27 June 2022 in Geophysics
High-resolution seismic imaging is crucial for deriving an accurate geologic section and ensuring successful petroleum exploration. However, traditional high-resolution imaging methods, such as least-squares reverse time migration (LSRTM) and migration deconvolution, usually involve high computational costs for approximating the inverse Hessian matrix with limited improvements in resolution. To obtain high-resolution migration images and reduce computing costs, we present a deep learning-based point-spread function (PSF) deconvolution method. We decompose the large Hessian matrix into small dispersed PSFs and design a convolutional neural network (CNN) to automatically predict the deconvolution operator of every single PSF. The deconvolution operator eliminates the PSF effect and balances the insufficient illumination of the acquisition geometry. To obtain an efficient CNN model for predicting deconvolution operators, we trained 2500 pairs of PSFs and their corresponding deconvolution operators collected from part of a model using the regular PSF deconvolution method. After training the network, we could predict all deconvolution operators in a few seconds for the following processing of migration images. The results on synthetic and field-data applications indicate that our method could provide a deblurred migration image comparable to that of the LSRTM approaches with significantly reduced computational and memory costs.
29 October 2021 in IEEE Transactions on Geoscience and Remote Sensing
Vertical seismic profiling (VSP) helps to derive high-resolution images around the instrumented borehole and is a cost-effective technique for CO 2 storage monitoring. In routine VSP data processing, P- and S-wave separation is a crucial step to extract independent single-mode waves for accurate imaging and interpretation. Conventional wave mode separation involves tedious, subjective, and non-reproducible manual interventions, especially when dealing with complex geology. To better automate the process, we propose a data-driven deep learning-based P- and S-wave separation method. Our method adapts a fully convolutional neural network that simultaneously extracts P- and S-potential data from multicomponent VSP measurements. To reduce the enormous computational cost in wave simulation while constructing training datasets with sufficient kinematic and dynamic variations, we introduce virtual wellbores where synthetic VSP data sampling wide variations in seismic kinematics and dynamics are recorded using only a dozen elastic wave simulations on a single velocity model. We qualify the separation results both directly in data space and in image space after reverse time migration (RTM). Generalization tests on various synthetic models and their corresponding RTM images demonstrate that the proposed strategy provides sufficient sampling of the high-dimensional data space and essentially ensures successful applications of the trained neural network to similar yet different geological scenarios.