Machine learning of a hysteretic hydrological signature: A case study with the catchment sensitivity function

Abstract:

The catchment sensitivity function has been widely utilized as a data-based tool to understand system scale flow dynamics. This function explains how sensitive a boundary outflux is to a change in water storage, which also allows us to model rainfall-runoff dynamics. The catchment sensitivity function is, in general, hysteretic, but it is yet unclear how to analyze and characterize its hysteresis. As a result, the catchment sensitivity function has usually been characterized by fitting a one-to-one relationship, neglecting hysteresis in the function. It limits our ability to learn more detailed catchment scale dynamics from the function and limits the applicability of the catchment sensitivity function when modeling catchments where the hysteresis is significant. In this study, we explore an opportunity of understanding and parameterizing the hysteresis using emerging deep learning tools. We utilize a machine learning tool to learn the hysteretic function from data. A preliminary result indicates that the machine learning tool can capture the hysteresis and can teach us what controls the hysteresis.