AI Model Documentation: Machine-Learning Prediction of Pavement Skid Resistance from 3D Surface Texture (FricPred Project)

This record contains the technical documentation of the finalized machine-learning model developed under the FricPred project ("Advancing Friction Prediction through AI-Powered Pavement Texture Analysis"; project code 09I03-03-V04-00041), together with the MATLAB analysis code used to train and validate the models, satisfying Deliverable No. 2 as defined in the project's grant agreement (Work Package 5): "Documentation of the finalized AI model, detailing its architecture, training methodology, performance metrics, and deployment guidelines." Summary The documented models predict two independent pavement friction measures — the British Pendulum Test Value (PTV) and the Traction Watcher One coefficient (TWO, µ) — from non-contact 3D surface-texture measurements acquired by two independent methods: high-resolution laser scanning (SRS) and stereo-photogrammetry. Four regression architectures (Linear Regression, Support Vector Regression, Random Forest, Gaussian Process Regression) were trained and compared under three validation schemes of increasing rigor (repeated random split, section-blocked 5-fold cross-validation, and Leave-One-Section-Out), on a dataset of 869 (PTV) / 848 (TWO) point-level observations across 45 / 44 real asphalt pavement sections near Žilina, Slovakia. The best-performing configuration — Gaussian Process Regression on the combined micro- and macro-texture descriptor set derived from laser-scan data — achieves R² ≈ 0.64–0.83 depending on validation scheme and target variable. Under the strictest section-wise validation (LOSO), it reaches R² ≈ 0.68 for PTV and ≈ 0.66 for TWO, whereas stereo-photogrammetry collapses to R² ≈ 0.23–0.27. The laser scanner consistently and statistically significantly outperforms stereo-photogrammetry as an acquisition method (p < 0.001 across all descriptor sets and both target variables). Contents Full documentation — architecture, training methodology, performance metrics for both PTV and TWO, and deployment guidelines Complete performance matrix — all four architectures × three validation schemes × both acquisition methods, for both target variables (documentation, Section 4.4) Analysis code (FricPred_Row_GroupK_LOSO_PTV.m, FricPred_Row_GroupK_LOSO_TWO.m) — the MATLAB pipeline used to train and validate every model reported here. Running either script regenerates the full machine-readable Model_Summary workbook (R² / RMSE / MAE per method, per validation scheme, per acquisition method). Related records 3D laser-scan (SRS) point clouds and friction measurements (PTV, TWO): https://doi.org/10.5281/zenodo.20606458 Stereo-photogrammetry orthophotos and height maps and friction measurements (PTV, TWO): https://doi.org/10.5281/zenodo.20608617 (Note: only raw, unprocessed surface measurements are shared in the above records — not the authors' specific texture-parameter processing pipeline — so that other researchers can apply independent processing choices and compare results.) This project was funded by the European Union — NextGenerationEU, through the Recovery and Resilience Plan of the Slovak Republic (call 09I03-03-V04, project 09I03-03-V04-00041), administered by the Research Agency (Výskumná agentúra) on behalf of the Government Office of the Slovak Republic.