Abstract:

The method of seismic data reconstruction is an important way to improve sampling density and obtain complete wavefield information in seismic signal processing. The data collected from field exploration is often incomplete or irregularly sampled due to limitations in surface conditions and cost control. Therefore, studying methods for reconstructing irregular seismic signals is of great significance. Based on the assumption that irregular seismic data can be regarded as random sparsity of regular complete data, the mathematical statistical distribution of the two data is highly consistent in the data interval and the performance in the sparse domain is extremely similar, this paper proposes a method of data reconstruction by using residual convolutional encoder-decoder to reduce the dimensionality of the data to the sparse domain and then increase it. Furthermore, this paper proposes amplitude consistency correction processing on the input data of the network to balance energy, in order to address the serious impact of inconsistency in amplitude depth and shallow layers caused by seismic attenuation, as well as significant differences in mathematical statistical distribution between different data, on the training and generalization of neural networks. The synthetic data example proves that the proposed method has higher accuracy than traditional two-dimensional prediction error filter and residual network interpolation methods, and is more efficient than traditional three-dimensional prediction error filter interpolation method. Different field land and ocean data examples have achieved good reconstruction results, proving that this method greatly increases the generalization ability of the network and reduces the training difficulty of the network.

Key words: residual convolutional encoder-decoder')">

residual convolutional encoder-decoder, amplitude consistency, irregular sampling, data reconstruction, seismic signal processing