Low Bit Rate 2D Seismic Image Compression With Deep Autoencoders Ana Paula Schiavon 1 , Jo˜ao Paulo Navarro 2 , Marcelo Bernardes Vieira 1 , and Pedro M´ario Cruz e Silva 2⋆ 1 Universidade Federal de Juiz de Fora, Juiz de Fora, Brasil apschiavon@ice.ufjf.br, marcelo.bernardes@ufjf.edu.br 2 NVIDIA, S˜ao Paulo, Brasil {jpnavarro,pcruzesilva}@nvidia.com Abstract. In this paper, we present a deep learning approach for very low bit rate seismic data compression. Our goal is to preserve perceptual and numerical aspects of the seismic signal whilst achieving high com- pression rates. The trade-off between bit rate and distortion is controlled by adjusting the loss function. 2D slices extracted from seismic 3D ampli- tude volumes feed the network for training two simultaneous networks, an autoencoder for latent space representation, and a probabilistic model for entropy estimation. The method benefits from the intrinsic charac- teristic of deep learning methods and automatically captures the most relevant features of seismic data. An approach for training different seis- mic surveys is also presented. To validate the method, we performed experiments in real seismic datasets, showing that the autoencoders can successfully yield compression rates up to 68:1 with an average PSNR around 40 dB. Keywords: Seismic Data Compression · Deep Autoencoders · Geophys- ical Image Processing · High Bit-Depth Compression. 1 Introduction The quality of acquisition sensors has been evolved significantly in the past years. This fact implies on higher resolution signals to process, to storage, and to transmit. The use of effective compressing algorithms plays an important role in seismic processing, aiming to deal with the substantial increase in data resolu- tion. Generally speaking, reliance on compression algorithms in terms of signal reconstruction is a concern in the field due to the dilemma of choosing loss- less methods, with perfect reconstruction, or lossy compression, with a greater reduction on storage with allowed reconstruction distortions. Typical compression methods benefit from the extensive oscillatory nature of the seismic data to model the algorithms. This leads to approaches involving ⋆ Authors thank CAPES, FAPEMIG (grant CEX-APQ-01744-15) for the financial support, and NVIDIA for the donation of one GPU as part of the GPU Grant Program.