Presentation Information

[AI-12]A Dual-Model Machine Learning Framework for Pavement Rutting and Condition Classification

Aaditya Ojha1, Jialu Yu1, *Shadi Saadeh1 (1. California State University, Long Beach (USA))

Keywords:

Pavement performance,Rutting prediction,Machine Learning,Pavement Condition Classification,Predictive Modeling

This study presents a dual-model machine learning framework using Long-Term Pavement Performance (LTPP) data to predict pavement rutting and classify overall pavement condition based on commonly available variables, including equivalent single axle loads (ESALs), pavement thickness, temperature, precipitation, and drainage. A K-Nearest Neighbors (K-NN) regression model was developed to estimate rut depth and demonstrated reasonable predictive performance, with a sum of squared errors of 205.03. The model performed particularly well for pavements exhibiting moderate rutting levels between 2 and 5 mm.In parallel, a Decision Tree classification model based on the CART algorithm was employed to categorize pavement condition as either “Good” or “Unsatisfactory” using International Roughness Index (IRI) thresholds. This classifier achieved an overall accuracy of 81.25% and a recall of 90.91%, indicating strong capability in identifying pavements in good condition. Feature importance analysis revealed pavement thickness, rut depth, and temperature as the most influential variables in condition classification.Both the K-NN regression and Decision Tree classification models exhibited consistent and reliable performance, highlighting their suitability for pavement performance assessment. The results demonstrate the effectiveness of integrating multiple machine learning approaches to support data-driven pavement management decisions. The study also underscores the potential of machine learning tools for enhancing pavement condition evaluation and forecasting. Future research should focus on expanding the dataset, improving variable definitions, and incorporating additional performance indicators to further enhance model robustness and applicability.