Automatic 3D horizon picking using a volumetric transformer-based segmentation network

Abstract

Seismic horizon picking is a critical step in seismic interpretation and is often labor-intensive and time-consuming, particularly in three-dimensional (3D) volume interpretation. We formulated the task of automatically selecting horizon surfaces from 3D seismic data as a 3D seismic image segmentation problem and developed a method based on a volumetric transformer network. The network uses 3D seismic subvolumes as inputs and outputs the probabilities of several horizon classes. Horizon surfaces can be extracted using postprocessing segmentation probabilities. Because the full annotation of a 3D subvolume is tedious and time-consuming, we utilize a masked loss strategy that allows us to label a few two-dimensional (2D) slices per training subvolume such that the network can learn from partially labeled subvolumes and create dense volumetric segmentation. We also used data augmentation and transfer learning to improve the prediction accuracy with the limited availability of training data. For two public 3D seismic datasets, the proposed method yielded accurate results for 3D horizon picking, and the use of transfer learning improved the accuracy of the results.