PRINCIPLES OF DIGITAL WIENER FILTERING*

The theory of statistical communication provides an invaluable framework within which it is possible to formulate design criteria and actually obtain solutions for digital filters. These are then applicable in a wide range of geophysical problems. The basic model for the filtering process considered here consists of an input signal, a desired output signal, and an actual output signal. If one minimizes the energy or power existing in the difference between desired and actual filter outputs, it becomes possible to solve for the so-called optimum, or least squares filter, commonly known as the “Wiener” filter. In this paper we derive from basic principles the theory leading to such filters. The analysis is carried out in the time domain in discrete form. We propose a model of a seismic trace in terms of a statistical communication system. This model trace is the sum of a signal time series plus a noise time series. If we assume that estimates of the signal shape and of the noise autocorrelation are available, we may calculate Wiener filters which will attenuate the noise and sharpen the signal. The net result of these operations can then in general be expected to increase seismic resolution. We show a few numerical examples to illustrate the model's applicability to situations one might find in practice.     

A COMPARISON OF STATISTICAL AND DETERMINISTIC WIENER DECONVOLUTION OF DEEP-TOW SEISMIC DATA*

Wiener ‘spiking’ deconvolution of seismic traces in the absence of a known source wavelet relies upon the use of digital filters, which are optimum in a least-squares error sense only if the wavelet to be deconvolved is minimum phase. In the marine environment in particular this condition is frequently violated, since bubble pulse oscillations result in source signatures which deviate significantly from minimum phase. The degree to which the deconvolution is impaired by such violation is generally difficult to assess, since without a measured source signature there is no optimally deconvolved trace with which the spiked trace may be compared. A recently developed near-bottom seismic profiler used in conjunction with a surface air gun source produces traces which contain the far-field source signature as the first arrival. Knowledge of this characteristic wavelet permits the design of two-sided Wiener spiking and shaping filters which can be used to accurately deconvolve the remainder of the trace. In this paper the performance of such optimum-lag filters is compared with that of the zero-lag (one-sided) operators which can be evaluated from the reflected arrival sequence alone by assuming a minimum phase source wavelet. Results indicate that the use of zero-lag operators on traces containing non-minimum phase wavelets introduces significant quantities of noise energy into the seismic record. Signal to noise ratios may however be preserved or even increased during deconvolution by the use of optimum-lag spiking or shaping filters. A debubbling technique involving matched filtering of the trace with the source wavelet followed by optimum-lag Wiener deconvolution did not give a higher quality result than can be obtained simply by the application of a suitably chosen Wiener shaping filter. However, cross correlation of an optimum-lag spike filtered trace with the known ‘actual output’ of the filter when presented with the source signature is found to enhance signal-to-noise ratio whilst maintaining improved resolution.     

PREDICTION ERROR FILTERS,WHITE NOISE AND ORTHOGONAL COORDINATES*

Optimum filters can be computed using orthogonal coordinates obtained from the eigenvalues and eigenvectors of the autocorrelation matrix. The method is used to obtain unit distance prediction error filters. The output of a unit distance prediction error filter when applied to the input wavelet is an impulse at zero time. The effect on the output of added white noise is easily obtained using the approach through the orthogonal coordinates. The added white noise results in output wavelets which are no longer impulses at zero time. The decrease in time resolution gives a filter that does not increase undesirable high frequency noise as much as filters computed without white noise. Orthogonal coordinates with little signal energy can be omitted from the filter computation resulting in output wavelets resembling those computed using added white noise.     

WAVELET SHAPING AND NOISE REDUCTION *

Approximate deconvolution by means of Wiener filters has become standard practice in seismic data-processing. It is well-known that addition of a certain percentage of noise energy to the autocorrelation of the signal wavelet leads to a filter that does not increase, or even reduces, the noise level on the seismogram. This noise addition will, in general, cause a minimum phase signal to become mixed phase. A technique is presented for the calculation of the optimum-lag shaping filter for a contaminated signal wavelet. The advantages of this method over the more conventional approach are that it needs less arithmetic operations and that it automatically gives the filter with the optimum combination of shaping performance and noise reduction.     

THE PERFORMANCE OF OPTIMUM STACKING FILTERS IN SUPPRESSING UNCORRELATED NOISE*

Optimum stacking filters based on estimates of trace signal-to-uncorrelated noise ratios are assessed and compared in performance with conventional straight stacking. It is shown that for the trace durations and signal bandwidths normally encountered in seismic reflection data the errors in estimating signal/noise ratios largely counteract the theoretical advantages of the optimum filter. The more specific the filter (e.g. the more frequency components included in its design) the more this is true. Even for a simple weighted stack independent of frequency, the performance is likely to be better than a straight (equal weights) stack only for relatively high signal/noise ratios, when the performance is not critical anyway.     

