A High-Resolution Velocity Inversion Method Based on Attention Convolutional Neural Network

Velocity model building is an indispensable part of seismic exploration, which can directly affect the accuracy of subsequent data processing. Traditional full-waveform inversion (FWI) is usually challenging to update the deep background velocity information. Moreover, deep learning (DL)-based velocity modeling efforts can face the problem of lacking generalization ability. Based on this, we propose an attention convolutional-neural-network-based velocity inversion (ACNN-VI) algorithm to update the deep layer background velocity and the reflection interface iteratively. First, we proposed a constantly iterative structure, which allows the initial model to keep iteratively close to the true model. Second, we proposed a convolutional neural network (CNN) based on an attention mechanism that can recover faults and layers efficiently. Furthermore, we propose a smooth strategy that enables the method in this article to be adapted to the case of an inferior initial model. Finally, our numerical tests prove that our method has good inversion results for different models.