1. Introduction

Part of the tropical atmospheric variability is organized in quasiperiodic oscillations such as the El Niño–Southern Oscillation (ENSO), the intraseasonal oscillation (ISO), or equatorial easterly waves, giving a possibility to understand, simulate, and predict some elements of tropical climate at the corresponding timescales. These tropical oscillations are, however, difficult to detect and characterize in a simple and comprehensive way because they are generally intermittent and mixed with the atmospheric variability on other timescales. The intermittence of these oscillations leads to a variety of spatial structures (e.g., amplitude distribution and lag between regions) and temporal characteristics (e.g., duration and dominant timescale) from one event to another. These properties characterize a particular mode that exists for a particular date. A given type of oscillation, such as the ISO, may be seen as a succession of such modes with time. In most cases, this variability must be considered typical of the oscillation and not a random noise hiding a hypothetical ideal structure. For this reason, even if a mean spatial structure may give useful condensed information, the analysis of this variability is also required to better understand the physical origin of the phenomenon.

Conventional methods, such as composite or cross-spectral analyses, are generally designed to extract only the mean structure of a given oscillation. For intermittent oscillations, this mean structure may be difficult to interpret because it is constrained by the strongest events that are not necessarily representative of the ensemble of events. If spatial distributions of amplitude and lag are very variable from one oscillation to another, this mean structure also may be influenced strongly by the reference region used to define the dates or the different phases of the oscillations for composite analyses or to characterize phase lags for cross-spectral analyses. Time–longitude or Hovmöller diagrams [see, e.g., Nakazawa (1988) or Rui and Wang (1990) for the intraseasonal oscillation] give useful information on the variability of a given oscillation. These diagrams give condensed information on the time evolution of the longitudinal structure (amplitude, propagation) of the perturbations. However, for these time–longitude diagrams the latitude band over which the signal [e.g., wind divergence, outgoing longwave radiation (OLR)] is averaged also must be specified. This may sometimes confuse the interpretation if the latitudinal structure of the perturbation is variable from one oscillation to another or if the main structure is not purely longitudinal.

Different structures or modes may be extracted by computing empirical orthogonal functions (EOFs). For example, interannual variability over the Pacific (Barnett 1985, 1991), teleconnection patterns in midlatitudes (Wallace and Gutzler 1981), and the ISO (Lorenc 1984;Lau and Chan 1985, 1986; Gutzler and Madden 1989) were analyzed using EOF analysis. The EOF analysis separates the time–space structure of a geophysical field according to a basis of orthogonal eigenvectors. Each eigenvector gives a spatial distribution of correlation or amplitude that represents the spatial structure of the mode. For complex EOFs (CEOFs), the eigenvector also gives the spatial distribution of relative phases (or lags), informing on the propagative or stationary character of the oscillation (Wallace and Dickinson 1972; Barnett 1983). However, for intermittent oscillations, different structures that are not necessarily orthogonal may be present in the time series. An analysis of long time series, compared with the timescale of the studied oscillation, may thus provide the first EOF (corresponding to the largest eigenvalue) with either an average spatial structure not representative of any of these modes or a spatial structure representative only of some of the strongest modes of the series. In many cases, because of the orthogonality constraint, the remaining EOFs (e.g., second, third, etc.) are not easily interpretable. This shortcoming of the EOF analysis exists for different types of EOFs, including the CEOF and the extended EOF (Weare and Nasstrom 1982). To avoid this problem, some attempts have been made to rotate the eigenvectors (rotated EOF) following different criteria such as the Varimax technique (Horel 1984). However, this approach is not always efficient, and the choice of a rotation solution is “only one of an infinity of alternative solutions” (Horel 1984) that give the impression that any spatial structure can be obtained. It is thus of interest to develop an approach that makes it possible to test the meaning of such a mean spatial structure.

