DAS-SPP: Transformer-based seismic phase picking on distributed acoustic sensing data

Advanced deep learning techniques are opening up new possibilities for seismic monitoring using distributed acoustic sensing (DAS) technology. In this study, we present a lightweight Transformer-U-Net architecture that combines the fine-grained spatial resolution of U-Net with the global contextual modeling capabilities of Transformers, to accurately pick P- and S-wave arrival times in DAS data. Our model is trained and validated using DAS recordings from the Campi Flegrei caldera in southern Italy, a region that is both geologically complex and seismically active. Even when trained on a limited number of manually labeled events, the proposed approach achieves high phase-picking performance by effectively combining local and global features. The model attains an F1-score of up to 90% for both P and S phases, outperforming state-of-the-art methods. To assess its robustness, we also evaluate the pre-trained model without prior fine-tuning on an independent DAS dataset from Iceland, where it maintains good performance despite different geological settings and acquisition set-up. Finally, we apply the model to several months of continuous DAS recordings, demonstrating its scalability and potential for real-time seismic monitoring.