Accurate and Efficient Salt Dome Segmentation in Seismic Imaging Based on Transformer Learning

Abstract This study explores transformer-based learning for accurate and efficient salt dome segmentation in seismic imaging, emphasizing transfer learning, data efficiency, and domain generalization. We adopt a Vision Transformer (ViT) architecture pretrained via masked autoencoding on over two million unlabeled seismic patches using the Seismic Foundation Model (SFM). Fine-tuning this model on as few as 1000 labeled slices yields high segmentation performance, achieving up to 95.7% accuracy and 0.91 mean Intersection over Union (mIoU) on unseen test data. Models fine-tuned on the full 4000-slice training set further improve generalization, reaching 98.3% accuracy and 0.96 mIoU. These results are validated on a curated out-of-distribution (OOD) benchmark that includes both real and synthetic seismic slices with varied geologic and amplitude characteristics. SFM-pretrained ViTs consistently outperform scratch-trained transformers and conventional convolutional neural network (CNN) baselines, even under limited supervision. Additionally, we show that applying slice-wise Z-score normalization during inference enhances model robustness under distribution shift with minimal computational overhead. Together, these findings highlight a practical and scalable approach to salt dome interpretation that combines foundation model pretraining with minimal labeling and simple adaptation strategies to achieve state-of-the-art performance, even across geologically and acoustically diverse seismic datasets.