This paper proposes an approach, the local mode analysis (LMA), that extracts and characterizes the different modes of a particular type of oscillation. The LMA breaks the condition of orthogonality between modes that succeed one another in time by applying a series of successive CEOF analyses on relatively short time windows of the whole time series. The choice of short time windows, analogous to the expected duration of a mode, avoids the mixture of different modes and thus gives a far more straightforward interpretation of the CEOF results. We select the CEOF analysis because it provides the distribution of amplitude and of relative phase on a single eigenvector. The CEOF is thus very efficient to describe any phase relationship, including propagative or stationary features, by a single component (a real instead of a complex EOF analysis requires two eigenvectors to describe a propagative signal). The proposed approach makes it possible to detect and characterize the most persistent modes of a time series. Note that the LMA, like the other analysis methods mentioned above, is based on the processing of a given time series representing the fluctuations of a given parameter in a given region and for a given timescale. In some cases, what we call different modes of an oscillation may actually characterize different types of oscillations with different physical origins that coexist in the same frequency range.

Section 2 of this paper contains a description and a justification of the LMA procedure. In section 3, the basic properties of this method are illustrated using an idealized signal. In section 4, the ISO of deep convection is studied using the LMA approach. Advantages and shortcomings of the LMA technique and its field of potential application for the study of various geophysical phenomenons are discussed in section 5.

2. Analysis procedure

The analyzed time series s(x, t) represents the time evolution of a given parameter for each region x at each time step t with 1 ⩽ x ⩽ N and 1 ⩽ t ⩽ T. The aim of the analysis is to extract characteristics of the oscillations of this field at a given timescale. In the following, we will show an example in which the signal s(x, t) is the OLR with a spatial resolution of 5° and a time step of one day. In this case, the analysis will be done to extract spatial structures at an intraseasonal timescale (20–80 days) in order to characterize the ISO over the Indian and the Pacific Oceans.

The original signal s(x, t) must generally be filtered in time to retain only the oscillations at the selected timescale. The use of a filtered and centered time series, noted S(x, t), is necessary to avoid a mixing of timescales that may perturb the detection of the modes. Also, the fluctuations must be quasiperiodic for the relative phase field to be interpretable in terms of propagation speed or, more generally, in terms of time lag between the perturbations over different regions.

The CEOF analysis in the frequency domain, introduced by Wallace and Dickinson (1972), is the basis of the LMA. The CEOF analysis is based on the computation of N real eigenvalues λ_n and complex eigenvectors Z_n(x) from the N × N cross spectral matrix ζ defined as

View Expanded

where * represents the complex conjugate, k is the harmonic number (for the whole spectrum k₁ = 1 and k₂ = T/2), and F̃(x, k) represents the Fourier coefficients defined as

View Expanded

For each region x, the summation over the harmonics [k₁, k₂] of the squared module of F̃(x, k) gives the variance of the filtered time series S(x, t). Since the spectral variance of this filtered signal is negligible outside a given temporal harmonic interval, k₁ and k₂ may be chosen to cover only the frequency range of the filtered signal. The consideration of this restricted interval only avoids useless computation in spectral domains with negligible variance but is not a correct filtering method. (If a filtering is required, it is more appropriate to first filter the entire time series and then to apply the LMA to this filtered time series.)

Each CEOF corresponds to a percentage of variance υ_n = λ_n/Tr, where Tr is a real number representing the trace of the hermitian matrix ζ or the total variance of the time series S(x, t) for the N regions. Since the eigenvectors are sorted by decreasing υ_n, the first eigenvector represents the largest percentage of variance. The complex eigenvectors Z_n(x) may be normalized to obtain

View Expanded

The complex spatial eigenvector Z_n(x) is used to compute the spatial distributions of amplitude A_n(x) and phase ϕ_n(x) of the mode using

View Expanded

The normalized spectrum (the principal component), giving the time–frequency characteristics of the mode, is defined as

View Expanded

with

View Expanded

To interpret the relative phase field, the spectrum P_n(k) must be characteristic of a quasi-periodic feature. If a characteristic timescale τ can be extracted from this spectrum, a difference of π in the relative phase of two regions represents a lag of τ/2 between the fluctuations of the two regions. Appendix A gives an alternative computation method of P_n and Z_n that saves computation time when the number of harmonics considered (k_s = k₂ − k₁ + 1) is smaller than the number of regions (N).

