Snow-to-liquid ratio prediction over the contiguous United States using machine learning

Winter storms impact populated regions across the continental United States (CONUS), resulting in travel and economic disruptions, loss of life, and damage to property and infrastructure. Snow-to-liquid ratio (SLR), or the depth of liquid water equivalent that results from a specified depth of snowfall, is used operationally to forecast snowfall amount during winter storms. However, no clear prediction method for SLR has been established and some operational SLR prediction methods were developed for specific regions, resulting in inaccurate snowfall forecasts in some cases.

Using 15 years of Community Collaborative Rain, Hail, and Snow Network (CoCoRAHS) manual snowfall observations and European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5) vertical profiles, we trained a CONUS deterministic prediction system for SLR. A random forest (RF) algorithm framework was developed for SLR predictions. The RF training and testing datasets were varied by region and snow climate which were developed using k-means clustering, gridded snowfall climatologies, and geographic boundaries. Quantitative snowfall forecasts (QSFs) were then produced by applying quantitative precipitation forecasts (QPFs) from NOAA’s High-Resolution Rapid Refresh (HRRR) and Global Forecast System (GFS) to the predicted SLRs. SLR forecasts and QSFs were also derived from several operational SLR prediction methods and compared with results from the RF. Verification was performed using the National Operational Hydrologic Remote Sensing Center (NOHRSC) snow analysis and the RF testing datasets. Depending on region and storm regime, SLR and precipitation forecast errors were examined as a function of temperature, wind, moisture, and elevation to determine storm environments that may have contributed to poor or accurate snowfall forecasts. The resulting prediction and verification system provides a template for improved snowfall forecasts at a CONUS-wide scale and improves upon current operational SLR prediction methods.