Abstract: Seismic facies prediction plays an important role in oil and gas exploration. Deep learning (DL) can improve the accuracy of seismic facies interpretation and meanwhile greatly reduce the workload of manual interpretation. Mainstream DL methods rely on a limited number of manually interpreted sections as labels, typically along Inlines and Crosslines of 3D seismic data. As a result, the trained neural network only extracts features along these two directions, leading to low-accuracy facies classification in the regions without manual interpretation labels. To improve DL-model generalization across the entire data volume, this paper incorporates a masking technique and proposes a facies prediction method based on Swin-Unet and masking to address data augmentation. A masker is employed at the network output to mask unlabeled seismic data and thus prevent the backpropagation of invalid gradients. This enables the DL algorithm to extract features in any direction, not limited to the Inline and Crossline directions, thereby realizing global spatial feature extraction to improve the overall accuracy of seismic facies prediction. Swin-Unet is a widely recognized network architecture that integrates U-net with Transformer. It combines the self-attention mechanism of Transformer with the capability of U-net for fine-grained information processing to capture long-range dependencies in an image. The tests using F3 and Parihaka datasets demonstrate the advantages of our approach over traditional DL methods.