Abstract:

Affected by geological conditions, acquisition environment and other factors, it is always impossible to obtain complete seismic data in the process of geological exploration, which seriously reduces the efficiency of subsequent geological interpretation work. With the development of computer GPU hardware and seismic data processing methods based on convolutional neural networks, more and more deep learning methods are applied to seismic data regularization. At present, such methods are usually limited to processing in the time domain, which often leads to too smooth reconstructed data and a lack of texture detail information. In this paper, a convolution neural network model with joint time-frequency domain characteristics is proposed. Through joint constraints in the time domain and Fourier domain of seismic data, the multi-dimensional distribution characteristics of seismic data in the time-frequency domain are extracted, and the weight of the joint loss function could be modified to adjust the attention of convolution neural network learning. The middle layer of the convolutional neural network is constructed by multi-level adjustable residual blocks to improve the ability of feature extraction. The number of residual blocks can be adjusted according to the needs of the task to balance the accuracy and efficiency of the network. Experiments show that the proposed method in this paper has a better detail retention effect and robustness than other methods.

Key words: seismic data regularization, convolutional neural network, time-frequency combination, deep learning, Fourier transform