The temporal evolution of the signal of the mode is given by

View Expanded

With this definition, the C_n are orthogonal:

View Expanded

The input filtered and centered signal S(x, t) is recovered by computing the real part of complex series:

View Expanded

The imaginary part of S̃(x, t) is the Hilbert transform of S(x, t) (Thomas 1969; Horel 1984). From this Hilbert transform, one can obtain a spatial distribution of amplitude A_t(x) and phase ϕ_t(x) at each time step t using

View Expanded

These distributions are what we call instantaneous spatial structures (ISSs) that have a meaning only for a quasiperiodic signal. It can be shown (appendix B) that for a CEOF analysis, the first eigenvector (i.e., largest percentage of variance) has a spatial structure that minimizes the distance relative to the ISS ensemble of the considered time series. For example, if the input signal has the form S(x, t) = a(x)b(t) cos[ωt + φ(x)], the whole series is a single mode, explaining 100% of the variance, which can be described at any time step by a single amplitude field a(x) [apart from a spatially uniform factor b(t)] and a single relative phase field φ(x) (apart from a spatially uniform phase shift ωt). Invariant ISSs thus have the same relative phase distribution and the same amplitude distributions apart from a constant factor. If the ISSs are almost invariant over the period of time considered, the first eigenvector will explain a large percentage of variance. If the ISSs are very variable with time, the percentage of variance of the first eigenvector will be small. In this case, the spatial structure will be either an average structure not representative of any of the modes present along the time series or a structure representative only of some of the most energetic modes.

A particular mode representing a given oscillation is present in a time series only if consecutive ISSs are similar over a sufficient time window (at least one oscillation). For a long time series, the spatial structure extracted from a mean analysis (e.g., composite, cross-spectral, EOF) is mainly affected by the strongest and/or the more reproducible ISSs. In order to avoid a mixing between nonorthogonal persistent modes that succeed one another in time, it is possible to compute a series of CEOFs on time windows sufficiently small enough to contain relatively uniform ISSs. This is the principle of the LMA.

For an original time series with T time steps, a series of M = (T − L)/ΔL + 1 CEOF analyses is performed where L is the number of time steps of the time window considered for each analysis and ΔL is the lag between two analyses. The value chosen for L depends on the expected duration of a mode. The sensitivity of the results with regard to the choice of L and ΔL will be discussed in the following sections. By definition, L will be relatively small, and it becomes necessary to weigh the time series using a Welch window to minimize end effects and to conserve as much as possible the spectral characteristics of the time interval under study. For the analysis number, m, the input time series will be

S_mx, tSx, tmLWt(13)

with 1 ⩽ t ⩽ L, 1 ⩽ m ⩽ M, and

View Expanded

An important effect of the Welch window is also to overweigh the ISSs on the central part of the time window relative to the extremities. The Welch window thus reduces the influence of possible strong oscillations at the beginning or end of the time window. These strong modes are detected in previous or later local analyses.

Only the first CEOF will be retained for each local analysis m; hereafter the mode represented by this first eigenvector Z^m₁(x) is called a local mode (LM). The variance percentage υ₁(m) of this first eigenvector will depend on the stability of the ISSs along the time window L (see appendix B). For a large variance percentage, the ISSs are almost invariant and the LM is thus persistent. The time evolution of the variance percentage may thus be used to reduce the number of modes. Each local maximum in the variance percentage series represents an interval for which the ISSs are more uniform than for nearby intervals. The selection of local maxima of the variance percentage makes it possible to reduce the M LM into a subgroup of principal modes (PMs).

3. LMA of an idealized signal

This section aims to highlight the basic properties of the LMA. To this end, the LMA is applied to an idealized signal representing five periodic modes that succeed one another in time over 120 regions (Fig. 1a). The signal S(x, t) is defined as

View Expanded

The variable W₆₀(t) represents the Welch window with a length of 60 points applied only to avoid abrupt changes between the modes. For this example, the LMA is performed with different lengths of the time window L = 60, 90, and 120 and a lag ΔL = 5, giving M = 65, 59, and 53 modes. To detect the variance percentage maxima for the different time window lengths, the time series is extended by 40 time steps before time step 1 and after time step 300 (Fig. 1).

The time evolution of the total variance Tr(ζ) for each analysis has strong fluctuations for short time windows but is nearly constant for a time window of 120 time steps (Fig. 1b). Nevertheless, the variance percentage of each LM always has strong and coherent fluctuations with maximum values for time windows centered at time steps t = 30, 90, 150, 210, and 270 (Fig. 1c). There are thus five PMs extracted for time windows centered on time steps corresponding to the maximum amplitude of the five input modes. For L = 90, these time windows, composed by more uniform ISSs, correspond to analysis number m = 6, 18, 30, 42, and 54 [Eq. (13)]. Note that maximum variance percentages are smaller for larger time windows because of the larger mixing with the other modes. For this idealized signal, the variance percentage of the five PMs is, however, larger than 0.9 for all considered time windows. For analyses corresponding to a transition between two modes (time window centered at t = 60, 120, 180, 240), the variance percentage is minimal because of the strongest mixing between two different successive modes at the beginning and end of the time window.

