APPLIED GEOPHYSICS,Vo1.7,No.4(December 201 0),P.31 5-324,9 Figures. DoI:l0.1007/s11770.010.0259—8 An iterative curvelet thresholding algorithm for seismic random noise attenuation★ Wang De-Li’,Tong Zhong-Fei ,Tang Chen’,and Zhu Heng’ Abstract:In this paper.we explore the use of iterative curvelet thresholding for seismic random noise attenuation.A new method for combining the curvelet transform with iterative thresholding to suppress random noise is demonstrated and the iSSUC iS described as a 1inear inverse optimal problem using the L l norm..Random noise suppression in seismic data is transfc}rmed into an L 1 norm optimization problem based on the curvelet sparsity transforil1. Compared to the conventional methods such as median filter algorithm,FX deconvolution, and wavelet thresholding the results of synthetic and field data processing show that the iterative curvelet thresholding proposed in this paper can suficientlfy improve signal to noise radio(SNR)and give higher signal fidelity at the same time.Furthermore,to make better use of the curvelet transform such as multiple scales and multiple directions,we control the curvelet threshoiding to further improve the curvelet direction of the result after iterative SNR. Keywords:curvelet transform,iterative thresholding,random noise attenuation lntroduction Seismic data processing plays an important role in seismic exploration.An successfu1 data processing is supposed to produce three aims:high signal to noise and multi—resolution analysis was applied successfully and widely in both seismic data processing and image processing.The wavelet transform is primarily suitable for isotropic singularities.Unfortunately,when anisotropy is taken into consideration.it can not give a satisfactory representation of anisotropic singularities radio(SNR),high resolution,and high ifdelity.The most important of which is SNR.So how to improve SNR has became the primary task of data processing.In order to achieve that aim,many methods have been developed such as median filters and FX decOnvOluti0n.The former is often used to eliminate linear and random noise and the latter to eliminate random noise.However,when the conventiona1 methods are applied jn denoising.they also such as the boundary and linear characteristics of digital pictures.It’s the reason that some degree of fuzzy phenomenon inevitably emerges when applying wavelet transforms to fusion,compression,or denoising of seismic signals(Huang,2007;Zhao,2007;Wu,2008). However,the discontinuous characteristics of edges or texture is the most important information of signals and it is necessary to develop a more effective method than wavelet transforn1. To further express the more universal characteristics reduce signal at the same time fNeelamani et a1.,2008). To overcome this problem,wavelet transform methods have been rapidly developed during the last decade of curve singularity in fl multi—dimensional signa1.the Manuscript received by the Editor June 1 7,201 0;revised manuscript received November 4,201 0. This work is supported ifnancially by the National Science&Technology Major Projects fGrant No.2008ZX05023-005—013) 1.College of Geo・Exploration Science&Technology.Jilin University,Changchun 1 30026.China. 2.CNOOC Research Institute,Beijing 1 0027,China. 3l5 Seismic random noise attenuation local ridgelet transform was developed and after that the curvelet transfoF1T1 which expresses the whole curve with loca1 straight lines was proposed based on it.The curvelet transform was first proposed by Candes and Donoho(1 999)and applied in some fields including image denoising.image fusion,deconvolution.and so on. Candes and Demanet(20021 and Candes and Guo(2002) developed a new framework of curvelet transfornl called second generation curvelet transform which improved the computation speed.Candes et a1.