AFTER-STACK MULTICHANNEL FILTERS WITHOUT MIXING EFFECTS *

Several types of multichannel filters have been introduced in the past with the purpose of rejecting, in a seismic section, coherent noise having a slope different from that of the signal. These filters, generally, tend to introduce a certain amount of mixing and therefore the output trace shows increased horizontal coherence. This is due to the model on which these filters are based, since the hypothesis is posed that the reflectors are continuous. This may be dangerous since it could lead to mistaken interpretations, for example when small faults or breaks are made to disappear in the output section. Other problems that could arise in the application of multichannel filters after-stack are space-aliasing and high-pass filtering. The former occurs when coherent noise is rejected with apparent Velocity V and frequency f_a=V/X, where X is the distance between traces. In this case, the signal also is distorted since it is rejected in the same frequency range. The high pass filtering effect occurs when the multichannel filter is designed to remove low coherent noise with high apparent velocity. In the paper a family of multichannel filters is presented based on a model of the seismic section such that minimum mixing effects appear. The filters are designed to give good results even in the case of low frequency and high velocity coherent noise. Some practical examples are shown.     

Microseismic events enhancement and detection in sensor arrays using autocorrelation‐based filtering

Passive microseismic data are commonly buried in noise, which presents a significant challenge for signal detection and recovery. For recordings from a surface sensor array where each trace contains a time‐delayed arrival from the event, we propose an autocorrelation‐based stacking method that designs a denoising filter from all the traces, as well as a multi‐channel detection scheme. This approach circumvents the issue of time aligning the traces prior to stacking because every trace's autocorrelation is centred at zero in the lag domain. The effect of white noise is concentrated near zero lag; thus, the filter design requires a predictable adjustment of the zero‐lag value. Truncation of the autocorrelation is employed to smooth the impulse response of the denoising filter. In order to extend the applicability of the algorithm, we also propose a noise prewhitening scheme that addresses cases with coloured noise. The simplicity and robustness of this method are validated with synthetic and real seismic traces.     

USE OF EXPECTED ERROR IN THE DESIGN OF LEAST-SQUARES OPTIMUM FILTERS*

The design of least-squares optimum filters is based upon minimizing a suitably defined error criterion. The expected value of this error is easily computable after the coefficients of the filter have been determined. When a particular filtering problem is specified, there are several parameters which are specifically not included in the optimization procedure. However, the magnitude of the expected error may be quite sensitive to these parameters. The examination of the relative values of the expected error for variations of these unspecified parameters may lead to a better definition of the filter problem. The parameters which are left unspecified by the general least-square filter definition include: 1. The addition of white noise to the signal autocorrelation to stabilize the filter behavior. 2. The specification of the shape of the desired output of the filter. 3. The specification of the lag between the desired output and the input. Examples are given showing the relationship between these parameters and the value of the expected error.     

SIGNAL ENHANCEMENT OF VIBRATORY SOURCE DATA IN THE PRESENCE OF ATTENUATION*

Amplitude spectra of input FM signals used in the vibratory source method of seismic exploration often show undesirable oscillations near the initial and terminal frequencies. These oscillations have an effect on the correlation background and distort the output signal. Considerable improvement in reducing the amplitude of these oscillations is obtained using a proper taper fuction. Attention is given to the relation between the tapering time and bandwidth of the spectrum. Analyses of the spectra of the received data from vibratory sources show considerable attenuation in comparison with the original field sweep. Since the matched filtering process will result in a series of waveforms which have the shape of the autocorrelation of the input signal, consideration is given to the autocorrelation function and its zero-lag coefficient of the FM signal in the presence of attenuation. A method has been developed which compensates for the attenuation and recovers the distortion of waveforms when the received data is correlated. The design of a waveform shaping filter for vibratory source data is given to reduce the influence of phase distortion on the received waveforms as well as to increase S/N ratio resolution. Parameters used for this filter are based on the properties of the FM signal and its autocorrelation function. Several examples from field data are presented to illustrate the methods. The results indicate that the use of the above techniques yields sections with good frequency resolution and improved S/N ratio.     

