Evaluating the Generalization Capacity of Volcanic Seismic Classification via Transfer Learning

Abstract

Effective volcanic monitoring depends on the timely detection and classification of seismic signals; however, the automation of these processes faces the challenge of site specificity, requiring extensive labeled databases for each individual volcano. To address this limitation, this study evaluates the generalization capability of deep learning models in seismic environments not included in the training phase. A transfer learning approach via fine-tuning was employed using EfficientNetB0, ResNet50, and DenseNet121 architectures. The training process utilized data from a single station, while generalization capacity was validated using data from two external volcanoes. The results demonstrate that although EfficientNetB0 possesses superior generalization capability, the levels of precision and specificity obtained at external stations remain insufficient for autonomous implementation. These findings suggest that fine-tuning pre-trained models is a promising strategy; however, it is concluded that models trained under a single-station scheme do not achieve the operational reliability necessary for deployment in contexts other than their origin.