AUTHOR=Wang Zhiwei , You Jiachun , Liu Wei , Wang Xingjian 

TITLE=Transformer assisted dual U-net for seismic fault detection

JOURNAL=Frontiers in Earth Science

VOLUME=Volume 11 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2023.1047626

DOI=10.3389/feart.2023.1047626

ISSN=2296-6463

ABSTRACT=Automatic seismic fault identification for seismic data is necessary for oil and gas resource exploration. The traditional manual method cannot meet the needs of practical application of massive seismic data. With the development of artificial intelligence technology, deep learning methods based on pattern recognition have become a research hotspot for seismic fault identification. Faced with this challenging task, some outstanding U-shaped neural networks (Unet) have been developed, but they still seem to have a gap to meet the strict requirements of fault prediction in a complex structure. Based on advantages of the convolution and transformer blocks, we combine a standard Unet with a transformer Unet, and propose a parallel dual Unet model, called dual Unet with transformer. To achieve a better accuracy of fault prediction, six loss functions (including Binary Cross Entropy loss, Dice coefficient loss, Tversky loss, Local Tversky loss, Multi-scale Structural Similarity and Intersection over Union loss) by using synthetic data are compared, based on three evolution metrics involving Dice coefficient, Sensitivity and Specificity, we find that the binary-cross entropy loss function is the most robust one. An example of comparing the prediction performance by using different Unet models on synthetical data is performed, which proves the outstanding performance of our established dual Unet model, verifying the practical application value. To test the practical feasibility of our proposed method, a real seismic data with rich fault system is utilized, predicted results prove that our proposed method is capability of predicting fault system more accurate than the well-developed U-net models, e.g., the classical Unet, without transfer learning, demonstrating that our proposed model possesses a potential wide application prospect.