Broadband reconstruction of seismic signal with generative recurrent adversarial network

Resolution is an important problem in seismic data processing and interpretation. High resolution seismic data can improve the accuracy of well-seismic tie, structural interpretation, attribute analysis, and lithology inversion. The resolution of seismic data depends on the main frequency and band width of seismic data, and the increasingly complex exploration objects require higher resolution algorithm to provide more abundant waveform and phase information. The practice shows that the signal to noise ratio of seismic data processing with traditional method will be affected while improving the resolution, thus reduce the reliability of seismic data. Based on the method and principle of generative recurrent adversarial network (GRAN), we proposes a method to generate pseudo sample sets which was constrained by geological layer and controlled with logging interpretation results, and train the GRAN model to realize the broadband reconstruction of seismic data. On the basis of theoretical discussion and fine calibration of well tie, the applicability of the new method is proved by blind well test. Finally, the method is applied to the seismic data processing of the actual work area and excellent results are obtained.