Data-driven methods for hydrologic inference and discovery

Abstract

Water flows in the Earth system are complex and difficult to quantify. Using data without recourse to an underlying physical theory has been a hallmark of hydrologic science for many years. Currently the expanding ability to handle large data sets using new methods has led to the development and use of sophisticated data-driven models with the ability to integrate physical theory into model architectures. My dissertation relies on theory-informed data-driven models (DDMs) to answer questions in hydrology. I first explore how theory can be integrated with DDMs. I apply and compare various methods to build DDMs to regionalize measured streamflow information to ungaged catchments where measurements are not available. I expand the work to address questions in sociohydrolgy–an emerging sub-discipline within the hydrologic sciences that seeks to integrate the physical and social aspects of hydrologic systems–and show how the amount of water used across the U.S. is related to both physical and social variables. Finally, I discuss how DDMs in general, and each of the problems that I address in my dissertation in particular, relate to various forms of logical inference and how the feedback between data analysis and established theory is connate to scientific progress.