In order to inspect the stability of the LMs from one analysis to another, we use the normalized distance d²(m) between the first eigenvector of the analysis m and the first eigenvector of the analysis m + 1. From Eq. (B5) this normalized distance is defined as

View Expanded

where

View Expanded

is the phase that minimizes the distance between the vectors. Such a minimization is necessary, since two eigenvectors that are identical except for a constant phase difference represent actually the same mode. The normalization gives a distance of 1 for orthogonal vectors and 0 for identical vectors apart from a constant phase difference. For consecutive LMs, this distance is large only for time windows at a transition between two modes (Fig. 1d). For time windows around a maximum variance percentage, the LMs are very stable for this idealized signal. This means that the time step ΔL between two analyses could have been augmented without changing the structure of the five PMs detected.

The spectra P₁(k) clearly illustrate the periodicity of the input signal for each of the five PMs (Fig. 1e). For a physical signal, this will make it possible to retrieve the spectral characteristics of the studied phenomenon. From the time evolution C₁(t) [Eq. 8], it is possible to define a mean characteristic timescale τ as

View Expanded

where

View Expanded

The variable η is the sum of the phase differences between two time steps weighted by the average amplitude of these two time steps and R is the number of time steps per time unit (i.e., the number of time steps per day for a characteristic timescale, or mean period, in days). Note that before this computation, it is necessary to modify the time series of the phase arg[C(t)] by adding or subtracting 2π for phase jumps larger than π. From this definition of the mean period, one can also define a deviation as

View Expanded

This deviation is zero for a harmonic signal and increases as the spectrum dispersion increases.

The PMs z^m₁(x) obtained with L = 90 for analyses m = 6, 18, 30, 42, and 54 [Eq. (13)] are presented in Fig. 2a. The five PMs obtained with L = 60 or 120 are identical to these modes. In Fig. 2a, the regional amplitude of the PMs is proportional to the length of the segments, and the relative phase is represented by a clockwise rotation for increasing time. Note that the phase is relative for a given mode but is independent from one mode to another. The modes of Fig. 2 are identical to the five input modes (Fig. 1a), confirming that a local maximum of the variance percentage can depict the different modes present in a time series. For example, the two PMs obtained for m = 40 and for m = 52 exhibit a wavelength of 40 and 60 regions, respectively, and an “eastward” propagation for the input signal.

A CEOF analysis over the whole time series (time steps 1–300) has been performed to compare the results to the LMA approach. The five modes extracted from the five first eigenvectors do not represent any of the input modes (Fig. 2b). Instead, the first two eigenvectors exhibit improper in-phase or out-of-phase standing oscillations between areas centered on regions 20 and 100. The three remaining eigenvectors are not interpretable. This result is due to both the constraints of maximization of the variance on the first CEOF and of orthogonality between the CEOF. This is not surprising, since the four modes that succeed one another in time in the input time series are not orthogonal.

4. Application to the ISO

The applicability of the LMA to the study of real atmospheric fluctuations is tested on the relatively well-known ISO (see Madden and Julian 1994 for a review of observational studies of the ISO). The analysis is performed on time series of the OLR coming from the National Oceanic and Atmosphere Administration, Advanced Very High Resolution Radiometer (NOAA AVHRR) measurements (Liebmann and Smith 1996). We consider daily mean values (day and night average) of the OLR for 5° × 5° regions over a large tropical area (32.5°N–32.5°S, 32.5°–237.5°E) for the period of June 1974–July 1993 (year 1978 is missing). The signal is first filtered with a recursive filter (Murakami 1979), having half-power for periods 20 and 80 days.

b. The ISO modes

The temporal evolution of the variance percentage (Fig. 3a), the total variance (Fig. 3b), and the mean period (Fig. 3c) of the ISO show some interannual and seasonal variations. The persistence of the ISO (Fig. 3a) is larger during Northern Hemesphere (NH) winter, with 10 of the 20 MPMs occurring between January and March. As previously depicted by Lau and Chan (1988), there is also a tendency for a smaller persistence and total variance for NH winter during ENSO years of 1977–78, 1982–83, and 1986–87. Note, however, that there are persistent modes during NH winter 1991/92. The most striking feature is the strong seasonal variation of the total variance of the OLR (Fig. 3b) over the considered region, with a clear maximum between January and March. There is also a larger NH winter variance after 1984 that leads to a larger seasonal variation on the variance.

