Improving the image quality of elastic reverse-time migration in the dip-angle domain using deep learning

Abstract

Elastic reverse-time migration (ERTM) is becoming increasingly feasible with the development of high-performance computing. It can provide more physical information on subsurface structures. However, the crosstalk artifacts degrade the imaging resolution of ERTM. To obtain high-resolution ERTM imaging, we have developed additional constraints through a convolutional neural network (CNN) in the dip-angle domain. This procedure can significantly improve the image quality of ERTM by recognizing the dominant reflection events and rejecting the crosstalk artifacts in the dip-angle domain. This method can be divided into the following three steps. First, we generate the dip-angle gathers of ERTM using Poynting vectors shot by shot. Then, we stack all the dip-angle gathers over all the shots. Finally, we adopt the CNN to predict the dip-angle constraint, which can suppress the crosstalk artifacts and enhance the ERTM image quality. The picking method using CNN is an end-to-end procedure that can perform automatic picking without additional human intervention once the network is well-trained. The numerical examples have verified the potential of our method.