Wavelet analysis is a relatively new technique that is an important addition to standard signal analysis methods.
Unlike Fourier analysis that yields an average amplitude and phase for each bannonic in a dataset. the wavelet transform produces an "instantaneous" estimate or local value for the amplitude and phase of each bannonic, This allows detailed study of nonstationary spatial or time-dependent signal characteristics.
The wavelet transform is discussed, examples are given, and some methods for preprocessing data for wavelet analysis are compared. By studying the ~on ofYanai waves in a reduced gravity equatorial model, the u~fulness of the transform is demonstrated. The group velocity is measured directly over a finite range of wavenumbers by examining the time evolution of the transform. The results agree weD with linear theory at higher wavenumber but the measured group velocity is reduced at lower wavenumbers, POSSIoly due to interaction with the basin boundaries.
1. Introduction
a. The wavelet transform
The wavelet transform has shown promise in a diversity of scientific fields, but to date it has not been much used in the oceanic and atmospheric sciences.
In part, this might be due to a lack of material discussing practical aspects of the technique. This article therefore includes an introduction to the wavelet transform as a tool for data analysis. For brevity, we shall confine our discussion to the transform of a scalar seriesf(t).
Wavelet analysis is based on the convolution off(t) with a set of functions gab(t) derived from the translations and dilations (and rotations in higher dimensions) of a mother wavelet g( t), where
a(>O) and b are real. Any set of functions gab(t) constructed from ( I ) and meeting the conditions outlined below are called wavelets. The convolution off(t) with the set of wavelets is the wavelet transform (WT)
This is known as the continuous wavelet transform since a and b may be varied continuously. Translation parameter b corresponds to position or time if the data is spatial or temporal, respectively. Dilation parameter a then corresponds to scale length or temporal period.
Equation (2) expands a one-dimensional time series into the two-dimensional parameter space (b, a) and yields a local measure of the relative amplitude of activity at scale a at time b. This is in contrast to the Fourier transform that yields an average amplitude over the entire dataset. Note, we have avoided the use of the words "wavelength" or "frequency" in our description of the WT. Though wavelets have a definite scale, they need not bear any resemblance to Fourier modes (sines and COSines). However, a co1TeSpondence between wavelength and scale a sometimes can be achieved, as discussed in section 3.
To see the limitation of standard Fowier analysis and the incentive for the development of wavelet analysis, consider the time series in Fig. la that changes frequency halfway through the measurement. Compare that to the signal in Fig. 1 b that is generated from the simultaneous presence of both frequencies. These two very different signals yield similar power spectra, shown in Figs. 1 c,d, both being dominated by the same two peaks. Without prior knowledge, it would be difficult to know which signal produced which spectrum, since information on signal evolution is lost during Fourier analysis. Variations of the Fourier transform have been used (e.g., Gabor 1946) in attempts to overcome this limitation, but have met only with qualified success.
The WT produces "instantaneous" coefficients and
therefore can yield information on the evolution of nonstationary processes.
The development of wavelet analysis began relatively recently with Morlet (1983). A later collaboration (Grossman and Morlet 1984) produced the continuous wavelet transform in one dimension based on the set of translations and dilations. Meyer ( 1985) extended the technique to n dimensions, and Murenzi (1989) included rotation. Three references are used for the introduction to wavelet analysis: Farge ( 1992), Ruskai et al (1992), and Combes et al (1989). Our goal in this article is not to give a complete review of the wavelet transform but to aid in its introduction to oceanographers and meteorologists. The references contain details on the history and uses of wavelets in a variety of scientific applications.
In the following we continue with the introduction to wavelet analysis. In section 2, a few methods for preprocessing data for wavelet analysis are discussed Two simple methods from Fourier analysis are tried and shown to be inappropriate as they distort the end regions of the WT. Another method, based on buffering the ends of the data with additional points, is shown to yield better results. Section 3 examines the dispersion ofYanai waves and demonstrates how quantitative information not available from Fourier methods can be obtained using the WT.
b. Wavelet selection
To be a mother wavelet both fonnally and in practice, g(t) must have the following properties.
(i) It must be a function centered at zero and in the limit as I t I→+∞, g( t) →0 rapidly. This condition
produces the local nature of wavelet analysis, since the coefficients T g{ b, a) are affected only by the signal in the cone of influence (COI) about t = b. In practice, the radius of the COI is the point I tl ='c beyond which gab{x) no longer has significant value. Usually, 'c cx: a, giving rise to conelike structures in the WT in certain cases. The COI of the endpoints is an inlportant consideration and will be discussed further in section 2.
( ii) Also, g{ t) must have zero mean. Known as the admissibility condition, this implies the invertability of the WT. That is, the original signal can be obtained from the wavelet coefficients using
c. Wavelet algorithms
There are several algorithms to implement a WT, and only two will be discussed here. The references detail other methods under development, particularly those involving discrete wavelets. Here, we discuss the continuous transform.
