Scholarlink

<p>Accurate lithology identification in lacustrine shale oil systems is essential for reservoir evaluation, but it remains challenging in the Chang 7<sub>3</sub> Sub-Member of the Yanchang Formation, Ordos Basin, because of volcanic-ash interference, thin interbedded lithologies, and the limited vertical resolution of conventional well logs. Electrical image logging provides high-resolution information on borehole-wall textures, but blank strips caused by incomplete wellbore coverage may degrade image quality and reduce the reliability of texture-based lithology classification. In this study, we propose an automated lithology identification workflow that integrates generative adversarial network (GAN)-based image inpainting, multi-scale texture feature extraction, and a CBAM-inspired feature attention stacking model. After GAN-based inpainting, texture descriptors were extracted using the gray-level co-occurrence matrix (GLCM), Tamura features, and local binary patterns (LBP). Correlation analysis and variance inflation factor analysis were then applied to remove redundant variables, resulting in six optimized texture descriptors as the final model inputs. To improve lithology discrimination, a lightweight CBAM-inspired feature attention module was introduced to adaptively reweight the six retained texture descriptors before classification. Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP) were used as base learners, while LightGBM served as the meta-learner in the stacking framework. The proposed model achieved an overall accuracy of 0.81 on the independent field test well, outperforming the best individual base learner by about 3 percentage points. The model showed particularly reliable performance for shale and tuff, indicating its ability to distinguish lithologies affected by volcanic ash and overlapping image textures. These results demonstrate that the proposed workflow provides an effective and automated approach for lithology identification in volcanic-ash-influenced lacustrine shale oil reservoirs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning for lithology identification in electrical image logging: A case study of the Chang 73 Sub-Member, Yanchang Formation, East Gansu region, Ordos Basin

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Jiaqi Li,
  • Hongping Bao,
  • Shilong Liu,
  • Pingping Shi,
  • Feiming Gao,
  • Liubin Wei,
  • Langlang Wang,
  • Hongyan Yu

摘要

Accurate lithology identification in lacustrine shale oil systems is essential for reservoir evaluation, but it remains challenging in the Chang 73 Sub-Member of the Yanchang Formation, Ordos Basin, because of volcanic-ash interference, thin interbedded lithologies, and the limited vertical resolution of conventional well logs. Electrical image logging provides high-resolution information on borehole-wall textures, but blank strips caused by incomplete wellbore coverage may degrade image quality and reduce the reliability of texture-based lithology classification. In this study, we propose an automated lithology identification workflow that integrates generative adversarial network (GAN)-based image inpainting, multi-scale texture feature extraction, and a CBAM-inspired feature attention stacking model. After GAN-based inpainting, texture descriptors were extracted using the gray-level co-occurrence matrix (GLCM), Tamura features, and local binary patterns (LBP). Correlation analysis and variance inflation factor analysis were then applied to remove redundant variables, resulting in six optimized texture descriptors as the final model inputs. To improve lithology discrimination, a lightweight CBAM-inspired feature attention module was introduced to adaptively reweight the six retained texture descriptors before classification. Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP) were used as base learners, while LightGBM served as the meta-learner in the stacking framework. The proposed model achieved an overall accuracy of 0.81 on the independent field test well, outperforming the best individual base learner by about 3 percentage points. The model showed particularly reliable performance for shale and tuff, indicating its ability to distinguish lithologies affected by volcanic ash and overlapping image textures. These results demonstrate that the proposed workflow provides an effective and automated approach for lithology identification in volcanic-ash-influenced lacustrine shale oil reservoirs.