5:15 PM - 7:15 PM
[SCG60-P05] Feature Extraction and Classification of Volcanic Earthquake Waveforms Using an Autoencoder
Keywords:Autoencoder, Clustering, Volcanic earthquake, Deep Learning, Machine Learning
In volcanic regions, frequent swarm earthquakes are often associated with volcanic activity. Monitoring these earthquakes is one of the important methods for evaluating current volcanic activity and provides insights into underground physical processes related to volcanic fluids, such as magma and hot water. Volcanic earthquakes can generally be classified into four types based on their characteristics: A-type earthquakes, B-type earthquakes, volcanic tremors, and explosion earthquakes. The frequency of these four types can vary from volcano to volcano, and sometimes further subdivisions or different earthquake types may exist. However, prior to an eruption, the occurrence of A-type earthquakes often marks the beginning, followed by an increase in B-type earthquakes and volcanic tremors, indicating a change in the type of earthquakes occurring before an eruption. Despite this, the exact relationship between earthquake occurrence and eruptions is still not clearly understood. Given this context, accurate classification of earthquake types is considered crucial for clarifying the relationship between volcanic earthquakes and eruption activity.
Currently, volcanic earthquake classification is based on visual observation of surface activity, which incurs significant human labor and time costs. Moreover, classification criteria can vary depending on the observer. Several machine learning models have been proposed to automate the classification of volcanic earthquakes; however, subjective labeling continues to impact the performance and interpretability of these models, as suggested by Suzuki et al. (2025) in their study using Transformer encoders. In this study, we used a Convolutional Neural Network (CNN) autoencoder for unsupervised feature extraction of volcanic earthquakes, and based on the obtained latent space, we applied k-means clustering for earthquake type classification. We used the same dataset as Suzuki et al. (2025), which consists of 38,690 volcanic earthquakes and 231,727 waveforms that occurred at Mount Asama from January 2003 to October 2022. All waveforms in this dataset had phase detection manually performed, with labels assigned as either A-type or B-type (hereafter referred to as the original catalog). In this study, we tested six different models with encoder and decoder layers set to 2, 3, and 4 layers, and latent space dimensions of 8 and 16. Clustering was performed with different numbers of clusters (k=2, 3). The results showed that the model with "2 encoder and decoder layers, k=2, and latent space dimension 8" achieved the minimum reconstruction error and the maximum silhouette coefficient as the clustering evaluation metric. The prototypes determined in this clustering result were a clear S-wave earthquake waveform typical of A-type earthquakes, and a waveform with unclear S-waves typical of B-type earthquakes. Therefore, we classified these clusters as A-type and B-type. When comparing this result with the original catalog, we found that, in the original catalog, all shallow earthquakes were classified as B-type, while in our results, many shallow earthquakes were classified as A-type. However, in areas thought to correspond to volcanic conduits, most were classified as B-type, which is consistent with the hypothesis that these earthquakes are caused by gas movement or magma bursts within the conduit.
In the future, we aim to refine the parameter tuning of the autoencoder to further reduce the reconstruction error, carefully review the clustering results and waveforms, and deepen our understanding of the relationship between volcanic earthquake features and volcanic eruptions.
Currently, volcanic earthquake classification is based on visual observation of surface activity, which incurs significant human labor and time costs. Moreover, classification criteria can vary depending on the observer. Several machine learning models have been proposed to automate the classification of volcanic earthquakes; however, subjective labeling continues to impact the performance and interpretability of these models, as suggested by Suzuki et al. (2025) in their study using Transformer encoders. In this study, we used a Convolutional Neural Network (CNN) autoencoder for unsupervised feature extraction of volcanic earthquakes, and based on the obtained latent space, we applied k-means clustering for earthquake type classification. We used the same dataset as Suzuki et al. (2025), which consists of 38,690 volcanic earthquakes and 231,727 waveforms that occurred at Mount Asama from January 2003 to October 2022. All waveforms in this dataset had phase detection manually performed, with labels assigned as either A-type or B-type (hereafter referred to as the original catalog). In this study, we tested six different models with encoder and decoder layers set to 2, 3, and 4 layers, and latent space dimensions of 8 and 16. Clustering was performed with different numbers of clusters (k=2, 3). The results showed that the model with "2 encoder and decoder layers, k=2, and latent space dimension 8" achieved the minimum reconstruction error and the maximum silhouette coefficient as the clustering evaluation metric. The prototypes determined in this clustering result were a clear S-wave earthquake waveform typical of A-type earthquakes, and a waveform with unclear S-waves typical of B-type earthquakes. Therefore, we classified these clusters as A-type and B-type. When comparing this result with the original catalog, we found that, in the original catalog, all shallow earthquakes were classified as B-type, while in our results, many shallow earthquakes were classified as A-type. However, in areas thought to correspond to volcanic conduits, most were classified as B-type, which is consistent with the hypothesis that these earthquakes are caused by gas movement or magma bursts within the conduit.
In the future, we aim to refine the parameter tuning of the autoencoder to further reduce the reconstruction error, carefully review the clustering results and waveforms, and deepen our understanding of the relationship between volcanic earthquake features and volcanic eruptions.