UHF (boundary layer) and VHF (troposphere–stratosphere) wind profilers have operated at Christmas Island (2°N, 157°W) in the central equatorial Pacific from 1986 to 2002. Observed profiles of winds are sparse over the tropical oceans, but these are critical for understanding convective organization and the interaction of convection and waves. While the zonal winds below about 10 km have previously shown good agreement with the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (RI), significant differences were found above a height of 10 km that were attributed to the low detectability of the wind signal in the profiler observations. Meridional winds at all levels show less agreement, with differences attributed to errors of representativeness and the sparseness of observations in the region. This paper builds on previous work using the Christmas Island wind profilers and presents the results of reprocessing the 17-yr profiler record with techniques that enhance the detectability of the signal at upper heights. The results are compared with nearby rawinsonde soundings obtained during a special campaign at Christmas Island and the RI, NCEP–Department of Energy (DOE) reanalysis (RII), and the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40). The newly processed profiler zonal and meridional wind observations show good agreement with rawinsonde observations from 0.5 to 19 km above sea level, with difference statistics similar to other studies. There is also significant improvement in the agreement of RI and RII reanalysis and profiler upper-level zonal and meridional winds from previous studies. A comparison of RII and ERA-40 reanalysis shows that difference statistics between the reanalyses are similar in magnitude to differences between the profiler and the individual reanalyses.
The sparseness of in situ wind observations over the tropical oceans makes wind profiler observations crucial for understanding weather and climatological processes as well as for validating forecast/analysis/assimilation models. The potential value of profiler data in improving reanalysis products over the central equatorial Pacific was demonstrated by Gage et al. (1988) who showed that the bias between operational analysis and profiler observations was reduced from 1–3 to 0.5 m s−1 with assimilated profiler observations. The value of assimilated wind profiler data in U.S. regional forecast models has been shown by Benjamin et al. (2004), where profiler data denial studies directly showed improvements from assimilated observations. In this paper we present the results of reprocessing a 13–17-yr record of UHF (boundary layer) and VHF (troposphere–stratosphere) wind profiler observed horizontal winds at Christmas Island (Kiritimati, Republic of Kiribati) in the central equatorial Pacific (2°N, 157°W). The aim of this work has been to remove errors (e.g., sea clutter) from the profiler data, combine UHF and VHF profiler observations, and extend the height coverage so as to provide a long-term dataset that can be used for studies of scales from circulations within individual convective systems to interannual scales over the central equatorial Pacific.
Profilers measure winds by determining the Doppler shift of turbulent refractive index irregularities that are transported by the flow (e.g., Gage 1990). Therefore, the ability of profilers to detect a signal is reduced during periods of low turbulence or in a dry atmosphere. Since the signal return is reduced as a function of range and the concentration of water vapor in the upper troposphere is often low, the detectability of a wind signal at upper levels is often low, limiting height coverage. The quality of profiler observations where the signal return is good has also been explored by a number of authors including May (1993) who compared VHF profiler winds with radiosondes in the Tropics and found RMS differences of 2.3 m s−1 (below 10 km). Most of this difference was attributed to variability of the wind field (e.g., Jasperson 1982). Adachi et al. (2005) found difference standard deviations of 1.2 m s−1 when comparing a boundary layer profiler to a collocated tower. Strauch et al. (1987) compared 1 month of winds obtained from coplanar–independent profiler beams and found difference standard deviations of about 0.9–1.3 m s−1. These were similar to those found in coplanar beam comparisons of an 8-yr VHF Christmas Island dataset by Schafer et al. (2003) below about 5 km; however, above this level the differences steadily increased. This may have been associated with the decreasing signal detectability with altitude or sea clutter.
