Published July 2022
Published in IEEE UFMG), Belo Horizonte, Brazil, Bruno A. A. Monteiro and Jefersson A. dos Santos, also from the Department of Computer Science and Mathematics, University of Stirling, in the United Kingdom, in addition to Hugo Oliveira, from the Institute of Mathematics and Statistics (IME) , University of São Paulo (USP), Brazil, has brought great advances in deep learning research for seismic image interpretation.
According to the researchers, current deep learning methods for interpreting seismic images require large amounts of labeled data and, due to strategic and economic interests, this data is not available in abundance. “In this scenario, seismic interpretation can benefit from self-supervised learning (SSL), relying on prior training without manually annotated labels within the target data domain and subsequent fine-tuning with few shots,” they described.
To demonstrate the potential of such an approach, the researchers conducted experiments with three classic context-based pretexting tasks: rotation, puzzle, and frame order prediction. “Our results for 1, 5, 10 and 20 shots showed significant improvement in average intersection over union (mIoU) measurements for semantic segmentation in most situations, outperforming the baseline method by 38% in the single scenario for the set of F3 data from the Netherlands, and 16.4% in the Parikka dataset from New Zealand. This gap widens further after ensemble modeling is performed. These experiments suggest that the application of SSL methods can also bring great benefits in seismic interpretation when little labeled data is available,” they concluded.