基于迁移学习的全岩光片有机显微组分识别与定量——以皖泾地1井下三叠统殷坑组烃源岩为例

烃源岩有机显微组分的识别、分类及定量在油气勘探和评价中是重要的研究内容之一。传统的全岩光片有机显微组分鉴定与定量仍采用人工识别与数点计数结合的方式,存在主观性强、人工定量工作强度大、效率低等问题。针对上述问题,本文以皖泾地1井下三叠统殷坑组烃源岩为例,借助于机器学习及图像处理技术,尝试建立了一个基于迁移学习的全岩光片显微组分图像识别与分类模型,并通过OpenCV图像处理库对模型的分类结果图像进行定量统计。结果显示,模型对研究区数据集的整体分类识别准确率可达84.32%,且通过OpenCV图像处理库对各显微组分定量的结果与人工数点法统计定量结果相近,表明该方法可以较客观快速地对大量同类型的烃源岩显微组分图片进行识别和定量,显著提高了全岩光片有机显微组分鉴定及统计效率。

Identification and Quantification of Organic Macerals on Polished Surfaces of Whole Rocks Based on Transfer Learning: A Case Study of the Hydrocarbon Source Rocks from the Early Triassic Yinkeng Formation

Identification, classification, and quantification of organic macerals in hydrocarbon source rocks are among the most important research contents in oil and gas exploration and evaluation. But there are problems such as strong subjectivity, high intensity of manual quantification and low efficiency in the traditional identification and quantification of organic macerals on polished surfaces of whole rocks due to its utilization of manual identification combined with point counting. In allusion to the above problems, the source rock of the Triassic Series Yingkeng Formation in the Well Wanjingdi-1 was studied. At first, an image recognition and classification model of maceral on polished surfaces of whole rocks based on transfer learning was established with the help of machine learning and image processing technology. Then, classification result images of the model were quantitatively counted using the OpenCV image processing library. The results reveal that the classification and recognition accuracy of the model as a whole for the data set in the research area can reach 84.32 %. Also, the quantitative results of each maceral obtained by the OpenCV image processing library are similar to the statistical quantitative results acquired by the artificial number point method. Evidently, the proposed method can objectively and quickly identify and quantify large numbers of maceral images of the same type of source rocks, significantly enhancing the identification and statistical efficiency of organic macerals on polished surfaces of whole rocks.