Scholarlink

Core images provide valuable geological information for characterization of subsurface deposits, and classification (i.e. lithology, mineralogy, alteration, etc.) increasingly relies on machine learning methods (i.e. CNNs and transformers) to extract value from core images; however, learning across different geological settings is rarely considered. This work leverages information from core images from multiple deposits to enhance classification with transfer learning and model fine tuning. Independently training base models on independent datasets (i.e. a single deposit or reservoir) restricts their ability to harness valuable insights from other datasets. We propose to use transfer learning with multiple geological datasets to enhance model performance and improve classification accuracy. Base models considered include (1) ResNet50 (2) a vision transformer (3) a novel attention fusion model based on self-attention between feature sets and (4) a novel concatenation model based on combining feature sets. Two transfer learning strategies are explored. Strategies differ based on how information is transferred between the models and include (1) weight matrices and (2) feature matrices. The transferability of knowledge between five geographically and geologically diverse core image datasets is explored, including (1) a South Australian copper deposit (2) a Canadian gold deposit (3) a Canadian oil-sands deposit (4) a US limestone reservoir and (5) an Australian and Maldivian carbonate deposit. The concatenation hybrid model with weight matrices transfer learning achieves the highest F1 score, accuracy increases from 10 to 50% compared to the base model. Intuition is that core images from other deposits/reservoirs would not help with classification, but that intuition is shown to be false and transfer learning is an effective way to improve core photo classification accuracy. This research transcends conventional constraints associated with core image classification using a single deposit through the implementation of transfer learning.

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

Development of a Transfer-Learning Core-Image Based Classification Approach Across Diverse Geological Settings

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Debastuti Ghosh,
  • Bo Zhang,
  • Nidhi Hegde,
  • Jeff Boisvert

摘要

Core images provide valuable geological information for characterization of subsurface deposits, and classification (i.e. lithology, mineralogy, alteration, etc.) increasingly relies on machine learning methods (i.e. CNNs and transformers) to extract value from core images; however, learning across different geological settings is rarely considered. This work leverages information from core images from multiple deposits to enhance classification with transfer learning and model fine tuning. Independently training base models on independent datasets (i.e. a single deposit or reservoir) restricts their ability to harness valuable insights from other datasets. We propose to use transfer learning with multiple geological datasets to enhance model performance and improve classification accuracy. Base models considered include (1) ResNet50 (2) a vision transformer (3) a novel attention fusion model based on self-attention between feature sets and (4) a novel concatenation model based on combining feature sets. Two transfer learning strategies are explored. Strategies differ based on how information is transferred between the models and include (1) weight matrices and (2) feature matrices. The transferability of knowledge between five geographically and geologically diverse core image datasets is explored, including (1) a South Australian copper deposit (2) a Canadian gold deposit (3) a Canadian oil-sands deposit (4) a US limestone reservoir and (5) an Australian and Maldivian carbonate deposit. The concatenation hybrid model with weight matrices transfer learning achieves the highest F1 score, accuracy increases from 10 to 50% compared to the base model. Intuition is that core images from other deposits/reservoirs would not help with classification, but that intuition is shown to be false and transfer learning is an effective way to improve core photo classification accuracy. This research transcends conventional constraints associated with core image classification using a single deposit through the implementation of transfer learning.