Using Machine Learning to Predict Road Temperatures

On average, approximately 5,000 lives are lost each year in car accidents caused by adverse weather conditions. It would be beneficial for the NWS to have a tool for predicting the road temperature in a timely manner such that they can more confidently communicate risks, including accumulating snow or ice accretion, to the public. Since NWP models do not explicitly forecast the temperature of the road, we propose a model which outputs a 1-hr probabilistic short-term forecast indicating the likelihood the temperature of a road surface is at or below freezing (0°C). The probabilistic road temperature model is developed using machine learning, specifically using a random forest. The random forest is trained using various surface-based variables from the 02-hour forecast HRRR model along with analyses from the Multi-Radar/Multi-Sensor (MRMS) reflectivity mosaics and QPE products. Verification is obtained by comparison of the predicted temperature to the observed road surface temperature provided by Road Weather Information System (RWIS) sites. Preliminary statistical results are encouraging with an average Brier Skill Score (BSS) of 0.65, an area under the receiver operating characteristics curve statistic of 0.989-0.993. Obtaining an accurate and reliable prediction of the road temperature is the first step in preventing weather-related road fatalities.