使用监督机器学习进行地震去聚类
Earthquake Declustering Using Supervised Machine Learning
AI总结：
本文提出一种基于监督机器学习的去聚类方法，利用ETAS模型生成合成数据并以最近邻距离为特征训练模型，在南加州和意大利地震目录上验证其优于传统NND和随机去聚类方法。
摘要：
地震目录去聚类是将事件簇与背景地震活动分离的过程，这是统计地震学、地震预测和概率地震危险性分析中的一项重要任务。文献中已引入多种去聚类方法，这些方法基于背景事件独立发生而事件簇由先前事件触发的假设运行。本文通过利用两种流行的统计方法，测试了监督机器学习（SML）在去聚类问题上的能力。首先，将传染型余震序列（ETAS）模型拟合到目标目录，并使用其参数生成合成地震数据，以复现目标目录的震级-空间-时间地震活动特征。接下来，计算每个模拟事件之间的最近邻距离（NND）度量，并将其作为特征用于训练SML算法。最后，将训练好的算法应用于合成测试数据的去聚类，然后应用于原始目标目录。结果表明，SML方法在测试数据上的表现优于基于NND的方法和随机去聚类方法，并且在应用于真实地震活动时，对背景事件和簇事件做出了更细致的区分选择。虽然SML技术的绝大部分预测能力似乎来自“第一”最近邻的NND值，但机器学习分析表明，通过加入额外的“后续”最近邻和差分震级特征可以提高预测精度。所开发的方法被应用于南加州和意大利的地震目录以进行去聚类。
英文摘要：
Earthquake catalog declustering is the procedure of separating event clusters from background seismicity, which is an important task in statistical seismology, earthquake forecasting, and probabilistic seismic hazard analysis. Several declustering methods have been introduced in the literature and operate under the supposition that background events occur independently while clusters are triggered by prior events. Here, we test the ability of Supervised Machine Learning (SML) on the declustering problem by leveraging two popular statistical methods. First, the Epidemic Type Aftershock Sequence (ETAS) model is fit to a target catalog and the parameters are used to generate synthetic earthquake data, which replicate the magnitude-space-time seismicity of the target catalog. Next, the Nearest-Neighbour Distance (NND) metrics are computed between each simulated event and used as features to train the SML algorithm. Finally, the trained algorithm is applied to decluster synthetic testing data and then the original target catalog. Our results indicate that SML method performs better than the NND-based and stochastic declustering methods on the test data and makes more nuanced selections of background and clustered events when applied to real seismicity. While the vast majority of the SML technique's predictive power appears to lie within the NND values of the ''first'' nearest-neighbours, a machine learning analysis reveals that predictive accuracy can be improved by additional ''next'' nearest-neighbours and differential magnitude features. The developed approach is applied to seismic catalogs in Southern California and Italy to decluster them.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
