Scholarlink

<p>Lithology identification in complex coal-bearing strata is often affected by smoothed well-log responses, shoulder-bed effects, and petrophysical overlap among transitional lithologies. To address these issues, this study proposes a gradient-guided spatiotemporal network, termed GG-STNet, for lithology identification from conventional well logs. Unlike conventional spatiotemporal architectures that mainly learn local and sequential features from raw log responses, GG-STNet uses first-order log gradients as boundary-sensitive cues in both multi-scale spatial feature enhancement and depth-context updating. The model was developed using 116 wells from the Zhengzhuang Block and independently evaluated using 56 cored wells from an adjacent area. On the in-area test set, GG-STNet achieved an accuracy of 0.907, an F1-score of 0.903, and a Kappa coefficient of 0.882. External blind validation yielded a mean well-level accuracy of 92.8% and a mean macro-F1 score of 0.890, suggesting reasonable adjacent-area performance within the same Carboniferous–Permian coal-measure succession. Case studies and geological consistency metrics indicate that GG-STNet improves boundary-sensitive lithology prediction in thinly interbedded and transitional intervals, although confusion remains between petrophysically similar lithologies such as coal and carbonaceous mudstone, and limestone and argillaceous limestone. These results suggest that gradient-guided spatiotemporal learning can be a useful strategy for lithology interpretation under comparable coal-measure logging conditions.</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

A gradient-guided spatiotemporal network for lithology identification in complex coal-bearing strata

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Shuangcheng Yang,
  • Xiaoying Lin,
  • Zhonghao Zhang,
  • Xiang Yu

摘要

Lithology identification in complex coal-bearing strata is often affected by smoothed well-log responses, shoulder-bed effects, and petrophysical overlap among transitional lithologies. To address these issues, this study proposes a gradient-guided spatiotemporal network, termed GG-STNet, for lithology identification from conventional well logs. Unlike conventional spatiotemporal architectures that mainly learn local and sequential features from raw log responses, GG-STNet uses first-order log gradients as boundary-sensitive cues in both multi-scale spatial feature enhancement and depth-context updating. The model was developed using 116 wells from the Zhengzhuang Block and independently evaluated using 56 cored wells from an adjacent area. On the in-area test set, GG-STNet achieved an accuracy of 0.907, an F1-score of 0.903, and a Kappa coefficient of 0.882. External blind validation yielded a mean well-level accuracy of 92.8% and a mean macro-F1 score of 0.890, suggesting reasonable adjacent-area performance within the same Carboniferous–Permian coal-measure succession. Case studies and geological consistency metrics indicate that GG-STNet improves boundary-sensitive lithology prediction in thinly interbedded and transitional intervals, although confusion remains between petrophysically similar lithologies such as coal and carbonaceous mudstone, and limestone and argillaceous limestone. These results suggest that gradient-guided spatiotemporal learning can be a useful strategy for lithology interpretation under comparable coal-measure logging conditions.