(2006)proposed two fast discrete realization methods based on the second generation curvelet transfornq which is easier and faster to perform than the foFmer discrete method and greatly decreases the redundancy of the traditionaI algorithm. For the present,the most representative applications of the curvelet transform are denoising and enhancing signals.In essence.a curvelet is a multi—scale 1ocalization of a ridgelet the basic steps of which are:First,make a wavelet transfornl of the signal and decompose it into a series of sub—band signals with different scales;and then, perform a loca1 ridgelet transform on every sub—band. the size of which can differ from the scale.The curvelet transfolrm is a mathematical transform based on the multi.scale ridgelet or local ridgelet transform which can describe obj ects with curve singularities.It combines the advantages of ridgelet and wavelet transfo・rms which are good at expressing the characteristics of lines or points, respectively,and uses the unique characteristics of multi— scale analysis(Huang,2007;Zhao,2007;Wu,2008). Herrmann and Verschuur(2004)and Herrman et al (2007)gave reports on curvelet—based multiple removal, curvelet.and wavelet—based AVO inversion.and so on. and proposed a new concept of curvelet—based seismic data processing.Huub and Masarten(2004)discussed pre.migration in the curvelet domain.Neelamani et a1. (2008)proposed a curvelet—based coherent and random noise attenuation method which denoised 3D seismic data with curvelet and wavelet transforms and the comparison showed that curvelet transforin is better than the wavelet transfoFin. In China,the application of the curvelet transform on seismic data processing is only in its early stage.Wang et a1.(2008)proposed a curvelet—based sparse constraint inversion method of primary—multiple separation.Wang et a1.f2009)studied a structure—oriented anisotropic Gaussian filter method and applied it to the denoising of 3D seismic time slices. In this paper,we apply the sparsity of the curvelet transform to express seismic data and describe the 316 1ssue of random seismic noise attenuanon as an optimaI problem with the 一norm.Then.we solve it with the iterative thresholding method proposed by Daubechies et a1.f2004).The comparison of the results of the iterative curvelet thresholding algorithm to that obtained by the traditional median filter algorithm FX decOnv0lutiOn algorithm,and wavelet thresholding algorithm shows that the iterative curvelet thresholding algorithm can achieve higher SNR and fidelity.Moreover.a control of curvelet direction on the result afler iterative curvelet thresholding is performed to further improve the SNR. ■-il he i sparsii dt y ot‘ cu rvel‘ et transtorm ‘ As a multi—scale analysis tool the wavelet transform can give an excel1ent non—linear approximation of a piecewise smooth 1 D function.which giveS it an enormous value for signal processing.GenerallY. servmg as an approximation method,it shows superior performance describing objects with point singularities in 1 D or 2D.When linear or surface singularities are taken into consideration.which is very common in practice.the wavelet transfoFin fai ls to give an optimal expression.It is not the perfect method to deal with image edges. The curvelet transform performs better describing edges and other abnormal points than traditional methods. It can achieve a precise rec0nstructiOn with very few toefficients.As the same as the wavelet transform, the curvelet transform is also a multi—scale transform including scale and location parameters as pans Of its structure.However,differing from the wavelet transforl-n, the curvelet transfo1"111 also includes direction parameters which enable it to have an appreciable advantage for directionality.Curvelet transforill iS also appropriate for direction and anisotropy.As a result,it has a more sparse reDresentati0n than the traditional Fourier transform and wavelet transfoFin for second order continuous and differentiable piecewise smooth functions. Because eurvelets can get a better sparsity than Fourier and wavelet transforrns and serve as a sparse base set in this paper,we apply it to random seismic noise attenuation with an iterative thresholding algorithm.Figure 1 shows reconstruction errors for one synthetic shot record comparison between Fourier,wavelets.and curvelets with different sparseness percentages.Based on this figure,we can draw a conclusion that the curvelet transform has the minimum recOnstruction error and is more suitable for seismic data than Fourier and wavelet transfoITnS. 