基于联合小波域深度学习的地震数据规则化方法

Seismic data regularization based on deep learning combining wavelet domain

数据规则化是地震资料处理的关键步骤之一,基于物理建模的传统方法计算量大且不具备广泛适用性。当前基于卷积神经网络的地震数据规则化方法通常局限在时域,尤其在低采样率条件下,重建数据过于平滑,纹理细节信息损失严重。小波分析具有多尺度、多方向的特性,更适于表示二维数据的纹理特性,可以聚焦地震数据信号的细节信息。为此,提出一种联合小波域的卷积神经网络模型,学习地震数据在时域与小波域的联合分布特征以逼近实际数据,将不规则地震数据重建问题转化为在卷积神经网络框架下各尺度不同方向分量的小波系数预测,重建规则化的地震数据;构建时域与小波域的联合损失函数,结合地震数据的整体分布和局部细节特征,约束网络模型,通过修正联合损失函数的权重调整卷积神经网络学习的注意力,提高重建地震数据信噪比。实验结果表明,与其他方法对比,该方法细节保持效果更好,对地震数据缺失位置不敏感,更具鲁棒性。

Data regularization is a fundamental step in seismic data processing,and the conventional method based on physical modeling requires massive computations and is not widely in use.At present,the regularization methods of seismic data based on convolutional neural networks(CNNs)are usually limited in the time domain,which leads to the problems of the excessively smooth reconstructed data and severe loss of texture details,especially at a low sampling rate. Wavelet analysis has the characteristics of multiple scales and multiple directions,which is more suitable to represent the texture characteristics of two-dimensional data and can focus on the details of seismic data signals.Therefore,a CNN model combining the wavelet domain is proposed to learn the joint distribution characteristics of seismic data in the time and wavelet domains and thus approximate the actual data.Specifically,the reconstruction of irregular seismic data is transformed into the wavelet coefficient prediction of components of different directions in each scale under the framework of a CNN to reconstruct regularized seismic data.A joint loss function in the time and wavelet domains is constructed,and by the overall distribution and local details of seismic data,the network model is constrained.The attention of CNN learning can be adjusted by the modification of the weight of the joint loss function to raise the signal-to-noise ratio(SNR)of the reconstructed seismic data.The experiments demonstrate that the proposed method can better preserve details compared with other methods,and it is insensitive to the missing location of seismic data and has good robustness.