Regional maxima of the mean variance of the PMs have comparable values for NH winters and summers (Fig. 4). This is also true for both the variance of the filtered signal and the maximum regional variance of the modes. The NH winter maximum is thus not due to stronger regional perturbation of the convection but mostly to a more spatially extended perturbation with amplitude maxima located over the Indian Ocean, northern Australia, and the western Pacific. For NH summer, the ISO is active only in the northern Indian Ocean and in the northwestern Pacific Ocean. Note that a large part of the difference between the variance of the filtered signal and the mean variance of the modes is due to the Welch window weighting that divides the variance by nearly a factor of 2. The remaining part of the difference is due to the reduction of the local variance for each mode because it represents the part of the local signal that is coherent at large scale. The distribution of the variance of the filtered signal is close to the distribution of the variance of the OLR in the 20–70-day band obtained by Lau and Chan (1988) for years 1974–86. The distribution of the variance of the modes has nearly the same shape, showing that most of the variability at this timescale is related to coherent perturbation of the convection at large spatial scale. The maximum regional variance of the modes has very large values compared to the mean variance. This shows the strong variability of the amplitude and location of the convective perturbation associated to the ISO and thus the variability of the structure of the different modes. The distribution of the maximum regional variance also shows that the maximum perturbation of the convection by the ISO is not located on the equator but rather around 10°–15° in the summer hemisphere. An interesting feature is the systematically smaller variance over land for southern India as well as for the different islands of the Maritime Continent. The only exception is northern Australia during NH winter, showing the strong interaction between the ISO and the Australian monsoon.

In agreement with Anderson et al. (1984) and Cadet and Daniel (1988), there is no evident seasonal variation of the period of the oscillation in the first half of the time series (Fig. 3c). The period exhibits, however, coherent seasonal cycles after 1988, with a shorter timescale of the oscillation during NH summer. Also, in agreement with Gray (1988), Lau and Chan (1988), and Kuhnel (1989), the timescale of the oscillation tends to be smaller during ENSO years (Fig. 3c). As for the amplitude and the persistence, the 1991–92 ENSO appears as an exception when compared to the three other relatively strong ENSO events. Note that the period represented in Fig. 3c corresponds to the mean period for the coherent perturbation of the convection at a large spatial scale and is thus more representative than any local spectrum.

1) Average modes

Average modes for each season (NH winter from December to May and NH summer from June to November) may be computed considering the mean cross-spectral matrix [Eq. (A4)] among all the time windows corresponding to the PMs (i.e., 45 PMs for winter and 41 PMs for summer). The first and second eigenvectors obtained from the seasonal mean cross-spectral matrix thus give an idealized average seasonal structure of the ISO. The LMA makes it possible to test the significance of these average modes in regard to the modes that actually succeed one another in time. To this end, the normalized distance [Eq. (A6)] between the average modes (i.e., the first and second eigenvectors of the mean analysis) and each PM (i.e., the first eigenvector representing a given PM) has been computed.

For the NH winter season, the first mean eigenvector (Fig. 5a) represents an eastward propagation that corresponds well to the mean structure of the ISO already reported in previous studies with, in particular, a phase opposition between the western Pacific Ocean and the Indian Ocean (Zhu and Wang 1993). Among the 45 PMs used to compute the mean structure, only 7 (15) modes have a normalized distance smaller than 0.3 (0.4) relative to this first eigenvector (Fig. 6a). A distance larger than 0.4 clearly means that the spatial structures are very different, as demonstrated by the two modes of winter 1989 that have equal distance of 0.4–0.5 relative to both the first and second eigenvectors. The mode of the first mean eigenvector is thus representative of, at best, 30% of the NH winter ISO modes mainly in the second half of the time series because of the larger amplitude after 1984. The second mean eigenvector for NH winter (Fig. 5b) represents an in-phase oscillation between the Pacific Ocean and Indian Ocean and is clearly the result of the orthogonality constraint. This second mean eigenvector does not represent any PM (Fig. 6a).

