1. Introduction

Argonne National Laboratory operates three 915-MHz radar wind profilers within its Atmospheric Boundary Layer Experiments (ABLE) facility (Wesely et al. 1997) in the Walnut River watershed in south-central Kansas. One of the profiler systems, called the Boundary Layer Radar System (BLRS), is similar to the system described in Carter et al. (1995) but predates it and is substantially more primitive in its data handling capabilities than the other two systems. Of particular relevance here is that the BLRS is more susceptible to signal contamination by migrating passerines because it has no built-in hardware or software algorithms to suppress intermittent clutter (such as signals from flying birds) in real time. The problem of bird interference is well known in the remote sensing community (e.g., Wilczak et al. 1995; Merritt 1995; Coulter and Holdridge 1995; Zbar 1996; Barth et al. 1996; Jordan et al. 1997) and continues to affect many profilers. Modern profilers, including the “other” two systems in the ABLE network, usually employ data retrieval algorithms based on special processing of individual Doppler spectra before spectral averaging takes place (e. g., intermittent clutter removal or a “bird algorithm”), which makes such profilers less affected by, but not immune to, bird interference. The BLRS and many other older profilers cannot examine individual spectra as they are collected;rather, a number of spectra are averaged together in frequency space before any analysis or inspection can be performed. Thus in these older systems any modern“front edge” algorithm cannot be implemented, which makes automatic removal of bird artifacts difficult. The primary objective of the present work is to develop a practical method of data processing to minimize bird distortion in hourly wind data obtained with the BLRS radar. The authors do not address the problem of bird interference in general, rather in one specific case of making useful measurements with the help of outdated equipment. The technique described, however, is applicable to any set of radar wind profiler data that contains the averaged spectra.

The magnitude of the problem is described in section 2 for data taken routinely in 1997–98 and more intensively during the Cooperative Atmosphere Surface Exchange Study (CASES-97) in May 1997 (LeMone et al. 1998), when rawinsonde launches were made from all three profiler sites within ABLE, and birds were actively migrating from south to north. The proposed procedure is described in section 3; section 4 deals with comparisons of in situ measurements with both the traditional and the proposed processing of BLRS data. (Henceforth we will refer to the proposed method as “new” and to the original BLRS processing scheme as “old.”) Section 5 is devoted to an example of application of both old and new methods to a BLRS profiler dataset obtained during a period of active bird migration.

2. The problem

The relatively short wavelength (33 cm) of 915-MHz radar wind profilers makes them very susceptible to strong reflections by songbirds; the reflections easily pass quality control (QC) procedures because the relatively broad beam ensures that the birds’ presence in the beam is well sampled. Migrating songbirds travel primarily in favorable winds; thus, winds can be systematically overestimated during southerly winds in spring and northerly winds in autumn, if the profiler is located along a preferred migration path. Wind direction values are affected as well because birds sometimes fly at an angle to, or even against, the wind in light wind conditions: they tend to be due south (springtime) or due north (fall). The fact that birds do not always travel with the prevailing wind makes it very difficult to remove their effects uniformly with a consistency test or QC procedure; instead, in the presence of persistent bird echoes, these methods are liable to (and do) choose the bird signal.

Figures 1 and 2 show the distribution of component differences (radar minus sonde) during three days in May 1997. Several things are immediately apparent. First, during daytime, the agreement between sonde and profiler winds is quite good at all sites. Second, during nighttime, the agreement between southerly components is good for profilers 1 and 2 (Beaumont and Whitewater, located approximately 60 km NE and NNW, respectively, of Oxford, the site of the BLRS). Third, the BLRS measurements frequently are about 4–8 m s⁻¹ larger than the sonde measurements during nighttime. A very similar distribution of differences (sonde minus profiler) was found prior to November 1996 for the profiler located near Lamont, Oklahoma (approximately 100 km SSW of Oxford), operated by the Atmospheric Radiation Measurement (ARM) Program (Coulter and Holdridge 1995). All the profilers mentioned are located along the same bird migration route, as a majority of passerines follow broad routes. [Other details on bird migration can be found elsewhere (e.g., Elphick 1995).] At the sites where wind artifacts from birds are not evident (i.e., Beaumont and Whitewater), profilers utilized a built-in procedure to inspect individual spectra and remove intermittent clutter signal (bird artifacts), apparently successfully, at least in these cases.