亡。一l。己 ∞coomJ10 J0JJ∞o0NlIeEJ0z Wang et a1. 。.lxll】=∑ , jJ =argmin l 亍:Ai .1】Ax-Yll: 占 (1) In equation(1),Y represents seismic data with noise, A=C represents the inverse curvelet transform.x represents the curvelet coefficients.s represents an arbitrarily small value.and f represents the final output. The method for solving the optimization problems is to invert a set of curvelet coefficients with a minimum L 1 Sparse percentage(%) Fig.1 The comparison of reconstruction error for three norm.The set of coefficients satisfies the condition that the 2一nornl of the misfit between the reconstructed data and original data is leSS than a certain noise level and f. the output after the inverse curvelet transforlTl,represents the data after denoising. The P problem is accomplished by an iterative thresholding method with descending threshold values. domains,the horizontal axis represents the percentage of preserved coeficifents and the vertical axis represents the normalized coefficients recOnstrucliOn error. Iterative curvelet thresholding denoising inq n U data l n 2D dataSimilar to wavelet transform,the curvelet transform is also linear.Therefore.a curvelet transform of a signal with noise is equivalent to a combination of the same process on signal and noise separately.Based on this In practice.it is very slow and the large amount of seismic data and the complexity of the curvelet itself make it more difficult to solve.Elad et a1.(2005 1 replaced the constrained optimization problem P with a series of simple unconstrained ones: y—Axll + ̄llxl1. The solution of equation(2)depends on the parameter which determines the degree of influence that the l— norm modulus has on the -norm error,i.e.,determining characteristic of curvelets,the basic processing steps are: First,perform a multi.scale curvelet transform of the signal with noise after preprocessing;then,extract the curvelet coefficients for the signal at every possible scale and remove the noise curvelet coe踊cients at the same time;and last.perforill an inverse curvelet transfonn tO recover the denoised signa1. the emphasis of the C-norm over the 2 data misift.The solution of P is reached by solving Pi by descending The curvelet transform has the ability to concentrate the energy of seismic signals to a few large curvelet starting from九 =sup { :Ily—Ai sup{}represents the upper limit function. where coeficifents,while white noise remains by itself with the same amplitude after a transforlTl on any orthogonal basis. So the values of a signal’s curvelet coeficients are relfatively greater than that of noise.The noise energy is dispersed and its amplitude is low.Thmugh curvelet thresholding with a appropriate thresholding value,we can achieve the aim of We use the iterative thresholding method proposed by Daubechies et a1.(2004)to solve the P problem and the sequence of iterations is: x= (x+A (y—Ax)). where is the soft—thresholding function removing the noise and retaining the signa1.This method can give us an optimal estimation and has wide adaptation. Especially with seismic data.the“concentration”ability of wavelet transfoFill,whose energy concentrates on relatively more coe确cients.appears weaker than the curvelet ( ):=sgn(x)・max(0, IX 1—1 1) where sgn{}represents the sign function. In the calculation process,choosing the threshold is very important.In practice,first,wc apply the curvelet transform to the original noisy data and sort into descending curvelet coeficifents.Then。