Convolutional Neural Network-Assisted Least-Squares Migration

Least-squares migration (LSM) is a data-fitting imaging approach seeking the seismic reflectivity image of the most accurate amplitude and optimal resolution. However, the high computational cost of LSM has hindered its broad application. In this study, we combine a convolutional neural network (CNN) with LSM to significantly improve the computational efficiency while retaining the imaging quality. Taking CNN as a "projector," we treat LSM as the "projection" from the ordinarily migrated images to the least-squares updated images. We conduct this CNN-assisted LSM in the shot gather domain using a Gaussian beam migration and the corresponding LSM. The training data for CNN consist of 10-15% of all shot gathers, with the Gaussian beam migrated shot gathers as the input and the LSM shot gathers as the target. After the training, the processing time for the remaining shot gathers took several minutes for 2D cases. The results from testing with the Sigsbee 2B synthetic dataset and a field marine dataset indicate the CNN-assisted LSM saved 80-90% of the computation time of the full LSM and achieved significantly higher image fidelity than that of the ordinary migration.