The simplest method is direct numerical integration.
Knowingf(t) and g(t), one can compute the transform at arbitrary points in parameter space using a discretized form of ( 2 ). The drawback of this technique is that it is time consuming. If one integrates over 0 < a ~ l and 0 < b ~ J, the integration time goes as lJ2.
An alternative is to exploit the convolution theorem and do the WT in spectral space:
This allows the use of optimized fast Fourier transform (FFT) routines and results in a much faster transform, with CPU time going as IJI~J. To use this method, g( cu) should be known analytically and the data must be preprocessed to avoid errors from the FFT algorithms. In particular, the discrete form of (5) will produce an artificial periodicity in the WT if f( t) is not periodic. This is demonstrated in the examples herein, and different methods for dealing with the problem are discussed Issues of aliasing and bias in FFT routines are well known and need not be discussed here.
2. Examples of wavelet analysis
The WT of the signal in Fig: la is a classic example in wavelet literature and will be examined with four different data preparation methods. All follow (5) and use the Morletwaveletg(t) = elcte-12/2. Method I does not involve any preparation of the data, and the transforDl is computed without considering the properties of FFr routines. The result is presented in Fig. 2, with the amplitude and phase shown separately. As in all the transforDls shown here. the vertical axis is the inverted scale Q, so the corresponding wavenumber increases on the vertical axis. The horizontal axis is time
b. The mesh in all the plates indicates the COI of the endpoints. Consider the first endpoint t = O. For b < r c, the wavelet has significant values at t < 0 where there is no signal. Hence, there is a steady degradation in the WT as b approaches O.
Regions where I T g( b, a) I are large indicate high correlation between the data and the wavelet. In Fig.
2, the modulus clearly indicates the abrupt change in the frequency of the signal by the shift of the large coefficients to a different scale. The phase indicates the cycling of the signal from -1r to 1r and allows for the location of wave crests in the signal. One can count the number of waves in a simple signal such as the one shown here. The branchings at larger a are due to the larger wavelets (large a) detecting more than one cycle.
Note the convergence of the phase lines toward higher frequency (smaller a) in the middle of the transfonn.
This indicates when the frequency shifted, and generally occurs at any singularity (sudden changes in frequency or phase) in the signal. Interpretation of phase plots for more complicated signals can be nontrivial.
In the WT phase, and to a lesser extent in the modulus, activity is indicated at many scales, though the signal is locally monochromatic. Although there is a "best" choice for the scale (a = O(» in the modulus, the nonzero correlations between the wavelets and data produce nonzero transfonn values at scales away from O(). This can lead to confusion when trying to determine which scales are present in the data; one method for dealing with this problem is discussed in the next section.
Note in Fig. 2 the curvature of the modulus at the end regions. The line of maximum wavelet coefficient bends to meet artificial periodic boundary conditions, falsely indicating frequency changes at the beginning and end of the signal. This demonstrates the generic problem when using ( 5 ) on nonperiodic data-the WT can induce a periodicity into the transfonn, greatly distorting the infonnation at the end regions. This is a critical problem when the initial and final signal characteristics are very different.
In an attempt to impose periodicity on the signal, method II preprocesses f( t) with a cosine window
N is the total number of data points. The resulting WT no longer has the false end regions, because it has effectively lost those areas. as seen in Fig. 3, which is clearly an unacceptable effect. A window with steep cutoffs would not eliminate as much data but would produce only a small correction to the distortion of the end regions.
Method III involves detrendingf(tJ and removing the mean, a standard procedure when computing FFfs.
The result is shown in Fig. 4. Again, errors in the wavelet coefficients appear as they did with method 1.
Though methods II and III are common methods for preparing data for Fourier analysis, they are inadequate for wavelet analysis.
Method IV, though developed independently by the authors, has already been investigated by other researchers (Bamier 1992, personal communication).
The data f( Ik) is buKered on either side with a tail that goes to zero, as indicated in Fig. 5. After the WT is complete, the regions corresponding to the tails are discarded, yielding Fig. 6. This method eliminates the problem of non periodic data and does not introduce significant distortions of the end regions. The length of the buffers are chosen empirically. When the buffers are too short, distortions as in Fig. 2 are found but lessen as the buffer length is increased. Various forms of the buff~ can be chosen. Here, we used a very simple form, but suggest for future applications matching not just the endpoint values but their derivatives as well. Even better would be a form that matches the endpoint frequency, amplitude, and phase, which is conceivable for simple signals, and would arguably eliminate most of the problems ~ted with the COI at the endpoints. Of course, if the end regions are unimportant, their WT can be ignored.