Schafer et al. (2003) investigated the quality of wind profiler observations using standard processing (described in the next section), and compared these to the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (RI). Based on coplanar beam comparisons, they found that wind profiler observations at Christmas Island were reliable only to about 10–12 km above sea level. Above this, low signal detectability impacted the results. At lower levels where the signal was determined reliable, there were differences with the reanalysis that were attributed to errors of representativeness and the lack of observational data available for assimilation over the tropical Pacific region. The largest differences were found in the meridional winds where low-level circulation was often in the opposite direction, with the vertical structure of the meridional circulation weaker in the reanalysis. Despite the good quality of profiler winds particularly at levels below about 10 km, reanalysis/assimilation models may still reject profiler data as errors of representativeness if they differ too much from the first-guess model (e.g., Kalnay et al. 1996; Boutier 2001). To address the low signal detectability at upper wind profiler range gates, Schafer et al. (2004) developed a coplanar spectral averaging (CSA) technique to be described in the next section. This method attempts to improve the signal detectability and spatial characteristics of the data by averaging wind profiler spectra from coplanar beams, and was shown to improve the height coverage of a UHF profiler and the agreement between the UHF profiler and VHF profiler located at Christmas Island.
The reprocessing of Christmas Island UHF and VHF data has utilized new techniques for increasing the signal detectability (using CSA applied to both UHF and VHF profilers), removing sea clutter, and improving quality control. In the next section we will discuss the processing of the 17-yr wind profiler observation record. In section 3 processed horizontal winds are first compared to rawinsonde data taken by NCAR Atmospheric Technology Division (ATD) in July and August 1994 to show the improvement in height coverage. This is followed in section 4 by a comparison with RI (Kalnay et al. 1996), NCEP–DOE reanalysis (RII; Kanamitsu et al. 2002), 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40), and an extension of Schafer et al. (2003) to profiler–reanalysis observations above 12 km and below 2 km to investigate impacts of the new processing. Our conclusions are summarized in section 5.
2. Methods and data
a. Profiler location and data
Christmas Island is located in a data-sparse region of the tropical Pacific at 2°N, 157°W. Figure 1 shows the locations of Christmas Island and rawinsonde launch sites within ±30° latitude and ±40° longitude of Christmas Island along with the reanalysis data grid. The closest rawinsonde site is Penrhyn over 1000 km south of Christmas Island. Collocated VHF and UHF wind profilers began continuous observation of winds and reflectivity at Christmas Island in April 1986 and March 1990, respectively, until October 2002 (Gage et al. 1994; Carter et al. 1995). Other than widely spaced rawinsondes which typically produce only one or two soundings per day, the largest source of wind data over the Pacific are the Quick Scatterometer (QuikSCAT) surface winds and satellite temperature and cloud drift wind vector data, which provide a critical source of upper-level data. While QuikSCAT data have been shown to agree well with Tropical Atmosphere Ocean (TAO) buoys in the Pacific region (Ebuchi et al. 2002) they are currently not assimilated into the reanalyses.
Wind profilers operate continuously measuring winds every few minutes, providing much higher temporal resolution than most observing platforms. Horizontal winds are measured using between two and five radial-directed beams (e.g., Strauch et al. 1984). Typically one beam is directed vertically and at least two other oblique beams are directed about 15° from the vertical in a zonal and meridional direction (e.g., directed 15° from the vertical to the east and to the north). Using this geometry the horizontal winds can be estimated from two oblique beams assuming zero vertical velocity or from three beams by accounting for vertical velocity using the vertical beam. By adding antenna-phase-shifting capability, many profilers have been upgraded to make measurements in five beam directions, with one vertical and four oblique beams directed, for example, to the north, south, east, and west. Using these systems, horizontal winds can be estimated using only the oblique coplanar beams, where the geometry automatically corrects for vertical velocity. Adachi et al. (2005) showed that oblique coplanar beam measurements produced better estimates in the boundary layer than those that used the vertical beam for vertical velocity correction because of the large spatial variability of vertical velocity. In addition problems with the radar system can be diagnosed by comparing coplanar beams, where winds measured by the independent oblique coplanar beams should mirror each other (Schafer et al. 2003). UHF wind profilers are commonly configured to measure winds from about 200 m above ground level (AGL) to about 5 km AGL while VHF profilers are configured to measure winds from about 2 to 20 km AGL, but the actual ranges of good data depend on system characteristics and atmospheric conditions. The most common processing applied to profiler observations is “standard processing,” which involves the calculation of moments from each beam’s radial velocity spectra after each dwell. Typically about six spectra are averaged by the system during each dwell (incoherent averaging). These radial wind estimates are then passed to a consensus routine that performs averaging and quality control by finding the largest group of estimates over the averaging period that are within a specified deviation of each other and then averaging these (Strauch et al. 1984). The consensus routine removes outliers, but the algorithm generally does not increase signal detectability and therefore may not extract all potentially available wind estimates at upper heights or when atmospheric scattering/reflection is weak (e.g., during dry atmospheric or low turbulence conditions). As a final step radial velocities from the beams are combined with appropriate scaling to give horizontal wind components.
