Development of a Machine-Learning-Based Tool to Predict Retroreflectivity of Pavement Markings Across the U.S.

Retroreflectivity plays a crucial role in pavement markings as it enhances nighttime visibility for drivers. Yet, owing to budget constraints, many U.S. state agencies rarely monitor the retroreflectivity of their markings, and instead, restripe markings based on visual inspection or fixed schedule. To address this issue, Federal Highway Administration (FHWA) announced a new final rule, effective September 6, 2022, that establishes a new minimum standard for pavement marking retroreflectivity. This rule also requires state agencies to implement a method within four years for maintaining pavement marking retroreflectivity at or above minimum levels. The ultimate objective of this research was to help state agencies comply with this new rule by developing a new decision-making tool for predicting retroreflectivity of pavement markings for up to three years using only measured initial retroreflectivity (at installation date) and other readily available project data (marking color, marking type, etc.). As such, data from National Transportation Product Evaluation Program (NTPEP) were retrieved and analyzed using six robust machine learning algorithms, and the algorithm yielding highest accuracy was selected to be incorporated into the proposed software. Findings indicated that random forest demonstrated best performance among the six algorithms. As expected, prediction accuracy decreased over time. Within the first year, random forest models successfully predicted retroreflectivity with a coefficient of determination (R²) of 0.97 and root mean square error (RMSE) of ± 31.12 mcd/m²/lux. Within the second and third years, R² and RMSE were 0.92/±43.7 mcd/m²/lux, and 0.91/±36.0 mcd/m²/lux, respectively, which are more accurate as compared with previous studies.