融入先验信息的深度学习地震子波振幅谱估计

Seismic wavelet extraction plays an important role in seismic data processing. The seismic wavelet is determined by the amplitude spectrum and phase spectrum; thus estimating the amplitude spectrum of seismic wavelets is a key step. The common methods for extracting amplitude spectra of seismic wavelets are subject to limitations on the reflectivity type and the seismic wavelet form, failing to achieve a wide application in actual seismic data processing. To obtain the accurate amplitude spectra of seismic wavelets, we proposed prior knowledge embedded deep learning. Firstly, given the prior information, i.e. smoothness of the seismic wavelet amplitude spectrum, the observed seismic data were preprocessed. Then, the smoothed amplitude spectrum of seismograms was input into a 12-layer deep neural network to obtain the amplitude spectrum of the seismic wavelet. Compared with traditional spectrum simulation, the proposed method can obtain an accurate estimation of the unimodal and non-unimodal amplitude spectra of seismic wavelets without any hypothesis about reflectivity. Moreover, it can avoid the parameter estimation in polynomial fitting methods and the calculation error induced by opera-ting band estimation. Specifically, it has superior anti-noise performance, and the advantage of this method has been proved by model examples and actual data tests.