Each profiler site has the option of storing wind spectra and/or moments. In the past this has been a decision based on data storage and transport capacity. Advantages of storing spectra include the identification of contamination of the wind observations by sea clutter and radio frequency interference, and the reprocessing of observations as new algorithms are developed. Aspects of the recorded data and subsequent processing techniques (described in the next sections) are shown in Table 1.
A number of processing strategies have been developed that improve data quality and recovery including statistical averaging of spectra from individual beams to gain the benefit of spectral averaging while also rejecting intermittent clutter (Merritt 1995), hardware improvements that enhance the processing of the raw time signal combined with time–height pattern recognition (Wilfong et al. 1999), and an objective analysis technique that addressed both quality control and missing data Carr et al. (1995). Other methods have focused on quality control using fuzzy logic (e.g., Cornman et al. 1998) and automated/user intervention quality control (e.g., Lambert et al. 2003). Schafer et al. (2004) developed a CSA technique combined with fuzzy logic quality control and applied this to a UHF boundary layer profiler to increase signal detectability. CSA involves averaging spectra over time from coplanar oblique beams (e.g., east and west beams). It was shown that CSA directly improved the number of good retrievals when compared to standard processing or averaging of spectra from single beams, both with and without subsequent quality control. CSA improved the agreement between collocated UHF and VHF profiler observations by increasing the height coverage of the UHF profiler, increasing the number of good velocity retrievals at each height, and reducing differences associated with spatial variability. Spectral averaging (either in post processing or during system incoherent averaging) and CSA generally do not increase SNR. They increase signal detectability by reducing the noise variability, thus, removing noise spikes that may otherwise overwhelm the signal.
b. VHF processing
Since CSA has been shown to directly improve height coverage and quality of velocity retrievals, CSA was used where five beam spectral observations were available. The method (as applied in this work) involves collecting 1 h of spectral data from both the west beam and the east beam (with each east beam spectrum reversed about the zero velocity axis). A median 1-h combined spectrum is then created and after five-point smoothing, the moments are calculated from the west–east combined spectrum. Scaling by the cosine of the beam elevation angle gives the hourly zonal velocity estimate. The same process is repeated for the south and north beams. The vertical beam is not used in this method as the geometry of the coplanar method inherently corrects for vertical velocity. Over 1 h, this method combines 14 spectra from coplanar beams, which consist of an average of 6 system-averaged spectra.
If only three-beam spectral observations were available, hourly horizontal winds were estimated from hourly median spectra of the oblique beams (SA; e.g., the hourly east beam spectra and hourly north beam spectra). As in the case for CSA processing five-point smoothing was applied to the spectra and the moments were estimated and scaled. Over 1 h, 11 spectra from an individual beam are combined, which consist of an average of 13 system-averaged spectra. Vertical velocity was neglected in this processing as hourly vertical velocities were found to be small (less than about 0.07 m s−1) and because spatial variability of the vertical velocity may not provide a good correction. In years where only moments were stored, and observations were only made in three beam directions, hourly wind estimates were calculated using hourly medians (MA) of the radial velocity estimates. In each processing case of CSA, SA, and MA, sliding hourly averages were created such that hourly estimates were spaced at the approximate 5–10-min period taken for the radar to cycle through all beam-pointing directions.