according to a pre.set curvelet coefficients percentage which was 317 transform.So curvelets have a better denoising result than wavelets for seismic data. To make use of curvelet’s sparsity for seismic data processing,we can describe the attenuation of random seismic noise as a constrained optimization problem: Seismic random noise attenuation preserved in the first loop to determine the threshold(the curvelet coefficients of the original data correspond to signal will be greater.On the contrary,the worse the denoising effect,the higher the signal fidelity.As a result.the calculation process threshold iS determined by the original data itself,number of iterations.and the pre— set preserved curvelet coemcients percentage in the first and last iterations. the preserved percentage)of the ifrst loop. We can choose a lower percentage in the first loop. for example 1%.because of the preferable sparsity of seismic data in the curvelet domain.After that. the threshold gradually steps down as the iterations proceed.We constantly update the curvelet coemcients by lowering the threshold until it reaches the original data curvelet coefficients which correspond to the pre set reversed curvelet coe临cients percentage of the last iteration.So iS varied with the number of iterations and final x iS the expected curvelet coefficients which relate to the signa1.The percentage of the 1ast iteration iS determined by the data SNR.We can choose a smaller Synthetic and field data examples Synthetic data examples The linear event modeI We choose SNR as the criteria to evaluate the quality of value when the data SNR iS high and otherwise choose a bigger value.A smaller value of the selected percentage denoising,which is defined as:SNR=20log 。【 IS,一S。 where So represents noise—free data and Sl represents noisy or denoising data.The SNR unit is dB. Trace number 80 100 provides a be ̄er denoising.However,the lOSS of effective Trace number 20 40 60 20 40 60 80 10O 0.2 0 2 0.2 童 C 童 C 墨0.4 墨 0.4 E 墨 o-4 0 6 0.6 0.6 0 8 0.8 O.8 (a) Trace number 0 (b) Trace number 20 40 (c) Trace number 60 80 1OO 0.2 0.2 0 2 耋 0.4 墨 0.4 O.4 0 6 0.6 0.6 0.8 0.8 0 8 (d) (e) (f) Fig.2 Original data(a),noisy data(b),result aRer median filter(c),FX deconvolution(d),wavelet thresholding(e),and iterative curvelet thresholding(f). 318 Wang at a1. The model that we use includes three linear events. Random noise is added to generate images with difierent SNR.Then.we perform four denoi sing methods parameters in several tests.From them.we see that the iterative curvelet thresholding method can obtain better results rSNR=9.1l 16 dB)in the lower SNR case.The big dip curvelets in the denoised profile caused by the original data with low SNR can be further attenuated by direction contro1. Figure 3 shows residual profiles of the four denoising known as median filter.FX deconvolutiOn.wavelet thresholding,and iterative curvelet thresholding on the data.The comparison of the results is shown in Figure 2. Figure 2 shows the original signal without noise(a), the low signal to noise ratio noisy data fSNR=一4.1 009 methods in the low SNR case.Obviously,the iterative curvelet thresholding method removes less signal and gets a higher fidelity.Also,we compare the SNR of the original data with different SNR and the data dB)(b),and the denoising results by the median iflter(c), FX deconvolution(d),wavelet thresholding method(e), and iterative curvelet thresholding method(n. From Figure 2 we see that the median filter does not after denoising using the four methods(see Table 1 1. From this table we see that the cOnventiOnal median ilter method improves the SNR by 4 to 6 dB.the fgive a satisfying result(SNR=1.7396 dB)in the X direction when the SNR is low.because there is still so much noise in the image.FX deconvolution does remove FX decOnv0lutiOn method improves the SNR by 6 to 1 2 dB.the wavelet thresholding method improves the SNR by 7 