Improvements to Modeling and Predicting Snow Distribution Using Machine Learning, Physics-Based Models, and New Observational Methods
Authors
Ryan Crumley1* (rcrumley@lanl.gov), Claire Bachand1, Katrina Bennett1, Colleen Iversen2
Institutions
1Los Alamos National Laboratory, Los Alamos, NM; 2Oak Ridge National Laboratory, Oak Ridge, TN
URLs
Abstract
Snowpack distribution in Arctic and alpine landscapes is highly variable and often occurs in repeating, year-to-year patterns due to local topographic, weather, and vegetation characteristics. Developing methods for monitoring and modeling complex Arctic snow distributions is key to understanding snow-vegetation interactions and subsequent impacts on permafrost. Recent advances in snow hydrology and machine learning (ML) have increased the ability to predict snowpack distribution using in situ observations and simple landscape characteristics (e.g., vegetation type, topographic position index, wind indices, elevation) that can be easily obtained for most environments. Here, the team presents novel methods to characterize snow-vegetation-topography interactions using ML, physics-based models, and new observational methods to determine snow depth and snow distribution in the subarctic Teller watershed on the Seward Peninsula in Alaska. First, researchers include results from a hybrid approach to couple a ML-based snow distribution pattern map with the physics-based snow process model, SnowModel. The team trained the ML algorithm on tens of thousands of snow survey observations that were collected over four winters during peak snow-water equivalent (Crumley et al. In review for Water Resources Research). The hybrid method more accurately depicted the spatial patterns seen in in situ observations and light detection and ranging datasets compared to SnowModel runs that did not incorporate the ML pattern map. Secondly, researchers developed a transferable ML model to predict snow depth from ground surface temperature (GST) measurements at daily temporal resolution (Bachand et al. In preparation). This model was trained on snow depth and GST data collected using distributed temperature profilers and performed well (R2>0.85) on the Seward Peninsula and elsewhere (e.g., Svalbard, Norway). These results demonstrate that ML can improve upon physics-based model results and enhance the capacity to observe snowpack characteristics in a changing Arctic environment going forward.