石油地球物理勘探2023，Vol.58Issue(6)：1299-1312,14.DOI:10.13810/j.cnki.issn.1000-7210.2023.06.001
深度神经网络三维地震资料断层解释损失函数对比
Loss function comparison for fault interpretation of three-dimensional seismic data based on deep neural network
摘要
Abstract
Fault interpretation is one of the key steps for seismic data interpretation.The rapid development of deep learning,represented by neural networks,has greatly improved the efficiency and accuracy of seismic fault interpreta-tion.The neural networks are trained by stochastic gradient descent optimization.The parameters of the network model are updated iteratively by using the loss function to measure the error of the model.The selection of the loss function is crucial for the seismic fault interpretation.In this paper,to interpret 3D seismic fault,we use the 3D U-Net model as the network structure and Adam as the optimizer to train the network with 3D synthetic samples.In terms of fault interpretation effects,we compare 10 loss functions including Balanced Cross-Entropy(BCE),Dice,Focal,Cosine,Log-Cosh Dice,Tversky,Focal-Tversky,Wasserstein,BCE-Dice,and BCE-Cosine.Normalization and data augmentation are applied to the trained data to mitigate the discrepancy between synthetic and field data.With the same network model,training parameters,and stopping criteria,we compare the convergence speed,calcula-tion efficiency,and anti-noise performance of the 10 loss functions on 3D U Net and analyze the fault prediction ef-fect by using actual seismic data of the F3 field from offshore Netherlands.The experimental results show that 3D U-Net trained by Tversky and focal-Tversky loss functions can predict fault with better continuity.When crossed or parallel faults are densely distributed,and adjacent fault features can influence each other,the 3D U-Net prediction faults trained by BCE,BCE-dice,and BCE-cosine loss functions are complete,clear,and rich in detail.The research can provide a reference for selecting appropriate loss functions in different scenarios for seismic fault interpretation.关键词
断层解释/损失函数/3D U-Net/数据增强Key words
fault interpretation/loss functions/3D U-Net/data augmentation分类
天文与地球科学引用本文复制引用
张苗苗,吴帮玉,马德波,王治国..深度神经网络三维地震资料断层解释损失函数对比[J].石油地球物理勘探,2023,58(6):1299-1312,14.基金项目
本项研究受陕西省自然科学基础研究计划面上项目"地震波震源波场本征正交分解模型降阶高效重建方法"(2023-JC-YB-269)和国家自然科学基金项目"微地震监测复杂结构偏移成像与速度建模研究"(41974122)联合资助. （2023-JC-YB-269）
