Horizon-based subsurface stratigraphic model building is a tedious process, especially in geologically complex areas where seismic data are contaminated with noise and thus are of weak and discontinuous reflectors. Seismic interpreters usually use stratal (proportional) slices to approximately inspect 3D seismic data along seismic reflectors yet to be interpreted. We introduce an artificial intelligence workflow consisting of three deep learning steps to provide a conditioned seismic image that is easier to interpret, a stratigraphic model that outlines major formations, and moreover a relative geologic time volume that allows us to automatically extract an infinite number of horizons along any seismic reflectors within a seismic cube. Depending on the availability of interpreters, the proposed workflow can either run fully unsupervised without human inputs or using sparse horizon interpretation as constraints to further improve the quality of extracted horizons, providing flexibility in both efficiency and quality. Starting from only seismic images and a few key horizons interpreted on very sparse seismic lines, we demonstrate the workflow to generate a stack of complete horizons covering the entire seismic volume from offshore Australia.
Read full abstract