Tracing the Central Italy 2016-2017 seismic sequence fault system - Insights from unsupervised Machine Learning and Principal Component Analysis

Aug 2, 2026

Upload by Anda

The study analyzes a dense deep-learning seismic catalogue from the Central Italy 2016-2017 seismic sequence to automatically identify and trace active, potentially hazardous faults and to study their spatiotemporal distribution and evolution. Unsupervised clustering methods such as HDBSCAN, DBSCAN, SOM, and OPTICS are combined with Principal Component Analysis to characterize clustered seismicity and derive a Principal Plane analogous to fault planes. Preliminary results demonstrate clustered, varying-density fault-related structures embedded in diffuse background seismicity and show how cluster-derived strike and dip can be compared with available focal mechanisms, using seismic-catalogue information alone.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Introduction

Data

Methods

Clustering workflow

Results

Tracing the Central Italy 2016-2017 seismic sequence fault system: Insights from unsupervised Machine Learning and Principal Component Analysis Ita hisa González Álvarez1*, Margarita Segou2, Brian Baptie 2 1 British Geological Survey, Keyworth 2 British Geological Survey, Edinburgh INTRODUCTION In this work, we investigate a rich deep learning seismic catalogue from the Central Italy 2016-2017 seismic sequence (Tan et al., 2021) with the aim of identifying and tracing active and potentially hazardous faults, as well as studying their distribution and evolution over the duration of the sequence. To do this, we tested a variety of unsupervised ML algorithms such as HDBSCAN, DBSCAN, SOM or OPTICS, which we used to design a completely automatic algorithm to identify clustered seismicity. We then combined it with Principal Component Analysis to analyse resulting clusters and relate them to active faults. Here, we present some of preliminary results from our preferred approach, which highlight the complexity of the fault system, as well as the potential of this method to successfully trace active faults using exclusively seismic catalogue information. DATA Tan et al. (2021) analysed 1 year of continuous data from the Central Italy 2016-2017 seismic sequence. They used the deep-neural-network phase picker PhaseNet to analyse waveforms from 139 seismic stations and build an enhanced seismic catalogue with over 900 000 earthquakes with moment magnitudes ranging from 0.5 to 6.2 (of which 72 000 contain focal mechanism information) and a magnitude of completeness of 0.5. Figure 1. Map view of the area of the 2016- 2017 Central Italy seismic sequence. Dark blue dots mark the location of the earthquakes included in the catalogue used in this study, dates ranging from 15 August 2016 to 15 August 2017. Light blue dots show the location of the three largest earthquakes in the sequence. Figure 2. Comparison of the non-cumulative frequencymagnitude distribution of earthquakes in the catalogue used in this study and the one elaborated by the INGV. Modiﬁed from Tan et al. (2021). METHODS Our work is based on the assumption that seismicity is clustered at, or near, faults. This led us towards densitybased clustering methods such as HDBSCAN, DBSCAN, or OPTICS, since earthquakes would tend to cluster tightly around fault planes. Therefore, by applying these methods to our enhanced catalogue, and extracting clusters that represent areas of high density of earthquakes, we can try to relate them to individual active faults. As the diagram below illustrates, we do this by combining HDBSCAN (McInnes et al., 2017) with our automatic parameter selection algorithm and Principal Component Analysis (PCA). The PCA of individual clusters allows us to deﬁne their Principal Plane (PP), which is the surface which explains the most variance of our data (depth and geographical coordinates of the earthquakes in the cluster, in our case) and would be analogous to the fault plane outlined by each cluster. From the PP, we can also obtain an equivalent to the fault’s strike and dip that we can compare with the focal mechanisms in our catalogue. • Minimum cluster size (mcs): smallest group of earthquakes that are considered a individual cluster. • Minimum number of samples (msam): number of earthquakes in the neighbourhood for a data point to be considered a core point. • Relative Validity (RV): relative score that allows comparing clustering results from different combinations of hyperparameters. It is an approximation to the Density Based Cluster Validity (DBCV) score. Input catalogue Temporal division HDBSCAN Parameter selection number of clusters Solution not unique? Maximum mcs Solution not unique? Maximum msam HDBSCAN clustering formcs = 5, 10, ..., 100 msam = 5, 10, ..., 100 Calculate Relative Validity (RV) score Extract clusterings with top 2% RV scores Look for clusterings with Minimum Unique

FAQs

What is the goal of tracing the Central Italy 2016-2017 seismic sequence fault system?

The work aims to automatically identify and trace active and potentially hazardous faults, and to examine how fault-related seismicity is distributed and evolves during the sequence.

Which unsupervised machine learning methods are used to identify clustered seismicity?

The approach tests density-based and clustering algorithms including HDBSCAN, DBSCAN, SOM, and OPTICS, then integrates them into an automatic procedure for clustered seismicity detection.

How does Principal Component Analysis help relate clusters to faults?

For each cluster, PCA defines a Principal Plane that captures the dominant variance in depth and geographic coordinates, providing equivalent strike and dip values that can be compared with focal mechanisms from the catalogue.

Aug 2, 2026

Upload by Anda