Seismic Facies Segmentation Via Mask-Assisted Transformer

Seismic facies segmentation plays a crucial role in identifying facies types based on the characteristics of seismic reflectors. The application of using convolutional neural networks (CNNs) in seismic facies segmentation is growing rapidly. However, CNN-based models face practical problems such as lacking training labels, relatively low efficiency, and underperformance in capturing seismic features. With the tremendous success in natural language processing and computer vision, the transformer with greater extracting multi-level representation abilities in sequences is now emerging in time-series forecasting. We propose to utilize the time-series segmentation transformer to resolve seismic facies segmentation. Moreover, an unsupervised generation mask workflow is introduced to assist the segmentation transformer function effectively by instructing the assignment of weights purposefully. By using 1% of data volume for a trace-by-trace mask-assisted transformer model training, we can reach 96% pixel accuracy in the blind testing set.