基于卷积神经网络的智能化地震数据插值技术

Intelligent seismic data interpolation via convolutional neural network

地震数据通常存在数据缺失问题,严重影响地震数据各个处理环节,需采用适当的手段对其重构.本文提出了一种基于深度学习卷积神经网络(CNN)的智能化地震数据插值技术.算法的关键在于构建一个适用于地震资料插值的CNN模型,该技术以缺失地震数据作为输入层,由卷积算法提取地震数据的特征信息,并通过池化层实现数据压缩降维,同时引入修正线性函数(ReLU)提高模型的非线性表达能力,再通过反卷积层恢复数据尺寸,最终搭建卷积自编码器模型(CAE),实现数据-数据的映射关系.该模型通过残差学习获得缺失数据特征并实现重构数据输出,与现有技术相比,该方法采用自监督学习方式,利用大量数据训练卷积自编码器模型,通过所得模型实现缺失地震道的数据重构.分别利用CAE模型及POCS插值技术对模型资料和实际数据进行插值,测试结果表明,CAE能有效实现地震数据插值,且与POCS方法相比具有更高的精度,验证了算法的可行性和有效性.

Data deficient, as a frequently asked question in seismic exploration, seriously affects the seismic processing, which needs to be reconstructed by appropriate methods. An intelligent seismic data interpolation technique based on deep learning Convolutional Neural Network(CNN) is proposed. The key of the algorithm is to build a CNN network suitable for interpolation of seismic data. The method use the decimated data as input layer, and extract the features of seismic data via convolution. In addition, pooling layer was used for data compression in which the training efficiency can be improved, and ReLU function was introduced into the model to enhance the nonlinear expression ability. Finally, the Convolutional Auto-Encoder model(CAE) was built for seismic data interpolation. Compared to prior methods, the method adopts self-supervised learning to train convolution auto-encoder after which the model can be used for seismic data interpolation. Here, CAE model and POCS interpolation method are used to interpolate the synthetics and field data respectively. The results show that CAE can effectively interpolate the decimated seismic data, and has higher accuracy than POCS method, which verifies the feasibility and validity of the algorithm.