Increase Resolution of the Seismic Velocity Model Through Seismic Data

Abstract

Summary Full Waveform Inversion (FWI) is an effective method to estimate high-resolution velocity models of the subsurface. Recently, a super-resolution (SR) method based on deep learning, dubbed M-RUDSR, has boosted the accuracy of FWI results. M-RUDSR is based on multi-task learning (MTL) and contains a global residual skip connection structure (R), an encoder-decoder structure of U-Net (U), and a dense skip connection structure (D). However, due to only the seismic velocity model and its edge images are employed, M-RUDSR does not make full use of high wave-number information contained in seismic data. Hence, we introduce the SR of seismic data and its edge images as extra auxiliary tasks of M-RUDSR. Specifically, the proposed method dubbed M-RUDSRv2 inherits the model structure of M-RUDSR and modifies its input and output to achieve better results. Step by step training method is applied to improve generalization ability and make full use of seismic data to improve the resolution of the seismic velocity model. The experimental results on synthetic examples and field examples demonstrate that the performance of M-RUDSRv2 is superior to that of MRUDSR in SR of the seismic velocity model.