Scholarlink

<p>Accurate prediction of pavement conditions is critical for resource allocation in transportation asset management. This study develops and evaluates machine learning classification models to predict pavement roughness classes, using a nationwide dataset from the U.S. National Highway System encompassing varied climates, traffic loads, and roadway types. Four algorithms: Logistic Regression, Support Vector Machine, Random Forest, and eXtreme Gradient Boosting (XGBoost) were compared, with XGBoost selected as the baseline due to its balanced performance. Across the four models, macro-averaged F1-scores ranged from 0.87 to 0.89, indicating comparable baseline performance on imbalanced data. To address class imbalance, four data augmentation methods: SMOTE, SMOTENC, CTGAN and TabDDPM were tested. The results indicate that augmentation did not improve performance. SMOTE and SMOTENC showed minor reductions in macro-averaged metrics, while CTGAN and TabDDPM yielded F1-scores comparable to the baseline. Confusion-matrix analysis shows that CTGAN slightly improved classification of the good class, whereas TabDDPM improved the Acceptable class. Synthetic data quality assessment confirmed that CTGAN and TabDDPM reproduced data distributions with reasonable fidelity but did not improve predictive outcomes. The findings suggest that classification models can reliably support pavement condition forecasting, while augmentation techniques offer limited advantages under the tested conditions in this case.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Pavement Condition Classification: Empirical Evidence from Machine Learning Models with Data Augmentation

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Ali Mansouri,
  • Narjes Shayesteh,
  • Abdolmajid Erfani

摘要

Accurate prediction of pavement conditions is critical for resource allocation in transportation asset management. This study develops and evaluates machine learning classification models to predict pavement roughness classes, using a nationwide dataset from the U.S. National Highway System encompassing varied climates, traffic loads, and roadway types. Four algorithms: Logistic Regression, Support Vector Machine, Random Forest, and eXtreme Gradient Boosting (XGBoost) were compared, with XGBoost selected as the baseline due to its balanced performance. Across the four models, macro-averaged F1-scores ranged from 0.87 to 0.89, indicating comparable baseline performance on imbalanced data. To address class imbalance, four data augmentation methods: SMOTE, SMOTENC, CTGAN and TabDDPM were tested. The results indicate that augmentation did not improve performance. SMOTE and SMOTENC showed minor reductions in macro-averaged metrics, while CTGAN and TabDDPM yielded F1-scores comparable to the baseline. Confusion-matrix analysis shows that CTGAN slightly improved classification of the good class, whereas TabDDPM improved the Acceptable class. Synthetic data quality assessment confirmed that CTGAN and TabDDPM reproduced data distributions with reasonable fidelity but did not improve predictive outcomes. The findings suggest that classification models can reliably support pavement condition forecasting, while augmentation techniques offer limited advantages under the tested conditions in this case.