Multimodel Evaluation of Pavement Marking Skid Resistance Using Supervised Machine Learning

Highway agencies aim to maintain sufficient friction at crosswalks to reduce vehicle skidding and pedestrian slipping. Although pavement surface friction has been extensively studied, the behavior of pavement markings under varying materials and environmental conditions remains less explored. This study develops, evaluates, and compares multiple supervised machine learning models to predict the skid resistance of pavement markings, expressed as the British Pendulum Number (BPN-TSR). The data set includes 1,092 laboratory and field observations covering four marking materials (preformed thermoplastic, thermoplastic, tape, and paint), two surface types (asphalt and concrete), three surface conditions (dry, wet, and icy), and two debris states (clean and dirty). Five algorithms were evaluated: random forest (RF), XGBoost (XG), support vector regression (SVR), Gaussian process regression (GPR), and multilayer perceptron (MLP). The data preprocessing stage involved label encoding, z-score normalization for SVR, GPR, MLP, and Pearson correlation to address multicollinearity. Class imbalance was managed using an 80/20 train–test split, five-fold cross-validation, and GridSearchCV for ensuring representative sampling, improving model accuracy, and enhancing generalization. Model performance was assessed using the coefficient of determination (R2), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Kolmogorov–Smirnov (K-S) tests, and residual diagnostics. Feature importance analysis identified surface condition, marking material, and test environment as the most influential predictors. The contribution of this study is a predictive framework for pavement markings’ skid resistance, providing a comparative evaluation of machine learning models, highlighting key influencing factors, and addressing the skid resistance research gap for such critical safety materials at crossings, an area with limited prior research. Practical deployment considerations and cost-benefit implications are also discussed to support real-world applications.