Exploratory analysis of volcanic seismic signals under limited labeled data using supervised CNN representation learning and unsupervised clustering

Volcanic earthquake monitoring plays a critical role in understanding subsurface processes and supporting hazard assessment in active volcanic regions. However, automated seismic analysis is often constrained by limited labeled datasets and class imbalance, reflecting realistic operational conditions. This study proposes an exploratory framework that combines supervised transfer learning–based convolutional neural network (CNN) representation learning with unsupervised clustering of unlabeled volcanic seismic signals. The labeled dataset comprises 315 events (250 volcanic tremor and 65 volcano-tectonic), which are used to train and evaluate the CNN model, while 72 unlabeled events are used in the downstream clustering stage. Time-frequency representations are used as CNN inputs to learn compact feature embeddings, which are subsequently clustered using K-Means and K-Medoids algorithms. Clustering performance is evaluated using Silhouette Score, within-cluster sum of squares, and between-cluster variance metrics. The highest Silhouette Score (0.4482) is obtained for k = 2, indicating clear separation between dominant signal families, while k = 4 reveals additional internal variability. Comparative analysis suggests that K-Medoids may provide improved robustness and cluster compactness relative to K-Means for higher-resolution partitions in this dataset. Rather than proposing a fully automated classification system, this work presents a reproducible earth science informatics workflow that supports structured exploratory analysis of heterogeneous seismic datasets under data-scarce conditions. The findings highlight the potential of supervised CNN-based representation learning combined with unsupervised clustering to complement expert interpretation and facilitate scalable volcanic seismic analysis.