DTSEC: A Seismic Event Classifier via Dedicated Dataset Fine-Tuning of a Pretrained Seismogram Foundation Model

Abstract Accurate identification of seismic event types is a key technique for improving earthquake monitoring capabilities and compiling high-quality seismic catalogs. In this study, we train a DiTing Seismic Event Classifier (DTSEC), which has been fine-tuned from the pretrained seismogram foundation model with over 100 million parameters to classify earthquakes, explosions, and collapses. DTSEC achieves an overall accuracy of 92.83% on the test set. Specifically, accuracies for earthquakes, explosions, and collapses are 96.35%, 91.23%, and 91.41%, respectively. Compared to models trained from scratch, DTSEC achieves strong performance using only 20% of the training data, demonstrating superior sample efficiency. To test regional adaptability, we apply DTSEC to a dataset from the Henan Seismic Network, China. Because of unique regional characteristics, the initial classification performance decreases. However, by adding a small number of local samples through joint fine-tuning, the results show that the performance of the model has been successfully improved in this region. This approach adapts the model to local features while preserving the generalization benefits from the large-scale global dataset. These results suggest that fine-tuning with limited local samples allows the model to adapt to regional characteristics effectively. This study presents a practical, deployable solution for seismic event classification in data-scarce regions.