Scholarlink

<p>Earthquakes cause major human and economic losses worldwide, with large-scale building damage overwhelming traditional inspection procedures. Rapid and reliable post-earthquake damage assessment is therefore crucial for effective emergency response and seismic risk management. This study explores the potential of supervised machine learning (ML) for predicting global building damage levels using over 27,000 inspection records from the 2003 Boumerdes earthquake in Algeria. Four classifiers—K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Decision Trees (DT), and Random Forests (RF)—were systematically compared. Bayesian hyperparameter optimization with Optuna, guided by five-fold cross-validation and the weighted F1-score, ensured robust model calibration under imbalanced class distributions. Model performance was assessed using weighted accuracy, precision, recall, specificity, F1-score, and Cohen’s kappa, supported by confusion matrix analyses. Results show that ensemble-based methods, particularly Random Forest, outperform other models by achieving higher predictive accuracy and improved recognition of rare yet critical classes such as severe damage and collapse. Feature importance analysis using mean decrease in impurity and permutation methods revealed the dominant influence of structural components. Overall, the findings demonstrate that optimized ML models offer scalable, objective decision-support tools to enhance post-earthquake assessment and guide seismic risk reduction.</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

Bayesian-optimized machine learning classifiers for post-earthquake building damage prediction

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Halima Belbahi,
  • Hichem Noura,
  • Mohamed Abed

摘要

Earthquakes cause major human and economic losses worldwide, with large-scale building damage overwhelming traditional inspection procedures. Rapid and reliable post-earthquake damage assessment is therefore crucial for effective emergency response and seismic risk management. This study explores the potential of supervised machine learning (ML) for predicting global building damage levels using over 27,000 inspection records from the 2003 Boumerdes earthquake in Algeria. Four classifiers—K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Decision Trees (DT), and Random Forests (RF)—were systematically compared. Bayesian hyperparameter optimization with Optuna, guided by five-fold cross-validation and the weighted F1-score, ensured robust model calibration under imbalanced class distributions. Model performance was assessed using weighted accuracy, precision, recall, specificity, F1-score, and Cohen’s kappa, supported by confusion matrix analyses. Results show that ensemble-based methods, particularly Random Forest, outperform other models by achieving higher predictive accuracy and improved recognition of rare yet critical classes such as severe damage and collapse. Feature importance analysis using mean decrease in impurity and permutation methods revealed the dominant influence of structural components. Overall, the findings demonstrate that optimized ML models offer scalable, objective decision-support tools to enhance post-earthquake assessment and guide seismic risk reduction.