【學術亮點】Explainable pavement surface condition classification using a TabNet-CatBoost hybrid machine learning framework
Technological Agriculture: Multi-Source Remote Sensing Image AI Recognition for Agriculture Greenhouse Gas Emission Assessment【Department of Civil Engineering / Yang, Ming-Der / Distinguished Professor】
科技農業：多元遙測影像AI辨識於農作溫室氣體排放評估【土木工程學系/楊明德特聘教授】
| 論文篇名 | 英文：Explainable pavement surface condition classification using a TabNet-CatBoost hybrid machine learning framework 中文：使用 TabNet-CatBoost 混合式機器學習框架進行可解釋的路面狀況分類 |
| 期刊名稱 | Case Studies in Construction Materials |
| 發表年份,卷數,起迄頁數 | 2025, 23, no. e05333 |
| 作者 | Shikur, Henok Desalegn; Yang, Ming-Der(楊明德)*; Kebede, Yared Bitew |
| DOI | 10.1016/j.cscm.2025.e05333 |
| 中文摘要 | 準確的路面狀況評估對於優化維護和確保交通網路安全至關重要，然而，先進機器學習模型的不透明性往往阻礙其實際應用。本研究的主要目標是開發並驗證一種新穎的、可解釋的混合框架，該框架能夠在保持透明度的同時準確分類路面狀況。該框架將基於注意力機制的TabNet模型（用於智慧特徵提取）與CatBoost分類器（用於穩健預測）相結合，使用一系列損傷和粗糙度輸入，包括車轍、縱向和橫向開裂、鱷魚皮開裂、塊狀開裂、邊緣開裂、坑洼和國際粗糙度指數。所提出的TabNet-CatBoost模型取得了卓越的性能，其宏F1得分為0.890，二次加權Kappa為0.917，顯著超越了傳統基準。事實證明，這種協同架構對於最具挑戰性的少數類別尤其有效，使關鍵失效狀況的召回率提高了一倍以上，從而增強了其在風險管理中的實用性。隨後的 Shapley 加法解釋 (SHAP) 分析證實，該模型的邏輯符合工程原理，並確定鱷魚皮開裂是主要預測因子（重要性為 22.7%）。由此產生的框架為路面狀況分類提供了一個強大且高度可解釋的工具，彌合了先進人工智慧與基礎設施管理中對可靠決策支援的實際需求之間的差距。 |
| 英文摘要 | Accurate pavement condition assessment is critical for optimizing maintenance and ensuring transportation network safety, yet the opaque nature of advanced machine learning models often hinders their practical adoption. The primary objective of this study was to develop and validate a novel, explainable hybrid framework that can accurately classify pavement condition while remaining transparent. The framework integrates an attention-based TabNet model for intelligent feature extraction with a CatBoost classifier for robust prediction, using a suite of distress and roughness inputs including rutting, longitudinal and transverse cracking, alligator cracking, block cracking, edge cracking, potholes, and the international roughness index. The proposed TabNet-CatBoost model achieved superior performance, with a macro F1-score of 0.890 and a Quadratic Weighted Kappa of 0.917, significantly surpassing traditional baselines. The synergistic architecture proved particularly effective for the most challenging minority classes, more than doubling the recall for the critical failed condition and thereby enhancing its practical utility for risk management. A subsequent Shapley Additive Explanations (SHAP) analysis confirmed that the model’s logic aligns with engineering principles, identifying alligator cracking as the predominant predictor (22.7 % importance). The resulting framework provides a robust and highly interpretable tool for pavement condition classification, bridging the gap between advanced AI and the practical need for trustworthy decision support in infrastructure management. |
| 發表成果與本中心研究主題相關性 | 目前發展的AI 模式以道路鋪面為溫度預測對象，可以應用在農地土壤分層溫度之預測。 |
