基于深度学习算法的致密储层薄片图像颗粒、孔隙智能表征方法研究  

Intelligent characterization of particles and pores in thin slice images of tight reservoirs based on deep learning algorithm

作　　者：王金焕 许承武 乔宏亮 唐露 刘天勇 曲端刚 徐坚 孟英杰 李乙鸿 WANG Jin-huan;XU Cheng-wu;QIAO Hong-liang;TANG Lu;LIU Tian-yong;QU Duan-gang;XU Jian;MENG Ying-jie;LI Yi-hong(Institute of Unconventional Oil and Gas,Northeast Petroleum University,Daqing 163000,Heilongjiang Province,China;No.2 Oil Production Plant of Changqing Oilfield Branch,PetroChina,Xi’an 710000,China;School of Chemistry and Chemical Engineering,Yulin University,Yulin 719000,Shaanxi Province,China)

机构地区：[1]东北石油大学非常规油气研究院,黑龙江大庆163000 [2]中国石油长庆油田分公司第二采油厂,陕西西安710000 [3]榆林学院化学与化工学院,陕西榆林719000

出　　处：《地质与资源》2025年第1期61-69,共9页Geology and Resources

基　　金：国家自然基金面上项目“原位加热下页岩储层孔-裂隙动态演化机制研究”(42172163).

摘　　要：在致密砂岩储层薄片图像分析中,针对传统方法的准确率不足和任务繁重等问题,采用结合了Transformer和卷积神经网络的TransUnet及Unet神经网络,用于颗粒、孔隙特征的高效表征.Unet、TransUnet在颗粒特征表征方面表现优异,实验数据显示Unet的交并比达到79.6%,召回率为87.3%,精确率为89.7%,TransUnet的交并比达到71.3%,召回率为86.1%,精确率为82.5%.实验图像显示,在局部像素差异较大的情况,TransUnet优于传统方法,证明其在紧密复杂颗粒分割的有效性.Unet在孔隙特征方面也表现出高效表征效果,其交并比、召回率和精确率分别为82.4%、84.3%和95.3%.实验还表明,虽然面孔率影响交并比,但模型整体仍保持高效率和准确性.这些结果充分说明深度学习方法,在复杂致密储层薄片图像的精确分割中效果显著,为非常规致密储层研究提供新思路,展现了其在地质学领域应用的巨大潜力.In the thin-section image analysis of tight sandstone reservoir,to solve the problems such as low accuracy and heavy work of traditional methods,TransUnet and Unet neural networks by combining Transformer with convolutional neural network(CNN)are used for efficient characterization of particles and pores.The TransUnet has excellent performance in particle characterization.The experiment shows that the intersection over union(IoU)reaches 0.86,with the recall rate of 0.824 and precision of 0.839,which is superior to traditional methods,proving its effectiveness in tight particle segmentation.The Unet shows efficient characterization of pores as well,with the IoU of 0.824,recall rate of 0.843 and precision of 0.953.Besides,experiment indicates that although porosity affects IoU,the model still maintains high efficiency and accuracy generally.These results fully demonstrate that deep learning method,especially TransUnet,is significantly effective in accurate segmentation of thin section images of complex tight reservoir,providing new ideas for the study of unconventional tight reservoir and showing its great potential in the field of geology.

关 键 词：深度学习 薄片分析 致密储层 粒度分析 TransUnet 

分 类 号：P618.13[天文地球—矿床学]