Machine Learning for Seismic Signal Analysis

Earthquake recognition and location are fundamental problems for seismological studies such as tomography. These analyses rely on identifying the onsets of the P and S waves originating from the events. The earthquake waveforms are characterized by changes in frequency and amplitude that help to discriminate the seismic events from the noise or anthropogenic signals. However, their shapes are influenced by many factors, such as the geology along the propagation path, reflections, and attenuations. An expert analyst can recognize P and S arrival times despite the waveform variability. However, until now, it has been hard to develop an automatic algorithm that can accurately and efficiently identify them. Generally, all the classical procedures available present several limitations, and still, the manual revision of the results is often necessary to guarantee the quality of the picks. Particularly challenging scenarios are the aftershock sequences and areas characterized by microseismicity. In the first case, the seismic rate can be so high that the waveforms of multiple earthquakes may be superposed, making it hard to identify the P onset of a subsequent event from the S coda of the previous one. Mi- croseismicity scenarios (ML < 1.5) are often characterized by low signal-to-noise ratio values. Hence, the earthquake signal and the P and S onset times are hardly distin- guishable from the background noise. In recent years, the number of sensors collecting data increased dramatically due to technological developments in data acquisition, transmission, and storage. Such data availability represents a precious opportunity to improve our understanding of the seis- mic release and the geometry of seismogenic sources. However, since the amount of data is too high, it is straightforward that the manual revision would require too many human resources. It is then necessary to develop new automatic algorithms. Recently, Machine Learning (ML) and Deep Learning (DL) methods have been suc- cessfully spreading in several fields and nowadays represent the forefront of research. They started to be also applied in seismology, notably to perform seismic wave onset picking. These algorithms can automatically identify and learn the characteristics of the data once they are exposed to a great number of examples (data-driven). This approach mimics the simpler learning mechanism of the human brain, making these algorithms very flexible so that they can be applied in several contests. These models also do not require the user to specify any parameter and, in the case of picking, allow to avoid the main problem of defining the characteristics to discriminate P and S onset times. As a LIST OF FIGURES 24 result, they may overcome the limitations that, until now, mainly hampered the devel- opment of picker algorithms. This thesis focuses on the application and development of DL pickers. At first, it provides a general introduction to DL approaches, particularly to convolutional neural networks, which are notably suitable for handling time series as seismograms. The state of the art of DL pickers already gathers several algorithms that successfully apply to the seismogram analysis and whose performance varies based on the application data (Chapter 1). These differences are due to the architecture but also to the training dataset. For example, suppose a DL picker has been trained mainly on waveforms of earthquakes with a specific epicentral distance. In that case, it will be more sensitive in recognizing similar signals than others generated by events that occurred closer or farther away from the recording station. These influences remark the dependencies of the algorithm abilities on the training dataset. They also highlight the necessity of creating high-quality benchmark repositories that collect carefully selected data representing a balanced, wide range of parameters that characterize earthquake waveforms (different magnitude, epicentral distance, signal-to-noise ratio). The public availability of these datasets would allow developers to train their architectures on the same data sets, allowing the following fair evaluation and comparison of the performance with other models. This thesis describes the steps to create two of these benchmark repositories: INSTANCE and Amiata DS (Chapter 2). INSTANCE collects waveforms of earthquakes that occurred in the Italian Peninsula and can be considered as repre- sentative of the Italian Sesimicity between 2005 and 2020. It also provides several noise traces. Although most benchmark datasets already available collect global data, IN- STANCE gathers waveforms recorded in a more limited geographical area. On one side, this characteristic may be seen as a limitation that hampers the generalization ability of a trained model. On the other hand, it allows focusing on different targets, such as the italian seismicity. Moreover, an algorithm trained on INSTANCE can be eventually applied to other contests after enhancing its abilities using Transfer Learning. Amiata DS collects data recorded in Mount Amiata, which is characterized by microseismicity. It has been mainly designed to enhance the abilities of an already trained model in iden- tifying small-magnitude events using the Transfer Learning (TL) technique. Besides the waveforms, the two datasets provide a rich set of metadata that can be used to target select the traces, as labels in supervised approaches, as training data, or for statistical purposes. As mentioned, the evaluation of the abilities of a DL picker can be complicated since results strongly depend on the application data. The fair comparison of the performance of several architectures is even more complex since they are usually trained on different datasets. In Chapter 3, two DL pickers (Generalized Phase Detection and Earthquake Transformer) are tested on two challenging scenarios: microseismicity data recorded in Mount Amiata and the aftershock sequence associated with the 4.5 Mugello mainshock of 2019. The first case analyses the data recorded by a local network deployed in Mount Amiata, which has hosted active geothermal power plants since 1960’. The data analysis investigates a small sequence of 56 events that occurred in the last 15 days of December LIST OF FIGURES 25 2018, with magnitudes ranging within −0.10 ≤ ML ≤ 1.06. Performances have been evaluated by comparing the results with an available reference catalog. Generally, the two algorithms show poor results, highlighting the difficulties in handling the data. The second application is to seismograms related to the Mw 4.5 Mugello aftershock sequence of 2019. The challenges of this scenario are due to the tight time-distance between events and their small magnitude. This seismic rate may determine the P onset of a new event to be superposed to the S coda of the previous one, making it hard to iden- tify. These peculiarities make Mugello an interesting test application for the two DL pickers. The data investigated are relative to December 2019. In this case, a different workflow has been applied than for the Amiata, allowing better discrimination between correct and false predictions. Results show the success of applying the algorithms in this scenario, improving the knowledge of the seismicity of the area and providing more complete automatic catalogs. The studies conducted so far highlight the difficulties of DL pickers in handling mi- croseismicity data. This issue has been tackled in Chapter 4 by applying the Transfer Learning (TL) technique. TL allows the enhancement of the abilities of an existing model by training it using a set of data with specific characteristics. In this case, this technique has been applied to the DL picker BasicPhaseAE. The existing model of Ba- sicPhaseAE, already trained on the ETHZ dataset (BPh ETHZ), has been enhanced through the training with seismograms collected in the area of Ritthershoffen, which hosts an enhanced geothermal field. Tests of the new network on Rittershoffen data show improved results compared to the original model. Performance on data extracted from Mount Amiata is poorer, probably due to the greater epicentral distances of the Amiata events compared to the Rittershoffen ones.