不同域下基于UNet的三维地震数据重建

3D Seismic Data Reconstruction Based on UNet in Different Domains

由于自然因素的限制以及经济条件的约束,采集到的地震数据通常存在地震道的缺失,需对缺失数据进行插值。然而,目前基于深度学习的方法大多在某一固定的域下完成地震数据的插值。为了对比不同域下地震数据重建结果的差异,本文基于UNet网络,对比了五种不同域下的地震数据插值方法,分别为:t-x(time-space)域,时间切片,f-x(frequency-space)域,频率域按炮切片以及三维方法。在仿真数据集和真实数据集上进行了实验,通过三维重建结果,二维切片,不同缺失率的地震数据重建信噪比以及五种方法的耗时对比,全面展示了几种方法在地震数据重建问题上的差异。实验结果显示,五种方法在一定程度上均可重建缺失数据,其中三维方法可以取得最高的重建精度,和基于f-x域的方法相比,重建信噪比提高了近2 dB;但三维方法以三维数据块作为输入训练网络,所以最为耗时,网络一轮的训练时间比其它方法多70~170 min。此外,对不同缺失率的地震数据重建结果显示,三维方法和基于f-x域的方法在重建高缺失率的地震数据时更有优势。

Due to the limitations of natural environments and the constraints of economic condition,acquired seismic data often contains missing traces,which seriously affects subsequent seismic data processing and interpretation.At present,most deep learning methods conduct the interpolation of seismic data within a specified domain.To compare the difference of seismic data reconstruction results in different domains,based on UNet network,this paper compares five seismic data interpolation methods in different domains:time-space domain,time slice,frequency-space domain,frequency-domain slicing by shot and 3D method.We conducted a series of experiments on synthetic and field data.The 3D reconstruction results,2D slice,SNR in seismic data reconstruction with different missing ratios and time consumption comparison of five methods comprehensively show the differences among various methods in seismic data missing problem.Experimental results indicate that all five methods are able to reconstruct missing data to some extent.3D method can achieve the highest reconstruction accuracy,and compared with methods based on the f-x domain,the reconstruction signal-to-noise ratio is improved by nearly 2 dB.However,because of using 3D data blocks as input to train the network,the 3D method is the most time-consuming,and one round of network training takes 70~170 minutes more than other methods.In addition,the reconstruction results of seismic data with different missing rates show that 3D methods and f-x domain based methods have more advantages in reconstructing high missing rate seismic data.