OPTIMUM DIGITAL FILTERING AND INVERSE FILTERING IN THE FREQUENCY DOMAIN*

Two distinct filters are developed in the frequency domain which represent an attempt to increase the resolution of fine structure contained in the signal whilst keeping the expected filtered noise energy within reasonable bounds. A parameter termed the White Noise Amplification is defined and used together with a measure of the deconvolved pulse width in order to provide a more complete characterisation of the filters. Each of the two main types of frequency domain filters discussed varies in properties with respect to a single adjustable parameter. This may be contrasted with a time domain Wiener filter which in general has three variables: length, delay and an adjustable noise parameter or weight. The direct frequency domain analogue of the Wiener filter is termed a gamma-Fourier filter, and is shown to have properties which span the range from those of a spiking filter with zero least square error at one extreme, to those of a matched filter at the other extreme of its variable parameter's range. The second type of filter considered—termed the modulated Gaussian filter—is similarly shown to be a perfect spiking filter at one extreme of its parameter range, but adopts the properties of an output energy filter at the other extreme.     

PARTIAL COHERENCE MATCHING OF SYNTHETIC SEISMOGRAMS WITH SEISMIC TRACES*

A crucial step in the use of synthetic seismograms is the estimation of the filtering needed to convert the synthetic reflection spike sequence into a clearly recognizable approximation of a given seismic trace. In the past the filtering has been effected by a single wavelet, usually found by trial and error, and evaluated by eye. Matching can be made more precise than this by using spectral estimation procedures to determine the contribution of primaries and other reflection components to the seismic trace. The wavelet or wavelets that give the least squares best fit to the trace can be found, the errors of fit estimated, and statistics developed for testing whether a valid match can be made. If the composition of the seismogram is assumed to be known (e.g. that it consists solely of primaries and internal multiples) the frequency response of the best fit wavelet is simply the ratio of the cross spectrum between the synthetic spike sequence and the seismic trace to the power spectrum of the synthetic spike sequence, and the statistics of the match are related to the ordinary coherence function. Usually the composition cannot be assumed to be known (e.g. multiples of unknown relative amplitude may be present), and the synthetic sequence has to be split into components that contribute in different ways to the seismic trace. The matching problem is then to determine what filters should be applied to these components, regarded as inputs to a multichannel filter, in order to best fit the seismic trace, regarded as a noisy output. Partial coherence analysis is intended for just this problem. It provides fundamental statistics for the match, and it cannot be properly applied without interpreting these statistics. A useful and concise statistic is the ratio of the power in the total filtered synthetic trace to the power in the errors of fit. This measures the overall goodness-of-fit of the least squares match. It corresponds to a coherent (signal) to incoherent (noise) power ratio. Two limits can be set on it: an upper one equal to the signal-to-noise ratio estimated from the seismic data themselves, and a lower one defined from the distribution of the goodness-of-fit ratios yielded by matching with random noise of the same bandwidth and duration as the seismic trace segment. A match can be considered completely successful if its goodness-of-fit reaches the upper limit; it is rejected if the goodness-of-fit falls below the lower one.     

Seismic directional random noise suppression by radial‐trace time–frequency peak filtering using the Hurst exponent statistic

Radial‐trace time–frequency peak filtering filters a seismic record along the radial‐trace direction rather than the conventional channel direction. It takes the spatial correlation of the reflected events between adjacent channels into account. Thus, radial‐trace time–frequency peak filtering performs well in denoising and enhancing the continuity of reflected events. However, in the seismic record there is often random noise whose energy is concentrated in certain directions; the noise in these directions is correlative. We refer to this kind of random noise (that is distributed randomly in time but correlative in the space) as directional random noise. Under radial‐trace time–frequency peak filtering, the directional random noise will be treated as signal and enhanced when this noise has same direction as the signal. Therefore, we need to identify the directional random noise before the filtering. In this paper, we test the linearity of signal and directional random noise in time using the Hurst exponent. The time series of signals with high linearity lead to large Hurst exponent value; however, directional random noise is a random series in time without a fixed waveform and thus its linearity is low; therefore, we can differentiate the signal and directional random noise by the Hurst exponent values. The directional random noise can then be suppressed by using a long filtering window length during the radial‐trace time–frequency peak filtering. Synthetic and real data examples show that the proposed method can remove most directional random noise and can effectively recover the reflected events.     

