基于BiX-NAS的地震层序智能识别--以荷兰近海地区F3数据为例  

Intelligent Seismic Stratigraphic Identification Based on BiX-NAS:A Case Study from the F3 Dataset in Netherlands Offshore Area

作　　者：陈建玮 陈国雄[1] 王德涛 徐富文 Chen Jianwei;Chen Guoxiong;Wang Detao;Xu Fuwen(State Key Laboratory of Geological Processes and Mineral Resources,China University of Geosciences,Wuhan 430074,China;The First Geological Brigade of Hubei Geological Bureau,Daye 435000,China)

机构地区：[1]中国地质大学地质过程与矿产资源国家重点实验室,湖北武汉430074 [2]湖北省地质局第一地质大队,湖北大冶435000

出　　处：《地球科学》2023年第8期3162-3178,共17页Earth Science

基　　金：国家自然科学基金面上项目(No.41972305);原创探索计划项目(No.42050103);地质过程与矿产资源国家重点实验室科技部专项经费资助(No.MSFGPMR2022-3)。

摘　　要：近些年来,深度学习方法在地震数据处理和解释领域得到了广泛关注和应用,其中大多数深度学习算法采用了端到端的深度卷积神经网络以实现地质体特征的提取与识别(如地层、断裂以及盐丘等).然而,这些算法往往含有数十万甚至百万的可训练参数,导致模型存在参数冗余、训练效率低等问题.为了解决上述问题,构建了一个轻量化的双向多尺度网络结构模型用于地震层序智能识别.该模型通过两阶段神经网络体系结构搜索算法(neural architecture search,NAS)剔除了双向多尺度网络结构的冗余连接,使得网络结构大幅简化,从而减少参数冗余,进而提高训练效率.采用荷兰近海地区的F3地震数据集对基于NAS算法简化的双向多尺度网络结构地层识别模型进行训练、验证和预测.结果表明:在实际的地层识别任务中,该轻量化模型的平均识别准确率达到了95.52%,并对远离训练工区的预测集具有良好的泛化性.此外,该模型的参数量仅为U形卷积神经网络(U-Net)模型的4.4%,在训练效率、模型参数量等方面优于前人的相关研究工作;并对地震振幅中的噪声干扰具有鲁棒性.因此,这些结果展现了BiX-NAS网络模型在实际地震地层自动识别中良好的应用前景.In recent years,deep learning methods have been widely focused and applied in the field of seismic data processing and interpretation,where most deep learning algorithms employ end-to-end deep convolutional neural networks for the extraction and identification of geological features(e.g.,stratum,fault,and salt dome).However,these algorithms often contain hundreds of thousands or even millions of trainable parameters,which lead to the model of parameter redundancy and low training efficiency.Therefore,a lightweight bi-directional multi-scale network is constructed for the intelligent identification of stratum.Specifically,the model eliminates the obvious redundant connections of the bi-directional multi-scale network structure through the two-stage Neural Architecture Search(NAS),which greatly simplifies the network structure,reduces the parameter redundancy,and improves the training efficiency.The Netherlands F3 dataset was used to train,verify and predict the simplified bi-directional multiscale network by the NAS.The results show that the average recognition accuracy of the lightweight model reaches 95.52% in the actual stratigraphic identification task,and it has well generalization to the prediction work area far from the training work area.In addition,the number of parameters of the proposed model is only 4.4% of the U-shaped convolutional neural network(U-Net),and which outperforms the previous related work in terms of training efficiency and the number of model parameters.It is also robust when processing noisy seismic data.Therefore,the BiX-NAS network model has good prospects for application in practical automatic seismic stratigraphic identification.

关 键 词：地层自动识别 深度学习 神经网络体系结构搜索算法 双向多尺度网络 

分 类 号：P628[天文地球—地质矿产勘探]