Department

Civil and Environmental Engineering

Document Type

Article

Publication Date

Spring 5-2026

Embargo Period

8-10-2026

Abstract

This study develops and validates maintenance-aware machine learning models for predicting the Highway Pavement Condition Index (HPCI) on the Korean expressway network. Multiple regression and tree-based models were trained and tested using the pavement condition surveys archived in the Highway Pavement Management System (HPMS). A stacking regressor that integrates random forest, gradient boosting, and extreme gradient boosting as base learners exhibited the most robust predictions. Performance metrics indicated that the stacking ensemble achieved a mean absolute error of 0.21, a root mean square error of 0.31, and a coefficient of determination exceeding 0.73 on the testing dataset. Also, the residuals revealed a near-zero mean, showed no significant bias based on one-sample t-tests, and demonstrated no evidence of autocorrelation, with Durbin–Watson statistics close to 2.0. Feature importance analysis showed that initial pavement condition, age-based features, and maintenance cycle strongly affect the outcomes, which suggests that the model should capture physically meaningful deterioration, recovery mechanisms rather than purely statistical correlations. A retrospective diagnosis demonstrated model’s ability to reproduce observed annual HPCI trajectories, including both gradual deterioration and sharp post-maintenance recovery. HPCI forecasts under a policy-based maintenance rule further supported model’s potential for identifying future maintenance needs and network-level condition evolution, which is an essential function of the HPMS.

Journal Title

Computer-Aided Civil and Infrastructure Engineering

Journal ISSN

1093-9687

Volume

47

Issue

2026

Digital Object Identifier (DOI)

https://doi.org/10.1016/j.cacaie.2026.100091

Download

Share

COinS