基于机器学习融合多变量参数的地震烈度评估研究

Research on earthquake intensity assessment based on machine learning with multi-variable parameter integration

  • 摘要: 地震烈度的快速精准评估可为极震区判定、伤亡预估、应急响应及物资调配等提供决策依据。为提高地震烈度评估的精度与鲁棒性,本研究基于日本K-NET强震动数据库,分别采用随机森林(RF)与极限梯度提升(XGBoost)两类集成学习算法,构建了融合多变量参数的地震烈度预测模型。通过采取全特征输入与关键特征输入两种策略,对比分析了两类模型在特征鲁棒性、预测精度、泛化稳定性及拟合优度等方面的性能差异。结果表明,全特征融合策略可显著提升烈度预测精度;XGBoost模型在全特征条件下拟合精度最优(测试集RMSE=0.242,R2=0.928),而RF模型在特征缩减条件下仍保持较高的预测稳定性(RMSE=0.262,R2=0.917),表现出更强的特征鲁棒性。应用于2008年汶川MS8.0地震记录,两种模型求出的全烈度区间平均预测误差均控制在0.5度以内,验证了模型对复杂地震场景的有效泛化能力。本研究为地震烈度快速评估提供了一个兼顾预测精度与特征鲁棒性的建模框架,也为多源地震动参数融合及震后应急响应提供了技术参考。

     

    Abstract:
    The assessment of post-earthquake seismic intensity constitutes a critical component of emergency response, disaster mitigation and resource allocation. Traditional approaches relying on single ground-motion parameters such as peak ground acceleration (PGA) or peak ground velocity (PGV) suffer from considerable uncertainties and limited robustness, particularly under complex seismic conditions. Furthermore, these methods fail to exploit the rich information contained in multi-dimensional ground-motion records. This study develops an integrated framework based on ensemble learning to incorporate multivariate ground-motion parameters and improve the accuracy and generalization of seismic intensity prediction.
    We compile a high-quality ground-motion dataset using strong-motion records from Japan’s K-NET network spanning 2015–2025. Raw accelerograms from K-NET undergo rigorous preprocessing, including baseline correction, band-pass filtering (0.1–10 Hz) and Arias-intensity truncation, to guarantee data quality and consistency. A total of 28 parameters are extracted from each processed waveform, covering amplitude metrics (e.g., PGA, PGV, peak ground displacement), frequency-domain attributes (e.g., predominant frequency, spectral acceleration at different periods), and cumulative energy indicators (e.g., Arias intensity, cumulative absolute velocity). This multi-dimensional feature set offers a comprehensive characterization of the damage potential of ground motion.
    To match the Chinese seismic intensity scale (GB/T 17742-2020), the original JMA seismic intensity labels are converted into Chinese instrumental intensity values. This conversion accounts for inherent differences in definition and calculation between the two standards, enabling the trained models to be directly applicable to China and supporting cross-standard comparison.
    Two modeling strategies are investigated in this work. The first is an image-oriented deep-learning pipeline. Time-series waveforms and Fourier amplitude spectra are converted into images and fed into a convolutional neural network (CNN) for automatic extraction of hierarchical deep features. The extracted features are then combined with random forest (RF) and extreme gradient boosting (XGBoost) classifiers for intensity prediction. This hybrid design inherits the CNN’s strength in spatial feature extraction and the favorable robustness of ensemble learners for regression tasks.
    The second strategy is parameter-driven, which directly uses the 28 derived parameters as inputs to RF and XGBoost. For both strategies, two feature schemes are evaluated: a full-feature scheme adopting all 28 parameters, and a key-feature scheme retaining only top-ranked features according to model importance scores. This comparison aims to strike a balance between prediction accuracy and model parsimony, and to examine model performance under feature reduction.
    The K-NET dataset is used for model training and validation, with root mean square error (RMSE), coefficient of determination (R2) and mean absolute error (MAE) adopted as evaluation metrics. The image-driven model achieves stable performance on the KiK-net test set: approximately 90% of K-NET test samples yield absolute intensity errors smaller than 0.5 units, an improvement of 18 percentage points relative to the baseline of 72%. Moreover, the model eliminates extreme outliers with errors exceeding 1.0 intensity unit, demonstrating the value of high-quality data and well-constrained site conditions for image-based learning.
    Among parameter-driven models, the XGBoost model trained on the full K-NET feature set yields the highest accuracy, with R2=0.949 and RMSE is about 0.24. The RF model delivers slightly lower accuracy on the full-feature set (R2 is about 0.928), yet exhibits outstanding stability after feature screening; it maintains R2=0.917 and RMSE=0.262 with only key predictors retained. This indicates that RF possesses stronger feature robustness, making it a viable alternative when computational resources are limited or only a subset of parameters is available. Feature importance analysis reveals that velocity-related parameters gain substantially higher weights after feature engineering and intensity-scale conversion, especially the effective peak velocity (EPV), whose importance rises from below 0.02 to 0.365. This result highlights the physical link between ground-motion velocity and seismic damage, which is frequently underestimated in empirical models that prioritize acceleration.
    We further test cross-standard generalization under the Chinese instrumental intensity scale (GB/T 17742-2020). The full-feature XGBoost model trained on K-NET data attains an even higher R2 of 0.959, verifying the validity of the scale conversion and the model’s adaptability to alternative intensity definitions. The practical performance is examined using 124 independent strong-motion records from the 2008 Wenchuan MS8.0 earthquake, which are excluded from training. The K-NET-trained RF and XGBoost models produce mean prediction errors within 0.5 intensity units across the full intensity range, and the maximum error for high-intensity samples (intensity≥VII) does not exceed 1.5 units. These results confirm the strong generalization and engineering applicability of the proposed framework for real earthquake events in China.
    In summary, this study presents a machine-learning framework for seismic intensity prediction that integrates multivariate ground-motion parameters. The full-feature XGBoost model trained on K-NET achieves state-of-the-art accuracy, while RF serves as a robust substitute under limited feature availability. Validation against Wenchuan earthquake data illustrates the framework’s capacity for rapid post-earthquake intensity mapping, emergency decision-making and disaster mitigation. This work also underscores the importance of velocity-based ground-motion parameters and the need for standard harmonization for cross-regional applications.

     

/

返回文章
返回