Sea clutter from beam sidelobes was found to be the most persistent problem in obtaining clear-air moments from the VHF spectra (Balsley et al. 1988). This is a problem that can potentially be made worse through spectral averaging as there is usually very little variation in the location of the sea-clutter peak in the velocity spectra. The VHF wind profiler observed signal is strongest at Bragg scatter wavelengths of about 3 m. This corresponds to an ocean wave velocity of about 2.1 m s−1. Two peaks appear in the velocity spectrum at about ±2.1 m s−1. The magnitude of the peaks is rarely the same, and while the spacing between the peaks should remain constant, there is a shift about the expected velocity value because the 3-m wavelength waves are superimposed on the mean ocean current. It was shown by Balsley et al. (1987) that the return from the sidelobes could be used to measure the direction and speed of the sea currents. When scaled by the cosine of the elevation angle the peaks in the spectrum occur at approximately ±10 m s−1. Examination of the time series shows that these velocities are also valid and common horizontal air velocities. In addition, when the time series showed a valid smooth wind transition to about 10 m s−1, the sea-clutter peak at −10 m s−1 also increased in power, often overwhelming the clear-air peak. The same occurred for valid wind velocity transitions in the opposite direction, yielding a positive feedback when wind velocity was similar to the wave velocity. The primary difference between these peaks is their width. The clear-air peak in the spectrum is wider than the sea-clutter peak because the air velocity varies more than the sea current at hourly time scales. Interpolation across the spectral peak was found to be a successful strategy for removing the clutter and preserving the clear-air velocity. If a velocity peak was found near the expected sea-clutter peak, a three-point interpolation in the spectra was performed centered on this peak (corresponding to a spectral velocity width of 0.7 m s−1). If the peak was narrow (e.g., associated with clutter), it was significantly attenuated. If it was broad (associated with clear-air velocity), there was almost no impact on signal power and the velocity was retained.
Since user-assignable operating parameters changed over the years of observation due to radar modifications, or improved knowledge of the required observations, the height spacing and maximum and minimum ranges are not constant over the observation period. Therefore, in order to create a dataset on a uniform height grid, a set of observation heights was chosen from 1998, where the profiler operating parameters were set to observe the largest height range (1.7–22.3 km at 480-m intervals). The closest observation within 0.5 km of each specified height was assigned to that height, that is, observations were assigned to specific heights and not interpolated. If at any time during the observation period there was no observation within 500 m of a particular height, that observation at that height and time was assigned a missing value flag.
c. UHF processing
The UHF profiler operated in two alternating observing modes. A low mode used pulse lengths of about 100 m and was configured to make observations from about 300 m to 5 km ASL, while the high mode used a pulse length of between 400 and 525 m and was configured to observe from about 400 m to 11 km ASL. The lowest heights in the low-mode observations often showed random peaks that could not be associated with clear air, while above 1 km, the high-mode observations produced better velocity estimates because of higher power in a longer pulse. As a result, only the high-mode observations were used in this study. Because the UHF and VHF profilers observe different pulse volumes, differences in the VHF and UHF observations may be related to differences in the size of the observed volumes. To create spectra more representative of the larger pulse length of the VHF observations, all median UHF spectra are calculated over 1000-m range bins at 200-m intervals beginning at 350 m. This approximates the 1-km pulse length of the VHF profiler.
