Download PDFOpen PDF in browserEvaluating the Impact of Trend-Smoothing on Pavement Condition Classification Models10 pages•Published: June 2, 2026AbstractMachine learning (ML) has become an increasingly important component of pavement management systems (PMS), where historical condition data are used to support maintenance and rehabilitation planning. However, these datasets frequently contain sensor noise and undocumented field activities, which can introduce abrupt and unrealistic improvements in condition scores. Such anomalies disrupt the expected gradual deterioration patterns and can reduce the predictive reliability of data-driven models. To address this, a trend-consistent adjustment method based on typical annual deterioration rates is applied to smooth the data and restore continuity. This study identifies the optimal degree of smoothing to apply during data preprocessing to maximize generalization accuracy when training data contain noise and unrecorded maintenance events. Model evaluation is performed using the original, uncorrected test data to reflect real-world prediction scenarios. Prediction performance improved significantly with correction, with the best model (Voting Classifier) reaching an F1-Score of 81.98% at the 10% correction level, representing a 7.35 percentage point increase over the raw data baseline. The optimal correction range was found to be 10-20% of imperfect sections corrected, confirming that light, selective smoothing balances trend fidelity with real-world variability better than raw or heavily smoothed data, producing a more reliable prediction of pavement deterioration.Keyphrases: artificial intelligence, data imputation, pavement deterioration, pavement management, predictive modeling In: Wesley Collins, Anthony Perrenoud and John Posillico (editors). Proceedings of Associated Schools of Construction 62nd Annual International Conference, vol 7, pages 713-722.
|
