter estimation and calibration, particularly for those properties that may vary over time and require the adoption of dynamic parameterization. Satellite-derived vegetation interannual variability serves as a case study in this dissertation, illustrating how the dynamic nature of vegetation can influence hydrologic responses. From the perspective of hydrologic modelers, these two parts of work enhance the hydrologic model’s realism by improving both its representation and parameterization, respectively. For water managers, a combination of surrogate model and the reservoir operation module enables integrated reservoir management modeling under different climate projection scenarios. Building upon the insights gained from the first two parts, the last part shows such application to translate hydrologic and climatic data into actionable strategies for water management. Utilizing 21 federal reservoirs in Texas as a case study, this part offers a framework for stakeholders to assess the effectiveness of current reservoir operation policies under future climate scenarios through the interactions among hydroclimatology, reservoir infrastructure, and operation policy.