Enhancing Winter Maintenance Decision Making through Deep Learning-Based Road Surface Temperature Estimation

State and local highway agencies spend substantial budgets repairing infrastructure damages caused by snow and ice every winter season. These infrastructure damages can be prevented by establishing effective winter maintenance strategies which require monitoring road surface temperatures (RST). Road weather information systems (RWIS) are a common source for obtaining data on both current and predicted RST. However, RWIS does not exist in many locations, making it challenging to monitor road surface temperature in locations where RWIS is not available. The objective of this research is to assess the applicability of deep learning models in estimating RST using weather forecasts from the National Weather Service which are provided in grid cells over the continental United States. The deep learning models were developed based on collected data on actual RST and associated weather forecasts from roadways in North Texas for two winter seasons. The accuracy metrics of the deep learning models were computed using mean absolute error, root mean square error, and R-squared. According to the results, the developed deep learning model can predict RST with a root mean square error of 1.71°C, mean absolute error of 1.35°C, and an R-squared value of 0.92. The proposed models can be incorporated into a winter maintenance decision support system, enabling the monitoring of the RST without dependence on RWIS.