基于压缩感知和深度学习的地震数据重建

在越来越复杂的勘探环境中,针对因地震数据不完整、不规则导致的重建结果不精确的问题,提出了一种基于压缩感知和深度学习的地震数据重建方法。首先,利用地震信号在Shearlet变换域内的稀疏性,在压缩感知框架下对数据进行预处理。然后,将预处理后的结果作为网络的输入数据,将完整的地震数据作为标签数据,对两者进行数据块处理并建立训练数据,利用卷积神经网络(CNN)实现地震数据端到端的重建。最后,基于训练后的网络模型获得最终重建结果。利用合成地震数据和实际地震数据的重建结果验证基于压缩感知和深度学习的地震数据重建方法的有效性。试验结果表明,在采样率相同的情况下,相比于压缩感知重建方法和深度学习重建方法,结合压缩感知和深度学习的重建方法能够更有效地恢复缺失数据,重建误差更小,在数据大量缺失的情况下,该方法也能够表现出较好的重建性能。

Due to the cost and limitations of the exploration environment,the raw data obtained from field seismic surveys is often incomplete and irregular,which affects subsequent processing accuracy.This study built a new seismic data reconstruction method based on compressed sensing and deep learning.The initial seismic data was reconstructed using compressed sensing in the data sparsity of the Shearlet domain.During the initial reconstruction an iterative thresholding algorithm and exponential threshold model were used to obtain an optimal solution.Thereafter,a convolutional neural network was used to complete the reconstruction.The initial reconstruction data was used as the network input data,the complete raw dataset was used as the label data,and the training dataset was used as the patch pair of the input and complete data.Finally,the trained network could be used directly for the reconstruction of incomplete datasets.This method combines model-and data-driven techniques to achieve the initial reconstruction and to complete the subsequent complex reconstruction process,respectively.The synthetic and real seismic data results show that the reconstruction method combined with compressed sensing and deep learning can recover the missing data information more effectively with the same sampling rate,has smaller reconstruction error,and can show good performance in the case of large missing data.The reconstructed data have the potential to improve the accuracy of subsequent seismic data processing.

基金项目：国家自然科学基金项目(41874133,U19B2008)资助。