Multi-scale residual network for seismic data denoising and reconstruction

Abstract

Seismic data processing algorithms greatly require noise-free and regularly complete sampled data. Seismic data denoising and reconstruction has become an important topic for the seismic data processing community. Existing denoising and interpolation algorithms based on Convolutional Neural Network have been treated separately as a single network in many previous applications. In contrast, we introduce a novel end-to-end trainable network for seismic data reconstruction, which focuses on solving both denoising and interpolation for seismic data jointly in this paper. Our algorithm is built upon an end-to-end framework that includes two sub-networks, known as denoising network and interpolation network, and each of them is conducted by an independent Convolutional Neural Network. Meanwhile, a multi-scale residual network, which can learn the complex features of clean and regularly sampled data in different scales and make these features interact with each other, is designed for the task of joint noise attenuation and interpolation more effectively. Numerical experiments on seismic dataset demonstrate the effectiveness of our proposed method in reconstructing clean and complete data. Moreover, we compare the proposed denoising and interpolation method to a recent state-of-the-art method.