基于深度学习的露头地层剖面裂缝自动提取  

Automatic fracture extraction from outcrops using deep learning

作　　者：吴思琪 王庆[1] 曾齐红[2] 张友焱[2] 刘远刚[1] 邵燕林[1] 魏薇[1] 邓帆 张昌民[1] WU Siqi;WANG Qing;ZENG Qihong;ZHANG Youyan;LIU Yuangang;SHAO Yanlin;WEI Wei;DENG Fan;ZHANG Changmin(School of Geosciences,Yangtze University,Wuhan 430100,China;CNPC Research Institute of Petroleum Exploration and Development,Beijing 100083,China)

机构地区：[1]长江大学地球科学学院,湖北武汉430100 [2]中国石油勘探开发研究院,北京100083

出　　处：《石油物探》2023年第2期245-257,共13页Geophysical Prospecting For Petroleum

基　　金：国家自然科学基金重点项目(42130813);湖北省教育厅科技项目(B2021040)共同资助。

摘　　要：目前我国油气勘探开发的主要目标在逐步转向碳酸盐岩油气资源和页岩油气资源,而碳酸盐岩储层以缝洞型为主,对野外露头地层剖面的裂缝提取能够为缝洞型储层研究提供直观、形象的结果。最常见的传统露头裂缝研究是耗时耗力的人工解译,为此提出基于深度学习的露头地层剖面裂缝智能化提取技术:在Mask R-CNN算法的基础上结合在线增广策略和注意力机制实现裂缝的自动化提取。为了证明Mask R-CNN改进算法对错误标签的容错能力,设计了含有不同程度错误标签的训练集对比实验,结果表明,当错误标签占比达到30%时模型精度下降较大;接着通过消融实验定量分析不同的改进算法对精度的提高程度;然后以人工提取结果为参照,与大津法(OTSU)、区域生长算法、UNet算法、DeepLabv3+算法等进行精度对比评价,结果表明,改进的Mask R-CNN裂缝识别方法提取的裂缝线性特征更准确完整,且正确率在97%以上,高于其它对比算法,验证了该方法的有效性;最后将改进的Mask R-CNN裂缝识别方法应用于四川盆地西南部峨边先锋地区野外露头地层剖面的裂缝提取,并对裂缝长度、密度、倾向和间距参数进行统计分析,定量分析裂缝分布特征,验证了该方法的可行性。本研究为加快石油勘探综合研究和目标精细评价,储层预测和露头地层剖面参数表征的自动化发展提供了依据。At present,the main targets of oil and gas exploration and development in China are gradually shifting to carbonate oil and gas resources and shale oil and gas resources,and carbonate reservoirs are mainly fracture-vuggy reservoirs.The extraction of fractures from outcrop formation profiles can provide intuitive and vivid results for the study of fracture-vuggy reservoirs.The most common traditional research on outcrop fractures is time-consuming and labor-intensive manual interpretation.Therefore,an intelligent fracture extraction technology based on deep learning is proposed.Based on the Mask R-CNN algorithm,the online augmentation strategy and attention mechanism are combined to realize automatic fracture extraction.In order to prove the error tolerance ability of the improved Mask R-CNN algorithm proposed in this study,comparison experiments were designed for training sets containing different degrees of error labels.The results showed that the accuracy of the model decreased greatly when the proportion of error labels reached 30%.Then,through the quantitative analysis of ablation experiments,different improvements can improve the accuracy of the algorithm.The results show that the accuracy of the model is improved after the introduction of attention mechanism and online enhancement.Then,the accuracy of manual extraction was compared with OTSU method,regional growth algorithm,UNet algorithm and DeepLabv3+algorithm.The results show that the linear features extracted by the proposed method are more accurate and complete,and the accuracy is more than 97%,which is higher than other comparison algorithms,and verifies the effectiveness of the improved Mask R-CNN method.Finally,the method was applied to extract fractures from outcrop profiles in the Ebian Xianfeng area,southwest Sichuan Basin.The parameters of fracture length,density,inclination and spacing were analyzed statistically,and the fracture distribution characteristics were quantitatively analyzed.The feasibility of the improved Mask R-CNN method was veri

关 键 词：露头地层剖面 裂缝自动提取 裂缝参数表征 卷积神经网络 深度学习 

分 类 号：P631[天文地球—地质矿产勘探]