For the NH summer season, the first mean eigenvector (Fig. 5c) represents a phase opposition between the northwestern Pacific Ocean and the northern Indian Ocean, with a northward propagation already described by Lau and Chan (1986). As for NH winter modes, the second mean eigenvector (Fig. 5d) is clearly a result of the orthogonality constraint. Among the 41 PMs used to compute the mean structure, only 3 (9) modes have a normalized distance smaller than 0.3 (0.4) relative to the first eigenvector (Fig. 6b). The mode of the first mean eigenvector is thus representative of, at best, 20% of the NH summer ISO modes. The second eigenvector is close to only one mode in September 1984.

2) Example of persistent modes not represented by the average modes

The mode of November 1985 is presented here as an example of a persistent mode (Fig. 3a) that can be extracted using the LMA but that does not resemble the first or second mean eigenvector obtained for either the summer or winter season (Fig. 6). This mode is characterized by an in-phase oscillation of the deep convection between the northern Indian Ocean and the South Pacific Convergence Zone (SPCZ) and a phase quadrature with the western equatorial Pacific (Fig. 7a). This particular structure of the oscillation stands for two–three oscillations (Fig. 8) and is thus one of the most persistent modes in the series. Its mean period is small compared to surrounding modes and is one of the smallest periods obtained for this season (Fig. 3c). The study of the physical origin of this mode is outside the scope of this paper; however, it is interesting to note that this mode appears at the end of the Indian monsoon and thus corresponds to the transition phase between the summer and winter regime above this region.

The PM of January 1992 (Fig. 7b) corresponds to an El Niño episode (Fig. 3c) and poorly resembles the first or second mean eigenvector obtained for either the summer or winter season (Fig. 6). The convective perturbation is weak over the Indian Ocean, and the eastward propagation of the perturbation that can be seen over the Pacific just south of the equator is very irregular. This mode exhibits an in-phase variation of the convection above Australia and the central Pacific and a phase opposition with the SPCZ region in the southwestern Pacific. The phase opposition between the central and western Pacific stands for at least two oscillations (Fig. 9). Over Australia, however, the timescale of the oscillation is shorter and the phase opposition is relative to the SPCZ stands for only one oscillation. The inhomogeneity of the period of the oscillation for the three main regions of activity is well depicted by the relatively large value (7.8 days) of the deviation of the period [Eq. (20)]. This value is smaller (2.4 days) for the more homogeneous mode of November 1985.

3) The eastward propagation during NH winter

Results from the LMA make it possible to inspect the propagation of the convective perturbation for the NH winter modes. This point is especially interesting because the propagative or stationary nature of the intraseasonal convective perturbation remains controversial (Zhang and Hendon 1997). In the average NH winter modes, both tendencies are clearly present (Fig. 5a), and it is interesting to study the variability of the propagative characteristic from one mode to another. Here, we consider all the modes (26 modes) obtained between January and March that correspond to months with maximum amplitude and persistence (Fig. 3). As shown in Fig. 4, maximum perturbations are located between the equator and 15°S. For each longitude x, the mean phase and the mean variance of each mode m over this latitude band was thus computed as

View Expanded

where the overbar represents the average over the 5° × 5° regions centered at latitudes 2.5, 7.5, and 12.5°S. The propagation time ΔL_n(x) at each longitude x (with a step of 5°) is thus computed by a finite centered difference as

View Expanded

where τ_m is the mean period in days for the mode m [Eq. (18)]. The values ΔL_m(x) correspond to the propagation time for 5° of longitude. The histogram of the ΔL_m values obtained from the 26 modes at each longitude is reported in Fig. 10a (values of ΔL_m larger than 5 are added in the first and last rows of the histogram). Values of ΔL_m are generally between 0 and 2 days (5°)⁻¹, showing a preferred eastward propagation with a speed larger than 3 m s⁻¹. Stable ΔL_m are obtained over the Indian Ocean with a propagation speed between 3 and 6 m s⁻¹ and over the Maritime Continent with a clear stationary perturbation between 120° and 140°E. Over the Pacific Ocean, the ΔL_m are more variable from one mode to another with, however, a tendency for a stationary perturbation between 190° and 220°E.

