[Paper Review] Interpolation and Denoising of Seismic Data using Convolutional Neural Networks

Sara Mandelli, Vincenzo Lipari+2 more|arXiv (Cornell University)|Jan 23, 2019

Seismic Imaging and Inversion Techniques53 references90 citations

TL;DR

The paper uses a U-net convolutional autoencoder to jointly interpolate missing seismic traces and denoise 2D shot gathers, showing competitive accuracy on synthetic and field data compared to traditional methods.

ABSTRACT

Seismic data processing algorithms greatly benefit from regularly sampled and reliable data. Therefore, interpolation and denoising play a fundamental role as one of the starting steps of most seismic processing workflows. We exploit convolutional neural networks for the joint tasks of interpolation and random noise attenuation of 2D common shot gathers. Inspired by the great contributions achieved in image processing and computer vision, we investigate a particular architecture of convolutional neural network referred to as U-net, which implements a convolutional autoencoder able to describe the complex features of clean and regularly sampled data for reconstructing the corrupted ones. In training phase we exploit part of the data for tailoring the network to the specific tasks of interpolation, denoising and joint denoising/interpolation, while during the system deployment we are able to recover the remaining corrupted shot gathers in a computationally efficient procedure. We consider a plurality of data corruptions in our numerical experiments, including different noise models and different distributions of missing traces. Several examples on synthetic and field data illustrate the appealing features of the aforementioned strategy. Comparative examples show improvements with respect to recently proposed solutions for joint denoising and interpolation.

Motivation & Objective

• Motivate the need for high-quality, regularly sampled seismic data for processing workflows.
• Propose a convolutional autoencoder (U-net) to reconstruct corrupted 2D seismic gathers via interpolation and/or denoising.
• Train the network on paired corrupted/clean patches to enable reconstruction of unseen data.
• Evaluate the method on synthetic BP-2004 and field datasets under various noise and missing-trace scenarios.

Proposed method

• Adopt a U-net convolutional autoencoder to learn a compact representation of clean seismic data.
• Operate in a patch-wise fashion on input gathers to handle large data volumes.
• Train with loss functions tailored to the specific task: denoising, interpolation, or joint denoising/interpolation, including a masked loss for interpolation.
• Post-process network outputs with gain-based denormalization and patch-wise reconstruction to form full gathers.
• Evaluate performance using S/N as the primary metric and compare with shaping regularization and f-x deconvolution baselines.
• Investigate patch size, overlap, and bursty missing-trace models to assess robustness.

Experimental results

Research questions

• RQ1 Can a U-net convolutional autoencoder learn to interpolate irregularly sampled seismic data and attenuate noise in 2D shot gathers?
• RQ2 How does patch size and overlapping strategy affect interpolation/denoising performance and computational efficiency?
• RQ3 How does the method perform under different corruption models, including random missing traces, AWGN, and bursty missing traces?
• RQ4 Is transfer learning viable when interpolating data acquired at different sampling rates?
• RQ5 How does the proposed CNN-based approach compare with traditional regularization and deconvolution methods for interpolation/denoising?

Key findings

• The U-net can reconstruct corrupted shot gathers with acceptable accuracy across missing traces, noise, and joint scenarios.
• Patch overlap improves S/N, with non-overlapping patches chosen for memory efficiency; patch size N=128 often yields good results.
• Interpolation alone achieves higher S/N with increasing performance up to a patch size of 128 and remains stable under bursty missing-trace models.
• Under random missing traces of 10–50%, S/N ranges from 32.8 dB to 18.8 dB depending on burstiness, indicating robustness to realistic corruption.
• For denoising, AWGN with initial S/N of -15.8 dB can be improved significantly (examples show gains over traditional methods).
• Transfer learning demonstration shows the model can interpolate data acquired at different sampling rates with reasonable performance.