运用机器学习方法提取Sentinel-2影像烃微渗漏异常——以Marsel探区为例  

Extraction of hydrocarbon microseepage information from Sentinel-2 remote sensing image using machine learning: A case of Marsel region

作　　者：伍仪霖 周子勇[1] WU YiLin;ZHOU ZiYong(College of Geosciences,China University of Petroleum,State Key Laboratory of Petroleum Resources and Prospecting,Beijing 102249,China;BGP INC.,China National Petroleum Corporation,Zhuozhou Hebei 072700,China)

机构地区：[1]中国石油大学(北京)地球科学学院,油气资源与探测国家重点实验室,北京102249 [2]中国石油集团东方地球物理勘探有限责任公司国际勘探事业部,河北涿州072700

出　　处：《地球物理学报》2024年第4期1330-1341,共12页Chinese Journal of Geophysics

摘　　要：油气藏的烃微渗漏是一种普遍存在的现象,遥感技术为油气藏的烃微渗漏地表检测提供了高效快捷的方法.传统的基于烃微渗漏机理的方法通过检测烃微渗漏引起的地表异常(植被、蚀变矿物等)响应以提取烃微渗漏位置,方法简单,但多解性强.本文以哈萨克斯坦Marsel探区为研究对象,提出了一种基于机器学习的遥感影像烃微渗漏异常提取方法.首先以研究区地表微生物检测结果为基础制作训练样本,为了对比不同样本学习结果,分别制作了斑块样本(patch sample)数据集和像元样本(pixel sample)数据集,在此基础上采用逻辑回归、支持向量机、随机森林、LeNet、AlexNet、GoogLeNet、ResNet算法构建两类数据集的学习模型.结果表明,对于经典机器学习算法,斑块样本最高准确率达0.833,像元样本最高达0.771;对于深度学习算法,斑块样本最高达0.782,像元样本最高达0.914.最后把这准确率最高的四种算法模型应用于哈萨克斯坦Marsel探区,并与地质地震资料进行对比,发现ResNet-18-1D对像元样本的预测结果与地震地质分析资料的对应性最佳,且准确率达0.914,Kappa系数达0.892.Hydrocarbon microseepage is a common phenomenon occurring in oil and gas field. Remote sensing technology provides an efficient and fast method for surface detection of hydrocarbon microseepage. The traditional mechanism-based methods detect the surface anomalies (vegetation, altered minerals, etc.) caused by hydrocarbon microseepage and extract the related information. The method is simple and easy to operate, but the result is some of ambiguity. This paper takes the Marsel area in Kazakhstan as a case, and proposes an machine learning-based method to extract hydrocarbon microseepage from remote sensing images. Firstly, the training samples were made based on the result of surface microorganism detection in the study area. In order to compare the learning results of different samples sets, two training sample datesets, i.e. the patch and pixel dataset were built. Subsequently, logistic regression, support vector machine, random forest, LeNet, AlexNet, GoogLeNet and ResNet algorithms were used to construct learning models for the above two kinds of datasets. The results show that the classical machine learning algorithm has the highest accuracy of 0.833 for patch samples and 0.771 for pixel samples;the deep learning algorithm has the highest accuracy of 0.782 for patch samples and 0.914 for pixel samples. Finally, the four algorithmic models with the highest accuracy were applied to the Marsel area in Kazakhstan and compared with the geological seismic data. It was found that the prediction results of ResNet-18-1D for the pixel samples corresponded best to the seismic geological analysis data with accuracy of 0.914 and Cohen's kappa coefficient of 0.892.

关 键 词：烃微渗漏 图像分类 机器学习 卷积神经网络 

分 类 号：P407[天文地球—大气科学及气象学] P631