to l 1 dB.and the iterative curvelet thresholding method improves the SNR by l1 tO l 4 the random noise f SNR=6.2656 dBt1 well but it also attenuates the low amplitude signal rthe middle event) greatly,which is undesirable.The wavelet thresholding method attenuates a great deal of random noise f SN R dB.So.the iterative curvelet thresholding method is a very effective denoising method.The comparison with the median fiIter and FX dec0nvOlutiOn methods is =5.4665 dB1 however.the attenuation of the low amplitude signalfthe middle event)and the other defects fobscure phenomenon on the edge of events)make the result poor.The results are acquired by choosing the best consistent with the results obtained by Neelamani et a1. (2008). Table 1 SNR of diferent data and denoising method in dB Trace number 2O 40 60 80 1o0 20 Trace number 40 60 80 100 20 Trace number 40 60 80 100 Trace number 20 4O 60 80 100 O 2 0 2 0 2 0 4 0.4 0 4 0.4 O.6 0.6 0 6 O.6 O.8 0.8 0.8 0.8 (a) (b) (c) (d) Fig.3 Residual profiles of four methods,median iflter(a),FX deconvolution(b),wavelet thresholding(c),and iterative curvelet thresholding(d). Directional control Because of the multidirectiOnal characteristic of the curvelet transform,the data in the time—space domain can be described as having more elaborate coeficients ifn the curvelet domain than in the wavelet domain.When we perform an inverse curvelet transform to convert 3l9 Seismic random noise attenuation curvelet coefficients back into the time.space domain. the profile will grow a feW weak amplitude curvelets in each direction.As the original data’S SNR decreases. the curvelets that lower the SNR of the denoising profile to the large dip curvelets in time-space domain)to be zero; (3)Perforii1 the inverse curvelet transform and output the result. Figure 4 shows the directional control result of Figure willincrease fFigure 20.Therefore to further improve the SNR.these curvelets should be removed.Because the dip of the signals iS smallin our mode1.we can a ̄enuate these curvelets by directional control in the curvelet domain. 2f and the residual profile(the diference of Figures 2f and 4a).We see that the profile quality has obviously improved and directional control hardly harms the signa1.Table 2 lists the SNR comparison before and after directiona1 contro1.It iS clear that the data SNR can be improved 2 to 3 dB by directional contro1. Trace number The process of directional control is: (1 1 Perform the curvelet transforill to the denoised data; (2)Set the directional COCfficients(which correspond Trace number 40 60 80 1O0 40 60 80 100 0_2 0.2 童 墨 0.4 墨 0-4 0.6 0.6 0.8 O 8 (a) (b) residual profile of before and after directional contro1. Fig.4(a)Directional control result of Figure 2f(b) Table 2 SNR comparison before and after directional control in dB Note:When the signal dip is small,directional control can work well and most poststack data will satisfy the condition.When the signal dip range is wide,however,narrow range directional control will remove the signa1.In this case,we should magniyf the directional control range to preserve the signal as much as possible. Shot record modeI To test the denoising effect of the iterative curvelet thresholding method on hyperbolic events,we perform it on one shot record from a layer mode1.For noisy data with different SNR,we perform median filter, 320 FX deconvolution,wavelet thresholding,and iterative curvelet thresholding to attenuate the noise.Figure 5 shows one shot record(a),noisy data with SN R=0.2457 dB(b)and the results of the four denoising methods. Figure 6 shows the corresponding residual profiles.We Wang at a1. see that the median filter method iS not able to perfolITn satisfactorily because there iS stil1 SO much residual noise.As a consequence of not taking directiona1 rotation into consideration.the SNR after denoising iS 5.9928 identical to the FX decOnvolutiOn with SNR=1 0.1 293 dB after denoising.However,the wavelet