Towards a Benchmark for Sedimentary Facies Classification: Applied to the Netherlands F3 Block

• Published in: Information Management and Big Data Vol. 1410; pp. 211 - 222
• Main Authors: Campos Trinidad, Maykol J., Arauco Canchumuni, Smith W., Cavalcanti Pacheco, Marco Aurelio
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Reservoir characterization; Seismic data; Seismic interpretation; Semantic segmentation
• ISBN: 9783030762278; 3030762270
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-030-76228-5_15

In this paper, we attempt to provide a new benchmark for image seismic interpretation tasks in a public seismic dataset (Netherlands F3 Block). For this, techniques such as data augmentation together with five different deep network architectures were used, as well as the application of focal loss function. Our experiments achieved an improvement in all evaluation metrics cited at the current benchmark. For instance, we managed to improve in 3.7%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.7\%$$\end{document} the pixel accuracy metric and 5.4%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.4\%$$\end{document} on mean class accuracy for a modified U-Net that uses dilated convolution layers in its bottleneck. In addition to this, the confusion matrices of each model are shown for a better inspection in the classes (sedimentary facies) where the greatest amount of misclassification occurred. The training process of almost all networks took less than one hour to converge. Finally, we applied Conditional Random Fields (CRF) as post-processing in order to obtained smother results. The inferences performed with the best topology, in an inline or section of the test set, is closer to achieving an interpretation at a human level.