Having this variability in mind, it is possible to compute an average ΔL(x) over the 26 modes. Since the phase information is more relevant if the regional amplitude of the mode V_m(x) is large, we have also computed the average ΔL(x) at each longitude with two different thresholds in the regional variance V_m(x) at 50 (W m⁻²)² and 100 (W m⁻²)². We thus obtained three sets for the longitudinal distribution of ΔL. To give a better reading of the results, it is convenient to compute the average lag in days L(ℓ) between the longitudes by integrating ΔL(x) between the western longitude (37.5°E) of the considered region and the longitude ℓ:

View Expanded

This lag has been centered relative to the lag at ℓ = 127.5°E, that is, the lag at 127.5° is zero. Note that this normalization does not mean that ℓ = 127.5° is a reference region for the computation of the lag. In fact, the lag was computed only from the local information given by Eq. (22) and the integration and normalization are performed only to give a better reading of the results.

Strong perturbation of the convection (Fig. 10b) and eastward propagation (Fig. 10c) are confined between 80° and 190°E, around the Maritime Continent. West of 60°E and east of 210°E there is a westward propagation of the convective perturbation. Even for this average over 26 modes, the propagation is not uniform between 80° and 190°E. In particular, as already shown on Fig. 10a, there is a stationary perturbation roughly between 120° and 140°E corresponding to the center of the Maritime Continent for which the mean perturbation of the convection is the highest (Fig. 10b). There are two breaks in the convective perturbation at the western and eastern limits of the Maritime Continent (110° and 140°E). Note that the stationary character roughly between 150° and 160°E is not significant because of the strong variability of the ΔL over these regions (Fig. 10a and the error bar on Fig. 10c). The lag L(ℓ) is not sensitive to the amplitude threshold over the Indian Ocean and the Maritime Continent. This shows a relative homogeneity in the propagation of either strong or weak convective perturbations for these longitudes. In contrast, the eastward propagation over the western Pacific is dependent upon the amplitude of the convective perturbation, with a propagative character farther east as the local amplitude increases and a stationary character (or even a westward propagation) (Fig. 10a) east of 150°E for the weakest modes.

These results show the importance of the Maritime Continent and Australia in the propagation and amplitude of the intraseasonal convective perturbation during NH winter. In most cases, the convection first developed over the Indian Ocean around 60°–70°E and then rapidly expanded west and east. This extended perturbation propagates eastward, has a maximum amplitude around 90°E, and decays rapidly at the longitude of the Indonesian archipelago. This perturbation triggers (or is associated with) a strong stationary fluctuation of the convection south of the Maritime Continent (Fig. 4), with an evident relation with the Australian monsoon. A few days later, a strong convective perturbation may be generated around 160°E, which propagates eastward to 200°–210°E. However, the intraseasonal convective perturbation over the Pacific has various propagative properties, including westward propagation or stationary oscillations.

5. Summary and conclusions

This paper presents a new approach to detect and characterize different modes of intermittent oscillations. This approach, the local mode analysis (LMA), is based on the detection of different modes that succeed one another in time along a time series. As mentioned in the introduction of this paper, what we call different modes of an oscillation may also characterize different types of oscillation with different physical mechanisms. The main objective of such an approach is to identify the different structures of a given oscillation in order to better understand its physical origin or to test the applicability of different theoretical hypotheses. This approach makes it possible to detect the dates for which especially strong and persistent oscillations are present and to characterize the structure of these modes. From a geophysical point of view, results of an LMA may be first used to assess the general behavior of a parameter in a given time frequency range by inspecting the time evolution of the variance, the percentage of variance, and the period. This makes it possible to detect any trends (e.g., interannual, seasonal) in the characteristics of the phenomenon. The knowledge of the different spatial structures given by the LMA makes it possible to separate different types of oscillation. A group of modes conforming to a given criteria such as a particular phase relationship between different regions (e.g., propagation, phase opposition) may be considered for a series of case studies or a composite analysis. The LMA also makes it possible to test the representativity of a mean structure in regard to actual modes that succeed one another in time.

