Deep Learning for Seismic Data Reconstruction: Opportunities and Challenges
KAUST DepartmentEarth Science and Engineering Program
Physical Science and Engineering (PSE) Division
Embargo End Date2021-12-03
Permanent link to this recordhttp://hdl.handle.net/10754/668227
MetadataShow full item record
AbstractNatural and instrumental conditions during field seismic survey lead to noise and irregularities in acquired seismic data. In this work, we explore challenges and opportunities related to denoising and interpolation of seismic data by deep convolutional neural networks. In particular, we apply three network configurations to field data and match them with suitable applications. We show that U-Net is beneficial for denoising applications while adversarial generative networks (GAN) are superior in interpolation tasks. Enhanced interpolation capability of GANs, however, comes at cost of increased uncertainty in the results and we raise awareness about this observation. In the end, we consider the pitfalls of conventional metrics and outline the requirements for data-driven approaches to be suitable in production applications.
CitationOvcharenko, O., & Hou, S. (2020). Deep Learning for Seismic Data Reconstruction: Opportunities and Challenges. First EAGE Digitalization Conference and Exhibition. doi:10.3997/2214-4609.202032054
SponsorsThe authors thank CGG for permission to publish. Special thanks to Henning Hoeber, Alex Clowes, Igor Mikhalev, Ewa Kaszycka, Gordon Poole, Sharon Howe, Jeremie Messud and our colleagues in CGG processing and imaging for the discussions and suggestions.
Conference/Event name1st EAGE Digitalization Conference and Exhibition