Upscaling and downscaling snow processes with machine learning in watershed models

Abstract

Hydrologic models play a vital role in understanding and predicting the movement of water within watersheds, providing essential insights for effective management and sustainability of water resources. However, watersheds exhibit significant heterogeneity in their landscape properties and complex responses to spatiotemporal variations in climatic inputs. This variability introduces a gap between the representation of physical processes at the point scale and their behaviour at the watershed scale, making it challenging to accurately capture the full complexity of the hydrologic cycle across different spatial scales. Bridging this gap requires the identification of effective scaling approaches tailored to capture the complexities across scales. Scaling approaches look to translate information from one scale to another, whether moving from a smaller to a larger scale (upscaling) or from a larger to a smaller scale (downscaling). Although various approaches in the literature have been applied to develop scaling methods for forcing variables, such as precipitation and temperature, and fluxes (e.g., evapotranspiration), there is a notable gap in deriving and applying scaling techniques for snow-related variables, such as SWE, snowmelt, or sublimation. Addressing this gap may help in improving hydrologic model accuracy in snow-dominated regions, where snow dynamics significantly influence water availability and watershed resources. The primary objective of this thesis is to develop, implement, and evaluate machine learning-based upscaling methodologies to aid in understanding the relationship between local-scale snow-related variables, landscape heterogeneity, and the large-scale hydrologic response of a catchment. Such methods are useful for effectively simulating the net impact of local variability in snow processes without resorting to fine-resolution models. A secondary focus of this research aims to identify the conditions under which emergent constitutive relationships specific to snow-related fluxes are (or are not) valid and to assess the transferability of these relationships. Finally, this work introduces a machine learning-based downscaling approach that refines large-scale mean model outputs into localized snow states and fluxes. Together, these scaling techniques explore the potential of machine learning to address challenges in hydrologic scaling specific to snow-related fluxes.up