Exploiting global information and local edge detail for full waveform inversion

Abstract

Data-driven deep learning full waveform direct inversion (DL-FWI) has emerged as an advanced technique for predicting subsurface structures. Popular approaches frequently encounter blurry edge pixels and inaccurate velocity values. Here, we propose an algorithm called TU-Net that captures both global information and local edge detail to address these issues. With respect to the network design, we incorporate a texture warping module (TWM) into the skip connections of the U-Net backbone. Due to the multi-scale feature extraction ability of TWM, our network is able to learn details in complex regions. With respect to the loss function design, we introduce the mixed pixel and edge (MPE) loss, which is a combination of the mean absolute error, the mean square error, and the edge-based losses. The newly proposed loss function balances the model’s focus on global pixel features with the local edge characterization, driving the network to produce high-quality edges. We apply the proposed approach on publicly available OpenFWI, SEG salt and Marmousi II datasets. Quantitative results demonstrate that TU-Net achieves better performance in terms of MSE, MAE, LPIPS, PSNR, UIQ, and SSIM than four state-of-the-art deep networks. The source code is available at github.com/fansmale/TU-Net.