Extraction of the Seismic Wavelet Based on Deep Neural Networks

Abstract

In reflection seismology, seismic wavelet estimation is of great significance for high resolution reflectivity inversion. The method for wavelet estimation can be classified into two categories. One is deterministic and the other is statistic. For the latter, a conventional method uses a spectrum fitting method to estimate the seismic wavelet. The commonly used methods are correlation-based method, the log-spectrum-averaging method and spectrum-shaping method. All of these methods assume that the reflection coefficient sequence (RCS) is white and the source wavelets are zero-phase which may not be valid under certain conditions. In this paper, we propose a new approach to obtain the seismic wavelet based on deep neural network (DNN). We compare the wavelet obtained by our method with the wavelet obtained by widely used spectrum modeling method. Then, the obtained wavelet is applied to perform the inversion of the RCS using the HPP algorithm. Compared with the conventional method, DNN can achieve a more accurate wavelet even if source wavelet is not zero-phase. The resolution of reflectivity inversion is significantly enhanced by using the obtained wavelet.