1. Introduction

This paper presents a 4-yr climatology of wavelike disturbances in the pressure field that propagate across a network of barometers near Champaign–Urbana, Illinois. These disturbances may include (i) events with periods of the order of minutes and wavelengths of the order of a few kilometers, generated by wind shear below the jet maxima (Herron and Tolstoy 1969; Mastrantonio et a1. 1976; Bedard et al. 1986; Hooke and Hardy 1975); (ii) disturbances generated by shear or other mechanisms lower in the atmosphere, including within the boundary layer, which generate and/or interact with turbulence and convection in the boundary layer and couple the boundary layer with the atmosphere above (Einaudi and Finnigan 1981; Clark and Thomas 1986; Hauf and Clark 1989); (iii) solitary waves and bores (Smith 1988; Fulton et al. 1990; Christie 1989, 1992); (iv) mesoscale disturbances with periods of the order of hours and horizontal wavelengths of the order of hundreds of kilometers, which interact with and sometimes trigger convection (Bosart and Cussen 1973;Uccellini 1975; Balachandran 1980; Stobie et al. 1983;Uccellini and Koch 1987; Ferretti et al. 1988; Schneider 1990; Ralph et al. 1993; and others). All these disturbances will be denoted generically as gravity waves throughout the paper.

Gravity waves play an important role in atmospheric dynamics because they can transport energy and momentum over large distances both horizontally and vertically. This role has been parameterized in general circulation models (GCMs). Topographically induced gravity waves have been shown to influence the strength and position of the polar jet and the temperature field in the troposphere (McFarlane 1987) and in the stratosphere and the mesosphere (Bacmeister 1993). Fritts and Lu (1993) and Hines (1997a,b) have developed a parameterization scheme that assumes the existence in the lower troposphere of a spectrum of upward propagating gravity waves that describes reasonably well the latitudinal and seasonal variations in the zonal winds and latitudinal temperature gradients of the upper mesosphere. See also Mengel et al. (1995), Mayr et al. (1997a,b), and Warner and McIntyre (1996).

At the mesoscale, gravity waves play the same fundamental role that Rossby waves play at the synoptic scale. Since they cover the entire range of spatial scales from frontal systems down to convective cells, they are the link between synoptic- and smaller-scale motions. They can have amplitudes of several hectopascals and they can induce rapid and sometimes major changes in otherwise well-forecast, synoptic-scale weather conditions. For example, Schneider (1990) described a disturbance that 1) produced a surface pressure deeper than a rapidly developing cyclone, 2) involved pressure tendencies (11 hPa in 15 min) that far exceed the criterion used for explosive cyclogenesis, and 3) enhanced precipitation rates in relatively narrow bands. Such disturbances create a substantial challenge for the research and forecast community. Thus, part of the difficulties that forecasters face is that these large mesoscale disturbances are not explicitly predicted by the current operational models. Even the climatology of these events is at present poorly known. Clearly, local weather forecasts could be helped considerably by a comprehensive documentation of these mesoscale waves, their characteristics, their areas of development, their source mechanisms, and the synoptic settings within which they develop and with which they interact. Mesoscale models with higher resolution have improved our understanding of the role of gravity waves at these scales (Schmidt and Cotton 1990; Pokrandt et al. 1996; Powers 1997). Indeed, the improved resolution of the current generation of operational models supports identification of gravity waves in the model output, although there are few observational studies with which to compare such results (Barnes et al. 1996).

