Post-stack seismic data compression with multidimensional deep autoencoders

Abstract

Seismic data are surveys from the Earth's subsurface with the goal of representing the geophysical characteristics from the region where they were obtained in order to be interpreted. These data can occupy hundreds of Gigabytes of storage, motivating their compression. In this work, we approach the problem of three-dimensional post-stack seismic data using models based on deep autoencoders. The deep autoencoder is a neural network that allows representing most of the information of a seismic data with a lower cost in comparison to its original representation. To the best of our knowledge, this is the rst work to deal with seismic compression using deep learning. Four compression methods for post-stack data are proposed: two based on a bi-dimensional compression, named 2D-based Seismic Data Compression(2DSC) and 2D-based Seismic Data Compression using Multi-resolution (2DSC-MR), and two based on three-dimensional compression, named 3D-based Seismic Data Compression (3DSC) and 3D-based Seismic Data Compression using Vector Quantization (3DSC-VQ). The 2DSC is our simplest method for seismic compression, in which the volume is compressed through its bi-dimensional sections. The 2DSC-MR extends the previous method by introducing the data compression in multiple resolutions. The 3DSC extends the 2DSC method by allowing the seismic data compression by using the three-dimensional volume instead of 2D slices. This method considers the similarity between sections to compress a whole volume with the cost of a single section. The 3DSC-VQ uses vector quantization aiming to extract more information from the seismic volumes in the encoding part. Our main goal is to compress the seismic data at low bit rates, attaining a high quality reconstruction. Experiments show that our methods can compress seismic data yielding PSNR values over 40 dB and bit rates below 1.0 bpv.