Abstract Adequate detection and classification of seismic events are crucial for understanding the internal status of a Volcano. Machine learning-based classifiers use different features from the time, frequency, and scale domains related to seismic events. Regarding power spectrum-based features, several methods can be used to compute such features. However, the more suitable method for analyzing volcanic activity is undetermined. This paper presents a study about the main frequency bands, which allows maximizing the performance metrics of an automated classifier for long-period (LP) and volcano-tectonic (VT) events based on parametric (Yule-Walker and Burg) and non-parametric (Welch and Multitaper) power spectrum density estimation methods. Feature selection using embedded (pruning) and wrapper (recursive feature elimination) methods was applied to select the main frequencies that maximize the balanced error rate of suitable classification algorithms, such as decision trees (DT) and support vector machines (SVM). Bootstrapping was used to estimate a confidence interval for the frequencies of the microearthquakes. An amplitude threshold difference of at least 3 dB was used to guarantee that possible frequency features that characterize each type of event do not overlap between classes. The method who achieved the worst overall performance was not considered by the voting strategy. A Dataset from Cotopaxi volcano was used to test the proposed classification schema. The best results show for DT classifier a total of 10 key frequencies, while for SVM classifier 39 key frequencies grouped in three main frequency bands, as main features to distinguish LP events from VT earthquakes. The best classification results were achieved by the Welch method with the DT and by the Multitaper method with the SVM classifiers. Furthermore, the study confirms that there is a frequency band above 40 Hz, which seems like a critical feature for the detection and classification of stages.

中文翻译：

---

寻找识别科托帕希火山微地震的可能频率：一种基于机器学习的方法

摘要 地震事件的充分检测和分类对于了解火山的内部状态至关重要。基于机器学习的分类器使用与地震事件相关的时间、频率和尺度域的不同特征。关于基于功率谱的特征，可以使用多种方法来计算这些特征。然而，更适合分析火山活动的方法尚未确定。本文介绍了关于主要频段的研究，该研究允许基于参数（Yule-Walker 和 Burg）和非参数的长周期 (LP) 和火山构造 (VT) 事件最大化自动分类器的性能指标。 （Welch 和 Multitaper）功率谱密度估计方法。使用嵌入（修剪）和包装器（递归特征消除）方法进行特征选择，以选择使合适分类算法（例如决策树 (DT) 和支持向量机 (SVM)）的平衡错误率最大化的主要频率。自举被用来估计微地震频率的置信区间。使用至少 3 dB 的幅度阈值差异来保证表征每种类型事件的可能频率特征在类别之间不重叠。投票策略不考虑整体表现最差的方法。来自 Cotopaxi 火山的数据集用于测试提议的分类模式。最好的结果显示 DT 分类器共有 10 个关键频率，而 SVM 分类器将 39 个关键频率分为三个主要频段，作为区分 LP 事件和 VT 地震的主要特征。最好的分类结果是使用带有 DT 的 Welch 方法和使用 SVM 分类器的 Multitaper 方法实现的。此外，研究证实有一个高于 40 Hz 的频带，这似乎是检测和分类阶段的关键特征。