Only sparse information exists on the climatology of gravity waves, that is, the statistics of their speed and direction of propagation, frequency of occurrence, amplitudes, temporal and spatial characteristic scales, and vertical structure. Some climatologies of wavelike disturbances in the troposphere, stratosphere, and mesosphere have been developed with data from rockets (Hirota 1984; Hirota and Niki 1985), balloons (Kitamura and Hirota 1989; Allen and Vincent 1995; Nastrom et al. 1997), lidars (Wilson et al. 1991), and radars (Sato 1994; Murayama et al. 1994). Other studies have been based on surface pressure data (Einaudi et al. 1989; Hauf et al. 1996; Koppel et al. 1999, and references therein). Since the pressure at the ground varies in response to changes in the mass of the overlying atmosphere and to vertical accelerations, it is thus a good indicator of gravity waves. In fact, pressure records are less noisy than temperature, wind, or humidity records because the inherent integration over height acts as a filter. Even so, pressure-based climatological studies of these disturbances are rare because the analysis is complex. Some of the difficulties are outlined in Grivet-Talocia and Einaudi (1998, henceforth denoted GE98).

In this work we present a climatology of mesoscale gravity waves based on pressure time series from a network, which had a diameter of about 50 km, of six or seven digital barometers. With this network we were able to investigate disturbances with temporal scales from tens of minutes to a few hours and spatial scales from a few tens to several hundred kilometers.

A network of at least three barometers may allow determination of the horizontal velocity of propagation of a wave event, which, in conjunction with the frequency content from analysis of the time series of the signal, provides information on the horizontal scales of the event. Since each of the atmospheric processes that generate wavelike disturbances has its own footprint in the surface pressure, different types of disturbances may require different kinds of analyses. When the pressure record is nearly monochromatic, as in the event analyzed by Einaudi and Finnigan (1981), the waveform can be efficiently obtained by a bandpass filter, which can be optimally designed using a standard FFT analysis. This technique assumes the decomposition of the event into a superposition of a few sinusoids, thus allowing a sharp frequency resolution. More often, the signal is complex and localized in time, such as short wave packets and solitary waves. Mesoscale phenomena such as thunderstorms go through a sequence of stages that leave various signatures in the pressure field and make the analysis of the records quite difficult. These cases are characterized by broad frequency bands due to their time localization. A set of bandpass filters could be used to extract all the energy of the disturbance from the time series. This process would lead to a set of smoothed waveforms in each frequency band, each one having little resemblance to the original disturbance. When processing each band independently, spurious correlations among the signals at the network stations would be introduced, producing erroneous calculated propagation velocities. In other words, a good determination of the event waveform is essential for obtaining accurate estimates for its speed and direction of propagation.

The wavelet transform method outlined in section 3 and described in more detail in GE98 satisfies this need, because it allows an accurate and efficient detection and extraction of the events in general and of the mesoscale events in particular. Wavelet transforms have recently been used in several studies of atmospheric dynamics (e.g., see the references in Lau and Weng 1995 and GE98). Hauf et al. (1996) applied wavelet analysis to barograph data, but their network was much smaller than ours and they analyzed only a small amount of data. Also, the various studies have used several different mother wavelets, appropriate to their particular database.

As with the standard FFT, the wavelet transform allows the decomposition of a signal into a set of basis functions, but these are localized in both time and frequency. This allows the time and frequency components of a disturbance to be considered simultaneously and minimizes the introduction of spurious correlations. GE98 also developed an algorithm that separates the events into two main classes, here called coherent events or incoherent events (henceforth abbreviated CEs or IEs, respectively), according to their degree of coherence across the network of barometers. The horizontal velocity can be estimated only for the coherent events. A precise definition of the criteria discriminating the two classes will be found in section 3.

We have estimated the depth of some of the events using data from the 50-MHz wind-profiling radar located at the Flatland Atmospheric Observatory (FAO). We find that most of the large-amplitude events extended up to at least 7.0 km, the highest altitude consistently observed by the radar.

We also attempt to relate our statistical results to meteorological features. The limitations and qualitative nature of such comparisons are described. They stem from the complex nature of weather and the difficulties in analyzing all the various weather data in an automatic process.

The barometer network and the data are described in section 2, while the wavelet method is outlined in section 3. The statistical results are presented and discussed in section 4 and are compared with meteorological events in section 5 and 6.

