Microearthquakes identification based on convolutional neural networks and clustering techniques

Microearthquakes are critical for understanding volcanic activity, leading to monitoring many volcanoes worldwide with seismic sensor networks. These networks generate a substantial amount of data, making visual analysis challenging. Consequently, researchers have focused on developing automatic microearthquake recognition systems over the past decades. A primary challenge with these systems is their reliance on labeled databases for training supervised learning models, where the output labels depend on the database labels. We propose using clustering algorithms in conjunction with a Fine-tuned Convolutional Neural Network (CNN) as a feature extractor to identify overlapping microearthquakes, and other types of microearthquakes withoutneeding labeled datasets. This methodology has two stages: The First stage relies on Transfer Learning, to specialize the CNN in microearthquake recognition. The Second stage uses the Fine-tuned CNN as a feature extractor. This methodology is applied to the Cotopaxi Volcano and validated in the Llaima Volcano. It uses unsupervised databases to find clusters of isolated events with similar characteristics to Long Period (LP), Volcano Tectonic (VT), Tremor (TRE), among others. Additionally, it identifies a cluster with overlapping microearthquakes. In the validation stage, 79 % of the VT events are associated to the same cluster without the need to adjust the Fine-tuned CNN. This test is performed on a dataset of a volcano never seen by CNN or Clustering algorithms. Normalized Entropy is used as a metric to verify the generalization of knowledge, the proposed work is compared with Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP). The proposed work obtains 0.04 lower uncertainty with respect to UMAP.