GNGTS 2018 - 37° Convegno Nazionale

GNGTS 2018 S essione 3.3 757 In order to properly test our algorithm, we split each dataset V H into training, validation and evaluation, using 18% of images for training, 7% for validation, and the remaining 75% for evaluation. Since we are working in patch-wise fashion, the process involves more than 7000 training patches, more than 2000 validation patches, and more than 30000 testing patches for each dataset V H . We evaluate the performances in reconstructing the gathers belonging to the evaluation set, using Signal to Noise Ratio (SNR) as accuracy metrics. Precisely, SNR is defined as: where σ 2 is the image variance. Fig. 3(b) shows results obtained for interpolating the image reported in Fig. 3(a). As a matter of fact, the trained autoencoder is able to almost perfectly reconstruct even the most corrupted images, minimally introducing any spatial artifact. Tab. 1 - Average SNR [dB] for U-net and MSSA. SNR [dB] V 10 V 30 V 50 U-net 23 dB 17 dB 11 dB MSSA 16 dB 9 dB 6 dB As state-of-the-art comparison, we tested the strategy known as Multichannel Singular Spectrum Analysis (MSSA) (Oropeza, 2011). This method is based on a rank-reduction algorithm exploiting Truncated Singular Value Decomposition for data reconstruction. We adopted the MSSA open-source MATLAB implementation provided by Chen (2016). Tab. 1 depicts the accuracies obtained over the 3 datasets V 10 , V 30 , V 50 , averaging the results for the images in evaluation set, comparing the proposed method (U-net) and MSSA. It is noticeable that our method improves over the MSSA solution. This is mainly due to the great capability of the U-net in learning features related to the seismic images during training. Moreover, the proposed method has a further advantage, which is the low computational effort in reconstructing a generic image. As a matter of fact, if a risible amount of time is needed for training the network parameters, the evaluation phase is very efficient: barely sec is enough for estimating each inpainted image Iˆ . Conclusions. In this paper, we proposed a method for reconstruction of corrupted seismic data, focusing on interpolation of missing pre-stack data traces in the shot-gather domain. In particular, our approach follows a completely innovative strategy with respect to common techniques exploited in the seismic processing pipeline. We exploited Convolutional Neural Networks for interpolating the missing data traces, which are easy to manage and computationally efficient. Moreover, we compared our results with state-of-the-art, showing significant performances. References Chen, Y. et al.; 2016: An open-source matlab code package for improved rank-reduction 3d seismic data denoising and reconstruction , Computers & Geosciences, 95, 59–66. Duijndam, A. et al.; 1999: Reconstruction of band-limited signals, irregularly sampled along one spatial direction, Geophysics, 64, 524–538, 1999. Gan, S. et al.; 2015: Dealiased seismic data interpolation using seislet transform with low-frequency constraint , IEEE Geoscience and remote sensing letters, 12, 2150–2154. Jin, Y. et al.; 2018: Seismic data denoising by deep-residual networks , in SEG Technical Program Expanded Abstracts 2018, Society of Exploration Geophysicists, pp. 4593–4597. Keys, R. G., and D. J. Foster; 1998: A data set for evaluating and comparing seismic inversion methods, Comparison of seismic inversion methods on a single real data set, 1–12.