基于M-DeepLab网络的速度建模技术研究  

Research on Velocity Modeling Technology Based on M-Deeplab Net

作　　者：徐秀刚[1,2] 张浩楠 许文德 郭鹏 Xu Xiugang;Zhang Haonan;Xu Wende;Guo Peng(College of Marine Geosciences,Ocean University of China,Qingdao 266100,China;Key Laboratory of Submarine Geosciences and Prospecting Techniques,Ministry of Education,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学海洋地球科学学院,山东青岛266100 [2]中国海洋大学海底科学与探测技术教育部重点实验室,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2024年第6期145-155,共11页Periodical of Ocean University of China

基　　金：科技部国家重点研究发展计划项目(2021YFE0108800);国家自然科学基金项目(42074140)资助。

摘　　要：本文提出了一种适用于速度建模方法的M-DeepLab网络框架,该网络将地震炮集记录作为输入,网络主体使用轻量级MobileNet,以此提升网络训练速度;并在编码环节ASPP模块后添加了Attention模块,且在解码环节将不同网络深度的速度特征进行了融合,既获得了更多的速度特征,又保留了网络浅部的速度信息,防止出现网络退化和过拟合问题。模型测试证明,M-DeepLab网络能够实现智能、精确的速度建模,简单模型、复杂模型以及含有噪声数据复杂模型的智能速度建模,均取得了良好的效果。相较DeepLabV3+网络,本文方法对于速度模型界面处的预测,特别是速度突变区域的预测,具有更高的预测精度,从而验证了该方法精确性、高效性、实用性和抗噪性。Seismic velocity modeling is the basis of seismic data processing and interpretation.Traditional methods require full human-computer interaction and a lot of time for picking up velocity spectra,which cannot be processed efficiently.With the continuous application of deep learning in the field of geology,deep learning velocity modeling has become a hot research topic.At present,DeepLabV3+network is widely used in the field of image processing,and the multiscale feature extraction of ASPP module in the network has good processing advantages for the construction of seismic velocity information.Therefore,based on the DeepLabV3+network,this paper proposes an M-DeepLab network framework applicable to velocity modeling methods,which takes seismic cannon set records as input and uses lightweight MobileNet in the network body to enhance the network training speed;and adds the Attention module after the ASPP module in the encoding session,and in the decoding session The speed features of different network depths are fused to obtain more speed features while retaining the speed information of the shallow part of the network to prevent network degradation and overfitting problems.Model testing proves that the M-DeepLab network can achieve intelligent and accurate velocity modeling for simple models,complex models,and complex models containing noisy data.Compared with the DeepLabV3+network,the method in this paper has higher prediction accuracy for the prediction at the velocity model interface,especially for the velocity mutation region,thus verifying the accuracy,efficiency and practicality and noise immunity of the method.

关 键 词：深度学习 速度建模 M-DeepLab网络 监督学习 

分 类 号：P631.4[天文地球—地质矿产勘探]