Mixing Unsupervised and Knowledge-Based Analysis for Heterogeneous Object Delineation in Seismic Data
P. Le Bouteiller, J. Charléty, F. Delprat-Jannaud, D. Granjeon and C. Gorini
Event name: 80th EAGE Conference and Exhibition 2018
Session: Seismic Interpretation I - Automation through AI, Machine Learning, Deep Learning
Publication date: 11 June 2018
Info: Extended abstract, PDF ( 604.05Kb )
Price: € 20
Seismic interpretation is increasingly supported by quantitative, partly-automated methods. For economic purposes, some geological objects (geobodies) are targeted for automated delineation in seismic data. Some of them are depicted by a variety of internal seismic facies. Geobody detection is often based on the supervised retrieval of a defined seismic facies. It usually assumes homogeneous properties inside the targeted object. Unsupervised classification algorithms are better suited for highlighting the variety of facies inside a seismic dataset. However, they are not designed to delineate geobodies in the processed data. We developed a method for heterogeneous object retrieval in 3D seismic data. The dataset is first clustered in an unsupervised way. Probability weights are then assigned to each cluster according to prior interpretation hypotheses. These two steps are parameterized on a small training set. They are then applied to a series of 2D sections throughout the whole dataset. Finally, specific criteria applied to the 3D-stacked processed sections allow to distinguish and select most-probable 3D geobodies. An application of our method to a real 3D seismic dataset for mass-transport deposit detection shows its efficiency. The method also advantageously keeps track of the prior uncertainty.