The seasonal variation in the occurrence of bird interference at Beaumont is illustrated in Fig. 3. In these cases the presence of birds was detected by a characteristic signature on reflectivity time–height plots from the profiler. (Usually, the reflectivity pattern contains multiple bright elements randomly distributed over certain time–height regions on the plot. Each outstanding element is an image pixel, defined by radar pulse dimensions and dwell time, with intensity about 15–30 dB higher than the atmospheric signal–noise picture.) It should be noted that an abundance of this bird signature on reflectivity plots indicates heavy contamination conditions, when the intermittent clutter removal procedure was unable to reject all bird signals. On the other hand, it is not necessarily true that all hourly wind data are corrupted; and again wind data degradation may occur in absence of pronounced bird signature on reflectivity plots. Total presence of bird signature in Beaumont profiler reflectivity plots amounts to 118 nights from 718 days of observation in 1997–98. Apparently, one can expect to find bird artifacts during 6–7 months, and inactive periods are about three months in winter and only two months in summer.

3. Method

We assume that the radar signal consists of three major components: atmospheric scattering, system noise, and intermittent clutter (reflections from birds, insects, etc.). We do not include other potentially serious sources of radar signal contamination—ground and sea clutter—because this problem has its own long history and it is not addressed in the present work. Both the atmospheric signal and the noise are assumed to be Gaussian, although with different spectral, time, and distribution parameters. The atmospheric signal has a narrowband spectrum and a correlation time on the order of tenths of a second; system noise has a uniform spectrum. An intermittent clutter signal has non-Gaussian statistics, typical for point targets, with correlation times much longer than those of atmospheric signals; in this case the reflection from birds usually is also the most energetic component.

The best method for eliminating bird artifacts is to analyze the original time series before any spectral processing takes place (e.g., Jordan et al. 1997). Unfortunately, the BLRS (and many older wind profiling systems) provides no access to data processing in its early stages. Time series measurement (time domain averaging), calculation of individual spectra, and spectral averaging over 20 s or more (frequency domain averaging) are performed automatically by hardware and/or software. Therefore, only the set of average spectra can be used to remove the undesired effects. Generally about 12–15 such spectra occur in a 1-h averaging period. The frequency domain averaging is performed to improve signal-to-noise ratios (SNRs) and, coincidentally, to maintain manageable amounts of data; however, this averaging has the disagreeable side effect of mixing all three signal components together, so that complete separation of principal components is no longer possible. The intermittent clutter signal loses its distinguishable non-Gaussian characteristics and acquires some Gaussian features as reflections from many objects are combined with parts of atmospheric and noise signals of similar frequency domain.

Merritt (1995), following Hildebrand and Sekhon (1974), proposed a statistical averaging method (SAM) that replaces the usual arithmetic average of spectra in frequency domain averaging with process that removes non-Gaussian portions of the signal. The basis of SAM is, first, that every spectral component of a Gaussian signal has an exponential energy distribution, and a set of estimates of this component should comply with an exponential distribution test (i.e., the variance must be equal to or less than the square of the mean). Second, the signal from bird reflections is usually much larger than the atmospheric scattering. Thus, for every spectral bin (component), the process allows only values that have the lowest power and also pass the exponential energy distribution test to be used in the spectral average. The resulting spectrum is presumed to contain only signals from atmospheric scattering and system noise.

We apply an algorithm similar to SAM to each set of averaged spectra (usually 15 per hour) to obtain a single spectrum representative for that hour. To maximize statistical significance for the exponential energy distribution test, we use all the spectra from three contiguous range gates (the range gate of interest plus one higher and one lower). Thus, we make a sacrifice in height resolution to increase the representativeness of the spectra. This step seems to be warranted because a bird reflection could contaminate data from several adjacent range gates (Wilczak et al. 1995). Although this procedure removes much of the contaminated portion of the signal, it also removes some part (hopefully minor) of the atmospheric signal.

