赵峦啸^1,2,,
刘金水³,
姚云霞³,
钟锴^1,2,
麻纪强^1,2,
邹采枫^1,2,
陈远远^1,2,
付晓伟^1,2,
朱晓军^1,2,
朱伟林^1,2,
耿建华^1,2,,
1. 同济大学海洋地质国家重点实验室, 上海 200092
2. 同济大学海洋资源研究中心, 上海 200092
3. 中海石油(中国)有限公司上海分公司, 上海 200000

基金项目: 国家重大科技专项"长江坳陷油气资源潜力评价"（ZX05027001-008）和国家自然科学基金面上项目（41874124）资助


详细信息

作者简介: 赵峦啸, 男, 1986年生, 副教授, 主要从事岩石物理和储层地球物理方面的教学和科研工作.E-mail:zhaoluanxiao@tongji.edu.cn

通讯作者: 耿建华, 教授, 主要从事地震数据处理、储层地球物理和地震岩石物理方面的教学和科研工作.E-mail:jhgeng@tongji.edu.cn

中图分类号: P631

收稿日期:2020-04-02
修回日期:2020-12-15
上线日期:2021-02-10

Quantitative seismic characterization of source rocks in lacustrine depositional setting using the Random Forest method: An example from the Changjiang sag in East China Sea basin

ZHAO LuanXiao^1,2,,
LIU JinShui³,
YAO YunXia³,
ZHONG Kai^1,2,
MA JiQiang^1,2,
ZOU CaiFeng^1,2,
CHEN YuanYuan^1,2,
FU XiaoWei^1,2,
ZHU XiaoJun^1,2,
ZHU WeiLin^1,2,
GENG JianHua^1,2,,
1. State Key Laboratory of Marine Geology, Tongji University, Shanghai 200092, China
2. Marine Resources Research Center of Tongji University, Shanghai 200092, China
3. CNOOC(China) Co., Ltd. Shanghai Branch, Shanghai 200000, China


More Information

Corresponding author: GENG JianHua,E-mail:jhgeng@tongji.edu.cn

MSC: P631

--> Received Date: 02 April 2020
Revised Date: 15 December 2020
Available Online: 10 February 2021


摘要
摘要:烃源岩的定量地震刻画对于勘探开发区块的优选、盆地油气资源量的估算都具有重要意义.陆相沉积环境下的浅湖或半深湖相的烃源岩横向变化快，其空间展布需要依靠钻井约束下的反射地震进行刻画，但是其地震弹性特征与岩性和有机质含量的映射关系呈现高度非线性化，因而很难利用传统基于地震岩石物理模型驱动的烃源岩地震预测方法进行有效刻画.本文以低勘探区的东海盆地长江坳陷为例，提出了一种在数据驱动的机器学习框架下，综合利用地质约束、钻井录井、测井、地球化学和叠前地震数据进行烃源岩的定量地震刻画的工作流程.其核心思想是利用随机森林集成学习算法对小样本数据表现优异的特征，以井位处的测井弹性数据（纵波速度和密度）、岩性、地球化学标定的总有机碳含量（TOC）为样本标签数据，在地质导向约束下通过随机森林算法生成学习网络，并将该网络与叠前地震反演结果相结合，采取先预测泥岩再预测总有机碳含量的"两步走"策略，完成对烃源岩空间分布及其非均质性的定量地震刻画，并对预测结果的不确定性进行评价.测试结果显示，随机森林算法相较于其他的机器学习算法能够更准确的识别陆相沉积地层的泥岩，并比传统的利用阻抗转化方法获得更可靠的总有机碳含量预测结果.
关键词: 陆相/
烃源岩/
地震预测/
机器学习/
随机森林/
岩性预测/
总有机碳含量

Abstract:Quantitative seismic characterization of source rocks is of great significance for well placements, reservoir potential evaluation, and estimation of oil and gas resources. Source rocks in the shallow or semi-deep lacustrine depositional environments are difficult to characterize using the conventional rock physics-driven seismic prediction methods due to the complex and highly non-linear relationship between seismic elastic characteristics and lithology and organic matter content. In this paper, through a case study demonstration in the Changjiang sag in the basin of eastern China sea, a data-driven workflow of quantitative seismic characterization of source rocks based on an integration of geological constraints, well logging, geochemistry and prestack seismic data under the framework of machine learning is proposed. The elastic properties of logging data (P-wave velocity and density), lithology and calibrated organic matter content at the well location are used as the labelled training data set. To reduce the uncertainty of source rock prediction, we suggest a strategy of distinguishing sand/shale first, then predicting the organic content. The trained network generated by the random forest algorithm is combined with the prestack seismic inversion results to map spatial distribution of source rocks under geological constraints. The results show that the random forest algorithm outperforms other machine learning algorithm in terms of accurately identifying the shale formation of the terrestrial sedimentary stratum. The random forest method is also capable of obtaining more reliable organic matter content in comparison with the traditional linear or nonlinear impedance-conversion method.
Key words:Lacustrine depositional environment/
Source rock/
Seismic prediction/
Machine learning/
Random forest/
Lithology prediction/
Organic matter content



PDF全文下载地址:

http://www.geophy.cn/data/article/export-pdf?id=608009fac23e67607ae103ed