Hybrid DAE-GAN Model with U-Net Architecture for Seismic Signal Denoising
DOI:
https://doi.org/10.21009/03.1401.FA08Abstract
Seismic data is important for geophysical studies, but it often faces interferences that complicate the analysis of underground structures. This research introduces a new method using deep learning to reduce noise in seismic recordings. It combines a Denoising Autoencoder (DAE) with a Generative Adversarial Network (GAN). In this method, a U-Net model serves as the Generator to create a noise-free signal from the contaminated input. A CNN-based Discriminator distinguishes between the generated and original signals. The Generator's loss function includes Mean Squared Error (MSE) for accuracy and Adversarial Loss for realistic features. The model was trained on the STEAD dataset and its performance evaluated with measures like Signal-to-Noise Ratio (SNR), RMSE, and PRD. Results show that this model improves SNR and produces a clean signal similar to the original both visually and spectrally. This approach could enhance automation and efficiency in preprocessing seismic data.
References
[1] S. M. Mousavi and G. C. Beroza, “A review of deep learning in seismology,” Rev. Geophys., vol. 60, no. 1, p. e2021RG000755, Jan. 2022, doi: 10.1029/2021RG000755.
[2] K. J. Bergen, P. A. Johnson, M. V. de Hoop, and G. C. Beroza, “Machine learning for data-driven discovery in solid Earth geophysics,” Science, vol. 363, no. 6433, p. eaau0323, Mar. 2019, doi: 10.1126/science.aau0323.
[3] O. Ovcharenko, V. Kazei, and T. Alkhalifah, “Deep learning for seismic imaging: Denoising, interpolation, and migration,” Geophys. Prospect., vol. 67, no. 7, pp. 1827–1840, Oct. 2019, doi: 10.1111/1365-2478.12839.
[4] W. Zhu and G. C. Beroza, “PhaseNet: A deep-neural-network-based seismic arrival-time picking method,” Geophys. J. Int., vol. 216, no. 1, pp. 261–273, Jan. 2019, doi: 10.1093/gji/ggy423.
[5] L. Zhang, L. Wen, and Y. Chen, “Seismic Fault Detection Using an Attention U-Net,” IEEE Geosci. Remote Sens. Lett., vol. 18, no. 11, pp. 1928–1932, Nov. 2021, doi: 10.1109/LGRS.2020.3015462.
[6] P. Vincent et al., “Extracting and Composing Features with Denoising Autoencoders,” in Proc. 25th Int. Conf. Mach. Learn., Helsinki, Finland, 2008, pp. 1096–1103.
[7] I. Goodfellow et al., “Generative Adversarial Nets,” in Adv. Neural Inf. Process. Syst., Montreal, QC, Canada, 2014, pp. 2672–2680.
[8] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, K. Wells, and A. F. Frangi, Eds., vol. 9351, Lecture Notes in Computer Science. Cham: Springer, 2015, pp. 234–241, doi: 10.1007/978-3-319-24574-4_28.
[9] O. M. Saad and Y. Chen, “Self-supervised seismic denoising using a Noise2Noise approach,” Geophys. J. Int., vol. 224, no. 1, pp. 319–332, Jan. 2021, doi: 10.1093/gji/ggaa443.
[10] Z. Zhang, Y. Lin, and Y. Li, “Self-Supervised and Physics-Guided Deep Learning for Seismic Random Noise Suppression,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–13, 2022, Art. no. 5904913, doi: 10.1109/TGRS.2022.3175823.
[11] S. M. Mousavi et al., “Stanford Earthquake Dataset (STEAD): A Global Dataset of Seismic Signals for AI,” IEEE Access, vol. 7, pp. 125828–125835, 2019, doi: 10.1109/ACCESS.2019.2938150.