A system noise level is then defined for the spectrum from each pointing direction, as described by Hildebrand and Sekhon (1974); all portions of the spectrum with power lower than the defined level are discarded from the spectral average. The result is a 1-h-averaged spectrum from each beam with most of the bird and noise signals removed.

The next step is to derive the wind speed and direction from the single, averaged spectra for the three beams. Because migrating birds normally fly with tail winds and the migration direction is generally northerly (spring) or southerly (autumn), we can apply an additional condition for the north–south wind component: if there are two well-defined signal regions, the bird signal should be farther from zero Doppler shift than the atmospheric signal. We look for two spectral regions of given width having maximum energy (major and secondary) and assume that the region closer to zero corresponds to an atmospheric signal and the other to reflections from birds. It is known that sometimes bird migrations occur against light winds. Present implementation of the method has no provisions for these rare occasions, but additional condition could be easily incorporated. In this case atmospheric and bird signals would be located at different sides of zero Doppler shift, with atmospheric signal being on the north side if it is for spring or the south side for autumn. For the west–east component no such conditions applies, so we use the major maximum as a region of atmospheric signal. Note that because the procedure as described alters the SNR and the spectral shape, the spectrum obtained cannot be used to estimate either signal energy or higher-order moments.

To summarize, there are three differences between the old and new BLRS processing schemes.

1. The old method uses hourly sets of averaged spectra to calculate moments and derive wind speed from these moments using arithmetic or consensus averaging; thus the final hour average is calculated in the“speed” domain. The new method derives hourly averaged spectrum with the help of SAM; that is, hourly averaging is made in the frequency domain.

2. The new method combines spectral data from three adjacent range gates to make a rejection procedure on the basis of an exponential test more reliable.

3. The new method applies special conditions to discriminate between atmospheric and bird signals in estimation of the Doppler shift from hourly averaged spectra.

In this instance, the proposed algorithm was implemented as a stand-alone routine to process prerecorded spectral data (postmortem processing); however, the algorithm could also be incorporated into online processing.

4. Comparison with in situ measurements

During the CASES-97 (LeMone et al. 1998) field campaign in May 1997, rawinsonde profiles of winds and temperatures were made every 90 min during five days at each of the three profiler sites. Because the bird contamination is most serious for the BLRS radar, we will limit further discussion to that profiler. The comparison for the BLRS was restricted to 75-m range gates between 310 and 1600 m. (Although BLRS data above this height are available, they are less reliable because of limitations on BLRS radar sensitivity.) The rawinsonde data were averaged according to the profiler range gates and were compared with 1-h-averaged radar values obtained with both the new and old procedures. This comparison (Tables 1 and 2) confirms that the radar data were contaminated during nighttime and also reveals that even the east–west component is susceptible to bird contamination, although to a much smaller extent than the north–south (N–S) component. Daytime data show the usual reasonably good agreement for this kind of comparison (e.g., Martner et al. 1993), with an average difference of less then 1 m s⁻¹ and a standard deviation of the difference of the order of 2 m s⁻¹.

Correlation plots for all range gates of the N–S component (Fig. 4) show that almost all nighttime data were contaminated. (All intensive rawinsonde operations took place during fair weather with predominantly south winds.) The new procedure improved the agreement of a significant part of the data and diminished the error in the remaining poor comparisons. Although the new procedure gave better results, they are still far from perfect (rms difference is 4.98 instead of 9.24 m s⁻¹; mean difference is −3.21 instead of −7.35 m s⁻¹). The averaged profile for the N–S component at night (Fig. 5) shows that the mean error of the new procedure tended to increase with height. The sensitivity of the profiler to bird artifacts also increased with height (Wilczak et al. 1995), and hence, the percentage of unrecoverable data increased and the averaged velocity became less and less reliable.

