Deep-learning for accelerating prestack correlative least-squares reverse time migration

Abstract

Prestack least-squares reverse time migration based on a correlative objective function denoted as PCLSRTM can retrieve a higher quality image of the subsurface than standard least-squares reverse time migration based on a waveform-matching objective function. However, to invert the optimal migration images of different shots via a gradient-based iteration scheme, PCLSRTM tends to require a huge computational overhead for a large number of reverse time migration (RTM) and reverse time de-migration operators. We introduce a convolutional neural network (CNN) framework to improve the computational efficiency in PCLSRTM. Specifically, the CNN architecture aims to reconstruct the optimal reflection image for a single-shot recording from the standard RTM image. Besides, the migration model can be used as an additional input of CNN architecture. Limited by the absence of true reflectivity of the subsurface as the label data in seismic imaging, we employ the standard PCLSRTM images of a small part of total shot recordings as the labels to train the constructed CNN model. There are two benefits for the well-trained CNN model. On the one hand, it can directly invert a comparable image quality with the PCLSRTM image at a low computational cost, when the migration model does not contain a strong lateral velocity variation. On the other hand, one can use the inverted image from the well-trained CNN model as the initial image of the PCLSRTM approach to accelerate convergence. Through synthetic experiments with the Marmousi-2 and SEG/EAGE salt models and marine field data, it can determine that our approach can build a reflection image with balanced amplitude and good continuity and merely requires about one-ninth to one-fourth of the computational costs of PCLSRTM.