Abdulrahman Danlami Isa , Haylay Tsegab Gebretsadik , Abdulrahman Muhammad , Hassan Salisu Mohammed , Ibrahim Muhammad Kurah , Adamu Kamaliddeen Salisu
{"title":"Porosity estimation using deep learning and ImageJ: Implications for reservoir characterization in Central Luconia Miocene carbonates","authors":"Abdulrahman Danlami Isa , Haylay Tsegab Gebretsadik , Abdulrahman Muhammad , Hassan Salisu Mohammed , Ibrahim Muhammad Kurah , Adamu Kamaliddeen Salisu","doi":"10.1016/j.marpetgeo.2025.107538","DOIUrl":null,"url":null,"abstract":"<div><div>Accurate porosity prediction is essential for effective reservoir characterisation, particularly in heterogeneous carbonate systems. This study investigates applying deep learning techniques to predict porosity from petrographic thin-section images of Miocene carbonate reservoirs in Central Luconia, offshore Malaysia. Four semantic segmentation models—UNet, SegNet, PSPNet, and Fully Convolutional Network (FCN)—were trained and evaluated, with UNet achieving the highest performance across all metrics: accuracy (0.990), precision (0.849), recall (0.940), F1 score (0.892), and Intersection over Union (IoU) (0.805). These results were benchmarked against traditional image analysis using ImageJ, where UNet predictions showed strong alignment, highlighting its reliability. SegNet also performed robustly, while PSPNet and FCN demonstrated lower predictive accuracy. The study further explored the impact of data augmentation, revealing performance degradation in SegNet and PSPNet due to distortion of micropore structures. A classification scheme based on UNet output identified porosity ranges, with 71.90 % of samples exhibiting <5 % porosity, reflecting a low-porosity-dominated system. A spatial porosity distribution map was also generated using UNet to visualise heterogeneity across thin-section samples. It is important to note that this map is a conceptual representation and does not reflect the actual porosity distribution of the broader reservoir. Instead, it serves as a hypothetical visualisation to enhance understanding of porosity characteristics and heterogeneity within the analysed thin-section images. This work demonstrates the advantages of deep learning over conventional techniques for pore-scale analysis. It offers a scalable framework for integrating artificial intelligence in quantitative reservoir quality assessment.</div></div>","PeriodicalId":18189,"journal":{"name":"Marine and Petroleum Geology","volume":"182 ","pages":"Article 107538"},"PeriodicalIF":3.6000,"publicationDate":"2025-07-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Marine and Petroleum Geology","FirstCategoryId":"89","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0264817225002557","RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"GEOSCIENCES, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
Accurate porosity prediction is essential for effective reservoir characterisation, particularly in heterogeneous carbonate systems. This study investigates applying deep learning techniques to predict porosity from petrographic thin-section images of Miocene carbonate reservoirs in Central Luconia, offshore Malaysia. Four semantic segmentation models—UNet, SegNet, PSPNet, and Fully Convolutional Network (FCN)—were trained and evaluated, with UNet achieving the highest performance across all metrics: accuracy (0.990), precision (0.849), recall (0.940), F1 score (0.892), and Intersection over Union (IoU) (0.805). These results were benchmarked against traditional image analysis using ImageJ, where UNet predictions showed strong alignment, highlighting its reliability. SegNet also performed robustly, while PSPNet and FCN demonstrated lower predictive accuracy. The study further explored the impact of data augmentation, revealing performance degradation in SegNet and PSPNet due to distortion of micropore structures. A classification scheme based on UNet output identified porosity ranges, with 71.90 % of samples exhibiting <5 % porosity, reflecting a low-porosity-dominated system. A spatial porosity distribution map was also generated using UNet to visualise heterogeneity across thin-section samples. It is important to note that this map is a conceptual representation and does not reflect the actual porosity distribution of the broader reservoir. Instead, it serves as a hypothetical visualisation to enhance understanding of porosity characteristics and heterogeneity within the analysed thin-section images. This work demonstrates the advantages of deep learning over conventional techniques for pore-scale analysis. It offers a scalable framework for integrating artificial intelligence in quantitative reservoir quality assessment.
期刊介绍:
Marine and Petroleum Geology is the pre-eminent international forum for the exchange of multidisciplinary concepts, interpretations and techniques for all concerned with marine and petroleum geology in industry, government and academia. Rapid bimonthly publication allows early communications of papers or short communications to the geoscience community.
Marine and Petroleum Geology is essential reading for geologists, geophysicists and explorationists in industry, government and academia working in the following areas: marine geology; basin analysis and evaluation; organic geochemistry; reserve/resource estimation; seismic stratigraphy; thermal models of basic evolution; sedimentary geology; continental margins; geophysical interpretation; structural geology/tectonics; formation evaluation techniques; well logging.