The difference between the old radar estimation and rawinsonde data for the N–S component at night has a distinctly bimodal shape (Fig. 6, top panel) with 87% of values beyond the ±2 m s⁻¹ interval around zero. The distribution of differences for the new procedure has maximum at 1 m s⁻¹; almost half of the cases exhibit an absolute difference less than 2 m s⁻¹. Daytime distributions (Fig. 6, bottom panel) have rather close parameters (width, mean, median, standard deviation, shape). Note that the BLRS data in Figs. 1 and 6 are based on different datasets. All valid values from three days of measurement were used in Fig. 1, whereas Fig. 6 is based on five days of measurement with limitation to layers within 310 to 1600 m and to cases when the sonde and both profiler estimates were available and valid.

In examining the results of the comparison of sonde versus profiler, one should keep in mind the inherent difference between these two techniques: remote sensing results imply considerable time-volume averaging, while a rawinsonde flight provides only a single “grab” sample. Even under the most favorable conditions, the two will not show perfect agreement.

5. Measurements during active bird migration

When no alternative method of measuring the wind profile (such as rawinsondes) is available to verify the presence of passerines, processing the data with both the old and new methods opens the possibility of detecting periods with significant bird contamination of the profiler signal. In favorable conditions (nighttime, migration season, southerly winds in spring or northerly winds in autumn) a difference (new minus old) between the two estimates greater than 3 m s⁻¹ with a sign opposite that of the N–S component is an indication of bird artifacts. The threshold value (3 m s⁻¹) is the sum of the daytime average difference between old and new estimates (0.37 m s⁻¹) plus the standard deviation of this difference (2.73 m s⁻¹); this threshold implies that in daytime the divergence of the two procedures is inherent and is not affected by external causes.

Figure 7 shows a month-long time series of the N–S wind component (new values) and the difference between new and old estimates (henceforward to be referred to as “the difference”). The night periods on the difference curve are highlighted by a heavy line. Data were averaged over the layer at 310–835 m, where the average error of the new procedure is not too large (less then 2 m s⁻¹). Before 7 June, there is a strong distinction between daytime and nighttime values of the difference, except when the wind was northerly (27, 28, 30 May). After 7 June the distinction is not evident; daytime and nighttime difference values are of the same order. This observation could be interpreted to indicate that bird migration was rare after 7 June. Northerly winds are not favorable to migration during spring, so on 27, 28, and 30 May, the differences are small. At the end of the period we see very good migration conditions, which are comparable with those of 21–23 May, with strong southerly winds; however, the difference values indicate no bird artifacts, apparently because the intensive migration period had ended (see also Fig. 3).

Day and night distributions of difference values for the periods before and after 7 June are shown in Fig. 8. Three curves in that figure are similar to each other (both in the daytime periods and the second nighttime period), indicating an absence of bird interference after 7 June. The fourth curve (for nighttime between 19 May and 7 June) has the familiar bimodal shape that implies two distributions mixed together; one consists of uncontaminated cases (northern winds, time around sunrise and sunset), while the other includes periods with significant bird migration. This clear distinction of distribution shapes supports the assumptions about an intense bird migration period and its end.

6. Conclusions

The method presented for processing spectral data of BLRS radar improved the accuracy of wind velocity measurements under the conditions of significant bird artifacts. Comparison with in situ measurements showed good agreement for daytime data (mean difference in the order of 0.2 m s⁻¹ for both components) and considerable improvement over traditional processing for nighttime data (mean difference diminished to 1.16 from 2.43 m s⁻¹ for the east–west component and to −3.21 from −7.35 m s⁻¹ for the N–S component when all heights were included; differences were less than 2 m s⁻¹ for heights less than 800 m). This method can be applied to any wind profiler dataset that includes the averaged spectral data from within averaging periods, such as the ARM Southern Great Plains profiler data prior to November 1996. Application of this method along with the traditional BLRS method enables the detection of periods when migrating birds degraded wind profiler measurements.