CORRECTING FOR COLOURED PRIMARY REFLECTIVITY IN DECONVOLUTION1

Statistical deconvolution, as it is usually applied on a routine basis, designs an operator from the trace autocorrelation to compress the wavelet which is convolved with the reflectivity sequence. Under the assumption of a white reflectivity sequence (and a minimum-delay wavelet) this simple approach is valid. However, if the reflectivity is distinctly non-white, then the deconvolution will confuse the contributions to the trace spectral shape of the wavelet and reflectivity. Given logs from a nearby well, a simple two-parameter model may be used to describe the power spectral shape of the reflection coefficients derived from the broadband synthetic. This modelling is attractive in that structure in the smoothed spectrum which is consistent with random effects is not built into the model. The two parameters are used to compute simple inverse- and forward-correcting filters, which can be applied before and after the design and implementation of the standard predictive deconvolution operators. For whitening deconvolution, application of the inverse filter prior to deconvolution is unnecessary, provided the minimum-delay version of the forward filter is used. Application of the technique to seismic data shows the correction procedure to be fast and cheap and case histories display subtle, but important, differences between the conventionally deconvolved sections and those produced by incorporating the correction procedure into the processing sequence. It is concluded that, even with a moderate amount of non-whiteness, the corrected section can show appreciably better resolution than the conventionally processed section.     

IDENTIFICATION OF SEISMIC REFLECTIONS USING SINGULAR VALUE DECOMPOSITION*

Singular value decomposition (SVD) is applied to the identification of seismic reflections by using two different models: the impulse response model, where a seismic trace is assumed to consist of a known signal pulse convolved with a reflection coefficient series plus noise, and the delayed pulse model, where the seismic signal is assumed to consist of a small number of delayed pulses of known shape and with unknown amplitudes and arrival times. SVD clearly shows how least-squares estimation of the reflection coefficients may become unstable, since a division by the singular values is required. Two methods for stabilizing this procedure are investigated. The inverse of the singular values may be replaced by zeros when they are less than a given threshold. This is called the SVD cut-off method. Alternatively, we may use ridge regression which in filter design corresponds to assuming white noise. Statistical methods are used to compute an optimal SVD cut-off level and also to compute an optimal weighting parameter in ridge regression. Numerical studies indicate that the use of SVD cut-off or ridge regression stabilizes the least-squares procedure, but that the results are inferior to maximum-likelihood estimation where the noise is assumed to be filtered white noise. For the delayed pulse model, we use a linearization procedure to iteratively update the estimates of both the reflection amplitudes and the arrival times. In each step, the optimal SVD cut-off method is used. Confidence regions for the estimated reflection amplitudes and arrival times are also computed. Synthetic data examples demonstrate the effectiveness of this method. In a real data example, the maximum-likelihood method assuming an impulse response model is first used to obtain initial estimates of the number of reflections and their amplitudes and traveltimes. Then the iterative procedure is used to obtain improved estimates of the reflection amplitudes and traveltimes.     

Incorporation of a Non-linear Image Filtering Technique for Noise Reduction in Seismic Data

Seismic noise is a fundamental part of seismic data which cannot be avoided when conducting any seismic survey. It consists of coherent and random noise. Noise removal or filtering is one of the major concerns in the field of seismic processing. In this paper, we introduce an image filtering technique based on a detection-estimation algorithm for Gaussian and random noise removal in seismic data, namely the trilateral filter, based on a statistic called rank-ordered absolute differences. The non-linear and adaptive behaviour of this filter makes it very robust in the presence of random and coherent noise, in addition to its computational simplicity and its ability to automatically identify noise in data. We have modified the strategy of trilateral filtering by adapting the rank-ordered absolute differences formula in order to extract the signal component. We have successfully used this filter for the removal of surface waves and random spiky noise from synthetic and field data. Results are very encouraging and show the superiority of this filter compared with other filters, particularly when used recursively.     

基于时变窄带滤波技术提取可控震源扫频信号方法研究

精密主动地震监测为我们主动探测地下介质结构, 并监视其动态变化提供了一条可能的技术途径。 由于精密控制震源释放的能量强度小, 随着传播距离的增加, 信号的快速衰减, 在离震源较远处有用的震源信号被掩盖在很强的噪声中, 这对于震相的识别与走时的拾取精度有着很大的影响。 本文设计了一种时变窄带滤波器进行更为精细的滤波, 期望进一步提高观测数据信噪比, 再结合匹配滤波方法实现主动震源信号的检测与波形变换。 另外, 由于不同震相有一定的到时差, 使用时变窄带滤波器提取一个震相波形信息时, 将压制其他震相的波形信号, 从而实现了震相分离的技术。 仿真计算与实际资料处理显示了该方法在提高观测资料质量、 震相识别分辨率及震相分离方面具有一定的有效性与优越性。 通过对广东省新丰江库区精密可控震源试验数据进行时变窄带滤波方法的处理, 在震中距为200 km左右处的台站记录中检测到主动源信号, 体现了对远台弱信号的提取能力。     