transform blurs the events edges and reduces the resolution.Moreover, this method also removes some signal(Figure 6c).From Figure 5f,we see that the iterative curvelet thresholding method obtains the highest SNR at I 6.0840 dB.The dB.ThiS method removes more shot record Sjg.:al (Figure 6a).The FX deconvolution method can obtain a relatively accurate result.with a SNR Of 1 0.9927 dB edges of the events seem distinct and the lost signal、 mainly concentrated on events with Iarge curvature.is but it also removes much of the signal fFigure 6b).The result from the wavelet thresholding method iS almost Offset(km) .the smallest among the four methods(Figure 6d). Offset(km)一∞一。El l 2 3 .Offset(km) 一协一 ELL 3 -2 .1 0 1 3 .2 .1 0 1 2 O.5 0.5 互 m E I'-- 墨 1 1 5 1.5 (a) Offset(km) 2 .(b) Offset(km) 3 .2 .1 O 1 2 (c) Offset(km) 0.5 0.5 耋 1 E 1 5 1.5 (d) (e) (D Fig.5 One shot record(a),noisy data(b),result after median iflter(c),FX deconvolution(d),wavelet thresholding(e),and iterative curvelet thresholding(f). O 、一 、一 E I--- 0 E I-- E (b) (C) (d) Fig。6 Residual profiles corresponding to the four methods of Figure 5. Median filter(a),FX deconvolution(b),wavelet thresholding(c),and iterative curvelet thresholding(d). 32l Seismic random noise attenuation Field data examples Figure 7 iS a reaI marine near offset profile.Although thresholding profiles.Obviously,the FX deconvolution seriously attenuates the scattered signals which maybe contain Very useful information in practice.AS far as the requirement of high fidelity in seismic data processing iS concerned.it seems not a suitable method. The residual profile denoised by the iterative curvelet thresholding method shows a feW scattered signals and the result iS relatively satisfactory. there doesn’t seem to be much noise in the profile, the events appear blurred.So we need to perform denoising.In addition,there are also some scattered waves(arrowed in the residual profile of Figure 9, which may contain very important information.Figure 8a shows the denoising result from FX deconvolution (after massive parameter testing).We see that the profile quality is improved and the scattered signal in the original data iS also attenuated.Figure 8b shows the denoising result of the iterative curvelet thresholding 0.2 Shol number 100 150 200 method.It iS clear that most of the random nojsc has been attenuated and the quality of the profile iS obviously improved.Far more important.the events become distinct and continuous and the scattered signal 0.6 C 罢0.4 in the original data has been preserved.So the method jS able to meet our high SNR and fidelity demand jn seismiC data processing.Figures 9a and 9b showS the residual FX deconvOlution and iterative curvelet Shot number Fig.7 Marine near offset profile Shot number 1O0 150 200 1OO 150 200 0.2 0.2 詈0.4 E 耋0.4 0.6 O.6 (a)FX deconvolution (b)Iterative curvelet thresholding Fig.8 Profiles after denoising by diferent methods. Shof number 100 150 200 Shol number 100 15O 200 0 2 0.2 奎 童 罂0.4 耋0.4 I’—-’ 0 6 0.6 (a)FX deconvolution residua Fig。9 Residual profiles 322 (b)IteratJ ve curvelet thresholding residual Wang et a1. Conclusions The curvelet transfornl,which belongs to the category of sparse functional representatiOn theory,is a new method of multi scale geometric analysis.Along with the advent of the fast discrete curvelet transfelrm algorithm,it is being applied more widely in seismic data processing.From this paper,we come to some cone Iusions (1 1 For seismic data,the sparsity of the curvelet transform which has both multi.scale and multi— directional characteristics performs better than the wavelet transform having only multi—scale and Fourier tran sform.The iterative thresholding method i s established based on the constrained sparse inversion. So the more sparsity the inverted parameters have,the higher quality result can be obtained.In theory,the curvelet—based iterative thresholding method will get the best results; (2)The