Machine Learning–Based Dual Prediction of Pavement Roughness and Condition Rating with Taylor and Information-Theoretic Validation

This study develops and validates a multioutput machine learning regression framework for the simultaneous prediction of the International Roughness Index (IRI) and pavement condition rating (PCR) for major district roads (MDRs) in Punjab, India. The dataset comprises 1,859 pavement segments characterized by six predictor variables: bituminous thickness, granular sub-base thickness, rut depth, crack severity, air temperature, and pavement surface temperature. Six regression models, namely linear regression (LR), support vector regression (SVR), K-nearest neighbors (KNN), random forest (RF), gradient boosting (GB), and neural networks (NN), were evaluated within a multioutput regression framework. Among these, GB achieved the best performance, with an R ² of 0.639 (root mean square error [RMSE] = 0.951; mean absolute error [MAE] = 0.625) for the International Roughness Index (IRI) and 0.998 (RMSE = 0.032; MAE = 0.007) for the pavement condition rating (PCR). The framework combines simultaneous prediction using a common set of predictor variables with a dual-layer validation strategy based on Taylor diagrams and scaled mutual information (IRI MI ≈ 0.71; PCR MI > 2.1), providing complementary statistical and information-theoretic evaluation. The predicted IRI and PCR values were interpreted using maintenance thresholds derived from Indian Roads Congress (IRC) guidelines to support pavement management decisions. The results demonstrate that reliable pavement condition assessment can be achieved using routinely collected survey data, providing a practical and scalable decision-support tool for maintenance planning in data-constrained road networks.