PHYSICS-INFORMED SELF-TRAINING LEARNING FOR SEISMIC IMAGING

Reverse time migration (RTM) may produce low-resolution images because the RTM image is the convolution of subsurface reflectivity and the Hessian matrix. Neural network can approximate the inverse Hessian matrix and then be used to predict subsurface reflectivity from the RTM image. However, seismic surveys cannot provide enough label to train a neural network with good generalization ability. To solve this problem, we propose a physics-informed self-training learning method to approximate the inverse Hessian matrix. We introduce the demigration and adjoint operators as physical constraints to generate dataset for the neural network and train a long short-term memory network in a self-training learning framework. The trained network is then used to deblur the RTM image and obtain the subsurface reflectivity. Imaging tests on the synthetic data of the Marmousi model illustrate that the proposed method can produce reflectivity with higher resolution and less noise compared with the data-driven method in case of small samples.