Norcia 2016 9-Station Seismic Dataset (Foreshock vs Aftershock)

The Norcia 2016 Foreshock–Aftershock 9-Station Dataset consists of seismic events recorded around the Mw 6.5 2016 Norcia earthquake mainshock. The dataset includes both foreshocks and aftershocks associated with the seismic sequence and is designed for foreshock/aftershock classification and seismic cycle analysis. Events were phase-picked using PhaseNet, recorded at nine seismic stations surrounding the Norcia fault zone. Along with waveform data, the dataset includes complete metadata, such as event time, station identifier, class label (foreshock or aftershock)s. This dataset was used in the EGU General Assembly 2026 study “Station-Level and Network-Wide SHAP Explanation of CNN Models for Seismic Cycle Monitoring: Evidence from Norcia 2016”. That work builds on evidence that fault-zone properties evolve during the seismic cycle and that seismic signals recorded before and after a mainshock contain distinguishable signatures. A CNN trained on RGB spectrograms from stations around the Norcia sequence demonstrated strong performance in distinguishing foreshocks from aftershocks. While multi-station models achieved high accuracy (≈97%), SHAP analysis revealed station-dependent variability and a reorganization of feature importance, suggesting that heterogeneous spatial sampling affects interpretability. Station-wise models produced more stable and physically coherent explanations, highlighting the role of propagation effects and motivating region-aware interpretability approaches. The dataset preprocessing pipeline and usage are available in the SeismoXAI repository.2016年诺尔恰前震-余震九台站数据集（Norcia 2016 Foreshock–Aftershock 9-Station Dataset）收录了2016年矩震级6.5级（Mw 6.5）诺尔恰主震前后记录的地震事件。该数据集涵盖该地震序列关联的前震与余震两类事件，专为前震/余震分类及地震周期分析研究打造。 所有地震事件的震相均通过PhaseNet（PhaseNet）拾取，数据来自诺尔恰断裂带周边的9个地震台站。除波形数据外，数据集还包含完整的元数据，例如事件发生时刻、台站标识以及事件类别标签（前震或余震）。 本数据集被用于2026年欧洲地球科学联合会（EGU）大会的研究论文"Station-Level and Network-Wide SHAP Explanation of CNN Models for Seismic Cycle Monitoring: Evidence from Norcia 2016"。该研究依托的核心证据包括：断裂带属性在地震周期内会发生动态演化，主震前后记录的地震信号存在可区分的特征。基于诺尔恰地震序列周边台站的RGB频谱图训练得到的卷积神经网络（CNN，Convolutional Neural Network），在区分前震与余震的任务中表现优异。尽管多台站模型的分类准确率高达约97%，但SHAP（SHapley Additive exPlanations）分析结果显示，不同台站间存在特征重要性的差异与重组现象，表明异质性空间采样会对模型可解释性产生影响。单台站模型则生成了更稳定且物理意义自洽的解释结果，凸显了地震波传播效应的作用，并推动了区域感知型可解释性方法的发展。 该数据集的预处理流程与使用方法可在SeismoXAI仓库中获取。