From a mathematical point of view, the method is based on the computation of a series of successive complex EOFs (CEOFs) on relatively short time windows of the whole time series. We have shown that the percentage of variance explained by the first eigenvector is a very useful parameter, since it characterizes the persistence of a given spatial structure of the oscillation. Selecting only the modes represented by a local maximum of the variance percentage, a subset of Principal Modes (PMs) may thus be extracted. This method appears to be robust, since the results are not very dependent on the size of the running time window. However, the application of the LMA requires some precautions for the spectral characteristics of the input time series and for the choice of the time window length.

• The input time series must be in a given timescale in order for the relative phase lag between regions to be interpretable. Filtering is thus often required. A spectral range that is too wide may also lead to a mixture of different processes at different timescales on a single mode. This last point may be tested on the deviation of the period (Eq. 20).

• The length of the “running” time window must be chosen in an interval that gives a sufficient spectral resolution for the lowest boundary and that avoids a mixture between modes for the largest boundary. The maximum length of the time window corresponds to the expected maximum duration of a given mode. The length of the time window may be adapted in this “reasonable” interval to get modes that are more or less persistent. A reduction of this length tends to add PMs that are less persistent.

The existence of a particular mode does not guarantee that its particular spatial structure is caused by a physical interaction between the different regions. It simply proves that such a spatial organization may exist with some persistence. A reproducible mode, that is, a spatial structure that appears on several PMs, is certainly more interesting than a single mode, from a meteorological point of view, as a particular mechanism to study or because it may be used to improve the forecast at the corresponding timescale. A reproducible mode may appear in a mean composite, cross spectral, or EOF analysis; however, such an average (single) mode does not always represent an oscillation that is actually present in a time series. For example, the mean ISO mode for the NH summer represents well some properties of the summer modes such as the phase opposition between the northern Indian Ocean and Pacific Ocean and the northward propagation of the convective perturbation; however, its idealized structure represents only very few actual ISO modes during NH summer.

The main finding concerning the ISO may be summarized as follows.

• Interannual and seasonal variations of the mean period and amplitude of the ISO show a tendency to have less persistent modes and a smaller period of oscillation during the El Niño years and during the summer.

• The maximum perturbation of the convection by the ISO is not located on the equator but rather around 10°–15° in the summer hemisphere. In addition, the local amplitude of the intraseasonal perturbation of the convection at these latitudes has similar values for summer and winter seasons.

• There is a strong variability of the spatial organization of the convective perturbation at the intraseasonal timescale. In particular, some persistent modes exhibit neither the phase opposition between the Indian and Pacific Oceans nor the eastward equatorial propagation that characterize the average modes. Mean modes for either summer or winter seasons are thus representative of only 20%–30% of the different local ISO modes.

Despite the various spatial structures of the intraseasonal convective perturbations, it is clear that there are preferred spatial organizations of these perturbations associated with the more persistent and the strongest modes. In particular, the eastward propagation over the Indian Ocean and the stationary oscillation over the Maritime Continent are strong characteristics of the ISO during the NH winter. Also, the eastward propagation over the Pacific Ocean appears only for the strongest perturbations of the convection above this region. This suggests that an ensemble of local conditions such as a particular SST distribution (related in particular to El Niño) or a phasing with the Australian monsoon are required to allow eastward propagation of a large convective perturbation from the Indian Ocean to the central Pacific. The fact that the Maritime Continent acts as a barrier for the propagation of the convective perturbation reinforces previous hypotheses, in agreement with observational studies (Jones and Weare 1996; Chen et al. 1996) and model analyses (Sperber et al. 1997), that this propagation may be due to local interactions between the ocean and the atmosphere. In this context, the tendency for a standing oscillation between the Indian and Pacific Oceans may be due to an additional large-scale interaction that exists for both winter and summer seasons.

The LMA approach presented in this paper appears to be a powerful tool to extract the main characteristics of intermittent atmospheric oscillations in a given timescale. This approach provides information on an ensemble of modes that actually succeed one another in time and is thus more adapted than conventional EOF analysis to study oscillations that have variable spatial structures. The LMA provides not only the time evolution of the persistence, the amplitude, and the period of the oscillations but also the spatial structure of each mode and its degree of similarity compared to average modes. This makes it possible to extract atypical persistent modes that may be of importance to better understand the behavior of the atmosphere. First results obtained by using the LMA to study the ISO are very promising and appear to be a good basis for further analysis of the physical origin of this phenomenon.