Automatic computation of relative geologic time volume using self-supervised learning

Abstract

Although relative geologic time (RGT) volume is an attribute highly utilized in seismic interpretation, accurate automatic prediction of RGT volume is very difficult. In this study, we developed the self-supervised learning-based algorithm, which can generate a RGT volume without labels. To replace the labels, we have proposed the new task using cycle-consistent tracking, which can train the machine learning network using only seismic images. The proposed algorithm has the advantage of generating self-supervision by itself and automatically generating RGT volumes without user's supervision. We have validated the developed algorithm using the Glencoe field data. The estimated results showed that the relatively reliable RGT volumes were predicted even in complex images containing discontinuous structures.