Answers are grounded in this paper and cite their sources.
Predictive modeling study reveals reliable road deterioration forecasting from sparse condition archives in national trunk networks, highlighting viable pavement management under data scarcity.
Many low- and middle-income countries must manage extensive road networks using temporally sparse condition data and incomplete maintenance histories, while prevailing pavement deterioration modelling paradigms assume data-richness in these dimensions. This structural mismatch limits the predictive utility of conventional Pavement Management Systems (PMS). To address this, a spatially validated, probabilistic hindcasting framework is proposed for characterizing pavement roughness evolution under such constraints, demonstrated on Kenya’s asphalt-surfaced national trunk road network. Using network-level surveys from 2018 and 2023, five-year International Roughness Index (IRI) class changes were modelled as an interval-based process conditioned on baseline (2018) condition, inventory, structural, and contextual attributes. Penalized logistic regression (PLR), decision tree (CART), and random forest (RF) models were trained across binary, three-class and six-class IRI formulations and evaluated under a spatially blocked nested cross-validation framework. Model performance was assessed using discrimination metrics (ROC-AUC, PR-AUC) and probability calibration diagnostics (Brier, Brier Skill, Expected Calibration Error, slope, and intercept). Post-hoc recalibration using Platt scaling and isotonic regression was applied selectively to correct miscalibration. Results show moderate-to-strong, spatially transferable discrimination (ROC-AUC = 0.66–0.87; PR-AUC consistently exceeding prevalence baselines), confirming effective ranking of deterioration risk from sparse condition archives, while CART exhibited discrimination collapse for some targets. Recalibration of PLR and RF effectively improved probability reliability. Retrospective consistency analysis incorporating attribute-change information confirmed temporal robustness of baseline models. Overall, the framework demonstrates that spatially transferable, uncertainty-aware deterioration forecasts can be achieved from limited data, providing a practical pathway for strengthening predictive PMS capability in data-scarce environments.
No takes yet. Share an insight, caveat, or question.
Odera et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: