基于少量剖面标注的三维地质异常体识别方法  

An identification method for a 3D geological anomalous body based on a small number of profile labels

作　　者：朱世龙 孙龙祥 朱剑兵[3] 袁晨辉 康宇 吕文君[4] ZHU ShiLong;SUN LongXiang;ZHU JianBing;YUAN ChenHui;KANG Yu;LÜWenJun(School of Computer Science and Technology,Anhui University,Hefei 230039,China;Institute of Artificial Intelligence,Hefei Comprehensive National Science Center,Hefei 230088,China;Shengli Geophysical Research Institute of SINOPEC,Dongying Shandong 257022,China;Department of Automation,University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]安徽大学计算机科学与技术学院,合肥230039 [2]合肥综合性国家科学中心人工智能研究院,合肥230088 [3]中国石化胜利油田分公司胜利物探研究院,山东东营257022 [4]中国科学技术大学自动化系,合肥230026

出　　处：《地球物理学报》2025年第3期1116-1129,共14页Chinese Journal of Geophysics

基　　金：国家自然科学基金面上项目(62273319);安徽省科学技术厅中央引导地方科技发展专项(202407a12020009);国家自然科学基金重点项目(62033012);中国石化科技攻关项目(PE19008-8);胜利油田重点科技项目(YKJ2201)资助。

摘　　要：地质异常体识别是地震相解释、油气勘探等地质分析任务中不可或缺的一环.传统的地质异常体检测方法需要大量的人工工作,在很大程度上依赖于专家的经验和专业知识.本文提出一种基于少量剖面标注的三维地质异常体识别方法,大大减少了人工干预并且可以达到较高的准确度.本文采用典型的语义分割网络,针对地质异常体的封闭性、连续性及可解释性等物理约束,设计了一种三维(3D)平滑操作.在地质异常体识别任务中引入半监督自训练方法,通过利用少量的有标签样本和大量无标签样本进行训练.在模型自训练的过程中通过3D平滑操作不断对伪标签进行修正,并利用伪标签进行监督学习.实验数据采用胜利油田某区块的真实三维地震数据,由专家对地质异常体进行标注.实验中划分了不同数量的标注样本,在测试数据上的平均F1和平均IoU值分别超过96%和93%,证明了该三维地质异常体识别方法的有效性.Geological anomaly identification is an essential component of geological analysis tasks such as seismic facies interpretation and oil and gas exploration.Traditional methods for detecting geological anomalies require a significant amount of manual work and heavily rely on expert experience and domain knowledge.In this paper,we propose a three-dimensional(3D)geological anomaly identification method based on a small number of annotated profiles,which greatly reduces manual intervention while achieving high accuracy.We employ a typical semantic segmentation network and design a 3D smoothing operation that incorporates physical constraints such as the closedness,continuity,and interpretability of geological anomalies.We introduce a semi-supervised self-training approach in the task of geological anomaly identification,which leverages a small number of labeled samples and a large number of unlabeled samples for training.During the self-training process,the pseudo-labels are continuously refined using the 3D smoothing operation,and the refined pseudo-labels are utilized for supervised learning.The experimental data consists of real 3D seismic data from a specific block in the Shengli Oilfield,with geological anomalies manually annotated by experts.The experiments involve different numbers of labeled samples,and the average F1 and average IoU values on the test data exceed 96%and 93%,respectively,demonstrating the effectiveness of the proposed 3D geological anomaly identification method.

关 键 词：三维地震数据 地质异常体识别 半监督学习 语义分割 自训练 

分 类 号：P631[天文地球—地质矿产勘探]