Tensor discriminant dictionary classification method for prestack seismic reflection patterns  

张量判别字典叠前地震反射模式分类方法

作　　者：Cai Han-Peng Jing Peng Yang Jun-Hui 蔡涵鹏;敬鹏;杨军辉(电子科技大学资源与环境学院,成都611731;电子科技大学信息地学中心,成都611731;四川数字交通科技股份有限公司,成都610000)

机构地区：[1]School of Resources and Environment,University of Electronic Science and Technology of China,Chengdu 611731,China [2]Information Geoscience Center,University of Electronic Science and Technology of China,Chengdu 611731,China [3]Sichuan Digital Transportation Technology Co.,Ltd.,Chengdu 610000,China

出　　处：《Applied Geophysics》2022年第2期197-208,307,共13页应用地球物理（英文版）

基　　金：supported by the National Natural Science Foundation of China(Nos.42130812,42174151,and 41874155).

摘　　要：The existing seismic reflection pattern classification methods need to convert multidimensional prestack seismic data into one-dimensional vectors for processing,which loses the characteristics of amplitude variation with offset/azimuth in the prestack seismic data.In this study,a tensor discriminant dictionary learning method for classifying prestack seismic reflection patterns is proposed.The method is initially based on the tensor Tucker decomposition algorithm and uses a tensor form to characterize the prestack seismic data with multidimensional features.The tensor discriminant dictionary is then used to reduce the influence of noise on the sample features.Finally,the method uses the Pearson correlation coefficient to measure the correlation degree of the sparse representation coefficients of different types of tensors.The advantages of the new method are as follows.(1)It can retain the rich structural features in different dimensions in the prestack data.(2)It adjusts the threshold of the Pearson correlation coefficient to optimize the classification effect.(3)It fully uses drilling information and expert knowledge and performs calibration training of the sample labels.The numerical-model tests confirm that the new method is more accurate and robust than the traditional support vector machine and K-nearest neighbor classification algorithms.The application of actual data further confirms that the classification results of the new method agree with the geological patterns and are more suitable for the analysis and interpretation of sedimentary facies.现有地震反射模式分类方法需将多维叠前地震数据转化为一维向量进行处理,丢失了叠前地震数据中振幅随偏移距/方位角变化的特征。本文提出了一种张量判别字典学习(TDDL)叠前地震反射模式分类方法。该方法首先基于张量Tucker分解算法,采用张量形式对具有多维特征的叠前地震数据进行表征,然后采用张量判别字典学习减小噪声对样本特征的影响,最后利用Pearson相关系数度量不同类型张量的稀疏表征系数的相关程度。新方法优势有:(1)能够保留叠前数据中不同维度上更丰富的结构特征;(2)调节Pearson相关系数的阈值能优化分类效果;(3)充分利用钻井信息以及专家知识经验标定训练样本标签。数值模型测试证实,相比于传统的支持向量机(Support Vector Machine,SVM)、K最邻近法(K-Nearest Neighbor,KNN)分类算法,新方法的准确度更高、鲁棒性更强。实际数据的应用进一步证实新方法的分类结果更加符合地质规律,更适用于沉积相的分析与解释。

关 键 词：Prestack seismic data seismic reflection pattern analysis TENSORS discriminative dictionary learning 

分 类 号：P631.4[天文地球—地质矿产勘探]