基于深度学习的日本强震远场触发中国东北地震预测
Prediction on Far-Field Triggering Earthquakes in Northeast China Induced by Japanese Strong Earthquakes Based on Deep Learning
ES评分 0
| DOI |
10.14075/j.jgg.2025.09.317 |
| 刊名 |
Journal of Geodesy and Geodynamics
|
| 年,卷(期) |
2026, 46(6) |
| 作者 |
|
| 作者单位 |
1 黑龙江省地震局, 哈尔滨, 150090; 2 中国地震局地震预测研究所, 北京, 100036
|
| 摘要 |
远场地震触发机制是地震预测研究的重要前沿方向。为探究日本强震活动与中国东北地区中强地震的时空关联, 本文构建了一种数据驱动的预测模型。基于1980-2024年USGS地震目录, 以日本M≥6.0地震为潜在触发源, 预测未来60 d内中国东北地区发生M≥4地震的概率。构建一个涵盖地震统计、空间分布、能量释放及余震序列等32维特征的日分辨率时间序列数据集, 系统对比长短期记忆神经网络(LSTM)、注意力机制增强LSTM(Attention-LSTM)、Transformer等深度学习模型与逻辑回归(logistic regression, LR)、随机森林(random forest, RF)等传统机器学习方法。结果显示, Attention-LSTM模型表现最优, F1分数达0.915, AUC值为0.661, 在地震预测专用的Molchan误差图分析中亦展现显著优势。该模型不仅可实现二元分类预测, 还可生成1 1 网格分辨率的空间概率分布图。研究验证了深度学习在揭示跨区域地震远场触发机制中的潜力, 可为中短期区域地震危险性评估提供新思路。
|
| Abstract |
Far-field earthquake triggering mechanisms represent a critical frontier in earthquake prediction research. To explore the spatiotemporal relationship between strong seismic activity in Japan and moderate-to-strong earthquakes in northeast China, this study develops a data-driven prediction model. Based on the USGS earthquake catalog from 1980 to 2024, and using M≥6.0 earthquakes in Japan as potential triggers, the model predicts the probability of earthquakes with M≥4 occurring in northeast China within the next 60 days. A daily-resolution time series dataset is constructed, incorporating 32-dimensional features covering earthquake statistics, spatial distribution, energy release, and aftershock sequences. Deep learning models including long short-term memory (LSTM) neural networks, attention mechanism-enhanced LSTM (Attention-LSTM), and Transformer are systematically compared against traditional machine learning methods such as logistic regression and random forest. Results indicate that the Attention-LSTM model performs optimally, achieving an F1 score of 0.915 and an AUC value of 0.661, and shows significant advantages in Molchan error diagram analysis, a method specific to earthquake prediction evaluation. The model enables both binary classification prediction and generation of spatial probability distribution maps at 1 1 grid resolution. This study demonstrates the potential of deep learning for revealing cross-regional far-field earthquake triggering mechanisms and provides new insights for short-to-medium-term regional seismic hazard assessment.
|
| 关键词 |
深度学习; 注意力机制长短期记忆网络; 远场地震触发; 地震时空预测
|
| KeyWord |
deep learning; attention-based long short-term memory network (Attention-LSTM); far-field earthquake triggering; spatiotemporal earthquake forecasting
|
| 基金项目 |
|
| 页码 |
710-717 |
高峰, 李美.
基于深度学习的日本强震远场触发中国东北地震预测 [J].
大地测量与地球动力学.
2026; 46; (6).
710 - 717.