地学前缘 ›› 2023, Vol. 30 ›› Issue (3): 208-220.DOI: 10.13745/j.esf.sf.2022.5.45

• 新技术、新方法研究实例 • 上一篇    下一篇

基于深度学习的页岩扫描电镜图像有机质孔隙识别与比较

陈宗铭1(), 唐玄1,2,*(), 梁国栋2, 关子珩2   

  1. 1.中国地质大学(北京) 能源学院, 北京 100083
    2.中国地质大学(北京) 自然资源部页岩气资源战略评价重点实验室, 北京 100083
  • 收稿日期:2022-04-25 修回日期:2022-05-20 出版日期:2023-05-25 发布日期:2023-04-27
  • 通讯作者: *唐 玄(1979—),男,博士,教授,博士生导师,主要从事页岩油气地质研究与CO2地质封存工作。E-mail: Tangxuan@cugb.edu.cn
  • 作者简介:陈宗铭(2000—),男,本科生,主要从事人工智能与页岩油气地质研究。E-mail: 1006192118@cugb.edu.cn
  • 基金资助:
    国家自然科学基金项目(41730421);国家自然科学基金项目(41972132);中央高校基本科研业务费项目(35832020051)

Identification and comparison of organic matter-hosted pores in shale by SEM image analysis—a deep learning-based approach

CHEN Zongming1(), TANG Xuan1,2,*(), LIANG Guodong2, GUAN Ziheng2   

  1. 1. School of Energy Resources, China University of Geosciences (Beijing), Beijing 100083,China
    2. Key Laboratory for Strategic Evaluation of Shale Gas Resources, Ministry of Natural Resources, China University of Geosciences (Beijing), Beijing 100083,China
  • Received:2022-04-25 Revised:2022-05-20 Online:2023-05-25 Published:2023-04-27

摘要:

将深度学习模型引入地质图像分析中,可以大幅提高工作效率,增加研究定量化程度,开拓图像研究新领域。本文以上扬子鄂西地区下寒武统牛蹄塘组页岩的离子抛光扫描电镜图像为例,通过对图片二值化等预处理后,利用Mask-RCNN、FCN和U-Net 3种深度学习模型对页岩中主要矿物、有机质及孔隙等进行识别,比较运行时间与识别结果的准确度,讨论了不同深度学习模型在地质图像识别和处理过程中的适用性和差异性。并优选效果最优的U-Net模型与JMicroVision、Adobe Photoshop等通用图像处理软件识别结果进行孔隙识别对比。结果显示:FCN模型能够基本识别图像中的主要矿物、有机质与孔隙,但对颜色相近的组分和裂缝识别效果较差;Mask-RCNN模型可识别分割性强的主要矿物,但对分辨率较低的孔隙和裂缝识别效果较差;U-Net模型对主要矿物、有机质及孔隙识别效率大大提高,在页岩地质图像识别方面具有优势。相较于通用图像处理软件,U-Net模型识别速度提高了300多倍。基于深度学习U-Net模型识别结果,研究区牛蹄塘组页岩孔隙结构类型可分为矿物内圆状孔、矿物间随机不规则孔、有机棱角状孔和有机密集微孔。基于足够数量电镜图片识别得到的孔隙结构参数对于实际储层分类评价具有参考价值。本实验为基于深度学习的页岩扫描电镜图像识别与分析提供了范例,对提高地质图像研究工作效率和推进油气智能化具有一定的借鉴意义。

关键词: 页岩, 黄铁矿, 裂缝, 有机质孔隙, U-Net, 深度学习, 扫描电镜图像

Abstract:

The introduction of deep learning models can greatly improve the efficiency of geological image analysis and thus increase the level of quantitative research. As an example, the Ar-ion polishing scanning electron microscope (SEM) images of shale samples from the Lower Cambrian Niutitang Formation in western Hubei, Upper Yangtze were analyzed using three deep learning models, Mask-RCNN, FCN and U-Net, to identify the minerals, organic matter and pores (basic tasks) after image pretreatment (binarization, etc.) We compared the running time and identification accuracy between the three models, and discussed the model applicability and model differences in geological image recognition and processing. In addition, we compared the best performance model, U-Net model, with the general image processing softwares (JmicroVision, Adobe Photoshop, etc.) in pore recognition. The FCN model performed well in the basic tasks, but could not distinguish the mineral components and fractures with similar colors; whereas the Mask-RCNN model could identify the main minerals with strong segmentation but not low-resolution pores and fractures. In comparison, the U-Net model greatly improved the efficiency of shale geological image recognition with an 300-fold increase in image recognition speed over the general image processing softwares. Applying the U-Net model, the pore structural types of the Niutitang shale of the study area can be divided into circular intra-granular mineral pores, random irregular inter-granular mineral pores, angular organic matter-hosted pores and dense organic matter-hosted micropores. The pore structural parameters obtained based on SEM image analysis of large enough sample size may be used for reservoir classification and evaluation. The example provided in this study may help improving the efficiency of geological image research as well as promoting artificial intelligence application in oil and gas research.

Key words: shale, pyrite, fracture, organic matter-hosed pore, U-Net, deep leaning, scanning electron microscope image

中图分类号: