Seismic Data Frequency Extension Based On Deep Learning

Due to the restriction of the source spectrum and the attenuation of seismic waves during the propagation process,seismic data are usually limited to the frequency band width,lack effective high-frequency or low frequency components.The resolution of seismic data is determined by the frequency and bandwidth of the data.The low frequency information of the expansion of seismic data can suppress the energy of the side loot,enhance the degree of layering ability,and improve the inversion accuracy.Increasing the dominant frequency and expanding high frequency information can compress the wavelet length,enhance the identification of thin layers,and improve the accuracy of constructing interpretation.This thesis is based on deep learning techniques to perform spectrum expansion.The research content mainly includes four parts.First,with a wide bandwidth,a weaker side lobe Wide band B-spline wavelet is synthesized by high resolution seismic data labels.Second,an encoder decoder network is built into convolutional neural network and circulating neural network according to the timespace characteristics of seismic data.Third,based on the composite different data sets,the network is trained,and different predictive results are obtained and compared.Fourth,the effectiveness of the method is verified by model data testing and actual data,and better results are achieved in ground data applications.Comparative analysis shows that the synthesis of training data can increase the accuracy of the prediction results using the reflection coefficient of the logging data and the wavelet of seismic data,and use the two-dimensional data sets to train the network model can improve the transverse continuity of the prediction data.In the actual data application,the high-frequency information that has occurred in the data after frequency extension.And the reliability of the method is verified by well-seismic contrast.