Digital rock physics analysis of core integrity using deep neural networks and computer vision: MR23C-0124

Alzayer, Z.B.; Fellgett, M.W.; Christodoulou, Z.; Williams, C.; Walsh, J.; Kingdon, A. ORCID: https://orcid.org/0000-0003-4979-588X; Hier-Majumder, S. ORCID: https://orcid.org/0000-0002-2629-1729. 2019 Digital rock physics analysis of core integrity using deep neural networks and computer vision: MR23C-0124. [Other] In: AGU Fall Meeting 2019, San Francisco, Ca, USA, 9-13 Dec 2019. San Francisco, USA, American Geophysical Union.

Before downloading, please read NORA policies.

Preview

Text
alzayer_et_al_2019_AGU_COMBINED.pdf
Download (4MB) | Preview

Official URL: https://agu.confex.com/agu/fm19/meetingapp.cgi/Pap...

Abstract/Summary

Core imaging and classification is an important step for generating a digital database for subsurface geology. The British Geological Survey collection contains cores from over 15,000 onshore and 8,000 offshore boreholes. Many cores are photographed at high-resolution creating an archive of over 100,000 core tray images containing between 1 m and 3 m of core per image. A crucial challenge in storage and classification of the digital data associated with these cores is the degree of damage or degradation which determines their suitability for further analysis by the scientific community. The large volume of core image data precludes manual phase segmentation and core quality determination from individuals. In this work, we present a new method using both deep learning and computer vision to automate the process. To test the feasibility of our technique, we use a small subset of 62 core tray images, captured with 3 light spectra (Red, Green, Blue). We use pre-trained neural networks to segment the image, which is followed by traditional computer vision techniques for edge detection. We also automate the process of calculating the number of fragments and area of each fragment present in each individual core image. Finally, we present an index for core integrity based on the output of these measurements. The work-flow demonstrates that deep neural networks and computer vision can be leveraged to quantify and non-intrusively assess geophysical properties at a large scale, using only a subset of the data, with open-source packages. This core quality index will allow users to quickly and consistently assess core condition, and in particularly degradation due to: transport, storage and previous sampling. By automating this process it is possible to quickly assess tens to hundreds of metres of core to identify areas suitable for sampling. It also provides semi quantitative information on how representative individual core samples are of bulk rock properties. This will improve integration between core analysis and other datasets, for example wireline logs.