Unsupervised Clustering of Seismic Scattering Features for Microseismic Event Classification in the Bossea Cave (NW Italy)

Geophysical and geological explorations are critical for traditional resource extraction, geothermal energy production, carbon capture and storage, hydrogen storage, natural catastrophe forecasting and assessment, and research of the shallow Earth. Geophysical technologies enable scientists to investigate beneath the surface of the Earth without the need for excavation. Microseismicity is one of the most practical geophysical monitoring methods which provides information of various subsurface phenomena in high space resolution in real time. In this work, we first explore applications of these microseismic studies in different fields. These methods can be applied at different scales, from local to global and show widespread potential in climate change evaluation and monitoring, adaptation, and mitigation tasks. Traditional microseismic analysis methods have limitations, and the dramatic increase in seismic data volume makes manual analysis increasingly challenging. A large number of machine learning-based methods have been proposed to address these limitations. This thesis provides a review of machine learning-based detection and classification strategies such as Support Vector Machines (SVM), Random Forest (RF), k-Nearest Neighbors (K-NN), Logistic Regression (LR), Artificial Neural Networks (ANNs), and Convolutional Neural Networks (CNNs) for microseismic monitoring. In the second part of the work, we used passive seismic data from Dabove et al. for clustering microseismic events by utilizing unsupervised machine learning techniques, recorded at the top of the Bossea Cave (NW Italy) (Dabove et al., 2023). We proposed a K-means approach that takes 12 scattering features as an input. Each scattering feature is computed as the maximum absolute value of the signal filtered in 5-Hz frequency bands using Morlet wavelets. The data was recorded continuously for more than seven months (from 2 December 2021 to 10 July). Furthermore, by calculating Silhouette score, 3 clusters was defined as optimal number of clusters for processing and the results were compared with the microseismicity trends obtained through a traditional STA/LTA detection algorithm and meteorological data (temperature and precipitation) to identify patterns. Also for further investigation, the code was run with 5 clusters too. Lastly, possible reasons were discussed for creation of these microseismic events in each clusters. These findings uncover valuable capabilities of unsupervised machine learning tools over traditional methods, opening novel views on characterization of microseismic responses.