Interpolation of regularly sampled prestack seismic data with self-supervised learning

PreviousNext No AccessSEG Technical Program Expanded Abstracts 2019Interpolation of regularly sampled prestack seismic data with self-supervised learningAuthors: Satyakee SenSribharath KainkaryamCen OngArvind SharmaSatyakee SenTGSSearch for more papers by this author, Sribharath KainkaryamTGSSearch for more papers by this author, Cen OngTGSSearch for more papers by this author, and Arvind SharmaTGSSearch for more papers by this authorhttps://doi.org/10.1190/segam2019-3215774.1 SectionsSupplemental MaterialAboutPDF/ePub ToolsAdd to favoritesDownload CitationsTrack CitationsPermissions ShareFacebookTwitterLinked InRedditEmail AbstractWe describe a framework for interpolation of broadband, prestack seismic data using a deep learning pipeline trained in a self-supervised manner. Unlike classification or segmentation tasks, image generation with convolutional neural networks (CNNs) is inherently more complex as the networks needs to learn the high dimensional distributions of broadband prestack gathers to be able to accurately reconstruct the missing traces. We highlight two main challenges specific to prestack seismic data for this task: (1) the choice of the loss function and (2) the lack of suitable training data in the form of image-label pairs to cast the problem as a supervised deep learning task. We show that a naïve implementation of standard loss measures like sample-wise L2 or L1 leads to cycle skipping issues at high frequencies and undesirable smoothness in the mid-frequency reconstruction. To resolve this, we use a perceptual loss function computed using a pretrained variational auto-encoder (VAE) that penalizes differences between the interpolated and the input gathers in the high dimensional feature space of these gathers. To account for the lack of high-quality labelled training data we use a self-supervised learning scheme where a generator network is trained to read in random noise and produce the desired super-resolution output.Presentation Date: Monday, September 16, 2019Session Start Time: 1:50 PMPresentation Start Time: 3:55 PMLocation: 214DPresentation Type: OralKeywords: machine learning, interpolation, algorithm, artificial intelligencePermalink: https://doi.org/10.1190/segam2019-3215774.1FiguresReferencesRelatedDetailsCited byRegeneration-Constrained Self-Supervised Seismic Data InterpolationIEEE Transactions on Geoscience and Remote Sensing, Vol. 61Fast hyperparameter-free spectral approach for 2D seismic data reconstruction30 August 2022 | Exploration Geophysics, Vol. 12Seismic Data Reconstruction via Recurrent Residual Multiscale InferenceIEEE Geoscience and Remote Sensing Letters, Vol. 19Attention and Hybrid Loss Guided Deep Learning for Consecutively Missing Seismic Data ReconstructionIEEE Transactions on Geoscience and Remote Sensing, Vol. 60Self-Supervised Learning for Efficient Antialiasing Seismic Data InterpolationIEEE Transactions on Geoscience and Remote Sensing, Vol. 60Self-supervised learning for anti-aliasing seismic data interpolationPengyu Yuan, Shirui Wang, Wenyi Hu, Prashanth Nadukandi, German Ocampo Botero, Xuqing Wu, Jiefu Chen, and Hien Van Nguyen1 September 2021 SEG Technical Program Expanded Abstracts 2019ISSN (print):1052-3812 ISSN (online):1949-4645Copyright: 2019 Pages: 5407 publication data© 2019 Published in electronic format with permission by the Society of Exploration GeophysicistsPublisher:Society of Exploration Geophysicists HistoryPublished Online: 10 Aug 2019 CITATION INFORMATION Satyakee Sen, Sribharath Kainkaryam, Cen Ong, and Arvind Sharma, (2019), "Interpolation of regularly sampled prestack seismic data with self-supervised learning," SEG Technical Program Expanded Abstracts : 3974-3978. https://doi.org/10.1190/segam2019-3215774.1 Plain-Language Summary Keywordsmachine learninginterpolationalgorithmartificial intelligencePDF DownloadLoading ...