Machine‑learning prediction of skid resistance by British Pendulum Tester from 3D pavement surface texture: assessing the relative contributions of micro‑ and macrotexture

Abstract

Skid resistance is strongly influenced by pavement surface texture, yet the texture–friction relationship is nonlinear and scale dependent. This study evaluates prediction of Pendulum Test Value (PTV) from 3D pavement-surface texture descriptors and quantifies the relative roles of microtexture and macrotexture. A dataset of 872 co-located texture–friction samples from 45 asphalt pavement test sections was used to train multivariable regression models on microtexture-only, macrotexture-only, and combined predictor sets. Across all inputs, nonlinear methods outperformed linear regression. On the combined set, support vector regression achieved the best accuracy (R2 = 0.82 ± 0.03; RMSE = 4.8 ± 0.3 PTV), followed closely by random forest (R2 = 0.78 ± 0.03; RMSE = 5.2 ± 0.4), compared with the linear baseline (R2 = 0.60 ± 0.04; RMSE = 7.0 ± 0.4). Under strict section-wise validation (leave-one-section-out and 5-fold group cross-validation), out-of-section performance decreases (for the RF reference model on the COMBINED set, R2 ≈ 0.46–0.50 and RMSE ≈ 8 PTV), highlighting limited transfer to unseen sections and the need for broader training data. Combining microtexture and macrotexture systematically improved performance over single-scale models, indicating complementary predictive information. Permutation-based interpretation with random forests identified Micro-Smrk2 as the most influential descriptor, with additional contributions from both texture scales. Group-level perturbation showed that microtexture is more critical for predictive accuracy on held-out data, while macrotexture provides non-negligible complementary signal.