comparison of the results obtained by traditional median filter,FX deconvolution,thresholding wavelet,and iterative curvelet thresholding methods has shown that iterative curvelet thresholding has the better resuits in random noise attenuation.It can ireprove SNR sufifciently and give higher signal fidelity at the same time.So it can be regarded as an effective tool to attenuate random noise; (3 1 Because of the multi.directional characteristic of the curvelet transform,when performing the inverse curvelet transform to curvelet coefficients which have been processed,small curvelets will emerge in every direction.We can perform directional controI to the denoised data to remove these curvelet.1ike artifacts and the SNR is further improved; f4)For the defect that iterative curvelet thresholding method removes some large curvature signals in hyperbolic events,in practice,we can perform NMO to the gather data before denoising and that will give better results.Furthermore,because of its huge redundancy the curvelet transform has a great computational cost。 So faster and more effective a1gorithms to solve the curvelet—based optimization problem using the L 1一nornq need to be developed; (5 1 It is important to further study the directional control of the curvelet transfornl to remove the coherent noise.Because,in essence,the directional contro1 in the curvelet domain is similar to dip filtering in the F—K domain but the direction information in the curvelet domain can be divided more elaborately.So dip filtering in the F—K domain can be completely achieved in the curvelet domain and the result wil1 also be better. ACknOwIedgementS The authors thank Felix J.Herrmann for the use of a Python interface to Unix-・like pipe・-based linear operators.We also thank the authors of CurveLab for providing their codes on the Web. References Candbs,E.,and Donoho,D.,1 999,Curvelets:A surprisingly effective non。adaptive representation of objects with edges.TN:Vanderbilt University Press,USA. Cand6s,E.,and Demanet L.,2002,Curvelets and Fourier integral operators:Compte Rendus del’Academie des Sciences.Paris,Serie I,336,395—398. Cand6s,E.,and Guo,F.,2002,New multiscale transforms, minimum total variation synthesis:Applications to edge— preserving image reconstruction:Signal Processing,82, l519一l543. Candes,E. and Donoho,D.,2004 New tight frames of curvelets and optimal representations of objects with C2 singularities:Comm.Pure App1.Math.,57(2),2 1 9—266. Cand6s,E. Demanet L. and Donoho,D.,2006,Fast discrete curvelet transforms:SI.AM Multiscale Modeling and Simulation.5.86 1—899. Daubechies,I.,Defrise,M.,and De Mol,C.,2004,An iterative thresholding algorithm for linear inverse problems with a sparsity constraint:Comm.Pure and Appt.Math..57.14l3一l457. E1ad,M.,Starck,J.L.,and Querre,P.,2005,Simulatneous cartoon and texture image inpainting using morphological component analysis(MCA):App1.Comput.Harmon. Anal,19.340—358. Hennenfent,G.,and Herrmann,F.J.,2004,Three.term amplitude.versus—of.fset fAVO)inversion revisited by curvelet and wavelet transforms:74th Ann.Internat. Mtg.,Soc.Exp1.Geophys.,Expanded Abstracts,21 1 2l4. Hennenfent,G.,and Herrmann,F.J.,2008,Simply denoise: Wavefield reconstruction via j ittered undersampling: Geophysics,73(3),V 1 9一V28. Hennenfent,G.,van den Berg,E.,Friedlander,M.P.,and Herrmann,F.J.,2008,New insights into one—norm 323 Seismic random noise attenuation solvers from the Pareto curve:Geophysics,73(4),23— 26. Wu,F.P.,2008,Study on curvelet transform and lts application in image processing:Masters Thesis,Jinan University. Zhang,H.L.,Zhang, C.,and Song,S.,2008,Curvelet Herrmann,F.J.,and Verschuur,E.,2004,Curvelet—domain multiple elimination with sparseness constraints:74th Ann.1nternat.Mtg.,Soc.Exp1.Geophys.,Expanded Abstracts.1 333—1 336. Herrmann,F.J.,Wang,D.L.,and Hennenfent,G.,2007, domain—based prestack seismic data denoise method:Oil Geophysical Prospecting,143(5),508—5 1 3. Zhao,X.,2007,Research