Seismic Data Compression by Variational Autoencoder with Hyperprior

The evolution of digitization and advancements in seismic data acquisition technology have led to an exponential increase in data volumes within the seismic industry. This surge in data has introduced significant challenges in storage, transmission, and computation. Despite the attempts of previous studies to explore seismic data compression, achieving a balance among the compression ratio, quality of decompressed data, and computational efficiency continues to pose a challenge. This study capitalizes on the recent advancements in deep learning to utilize a variational autoencoder incorporated with a hyperprior. The aim is to simultaneously optimize the reconstruction accuracy and compression ratio in a comprehensive, end-to-end approach. Our experimental analyses demonstrate the efficacy of the introduced compression model on both pre-migration and post-migration seismic datasets.