A machine learning–based computational framework for road performance assessment in Developing Countries

Abstract

Road infrastructure assessment in developing countries remains fragmented. Existing methods focus on isolated dimensions such as pavement condition or geometric design, limiting their utility for comprehensive planning. This study develops the first holistic road quality assessment framework, applying knowledge distillation techniques to Kenya's national road dataset (260,773 segments). Six infrastructure dimensions—surface condition, material type, geometric design, facilities, usage patterns, and functional classification—were integrated into a universal scoring system. The Random Forest teacher model (R² = 0.93; MAE = 0.0051) was successfully distilled into an interpretable polynomial regression formula. Results show that surface condition has the highest influence (37.3%), followed by infrastructure facilities (20.0%), usage patterns (14.6%), and functional classification (11.6%). The framework captures the heterogeneity of Kenya's road network, where 85.1% of roads are unpaved. This work represents the first synthesis of multiple infrastructure criteria into a unified, quantitative formula. The distillation of complex machine learning into an interpretable polynomial equation enables consistent, reproducible road quality scoring. The methodology is transferable—other countries can adapt the six-dimension framework by recalibrating weights to reflect local infrastructure priorities. This provides an evidence-based tool for infrastructure planning, maintenance prioritization, and policy decision-making. Countries with existing road inventory systems can implement this framework by retraining the model on their data to derive context-appropriate dimensional weights.