.png){kind=icon}
Return
An explicit multi-granularity input modeling method for seismic phase picking
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.116339.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Seismic phase picking is a fundamental step in seismic monitoring; however, existing deep-learning-based methods often experience performance degradation under cross-regional distribution shifts and low signal-to-noise ratio (SNR) conditions. A key reason is that such methods fuse multi-scale features implicitly within a single input representation, which limits their robustness to distribution shifts and noise. In contrast, multi-granularity modeling constructs multiple temporal-resolution representations explicitly at the input-representation level and models their interactions. To address this problem, we propose an explicit multi-granularity input modeling method (MGPP) for seismic phase picking. The proposed model performs joint modeling of local transient features and global temporal structure at the input-representation level. Specifically, it integrates channel-adaptive patch embedding, multi-granularity feature extraction, cross-granularity information interaction, and adaptive fusion within a unified collaborative modeling framework. In the cross-dataset setting, where models are trained on Stanford Earthquake Dataset (STEAD) and tested on Pacific Northwest dataset (PNW), the proposed model achieves precision values of 0.992 and 0.986 for primary wave (P-wave) and secondary wave (S-wave) picking, respectively. Additional low-SNR experiments show that our model attains the highest precision across all SNR intervals; in particular, under the 5–10 decibel (dB) condition, it improves P-wave and S-wave precision by 2.36% and 3.16%, respectively, over the second-best model, the multi-scale feature fusion network (MFFN). These results support improved stability and robustness under the cross-dataset and complex-noise conditions evaluated in this study.
Keywords:
Seismic phase picking
Deep learning
Explicit multi-granularity input modeling
Cross-granularity interaction
Cross-dataset robustness
Low signal-to-noise ratio robustness
Journal
IF:
8
Papers:
5.7K
Citations:
3.5W
Organization
No organization information available
Cited Papers
No cited papers available