EUVIMEDEUVIMED entdecken
Crossref

Research on an automatic seismic first-arrival picking application based on deep learning semantic segmentation

Daicheng Peng, Bo Deng, Guilong Gao, Xiurong Li, Gang Wang, Min Yin, Xiong Ma

Journal of Seismic Exploration · 21.05.2026 · 026120051

DOI: 10.36922/jse026120051 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Seismic Exploration, online erschienen am 21.05.2026. Verfasst von Daicheng Peng und 6 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

First-arrival picking is a fundamental step in seismic data processing, directly influencing the accuracy of subsequent procedures such as static correction, near-surface velocity modeling, and seismic imaging. However, under complex near-surface conditions and low signal-to-noise ratio environments, conventional automatic picking methods often suffer from limited accuracy, poor continuity, and strong dependence on data quality, necessitating extensive manual intervention. To address these challenges, this study proposes a semantic segmentation-based deep learning framework for automatic first-arrival picking, reformulating the task as a pixel-level classification problem on two-dimensional time–trace seismic records, enabling holistic modeling of wavefield characteristics before and after the first-arrival. Within this framework, three representative architectures—U-Net, MobileNet-U-Net, and Attention-U-Net—were constructed and systematically evaluated from the perspectives of standard design, lightweight optimization, and attention enhancement. Synthetic seismic datasets generated via finite-difference forward modeling were used for training, and data augmentation strategies were introduced to improve robustness under noise contamination and missing-trace conditions. The trained models were validated on both synthetic test data and real seismic data from the Loess Plateau. The results demonstrate that all three models effectively performed automatic first-arrival picking, with the Attention-U-Net achieving the best overall performance, exhibiting superior accuracy, continuity, and robustness, particularly under strong noise and severe trace-loss conditions. Compared with conventional methods, the proposed approach produced more reliable and stable picking results and significantly reduced the need for manual correction. These findings indicate that semantic segmentation-based deep learning methods, especially those incorporating attention mechanisms, provide an effective and practical solution for high-precision first-arrival picking in complex seismic environments.

Zitieren

APA

Peng, D., Deng, B., Gao, G., Li, X., Wang, G., Yin, M., & Ma, X. (2026). Research on an automatic seismic first-arrival picking application based on deep learning semantic segmentation. Journal of Seismic Exploration, 0(0), 026120051. https://doi.org/10.36922/jse026120051

Vancouver

Peng D, Deng B, Gao G, Li X, Wang G, Yin M, et al. Research on an automatic seismic first-arrival picking application based on deep learning semantic segmentation. Journal of Seismic Exploration. 2026;0(0):026120051. doi:10.36922/jse026120051

BibTeX
@article{Peng2026,
  author = {Peng, Daicheng and Deng, Bo and Gao, Guilong and Li, Xiurong and Wang, Gang and Yin, Min and Ma, Xiong},
  title = {Research on an automatic seismic first-arrival picking application based on deep learning semantic segmentation},
  journal = {Journal of Seismic Exploration},
  year = {2026},
  volume = {0},
  number = {0},
  pages = {026120051},
  doi = {10.36922/jse026120051}
}
Kostenlos starten →