2. Experimental setup

This section describes the experimental setup and the dataset that will be processed in the following. Figure 1 shows the network of eight digital barometer sites that were used in this study, although only six or seven of them were operational at the same time. The central station, FLA, is located at FAO at 40.05°N, 88.38°W, near Champaign–Urbana, Illinois. This network differs from previous barometer networks in three respects. First, it is much larger, spanning about 50 km. Second, the terrain in and around the network is extremely flat, with the altitude difference between the highest and lowest stations, WIL and TUS, being only 25 m. Third, the barometers are digital rather than analog, which permits study of long-period fluctuations. The barometers are located in offices to shield them from wind and rapid temperature fluctuations. Each station instantaneously samples the pressure every 10 s and archives a block average of 12 samples every 120 s, which is then the effective sampling interval. The resulting precision is about 10⁻² hPa. Once a day the 120-s means are transferred by telephone to a computer database at the National Oceanic and Atmospheric Administration (NOAA) Aeronomy Laboratory in Boulder, Colorado.

At the beginning of 1991 the network consisted of six stations (lacking ISW and WIL). On day 336 of 1991 station ISW was installed, on day 272 of 1992 station ALP was shut down, and on day 18 of 1993 that instrument was relocated to WIL. Thenceforth the network was not changed. Therefore, the number of stations that recorded data simultaneously was usually seven but sometimes only six.

Data were analyzed from 15 quarters spread over five calendar years, from the beginning of 1991 to the end of 1995. Several problems, including interruptions of the telephone links and malfunctioning of the data recording system, reduced the size of the usable dataset. The years 1991 and 1992 are complete, while only parts of 1993, 1994, and 1995 are available as shown in Table 1. The quarters are defined by the day intervals in column two, while the months in column three are only nominal.

Since the configuration and separation of the stations limits the spatial and temporal scales of disturbances that can be correctly identified and processed, as will be shown later, the present study concentrates on mesoscale disturbances. Synoptic and tidal fluctuations are eliminated by preprocessing the raw pressure time series with a highpass filter with a cutoff at six hours. By repeating the entire analysis with data highpass filtered with a cutoff at eight hours we found that the results are insensitive to the cutoff period. The highpass-filtered pressure time series will be denoted henceforth by f_i(t), where the subscript i denotes the stations, with i = 1 for FLA.

3. Data processing

This section describes the data processing technique used in this paper, which allows the iterative detection, extraction, classification, and analysis of pressure events from the network of barometers. The technique is based on a selective reconstruction of the waveform of an event that is coherent through the network. This is achieved through the partial inversion of the wavelet transform of the pressure data at the different stations combined with cross-correlation techniques. In a second stage, the arrival times of the wave event at the network stations are used to determine its speed and direction of propagation.

The wavelet algorithm used in this paper is described in detail in GE98, so only a summary of the main steps and their theoretical justification will be given here. Many excellent introductory books on the wavelet transform have already been published (Daubechies 1992; Chui 1992; Meyer 1992). We will give here only the few definitions that are strictly necessary to understand the data processing technique. Note that in the present paper the notation for the wavelet transform has been simplified with respect to GE98.

Let us consider a function ψ(t) that is localized both in time and frequency. In other words, both ψ(t) and its Fourier spectrum ψ̂(ω), where ω is the frequency, have significant magnitudes only on a finite interval of the time and frequency axis. The function ψ(t) is denoted mother wavelet. A set of dilated and translated versions of ψ is derived through

View Expanded

The parameter b shifts the wavelet in time, while a, usually called the scale, modifies its dominant period T (or equivalently its dominant frequency, which is inversely proportional to a). Since the functions ψ_b,a(t) have a constant shape, they preserve the time–frequency localization of the mother wavelet. The time localization is tuned by the parameter b, and the frequency localization by the parameter a. When the signal f(t) is real and when the mother wavelet is progressive, that is, its Fourier spectrum is identically zero for negative frequencies, see Grossmann et al. (1989), we can restrict the analysis to positive values of a. The resulting expressions for the wavelet transform and its inverse are (Daubechies 1992)

