Vision transformer with affinity similarity model for interactive seismic interpretation

Seismic interpretation, particularly tasks such as salt body segmentation, has significantly advanced through deep learning methods. While traditional convolutional neural network (CNN)-based approaches have shown effectiveness, recent developments in Transformer architectures, notably Vision Transformers (ViTs), provide compelling new alternatives. This study introduces a novel weakly supervised Vision Transformer-based approach integrated with a long-term affinity similarity memory mechanism specifically designed for seismic interpretation. Initial ViT pre-training on synthetic seismic data establishes robust baseline geological object recognition. Subsequently, the affinity-based memory model training significantly enhances spatial continuity, crucial for accurately interpreting continuous geological structures across seismic volumes. Distinct from prior approaches, our method uniquely applies an affinity-based memory propagation technique specifically adapted for seismic inline propagation, substantially improving prediction continuity and considerably reducing manual annotation efforts.