Seismic Fault Identification Based on Vision Transformer

To address the limitations of traditional convolutional neural networks (CNNs) in seismic fault identification—such as restricted local receptive fields, limited capability for modeling long-range structural correlations, and low sensitivity to small or subtle faults—this study proposes a seismic fault identification framework based on a Vision Transformer (ViT) architecture combined with self-supervised pretraining and transfer learning. Self-supervised pretraining is first conducted on large volumes of unlabeled three-dimensional seismic data to learn general representations of geological structures, thereby reducing the dependence on manually labeled samples. The pretrained ViT model is subsequently transferred to the fault identification task and systematically compared with a conventional U-Net architecture. Experiments on a publicly available synthetic seismic dataset show that the ViT-based model achieves improved fault localization accuracy, spatial continuity, and robustness to noise compared to U-Net. Application to real 3D seismic data from an oilfield further demonstrates that the proposed method is capable of detecting a larger number of small-scale faults with enhanced structural continuity, highlighting its applicability in structurally complex settings. The results suggest that Transformer-based global modeling provides an effective alternative for automated seismic fault interpretation.