GNGTS 2019 - Atti del 38° Convegno Nazionale

566 GNGTS 2019 S essione 3.1 • Joint interpolation and denoising We consider the joint presence of AWGN and uniformly distributed missing traces. We add noise leading into 3 dB , 0 dB , and - 3 dB of S/N and delete a percentage of 10%, 30% or 50% of the traces for each gather. This way, we generate 9 different datasets. Table 2 resumes the average results on evaluation, considering all combinations of missing traces and additive noise variances. Conclusions We proposed a CNN based method for noise attenuation and interpolation of missing data traces in the shot-gather domain. We tested the CNN on random noise cases with Table 2 - Results of joint denoising/interpolation for AWGN and uniformly distributed missing traces. S/N= -3dB S/N= 0dB S/N= 3dB α =10 12.2 dB 13.8 dB 15.6 dB α =30 11.5 dB 12.9 dB 14.4 dB α =50 10.4 dB 11.6 dB 12.9 dB Field data example. As field data example, for the interpolation task, we exploit the well- known Mobil Avo Viking Graeben dataset (1001 marine shot gathers of 128 traces). We build 3 different datasets from the original one, deleting 10%, 30% and 50% of the traces. We split each dataset into 250 gathers for training and validation and leave the remaining to evaluation set. Then we extract 129 overlapped patches with size 128×128 . In term of S/N we obtained 25.7 dB , 20.5 dB and 16.7 dB respectively, for increasing percentages of missing traces. To test the effectiveness of the reconstruction procedure evaluation we performed Kirchhoff migration of the original dataset, of the dataset with 50% randomly missing traces and of the interpolated dataset. The images depicted in Fig. 3 show that, in reconstructed data, noise is attenuated, and loss of continuity is almost perfectly recovered. As a comparison, we migrated also the dataset interpolated trough nonlinear shaping regularization (Chen et al. 2015) which shows some spurious events and more data leakage. Fig. 3 - Migrated section of the Viking dataset. The right images show the migration of the original data (top) and the decimated data with 50% randomly missing traces (bottom). The central images show the migration of the data recovered by U-net (top) and images and shaping regularization (bottom). The images at right shows the corresponding residuals.