View Expanded

where ℜ denotes the real part. Note that we use here the L¹ normalization instead of the more standard L² normalization, because it allows a simple link (independent of the parameter a) between the amplitude of the disturbances and the magnitude of the wavelet transform (see GE98). The only necessary condition for these equations to be true is the so-called admissibility condition (Daubechies 1992),

View Expanded

which essentially requires that the mother wavelet be a bandpass filter. The mother wavelet used in this paper is the analytical signal associated to the fourth derivative of a Gaussian (see GE98). This progressive wavelet has been chosen due to its good time–frequency localization properties.

Equation (3) means that the signal f(t) is exactly decomposed into a superposition of basis functions that are localized in both time and frequency. Therefore, the magnitude of the wavelet transform |W(b, a)| indicates the presence in the signal f(t) of a particular period T (proportional to a) at a particular time b. This localization in time is the advantage of the wavelet transform with respect to the Fourier transform. We will show that the wavelet decomposition is indeed a suitable tool for the study of disturbances such as solitary waves or short wave packets.

The inversion of the wavelet transform in Eq. (3) is achieved by integrating over the whole semiplane Σ = {(b, a)|a > 0}. If, however, the characteristics of a particular feature or event are of interest, then W(b, a) can be considered over only a smaller domain Ω ⊂ Σ. This allows study of the time–frequency components of that particular event separately from other events and geophysical noise. The waveform p(t) of the event can then be extracted using Eq. (3) by integrating a and b over only Ω. The problem is then to develop an automatic procedure to determine Ω from f(t), so that the extracted signal includes only the event under investigation and is identically zero elsewhere. As a result, the process of extracting p(t) can be interpreted as a nonlinear adaptive filter, expressed by

View Expanded

The method of determining Ω is illustrated by application to the two-day pressure time series f₁(t) observed at FLA shown in Fig. 2a. The corresponding contours of |W(b, a)| are shown in Fig. 3a as a function of time and frequency or period. The large amplitudes of the contours centered at days 304.6 and 305.6 suggest that these may represent events. The simultaneous contours of |W(b, a)| at URB shown in Fig. 3b are well correlated with FLA during the events, but not outside of the events. The poorly correlated values outside of the events are due to some kind of local geophysical noise.

Thus, an event is identified by a maximum of |W(b, a)| at (b_c, a_c) with a magnitude larger than A(a_c). The domain Ω of the event is then determined by extending in all directions from (b_c, a_c) until the first directional minimum of W(b, a) is encountered. This method includes most of the energy of the event while minimizing inclusion of fluctuations due to other events or geophysical noise.

The wavelet inversion through Eq. (5) leads to the extracted waveform of the event p₁(t), shown in Fig. 2b. Figure 3c shows its time–frequency components. Figures 2c and 3d show the waveform and the time–frequency components of the residual f₁(t) − p₁(t).

The domain of an event at FLA, denoted Ω₁, is determined first. Then for each other station i the corresponding domain Ω_i is determined under the hypothesis that the event is a coherent disturbance propagating with small distortions through the barograph network. A first estimate of Ω_i at station i is obtained by time shifting Ω₁ by the propagation delay time between the two stations, Δb_opt, that maximizes the cross-correlation function

View Expanded

where W_i(b, a) indicates the wavelet transform of the pressure signal f_i(t) at station i. This function selects only those time–frequency components included in Ω₁. Then the coordinates of the maximum of |W_i(b, a)| within the domain Ω₁ translated by Δb_opt are found and Ω_i is determined by the steps already described for the reference station. This selective cross-correlation procedure ensures that the same event is identified at each station, but it permits inclusion of local distortions in the corresponding extracted waveform p_i(t). This is evident in Fig. 5, where the waveforms p_i(t) of the event in Fig. 2 are plotted for all six stations that were operating at that time. Although the main shape of the disturbance is the same at each station, the details are different.