A frequent comment about wavelet analysis is that it is difficult to obtain more than qualitative information (such as detecting the frequency change in the preceding example). This may be so in some cases; in others it is possible to obtain quantitative information that would be difficult if not im~ble to obtain by traditional Fourier analysis. As a practical illustration of wavelet analysis we will directly measure the dispersion ofYanai waves in an ocean model. The results indicate general agreement with linear theory except near the eastern boundary where the wave propagation appears to slow. The spectral ocean model used to create the data is described by Kelly ( 1992).
3. Dispersion of Yanai waves
Instability waves in the equatorial regions of the world's oceans have been under study since the mid1970s. Interest began following the observations made
during the GARP (Global Atmospheric Research Project) Atlantic Tropical Experiment ( GATE) in the summer of 1974 (Duing et al 1975). Occurring in a narrow frequency band, with periods around 25 days and zonal wavelengtlis of about 1000 kIn, their structure has been shown to be dynamically similar to Yanai waves (mixed Rossby-gravity waves) both in the observations (Weisberg and Horigan 1981; Tsai 1990) and in the numerical models (Cox 1980; Kindle and Thompson 1989; Woodberry et al 1989). The waves propagate westward and upward with a group velocity that is eastward and downward ( Weisberg et al 1979; Cox 1980). The zonal phase and group velocities are about 33-73 cm s-J and 16 cm S-I, respectively (Weisberg et al 1979). Similar oscillations were observed along the equatorial front in the eastern tropical Pacific using satellite imagery (Legeckis 1977).
As a result of these observations, it was hypothesized that they were generated by a meridional shear instability between the westward-flowing South Equatorial Current (SEC) and the eastward-flowing North Equatorial Countercunent(NECC)(Pbilander 1976,1978).
A more recent study by Weisberg and Weingartner ( 1988) showed that the generation region of the waves is slightly to the south of the boundary between the SEC and the NECC. Nevertheless, there is a general agreement that for the eastern tropical Atlantic and Pacific the waves are caused by a latitudinal shear instability. A different mechanism bas been proposed for the western Indian Ocean where the waves are more likely to be boundary forced (Kindle and Thompson 1989; Woodberry et at 1989; Moore and McCreary 1990 ). A hypothesis that the waves in this region result from cross-equatorial wind stress bas been presented, with the preferred frequency selection around 2S days being due to the dispersive properties ofYanai waves (Kelly 1992). We now discuss how the wavelet transform was useful in supporting this idea .
The linear dispersion relation for Yanai waves (Moore and Philander 1977) is
where k is the wavenumber, q is the frequency, and c = (g'H)1/2, From (7), Yanai waves with westward phase speed have eastward group velocity with Cg in- creasing as k - 0 - so the longer waves propagate faster.
This is reflected in the pattern of the measured antisymmetric height field
which shows a wavelength increasing eastward across the basin. Since the h(x, y) field for a Yanai wave is antisymmetric about the equator, measurements of h,4 help isolate the relatively small-amplitude Yanai waves from the large-amplitude symmetric structures (e.g., Kelvin waves). Cross-longitudinal measurements of h,4(x, y = +3°) are shown for 80,140, and 200 model days in Fig. 7. It is clear that wavenumber depends on longitude and that the waves are propa~ting energy eastward. A similar wave pattern was found in a reduced gravity model by Kindle and Thompson ( 1989).
Though power spectra would indicate which wavenumbers are present, no understanding of the spatial characteristics could be achieved.
In contrast, the WTs clearly show the structure ofthe wave field and its development in time. Figure 8(see p. 2861 ) shows the modulus of the transform at80,140, and 200 days, which indicates the waves as atongue of high values stretching progressively fartheracross the basin.The vertical coordinate is the Fouriermode that corresponds to wavelet scale (Appendix ).To obtain quantitative results from the transforms,the modulus is scanned along each scalc for local max-imum, revealing the “spine of the dispersion tongue.( The result of a scan over the entire transform issometimes called a "skeleton." ) Since it is at these localmaximum that the maximum correlation of the wave-let and f) occurs, we assume it is the motion of themaximum points that best indicate the dynamics.Thepoints in Fig. 9 are the maximum points correspondingto the dispersive tongues in the modulus of severalwTs Using these results, we can now directly obtainthe range of scales involved and the distance that theyhave propagated. Measuring the displacement of thespine at cach scale over time yields the group speed(a). Note in Fig. 9 that the total displacement of thelines is greater at the smaller wavenumbers,indicatinglarger group velocity.
