Machine-Learning-Based Prediction of Skid Resistance from 3D Pavement Surface Texture: Identifying an Effective Texture Depth

Road surface friction is influenced by pavement texture across multiple spatial scales. This study examines how well the British Pendulum Tester (BPT) Pendulum Test Value (PTV) can be predicted from 3D pavement texture descriptors derived from a static road scanner, with a particular focus on the effect of vertical truncation depth. Forty‑five asphalt surfaces were measured in situ. For each surface, BPT was performed at three locations (n = 135 point‑level samples). From each measured height map, depth‑truncated topographies were generated by clipping the surface below a truncation depth d to approximate the load‑bearing contact zone. Micro‑ and macro‑texture components were separated using an areal Gaussian filter (cut‑off wavelength 0.50 mm), and a set of 98 micro and 137 macro texture descriptors was computed. Two regression models were evaluated: Random Forest (RF) and Gradient Boosted Trees (GBT). Dimensionality was reduced by train‑only feature pre‑selection (top N features) using Spearman rank correlation with PTV. Model performance was quantified using pooled out‑of‑sample coefficient of determination (R²) under group‑aware validation schemes: grouped k‑fold cross‑validation (groupKfold; groups = surfaces) and leave‑one‑surface‑out (LOSO). Across models and validation schemes, predictive performance consistently peaked at d ≈ 1.4–1.6 mm (pooled R² up to ≈0.31 in groupKfold and ≈0.47 in LOSO for point‑level modelling). A surface‑aggregated variant (per‑surface averaging of three measurements) yielded comparable performance (pooled R² up to ≈0.32 in groupKfold and ≈0.42 in LOSO). Overall, the results suggest that depth‑limited texture descriptors contain measurable, albeit moderate, predictive information about friction, and that selecting an effective truncation depth (around 1.5 mm) is critical under BPT conditions for building more generalizable texture-friction models.