A refined estimate for the delay times

t_it_it₁iN_s(8)

between each station i and FLA is now obtained as the value of τ that maximizes the cross-correlation function

View Expanded

There are two considerable advantages in calculating the Δt_i from the extracted p_i(t) rather than attempting to do so from the highpass-filtered f_i(t). First, this minimizes the influence of fluctuations that do not belong to the event under consideration. Second, the integration interval is automatically determined by the finite duration of the p_i(t). If the f_i(t) were used instead of the p_i(t), then the integration domain would be undefined because the duration of the event is not known a priori. A measure for the coherency through the network is given by the mean of the cross-correlation coefficients, C = 〈C_1i(Δt_i)〉.

As a result of this procedure, events are separated into two different classes.

• Coherent events (CEs; called class 1 in GE98): A CE is an event with a large mean cross-correlation coefficient (C > 0.75) and a small fitting error (δt_rms < 3). Thus, a CE is a well-identified event that propagates coherently across the network and for which the propagation velocity (speed and direction), and wavelength, as well as the duration, amplitude and period band, can be determined.

• Incoherent events (IEs; called class 2 in GE98): An IE is an event with small C and/or large δt_rms, that is, not satisfying the criteria described above. Such an event can be a localized disturbance that does not propagate through the entire network, a wave packet with a wavelength that is small with respect to the dimension of the network, or an event with excessive superimposed geophysical noise. For an IE, only the duration, amplitude, and period band can be determined.

The time-averaged wavelet spectrum E_i(a) at each station i is defined as

View Expanded

A mean wavelet spectrum for each event is then obtained by averaging E_i(a) over all stations to form

EaE_ia_i(12)

The dominant period T_m of each event corresponds to the value of a at the maximum of E(a). For CE the horizontal wavelength λ is also estimated as λ = υT_m, where υ is the phase speed. The value of T_m is of course only indicative, since a nonmonochromatic disturbance is characterized by a band of periods centered around T_m. This period band is here defined as the interval including the central 50% of the total energy in E(a). Similarly, the duration of an event is here defined as ΔT = t_e − t_s, where t_s and t_e are the times at the 1% and 99% levels of the cumulative instantaneous energy;thus, ΔT includes 98% of the total event energy. The amplitude of each event is simply defined as the average of the maximum peak-to-peak fluctuations at the different stations.

The procedure described above permits extraction and study of a single event. The event with the largest amplitude at FLA was detected first, its reconstructed waveform p(t) was subtracted from the highpass-filtered pressure trace f(t) at all stations, the next largest event was detected, and the extraction procedure was iterated until all events above an amplitude threshold (derived in GE98) were processed. It is shown in GE98 that the application of a number of statistical consistency tests to the iterated extraction procedure ensures that wrong identifications are avoided. Thus, all significant events were extracted and global statistics and climatological results were obtained.

For example, when the analysis procedure described above was first applied to the f_i(t) from day 304.5 to 306.5 1991 (f₁(t) for FLA is shown in Fig. 2a), a coherent event was found centered at 305.6. The reconstructed waveforms are shown for all six stations in Fig. 5. However, when the procedure was applied to the residuals (see Fig. 2c) no event was found at 304.6 because the results did not satisfy the criteria on C and δt_rms.

The described method is very powerful for the study of disturbances that are localized in time. However, there are some disadvantages for the analysis of some other kinds of disturbances, such as long-lasting, nearly monochromatic, short-period wave trains. Sharp time localization is not critical for such events, but a precise definition of the period is needed. This could be estimated from the wavelet period spectrum E(a), but with low precision. In addition, the wavelet reconstruction includes all the significant time–frequency components that are in the extracted waveforms. If longer-period amplitude modulations are present with amplitudes that are comparable with or larger than the amplitude of the shorter-period event, then the inferred shorter-period phase velocity may be incorrect. Such problems could be avoided by using a sharp bandpass filter centered on the frequency determined by the FFT, with sacrifice of temporal resolution. In any case, the timescales that are investigated in this paper are so long that a wave packet with many cycles would last for several days; such a disturbance was never detected in our pressure records. When a smaller-size network is analyzed or when this data processing technique is applied to other datasets of different nature, such as seismic signals or ocean waves, more care should be taken for the mentioned cases.

