Elimination of seismic random noise based on the SW statistic adaptive TFPF
Lin Hong-Bo,Ma Hai-Tao,Li Yue,Shao Dong-Yang
DOI: https://doi.org/10.6038/cjg20151218
2015-01-01
Abstract:Owing to complex properties of random noise in raw data in metal mine and low signal-to-noise ratio (SNR), it is extremely difficult for conventional denoising methods to obtain expected filtering results. Time-frequency peak filtering (TFPF) is an effective method to eliminate seismic random noise in seismic data at low SNR. However, the selection of window length of TFPF significantly affects the performance in signal preserving and seismic random noise attenuation. The conventional TFPF using a fixed window length usually obtains unbiased signal estimation by using a short window length, but it leads to relatively poor performance of seismic random noise attenuation. Therefore, it is crucial to adapt the window length for TFPF according to the characteristics of signal and noise, respectively. Taking statistical property of seismic random noise into account, we propose a Shapiro-Wilk (SW) statistic based adaptive time-frequency peak filtering (S-TFPF) to suppress seismic random noise in seismic data at low SNR. The SW test, a statistical method for the measurement of Gaussianity of time series, is introduced into TFPF method. Based on the assumption that seismic random noise usually is white Gaussian noise and seismic signals are non-Gaussian, the SW statistics of seismic random noise are different from those of seismic signals. Therefore, the seismic signals in seismic data can be identified by means of the SW statistics. Furthermore, Gaussianization of seismic data is done by applying a band-pass filter to seismic data, which makes complex seismic random noise Gaussian and keep seismic signals. As a result, the accuracy of identification of valid signals under complex seismic random noise is improved based on SW statistics. Then, adaptively adjusting window length of S-TFPF is implemented based on the SW statistics. In this algorithm, the window length of S-TFPF in the signal-dominant segment are set according to the frequencies of signals to preserve signals, whereas the window length of S-TFPF for noise-dominant segment increases with the variance of noise increasing, so as to completely eliminate seismic random noise. The Gaussianity of seismic noise data is investigated by SW test and the performance of new method is analyzed on synthetic data and field data. The SW test result show that most seismic random noise are non-Gaussian noise and their SW statistics are lower than but close to the SW statistic of ideal Gaussian noise. The significant difference of the SW statistics exists between random noise and seismic signals. However, the difference of SW statistic of noisy seismic data decreases, because signals are contaminated by seismic random noise and properties of seismic random noise are complex. After preprocessing seismic data by means of Gaussianization, the SW statistics of seismic random noise becomes closer to 1 and the SW statistics of seismic signals slightly decrease, which leads to an accurate segmenting of seismic signal and seismic random noise. Then the adaptive window length of the S-TFPF is obtained based on the SW statistics and apply to processing synthetic and field seismic data. The results show that the S-TFPF method better keeps the amplitude and frequency component of filtered seismic signals than the TFPF. Furthermore, the filtered seismic data obtained by the S-TFPF has higher SNR and lower mean square error comparing with the TFPF. Application to the field data shows that the filtered seismic data by using S-TFPF has less background noise and more continuous seismic events. The proposed method improves the adaptability of window length of the TFPF using SW statistics of seismic data. In the new method, the window length can be adapted at different segments of seismic data according to characteristics of seismic signals and statistical property of seismic random noise, respectively, thus reducing the bias of seismic signal estimation and improving denoising performance of the TFPF. The results of synthetic and field data demonstrate the practicability and effectiveness of the S-TFPF method.