Applications of Savitzky-Golay Filter for Seismic Random Noise Reduction

This article utilizes Savitzky–Golay (SG) filter to eliminate seismic random noise. This is a novel method for seismic random noise reduction in which SG filter adopts piecewise weighted polynomial via leastsquares estimation. Therefore, effective smoothing is achieved in extracting the original signal from noise environment while retaining the shape of the signal as close as possible to the original one. Although there are lots of classical methods such as Wiener filtering and wavelet denoising applied to eliminate seismic random noise, the SG filter outperforms them in approximating the true signal. SG filter will obtain a good tradeoff in waveform smoothing and valid signal preservation under suitable conditions. These are the appropriate window size and the polynomial degree. Through examples from synthetic seismic signals and field seismic data, we demonstrate the good performance of SG filter by comparing it with the Wiener filtering and wavelet denoising methods.     

Enhancement of Deep Seismic Reflections in Pre-stack Data by Adaptive Filtering

—Adaptive filters offer advantages over Wiener filters for time-varying processes. They are used for deconvolution of seismic data which exhibit non-stationary behavior, and seldom for noise reduction. Different algorithms for adaptive filtering exist. The least-mean-squares (LMS) algorithm, because of its simplicity, has been widely applied to data from different fields that fall outside geophysics. The application of the LMS algorithm to improve the signal-to-noise ratio in deep reflection seismic pre-stack data is studied in this paper. Synthetic data models and field data from the DEKORP project are used to this end.¶Three adaptive filter techniques, one-trace technique, two-trace technique and time-slice technique, are examined closely to establish the merits and demerits of each technique. The one-trace technique does not improve the signal-to-noise ratio in deep reflection seismic data where signal and noise cover the same frequency range. With the two-trace technique, the strongest noise reduction is achieved for small noise on the data. The filter efficiency decreases rapidly with increasing noise. Furthermore, the filter performance is poor upon application to common-midpoint (CMP) gathers with no normal-moveout (NMO) corrections. Application of the two-trace method to seismic traces before dynamic correction results in gaps in the signal along the reflection hyperbolas. The time-slice technique, introduced in this paper, offers the best answer. In this case, the one-trace technique is applied to the NMO-corrected gathers across all traces in each gather at each time to separate the low-wavenumber component of the signal in offset direction from the high-wavenumber noise component. The stacking velocities used for the dynamic correction do not need to be known very accurately because in deep reflection seismics, residual moveouts are small and have only a minor influence on the results of the adaptive time-slice technique. Noise reduction is more significant with the time-slice technique than with the two-trace technique. The superiority of the adaptive time-slice technique is demonstrated with the DEKORP data.     

A SEMBLANCE-GUIDED MEDIAN FILTER1

Reiter , E.C., Toksoz , M.N. and Purdy , G.M. 1992. A semblance-guided median filter. Geophysical Prospecting 41 , 15–41. A slowness selective median filter based on information from a local set of traces is described and implemented. The filter is constructed in two steps, the first being an estimation of a preferred slowness and the second, the selection of a median or trimmed mean value to replace the original data point. A symmetric window of traces defining the filter aperture is selected about each trace to be filtered and the filter applied repeatedly to each time point. The preferred slowness is determined by scanning a range of linear moveouts within the user-specified slowness passband. Semblance is computed for each trial slowness and the preferred slowness selected from the peak semblance value. Data points collected along this preferred slowness are then sorted from lowest to highest and in the case of a pure median filter, the middle point(s) selected to replace the original data point. The output of the filter is therefore quite insensitive to large amplitude noise bursts, retaining the well-known beneficial properties of a traditional 1D median filter. Energy which is either incoherent over the filter aperture or lies outside the slowness passband, may be additionally suppressed by weighting the filter output by the measured peak semblance. This approach may be used as a velocity filter to estimate coherent signal within a specified slowness passband and reject coherent energy outside this range. For applications of this type, other velocity estimators may be used in place of our semblance measure to provide improved velocity estimation and better filter performance. The filter aperture may also be extended to provide increased velocity estimation, but will result in additional lateral smearing of signal. We show that, in addition to a velocity filter, our approach may be used to improve signal-to-noise ratios in noisy data. The median filter tends to suppress the amplitude of random background noise and semblance weighting may be used to reduce the amplitude of background noise further while enhancing coherent signal. We apply our method to vertical seismic profile data to separate upgoing and downgoing wavefields, and also to large-offset ocean bottom hydrophone data to enhance weak refracted and post-critically reflected energy.