The relation between the wavenumber k of the Fourier modes and the scale a is not fully understood and may not be useful with irregular wavelets. However, in the case of the Morlet wavelet a relation can be obtained by transforming a monochromatic signal as dis
cussed in the Appendix. Such a transform would have a structure as in FIg. 6 but with a single maximum of the modulus in a. The line of maximum correlation yields a conversion from wavelet scale to Fourier wavelength. For the Morlet wavelet with c = 5 (used in the WT in this article), the wavelength is given by A ~ .l.2a.1lUs allows us to compare the measurements of propagation speed with the theoretical speed at conventional wavenumbers.
The results of the measured Cg are compared with theoretical values in Fig. 10. There is general aareement with the linear theory in the range &om approximately k = -0.005 rad km-1 to k = -0.02 lad kin-I. (Note, we deal only with k < 0.) For smaller k the measured Cg diverges from the theoretical value, possibly due to the finite size of the basin. When the maxima for all the days under study are plotted together, one can see a sudden decrease in the propagation speed of wavelengths 8bove 1600 kin when they near the eastern boundary. This scale corresponds to the value of k where the drop-off in Cg occurs. To test this we ex.
paDded the basin domain from IS 000 to 22 SOO km, while keeping the same number of modes in the spectral model. In the longer domain there is no drop-off of the measured group velocity ~t lower wavenumber since the longer waves do not have time to encounter the eastern boundary during 200 model days.
As mentioned in Kelly ( 1992), in the western region of the basin Yanai waves exist only in a limited range of k. Examining Fig. 9, one can see a gradual flattening of the curves in the western basin as time progresses, indicating a narrowing of the range of wavenumbers.
The waves remaining in the west are several hundred kilometers in wavelength, as observed in the Indian Ocean. This suggests that the narrow range of periods found in the ocean may be due to the dispersion of Yanai waves; the longer, faster waves leave the western region after roughly 100 days.
4. Summary The wavelet transfonn promises to be a useful tool in oceanography and meteorology. For the purpose of data analysis, the continuous transfonn in spectral space is useful and efficient. However, standard methods for preprocessing data for Fourier analysis are insufficient for wavelet analysis. The best method examined here is that of buffering the ends of the signal with points that smoothly go to zero. The region of the transfonn corresponding to these points is then discarded after the transfonn. Without this buffering, a signal whose properties are different near its ends will result in a WT that has been forced to periodicity at all scales through a distortion (in some cases severe) of the end regions. The greater the aperiodicity of the signal, the greater the distortion.
We demonstrate the usefulness of the WT by examining the dispersion ofYanai waves. The transfonn modulus clearly reveals the propagation of the different wavelengths across the basin. By scanning the modulus, Cg( k) is measured directly and shown to agree with linear theory, though there is a reduction in propagation speed of the longer wavelengths. The narrowing of the range of wavelengths in the western region observed supports the hypothesis that the narrow range of frequencies observed in the western equatorial oceans is a consequence of Yanai wave dispersion.
These results could not be obtained using standard Fourier techniques.
The use of wavelets goes beyond simple data analysis, and a full discussion is beyond the scope of the paper, but a couple of references might suggest possible future applications. An extension of wavelet analysis, called multiresolution analysis. adds the scaling /unctions
The addition of the scaling functions to the set of wavelets allows for the decomposition of functions that are not square integrable.
Frequently used with discrete wavelets [following the notation of Farge (1992)] \lIifix) that are defined by \lIij(x) = 2j/~2jx - i), the reconstruction equation forf(x) is then
Of more direct interest to the reader might be recent attempts to model two- and three-dimensional fluids in wavelet space. A representation of the three-dimensional Navier-Stokes equation in wavelet space is given by Meneveau ( 1991 ), with the emphasis on reducing the number of degrees of freedom necessary for modeling fluid turbulence. Zimin (1981), antecedent to the formal development of wavelets, studied the Navier-Stokes equation in terms of basis functions localized in space and time. In one of the few papers based on the work of Zimin that has been translated into English, Aristov and Frick (1988, 1989) present a study of convection in a rotating fluid like previous hot topics in science, the WTshows much promise but solid results have been difficult to achieve. The works of Meneveau (1991) and Zimin ( 1981) are far from final results; several more years are probably needed for those subjects to mature.
However, for applications such as described in section 3 of this article, the WT is sufficiently developed that it can now be used with confidence as a tool in the oceanic and atmospheric sciences.
Acknowledgments. SDM is supported by the ONR Distinguished Educator Postdoctoral Fellowship to Dr.
James O'Brien. The theoretical work of the Mesoscale Air-Sea Interaction Group is CUITently being supported by the Physical Oceanography Branch of the Office of Naval Research and the Ocean Processes Branch of NASA. Discussions on wavelet analysis with Dr. Bernard Barnier were useful. The color plates were created using the Ferret package, provided by Dr. D. E. Harrison and Steven Hankin. Thanks also to Mark Verschell for his on-site help with the Ferret routines.
扫一扫
354
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
