Machine Learning Approaches to Estimate Road Surface Temperature Variation along Road Section in Real-Time for Winter Operation

Abstract

Monitoring road surface temperature (RST) is crucial to establish winter maintenance strategies for traffic safety and proactive congestion management. Public agencies have conventionally relied on mathematical models to predict road conditions. Typically, those models employ data collected from fixed environmental sensor stations sporadically located over a wide network and estimate parameters that are specific to a site. In addition, taking interactions among meteorological, geographical, and physical road characteristics into a model is almost impossible. This study proposes a new and practical framework that can estimate an RST variation model via an off-the-shelf Classification Learner application embedded in the MATLAB machine learning tool. To develop the model, this study uses climatological information, vehicular ambient temperature data from a probe vehicle, and road section information (i.e., basic section, bridge section, tunnel section). The performance of the developed models is then compared with actual RSTs measured from a thermal mapping system. The final evaluation found the estimated RST variation along road section and observed ones compatible, indicating that the proposed procedure can be readily implemented. The proposed method can help public agencies develop both reliable and readily transferrable procedures for monitoring RST variation without having to rely on data collected from costly fixed sensors.