Learning models for predicting pavement friction based on non-contact texture measurements: Comparative assessment

Ensuring highway safety relies heavily on pavement friction resistance. To enable network-level pavement skid resistance monitoring and management, this study proposes a non-contact three-dimensional laser surface testing method to obtain detailed aggregate surface data. The existing contact-based skid resistance measurement methods suffer from poor reproducibility and repeatability, hindering their application for network-level management. In this research, traditional multiple linear regression and four machine learning methods, support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), and convolutional neural network (CNN), are utilized to evaluate and predict pavement frictional performance. To assess the proposed methods, data from 45 pavement sites in Oklahoma, including 6 major preventive maintenance (PM) treatments and 7 typical types of aggregates, are collected. Parallel data acquisition is conducted at highway speeds using a grip tester and a high-speed texture profiler to measure pavement skid resistance and surface macro-texture, respectively. Aggregate properties are captured in 3D using a portable ultra-high-resolution 3D laser imaging scanner, leading to the calculation of four types of 3D aggregate parameters characterizing the micro-texture of aggregate surfaces. The relationship between pavement surface friction and texture is explored using machine learning models. The results reveal that the random forest and gradient boosting decision tree models exhibit the highest accuracy, SVM and CNN perform moderately, while the traditional linear regression method fares the worst. By assessing the importance of the 38 parameter variables, the most critical 21 variables were selected for model development. Test results demonstrate that the GBDT model exhibits the best predictive performance, with an explanatory capability of 87.4​% for road friction performance. The findings demonstrate the feasibility of replacing contact-based pavement friction evaluation with non-contact texture measurements, offering promising prospects for a network-level pavement skid resistance monitoring and management system. • Non-contact 3D laser scanning proves to be an effective technique for evaluating pavement friction performance. • GBDT model achieved the best accuracy, with an R 2 of 87.4% on the test set, surpassing both RF and SVM models. • Total AADT, texture spatial ratio, temperature, and homogeneity index were identified as key predictive variables. • Optimizing input variables improves prediction accuracy and reduces model overfitting.