[S-CG54] New Insights of Fluid-Rock Interactions: From Surface to Deep Subduction Zone

[SCG54-P03] Depth direction Super-resolution Framework for Rock CT Images Based on Deep Learning

*Ryogo Kagawa¹, Atsushi Okamoto², Toshiaki Omori¹ (1.Kobe University, 2.Tohoku University)

Keywords:Super-resolution, Deep learning, Data-driven approach, Rock CT images

In recent years, rock CT images obtained from geological studies have been paid much attention in earth and environment science to understand geomaterials and underground resources in earth. However, rock CT images have a low-resolution in depth direction due to time and environmental constraint while three-dimensional structure is important to understand geological properties. In order to reveal earth geological structure, it is important to establish a super-resolution method for rock CT images in the depth direction.
In this study, we propose a super-resolution method for estimating three-dimensional rock CT images by deep learning. Rock CT images used in this study are obtained from Oman drilling project and consist of serpentinite, hydrolite, magnetite, and Cr-rich spinel. We construct our super-resolution method by applying frame interpolation concepts based on temporal image flows reflecting dynamical properties of image data. In the proposed method, we estimate an unobservable rock CT image by spatially neighboring CT images by depth directional interpolation. We apply optical flows between observable multiple input CT images to generate an intermediate CT image at unobservable depth. We show the effectiveness of the proposed method by comparing intermediate CT image estimated by our proposed method with intermediate one estimated by conventional method qualitatively and quantitatively.