Original Development of a Long-Term Pavement Performance Prediction Model for Tropical African Road Environments Using Machine Learning

Long-term pavement performance prediction in tropical African road environments differs fundamentally from temperate climate contexts due to seasonal extremes, sub-grade moisture seasonality, and different traffic composition, yet existing performance prediction models such as AASHTO-MEPDG and HDM-4 are calibrated for non-tropical conditions. This study develops an original machine learning-based pavement performance prediction model calibrated for tropical African conditions using Nigerian pavement performance data. A comprehensive pavement performance database was assembled from 22 years of condition monitoring data on 340 pavement sections across all six Nigerian geopolitical zones, covering diverse climate zones, traffic loadings, construction standards, and maintenance intervention histories. Features for the predictive model include pavement structural indicators, climate variables (cumulative rainfall, dry season temperature maxima), traffic loading spectra, subgrade CBR, and maintenance intervention records. Ensemble machine learning methods (gradient boosting, random forest, and deep neural network architectures) were trained, cross-validated, and compared against HDM-4 predictions. The best-performing model (gradient boosting ensemble) achieves RMSE of 3.8 PCI units for 5-year performance prediction, compared to HDM-4's RMSE of 9.2 for the same dataset. The model identifies sub-grade moisture seasonality and initial IRI as the two strongest performance predictors in the Nigerian context, differing from the axle load dominance in temperate model calibrations. The study contributes an original, empirically grounded pavement performance prediction capability for tropical African road engineering.

Keywords: pavement performance prediction, machine learning, tropical pavement, HDM-4, Nigeria