Deep Seismic Compression
J.P. Navarro, A.P. Schiavon, M. Vieira and P.M. Silva
Event name: 81st EAGE Conference and Exhibition 2019
Session: Deep Learning and Data Analytics - Seismic Applications I & Methods and Applications II
Publication date: 03 June 2019
Info: Extended abstract, PDF ( 1.81Mb )
Price: € 20
In this paper, we present a deep learning approach for lossy compression of 3D post-stack seismic data. The network is feed with two-dimensional 32-bits slices from the original volume, training at the same time two auto-encoders, one for latent space representation, and other for entropy estimation. The bitrate of the compressed volume is controlled by hyper-parameter tuning. The method benefits from the intrinsic characteristic of deep learning methods and automatically captures the most relevant features of the data. It presents impressive results in data reconstruction with high PSNR values at low bitrates. High compression rates (up to 45:1) are obtained, training specialized networks for single datasets. We validate the results using different raw datasets (without any pre-conditioning) from SEG Open Data repository.