The research found that spatial distribution depends most heavily on vegetation, elevation and landscape features, such as stream banks and benches -; areas of topographic variability where shrubs grow and snow accumulates.
Based on random-forest machine learning, the statistical model characterizes the spatial pattern of the end-of-winter snow distribution and identifies the key factors controlling the spatial distribution. The model also predicts the snow distribution for the local study sites and can be generalized across the region.
Bennett said the analysis will be useful in validating physically based permafrost hydrology models, such as the Advanced Terrestrial Simulator developed at Los Alamos. The work will also help validate and provide improved snow redistribution representation in the land surface model within the Department of Energy's Energy Exascale Earth System Model.
"Ultimately, it will increase our understanding of changing hydrology, topography and vegetation dynamics in the Arctic and sub-Arctic," Bennett said.
Seasons in the Snow
The multi-institutional research team, which included members from Los Alamos, University of Alaska Fairbanks, Lawrence Berkeley National Laboratory, Oak Ridge National Laboratory and University of Wisconsin–Madison, conducted snow surveys in the spring months of 2017–2019 at two small sites on the Seward Peninsula.
"We want to gratefully acknowledge Mary's Igloo, Sitnasuak and Council Native Corporation for their guidance and for allowing us to conduct our research on their traditional lands," Bennett said.
The field work focused on collecting end-of-winter snow-depth and snow-density measurements to calculate the amount of water contained within the snowpack. Those measurements characterize the impacts of snow cover on water and temperature better than snow-depth measurements do.
To create a model of snow distribution, the team estimated landscape factors for topography, vegetation and wind, and then quantified their impacts on snow distribution using three statistical models.