4. Results

5. Relation to meteorological features

This section presents an attempt to discover the causes of the events by comparing the propagation speeds and directions of CEs and nearly simultaneous meteorological features. Three kinds of features have been considered: precipitation; surface boundaries, defined in section 5b; and dynamical instabilities at altitudes above 1 km inferred from radiosonde balloon wind and temperature profiles from the Peoria, Illinois, radiosonde station, about 130 km NW of FAO.

The present comparisons must be considered only semiquantitative and preliminary. We have analyzed pressure data automatically using a robust and relatively simple algorithm. We do not believe that similar algorithms can be applied to the analysis of other meteorological data from meteorological radars, wind profilers, surface stations, rawindsondes, etc. As a result, the establishment of the relation of gravity waves to meteorological features would require a case study approach.

6. Conclusions

We have presented a climatology of large-amplitude disturbances using data from a network of sensitive digital barometers located around the Flatland Atmospheric Observatory near Champaign–Urbana, Illinois. The data cover 15 quarters from 1991 to 1995. The characteristic temporal and spatial scales of the processed data are limited by two factors. First, the size of the network, with distances between stations ranging from 5 to 50 km, does not allow a full analysis of events with periods smaller than ≈30 min. Second, a preprocessing highpass filter was applied to the data to remove fluctuations with periods longer than 6 h. Consequently, the characteristic horizontal scales range from a few tens to a few hundreds of kilometers.

Typical waveforms are localized in time, as shown in Fig. 6. A powerful and selective filtering technique based on the wavelet transform is utilized. This method provides objective localization of a disturbance and an accurate extraction of its waveform at each station through the selective inclusion of its time–frequency components. The application of a cross-correlation technique to the extracted waveforms is then used to determine the horizontal speed and direction of propagation for those events that are coherent through the network. This class of events, denoted CEs (coherent events), is characterized by mean cross-correlation values larger than 0.75. Another class of events, denoted IEs (incoherent events), includes disturbances with worse correlation or large variability in the estimates for speed and direction, and it also includes those coherent waves whose wavelength is too short relative to the size of the network. For these last disturbances an estimate of the propagation velocity was not determined due to the complication of aliasing.

Only disturbances with amplitudes larger than a threshold function are considered. This threshold, a function of the dominant period, was obtained by averaging over the entire dataset. Typical amplitude values for the detected events, both CEs and IEs, ranged from about 0.3 hPa at 30 min to about 1 hPa at 4 h.

There are some limitations in the analysis method utilized. First, it is not appropriate for the study of nearly monochromatic waves lasting for several periods. In that case standard FFT techniques could be used. Second, it cannot track disturbances whose speed and direction are strongly varying functions of time and/or space. In cases such as arc-shaped disturbances, time-to-space conversion objective analysis can be used (Koch and O’Handley 1997). Third, it identifies as separate those events that might have been generated by a single phenomenon. In the case of a thunderstorm, for example, it cannot recognize that its various dynamical stages have different signatures in the pressure field, even though they are related to a common origin. Fourth, it cannot separate waves with the same period content that occur at the same time. Only events with distinct and separate signatures in the time–frequency plane can be processed with accuracy. On the other hand, the presented method is not limited to pressure time series analysis, but can be used to analyze geophysical data of various nature from any network of at least three instruments. The network should be sized according to the range of wavelengths, periods, and propagation speeds of the disturbances to be investigated. Of course, the wavelet analysis and filtering tools must be tuned to the data being processed. For a more detailed discussion we refer to GE98 and to the references therein.

