Snowfall-driven topographic evolution: impacts on snow distribution patterns
Date
2023
Journal Title
Journal ISSN
Volume Title
Abstract
This study develops a scalable meteorologically independent snow accumulation model to better estimate snowpack depth using an enhanced representation of actual processes. Current snow accumulation models incorporate bare or snow-free surface properties derived from elevation, aspect, vegetation, and prevailing wind characteristics to determine the drivers of snow distribution yet neglect to consider how subsequent snowfalls can reshape the initial terrain conditions. We hypothesize that a snow depth model that accumulates snowfall while accounting for the antecedent snow-affected surface characteristics is more representative of natural processes and will therefore yield more accurate depth estimates than models that reference a snow-free topographic surface. To address this premise, the research explores (1) conducting a sensitivity analysis to evaluate the behavior of both models, (2) determining the differences between the two snow accumulation modeling approaches, and (3) assessing each model's performance in different location, scale, and temporal resolution conditions to determine their resiliency and transferability. Terrestrial LiDAR was employed at two field sites following snow deposition events and captured a range of spatial extents and resolutions. The Upper Piceance Creek (UPC) site near Meeker, CO covered approximately 10 m2 at centimeter resolution; the Izas Experimental Catchment in the Spanish Pyrenees covered 1 km2 at meter resolution. A regression tree machine learning model was utilized to estimate snow depth based on 14 topographic features. This process engaged in two mechanisms: 1. Static method, where snow depth (dst) determined from the bare earth digital terrain model (ds0) was estimated with snow-free topographic features and 2. Dynamic method, where snow depth (dst) determined from the previous snow surface height (dst-1) was estimated with the dst-1 snowfall affected surface. The analyses found that the models were resilient to changes in training allocations under a random sampling method, but sensitive to both the prevailing wind direction used for feature creation and the overall resolution used to represent surface features. The primary difference between the static and dynamic models for snow depth estimates was the number of features used and their relative importance. The static method had a higher overall median importance and relied mainly on Directional Relief and Relative Topographic Position for snow depth estimates, while the dynamic method displayed lower overall median importance but utilized more surface features over a single accumulation season. The dynamic method outperformed the static method at UPC by approximately 0.07 in a Nash-Sutcliffe efficiency comparison, and only 0.01 at Izas Experimental Catchment suggesting issues with process-scale representation of snow accumulation at the Izas site.
Description
Rights Access
Subject
snow
accumulation modeling