Compressão de dados sísmicos pós-pilha usando uma rede adversária generativa com função de perda composta

Abstract

Seismic data provide structural and static information of the region where they were acquired, used to determine regions to explore oil and gas. The improvements in the acquisition methods, especially in the high quality of the sensors, have also increased the size of the seismic data. The motivation for compressing them comes from the demand for hundreds of terabytes to transmit and store seismic data. This work presents a method for three-dimensional (3D) poststack seismic data compression integrating a 3D convolutionbased autoencoder to a generative adversarial network (GAN). The main challenge of the 3D convolutional autoencoder is how to exploit volumetric redundancy, keeping the latent representation dimensions. The proposed method is based on a convolutional neural network for seismic data compression called 3DSC. The main hypothesis is that 3DSC architecture can be improved by adversarial training. Thus, a new 3D-based seismic data compression method (3DSC-GAN) is proposed by coupling the 3DSC network to a GAN. The decoder module is used as a generator of poststack seismic data integrated with a discriminator module to better exploit the volumetric redundancy of the 3D data. Also, a new fashion to calculate the distortion for multidimensional data is proposed, such as seismic data. Since the generic functions ignore the 3D data structure and consider it as a 1D vector, the idea is to apply a different loss function for each axis, for a reduction in dimensionality that better captures the error according to its magnitude. For this purpose, an extensive study is performed to analyze the possible combinations of functions for the 3D poststack seismic data compression problem. Results indicate that the 3DSC-GAN method outperforms previous ones for very low target bit rates, increasing the peak signal-to-noise ratio (PSNR) with high visual reconstruction quality. In addition, the experiments using the new distortion function show that it benefits the network learning process, generating a superior reconstruction compared to methods that use PSNR as a distortion function, in quantitative and qualitative terms.