The statistics presented in this paper show that the frequency of occurrence of events has a strong seasonal dependence, with a maximum in fall and winter and a minimum in summer. Of the total analyzed time, CEs occurred 20%–21% in fall and winter and 12% in summer, while all events occurred 34% in both fall and winter and 23% in summer. The percentages above are smaller than those given by Einaudi et al. (1989), who used a network of eight microbarographs deployed within a 500-m radius near Erie, Colorado. One should note, however, that (i) Erie is about 25 km east of the foothills of the Colorado Rocky Mountains where the level of mesoscale activity is greater than at locations away from mountains (Canavero and Einaudi 1987); (ii) the size of the Erie network and the mechanical response of the microbarographs filtered out disturbances with periods longer than about 20 min, so that there was essentially no overlapping of characteristic temporal scales with the present study; and (iii) the present analysis imposes rather stringent conditions on the identification of the truly coherent disturbances.

The histograms of periods show a distribution that covers the whole range of investigated timescales. Both the number of CEs and their frequency of occurrence increase with increasing period. The horizontal wavelength histograms show that almost all wavelengths were in the range between 40 and 600 km. The histograms show a decay of the number of occurrences for periods longer than about 4 h and horizontal wavelengths longer than about 350 km due to the preprocessing highpass filter applied to the data.

The speed histograms show a peak in the 25–30 m s⁻¹ range, well within typical values of the jet stream, while the histograms of direction indicate that the overwhelming majority of the disturbances moves in the eastward direction ±90°, that is, in the general direction of the jet streams and jet streaks.

The identification of the source of a disturbance is often difficult even in the context of a detailed case study, since the source may not be in the neighborhood of the observation site and since the weather in the area of generation is often quite complex. Furthermore, tracking a disturbance far away from the observation site may require data recorded at standard surface stations. These unfortunately often have poor temporal resolution. We have nevertheless attempted to relate the observed events to meteorological features in the neighborhood of the observation network. These are some of the results.

1. The CEs were correlated only weakly if at all with the surface meteorological features that were investigated. However, 27% of the CEs occurred within 6 h of a boundary passage. Most of the CEs preceded the associated boundary passages or crossed the network at the same time. There was no significant correlation between the directions of propagation, but coherent pressure events generally moved faster than the associated weather events. This does not necessarily imply that the motion of a boundary passage is itself the source of the disturbances; rather, they may both be simply a manifestation of a baroclinic environment that is conducive to the generation and propagation of the waves as discussed earlier.

2. Several disturbances occurred without any charted boundary passage or precipitation activities.

3. There were some intervals lasting several days during which no events were detected. During these intervals there were also no significant meteorological features identified in the surface charts.

4. Precipitation was present with 73% of the large-amplitude CEs.

5. About 40% of the large disturbances presented features typical of solitary waves. We included in this category event types I–IV reported in Fig. 6. They occurred with the general characteristics discussed in 1–4 above. Their largest number of occurrences had periods just below 4 h with a broad distribution in horizontal scales ranging from 30 to 400 km. While none of the cases identified in this study reached the amplitude of that presented by Schneider (1990), this analysis confirms that this particular type of disturbance appears to be an important component of the dynamical behavior of the atmosphere at the mesoscale.

Although an automatic analysis of the wind profiles obtained from a 50-MHz Doppler radar at the Flatland Atmospheric Observatory was not possible for the reasons mentioned in section 4, we performed a qualitative analysis of the depth of the disturbances with amplitude larger than 2 hPa. The results show that in most cases the events also have a strong signature in the wind fields and that this signature is visible at least up to 7 km AGL, which was the height limit dictated by the radar sensitivity. Thus, these disturbances involve the entire troposphere.

A detailed analysis of a few cases involving large-amplitude events is in progress to determine more accurately their source mechanisms and the synoptic environment conducive to their propagation.