Reconstruction of historical pavement condition records using temporal transfer learning under non-random missing data

Abstract

Incomplete historical pavement condition records, particularly when missing non-randomly, challenge pavement performance analysis and decision support. Models trained on selectively observed data distort deterioration estimates, weakening life-cycle planning credibility. Moreover, existing pavement condition imputation approaches focus on filling recent gaps rather than the inverse problem of reconstructing earlier conditions from later surveys. Explicit treatment of missing-not-at-random (MNAR) data mechanism also remains limited. This study proposes and validates a hindcast-based imputation framework for reconstructing missing historical pavement roughness by transferring models trained on recent data backward in time through principled calibration. Using international roughness index (IRI) data for two survey waves on the Kenya national trunk road network in 2018 and 2023, the framework integrates three complementary base learners (random forest, XGBoost, and neural networks) with post-hoc ensembles of the best-performing models. Performance is evaluated using spatial cross-validation to ensure geographically realistic assessment. Temporal non-stationarity between survey waves is addressed through drift calibration, incorporating a modelled MNAR weighting scheme based on network co-variates. Beyond point accuracy, the framework emphasizes systematic uncertainty diagnostics through prediction-interval coverage, width, and residual structure. Across all algorithms, temporal calibration emerges as the dominant driver of performance improvement, eliminating systematic bias and yielding stable post-calibration accuracy (MAE ≈ 2.7 IRI units; NRMSE ≈ 0.17). While aggregate error metrics differ little across models, uncertainty evaluation reveals meaningful variation in interval behaviour, demonstrating that credible imputation requires assessing dispersion in addition to central tendency. Sensitivity analysis shows that MNAR weighting materially strengthens the interpretability of calibration and uncertainty by aligning inference with the documented selective observation process. The framework demonstrates how combining temporal drift correction, non-random missing data treatment, spatially robust evaluation, and explicit uncertainty assessment enables credible reconstruction of historical pavement condition, offering a practical foundation for network-level performance analysis and pavement management in data-constrained settings.