Applications of information theory and machine learning for hydrologic modeling

Abstract

An explosion of new data sources, expansion of computing resources, and theoretical advancesin data science have spurred the rapid adaptation of data-driven methods in earth system science, including hydrology. In this dissertation I will describe three applications of data-driven methods with applications to hydrologic modeling. In chapter 2 I present a framework for hydrologic model intercomparison which examines process interactions within a process-based hydrologic model (PBHM). I show that taking a more holistic approach can shed light into the functioning of these complex models. In chapter 3 I couple machine learned representations of turbulent heat fluxes into a PBHM, and show that neural networks can provide better predictions and transferability than the process-based equations that are used in PBHMs. Building on this, in chapter 4 I use explainable AI (XAI) methods to examine what the neural network has learned. I find that the neural network is able to learn physically plausible relationships and can identify how to partition between latent and sensible heat fluxes based only on short-term temporal data. I also show how we can use XAI to examine what neural networks have learned between sites.This method can uncover that certain sites can be used as predictors for many other sites, as well as that site specific traits such as vegetation type play a large role in the neural network’s ability to generalize to sites it was not trained on. Finally, based on the findings of these three applications I discuss in Chapter 5 how data-driven techniques in general can contribute to improved hydrologic understanding