A Convolutional Autoencoder Method for Simultaneous Seismic Data Reconstruction and Denoising

Abstract

Petroleum geophysical exploration is based on seismic data and has been widely affected by deep learning technology in recent years. As a consequence of the high efficiency and nonlinear fitting ability of deep learning models, we propose an improved convolutional autoencoder (CAE) method to achieve simultaneous reconstruction and denoising of seismic data. The architecture of the improved CAE is based on group convolution and inception structures, which have powerful feature extraction capabilities for seismic data. The CAE method regards the reconstruction and denoising of seismic data as a feature extraction process of the target seismic signals; this enables the method to simultaneously reconstruct the seismic signal accurately and suppress the random noise mixed in the seismic data. During the training of the CAE, the mean absolute error (MAE) loss function and Adam optimization algorithm were used. Because this is a data-driven method and does not require the threshold to be input manually, it can process a large amount of seismic data quickly and intelligently. Synthetic and field data examples demonstrate the effectiveness of the CAE method.