and application of image denoising method based on curvelet transform:Masters Thesis,Shandong University of Science and Technology. Seismic data processing with curvelets:A multiscale and nonlinear approach:77th Ann.1nternat.Mtg.,Soc.Exp1. Geophys.,Expanded Abstracts,2220—2224. Herrmann,F.J.,and Hennenfent,G.,2008,Non.parametric seismic data recovery with curvelet frames:Geophysical Journal International,173(1 1,233—248. Huang,W.,2007,Research on curvelet transfoFin and its applications on image processing:Masters Thesis,Xian Deli Wang,Professor at Jilin University.He received a BS(1 9951 and an MS University of Technology. Huub,D.,and Maarten,V.,2004,Wave.character preserving pre—stack map migration using curvelets:74th (1 998)in Applied Geophysics from Changchun College of Ann.Internat.Mtg.,Soc.Exp1.Geophys.,Expanded Abstracts.96 l一964. Neelamani,R.,Baumstein,A.I.,Giliard,D.G.,and Hadidi, M.T. and Soroka,W.L.,2008,Coherent and random Geology and a PhD(2002)in Geophysics from the College Of GeOexDlOratiOn Science and Technology,Jilin University. From 2006 to 2007 ng was noise attenuation using the curvelet transform:The Leading Edge,27(2),240—248. Wang,D.L.,Saab,R.,Yilmaz,O.,and Herrmann,F.J., a visiting associate professor at the Seismic Laboratory for Imaging and Modeling, Department of Earth and Ocean Sciences,University of British Columbia.Canada.His research interests include high.resolution seismic data processing.seismic modeling,and parameter inversion in anisotropic media.Now his research focuses on primary—multiple separation,seismic denoise,and seismic inversion for anisotropic media.He is member of SEG. 2008,Bayesian wavefield separation by transform— domain sparsity promotion:Geophysics,73(5),A33一 A38. Wang,W.,Gao,J.H.,Li,K.,Ma,K.,and Zhang,X.,2009, Structure.oriented Gaussian filter for seismic detail preserving smoothing:Image processing:1 6th IEEE International Conference.60 1—604. 324 应用地球物理(英文版) 2010年第7卷第4期 中文摘要 南海重磁异常特征及火成岩分布//Grayity and magnetic anomalies field characteristics in the South China Sea and its application for interpretation of igneous rocks,李淑玲 ~,孟小红 ~,郭良辉 ~, 姚长利 ’。 ,陈召曦 ~,李和群 ,Applied Geophysics. 7(4),P.295—305. (1、地质过程与矿产资源国家重点实验室(中国地 质大学,北京),北京100083; 2、地下信息探测 技术与仪器教育部重点实验室(中国地质大学,北 京),北京100083;3、中国地质大学(北京)地球 物理与信息技术学院,北京100083) 摘要:南海火成岩油气藏具有广阔的勘探前景,综 合利用地球物理方法圈划与识别火成岩体、研究火 成岩分布是火成岩油气藏研究的基础。针对南海重 磁场特征,采用低纬度、变倾角化极技术进行了磁 异常化极处理,利用优选延拓方法实现重磁异常分 离并提取南海海域浅部火成岩重磁异常信息,利用 磁异常三维相关成像给出南海火成岩的三维空间等 效分布,在重磁梯度突出局部异常边界信息的基础 上,通过梯度加权的重磁相关分析勾画不同类型火 成岩的平面展布,火成岩的分布特征显示出受地壳 深部结构及断裂构造的控制与影响。 关键词:南海,重磁场,低纬度化极,优选延拓, 火成岩分布 南海东北部构造及块体运动指向的地震相响应// Seismic facies response of Tectonics and Direction of block movement in the northeastern South China Sea,陈洁 ,钟广见 ,刘少华。,Applied Geophysics, 7(4).P.306—314. (1.广州海洋地质调查局,广州510760;2.中科院 地质地球物理所,北京1 00029) 摘要:南海的构造与演化与资源环境等关系密切是 本文研究重点。本文针对南海的东北部构造及其块 体构造方向,利用所采集的区域地震剖面,通过解 析地震相与构造及其演化的关系,提出以下观点: (1)构造分区特点明晰,可划分为五个不同构造单 元,构造单元之间既有联系又相互独立; (2)南海 沉积盆地无论表现为拉张一弱挤压一强挤压的何种 构造格局,其区域构造应力场是统一的: (3)首次 发现反射地震剖面上显示出两个浅俯冲点。每个块 体构造层呈手风琴风箱式折曲并向东聚敛,体现沉 积盆地从发育、成长、结束、消亡不同阶段在南海 的表现,其块体俯冲方向以及块体包络区域性倾伏 方向均与区域应力场方向一致。 关键词:南海扩张,浅俯冲,单方向漂移,挤压变 形,手风琴风箱式地震相外形 Curvelet阈值迭代法地震随机噪声压制//An iterative curvelet thresholding algorithm for seismic random noise attenuation,王德利 , 仝中飞 ,唐晨 ,朱恒 ,Applied Geophysics,7(4),P.315—335. (1.吉林大学地球探测科学与技术学院,长春 130026:2.中海油研究总院,北京100027) 摘要:本文将近些年发展起来的多尺度分析技 术一一Curvelet变换与求解优化反演问题的阂值迭 代法相结合,研究了基于Curvelet变换的阈值迭代 法在地震数据随机噪声衰减中的应用。充分利用了 Curvelet变换对地震数据表示的稀疏性,提出将地 震数据随机噪声压制问题转化为基于Curvelet稀疏变 换的Ll范数最优化问题,并采用前人提出的阈值迭 代法求解。通过与常规的中值滤波、FX反褶积和小 波阈值法去噪方法对比,理论合成数据和实际数据 试算表明,Curvelet阈值迭代法去噪法具有优势,该 法不仅能够获得较高的信噪比,而且对有效信号的 损失较小。为充分利用Curvelet的多尺度、多方向特 性,提出了在Curvelet阈值迭代法去噪结果的基础上 再进行方向控制,进一步提高了数据信噪比。 关键词:Curvelet变换,阈值迭代法,随机噪声衰减 基于第二代Curvelet变换的面波压¥,J//The surface wave suppression using the second generation curvelet transform,郑静静 ,印兴耀 ,张广智‘,武国虎 ,张 作胜 ,Applied Geophysics.7(4),P.325—335. (1.中国石油大学(华东)地球资源与信息学院, 山东东营257061;2.胜利物探公司国际部,山东东 营,257084) 399