<  Back to the Polytechnique Montréal portal

Advanced Seismic Processing Supported by Deep Learning for Subsea Permafrost Characterization

Jefferson Bustamante Restrepo

Ph.D. thesis (2023)

[img] Restricted to: Repository staff only until 24 April 2025
Terms of Use: All rights reserved
Show abstract
Hide abstract

Abstract

«ABSTRACT: A distinctive feature of the Arctic region is the extensive presence of permafrost, defined as subsurface earth material frozen for at least two years. Permafrost degradation owing to rising global temperatures is concerning because it can release carbon stores, increasing atmospheric carbon and contributing to global warming. Therefore, it is crucial to monitor the distribution and degradation of permafrost to predict its impact on global warming. While geophysical techniques effectively assess onshore permafrost conditions, their ability to accurately characterize subsea permafrost zones remains limited. Seismic processing can distinguish between ice-bearing and ice-free subsea sediments by the velocity increase induced by the ice content. However, conventional processing faces challenges due to the lack of recognizable refractions at the base of the permafrost layer and interference between different types of seismic arrivals. In particular, conventional techniques are not precise enough in determining the vertical distribution of permafrost, which hampers our understanding of permafrost distribution on a regional scale. In contrast, deep learning approaches bypass the need for segregating different arrivals by utilizing recorded data to establish a direct mapping between seismic images and subsurface structures. Deep learning can potentially leverage unconventionally processed arrivals to better resolve the subsurface. This thesis introduces a deep learning methodology for the effective large-scale characterization and mapping of seismic velocity and attenuation variations caused by different subsea permafrost conditions. To achieve our objective, we have formulated a three-step approach. Firstly, a sensitivity analysis has been conducted to determine the suitability of employing seismic surveys for detecting changes in permafrost properties. Our findings suggest that a combined approach integrating seismic data after applying multiple domain transformations can improve the inversion approach in discerning changes in permafrost distribution. Secondly, a deep-learning methodology has been devised to invert seismic properties from recorded data collected within permafrost zones. Building upon results from the sensitivity analysis, we designed a multi-input, multi-output neural network that inputs seismic data from distinct domains and outputs inverted seismic velocities and attenuation values. The results demonstrate the capability of the network to accurately image the subsurface velocity and attenuation model in permafrost zones. Lastly, we evaluated the proposed deep learning approach, using two regional-scale seismic surveys conducted in the Canadian Beaufort Sea to identify subsea permafrost in the study area. To effectively map permafrost distribution, we propose adopting predetermined thresholds on inverted parameters. This enables classification into three categories: ice-free, ice-bearing, and ice-bonded permafrost.»

Résumé

«RÉSUMÉ: Une caractéristique distinctive de la région arctique est la présence étendue du pergélisol, défini comme un matériau terrestre sous-surface gelé depuis au moins deux ans. La dégradation du pergélisol due à l’augmentation des températures est préoccupante car elle peut libérer des réserves de carbone, augmentant ainsi la quantité de carbone dans l’atmosphère et contribuant au réchauffement climatique mondial. Il est donc crucial de surveiller la distribution et la dégradation du pergélisol pour prédire son impact sur le réchauffement climatique. Alors que les techniques géophysiques évaluent efficacement les conditions du pergélisol sur terre, leur capacité à caractériser avec précision les zones de pergélisol sous-marin reste limitée. Le traitement sismique permet de faire la distinction entre les sédiments sous-marins contenant de la glace et ceux sans glace en exploitant la vitesse sismique de la glace dans le pergélisol sous-marin. Cependant, le traitement conventionnel est confronté à des défis en raison de l’absence de refractions reconnaissables à la base de la couche de pergélisol et de l’interférence entre différents types d’arrivées sismiques. En particulier, les techniques conventionnelles ne sont pas assez précises pour déterminer la distribution verticale du pergélisol, ce qui entrave notre compréhension de la distribution du pergélisol à l’échelle régionale. En revanche, les approches d’apprentissage profond contournent la nécessité de séparer les différentes arrivées en utilisant les données enregistrées pour établir une corrélation entre les images sismiques et les structures sous-surface. L’apprentissage profond peut potentiellement exploiter des arrivées traitées de manière non conventionnelle pour mieux résoudre le sous-surface.»

Department: Department of Civil, Geological and Mining Engineering
Program: Génie minéral
Academic/Research Directors: Gabriel Fabien-Ouellet and Mathieu J. Duchesne
PolyPublie URL: https://publications.polymtl.ca/56995/
Institution: Polytechnique Montréal
Date Deposited: 24 Apr 2024 08:12
Last Modified: 24 Apr 2024 21:59
Cite in APA 7: Bustamante Restrepo, J. (2023). Advanced Seismic Processing Supported by Deep Learning for Subsea Permafrost Characterization [Ph.D. thesis, Polytechnique Montréal]. PolyPublie. https://publications.polymtl.ca/56995/

Statistics

Total downloads

Downloads per month in the last year

Origin of downloads

Repository Staff Only

View Item View Item