Data-augmented explainable AI for pavement roughness prediction

Abstract

Effective pavement management systems rely on accurate predictions of pavement conditions to guide strategic decisions about maintenance and rehabilitation projects. Although recent studies have explored various artificial intelligence-based methods for predicting pavement roughness, notable gaps remain in the literature. Existing studies often use homogeneous data from similar climates and pavement types and overlook imbalances in historical pavement condition data. They also treat machine learning models as black boxes, relying on static feature rankings that miss complex relationships between inputs and predictions. This paper bridges these gaps by applying an explainable AI framework, enhanced with data augmentation, to a diverse and comprehensive dataset of pavement conditions. The proposed approach enhanced performance across a comprehensive set of metrics, reducing RMSE by 28.28 %, RSR by 36.92 %, and WMAP by 33.74 %, while increasing R-squared by 7.28 % and VAF by 6.61 %. Explainable AI analysis provided practical insights, enhancing model applicability and supporting informed maintenance decisions.