A number of problems were discovered with the UHF observations, most of which could be corrected. Radio frequency interference (RFI) was observed in the UHF spectra. These RFI peaks occurred outside the normal range of air velocity and so could easily be removed from the spectra through interpolation across the affected velocity range. Erroneous velocities associated with failing hardware were detected and removed by identifying large differences in average signal power between north and east beams or by comparison of coplanar beams during a five-beam operation. After processing spectra over height and for interference, hourly zonal and meridional wind estimates were calculated in the same way as described for three- and four-beam VHF spectral processing. Vertical velocities were again neglected (during the period of three-beam operation) as the profiler appeared to have problems estimating them. The spectral processing and cluster algorithm rejected most observations that were affected by precipitation and the hourly averaged vertical velocity is expected to be small because of the lack of any significant terrain.
d. Quality control of VHF and UHF data using a cluster algorithm
Once the hourly horizontal winds were estimated and assigned to appropriate heights, quality control was facilitated using a cluster algorithm. A cluster algorithm was developed that checked every hourly wind estimate at every height. If observations over a day clustered into a velocity range that was small compared to all possible velocities (i.e., the full scale of velocities that could be expected if moments were calculated using noise), it was assumed that a signal was present on that day. The cluster algorithm depends on four parameters (described in Fig. 2), and compared each hourly wind estimate to all hourly winds for a day centered on the observation time at the same height and the two heights immediately above and immediately below. If more than 25% of the hourly estimates in the time–height block (1 day over 5 height ranges) were within ±8 m s−1 of the estimate being checked, then the estimate was accepted, otherwise it was marked as a missing value. Each estimate was checked independently of all others, such that an estimate falling into the comparison block that was previously rejected was still included in subsequent comparisons. Based on the velocity range in the velocity spectra, only about 18% of observations would be expected to be within ±8 m s−1 of each other if they represented random noise. This cluster algorithm, which is based only on velocity, was found to be the most effective quality control because large variability in signal strength associated with atmospheric conditions and instrumental drift made other quality control methods such as fuzzy logic or thresholding ineffective. This method extracted mostly usable data even in periods where the signal was intermittent at high altitudes.
The parameters for the cluster algorithm were chosen by visually examining the performance on time series at heights where there was significantly more spread in measured velocities because of low or no detectable signal. The cluster algorithm was found to extract a clear signal in the time series that was realistic when compared to time series at lower altitudes and when compared to the known seasonal cycle. Sensitivity to the choice parameters was examined by choosing two criteria that could be used to evaluate the performance against both the RII and ERA-40. The first is that the algorithm should give the lowest zonal and meridional RMS difference between the reanalyses and profiler winds, while the second criterion was that the algorithm should discard the least amount of data. We emphasize that this step in no way directly incorporates information from the reanalysis, but instead serves to minimize differences between them, while retaining the maximum amount of data. The reanalysis was chosen as “truth” because there are no other long-term time–height observations near Christmas Island covering at least 1 yr and all seasons. The distance to the nearest rawinsondes is so large that the regions are governed by different seasonal cycles and dynamical processes, making comparisons difficult. In addition it was shown by Schafer et al. (2003) that the profiler and RI zonal winds below 10 km do agree reasonably well.
Seventeen cluster algorithm parameter sets were run on 1 yr of data from 1994 (Fig. 2). Also shown are processing results using the cluster parameter set 8 followed by quality control using the five-beam verification described in the next section (test 0). The zonal and meridional RMS differences were calculated between the reanalyses and profiler for all heights from 2 to 19 km over all values accepted by the cluster algorithm for the year, giving 17 values of zonal RMS difference (ZRMS) and 17 values of meridional RMS difference (MRMS). The percentage of data points rejected (NR) for each cluster algorithm parameter set was recorded. These 17 sets of 3 values were then normalized between 0 and 1 and plotted in Fig. 2 (for RII only). It can be seen that the lowest number of values rejected and the largest RMS difference occurs for parameter set 1 (loosest criteria), while the maximum number of points rejected (almost 100%) and lowest RMS difference occurs for parameter set 6 (tightest criteria). To determine the optimum set, the normalized values were combined. The zonal and meridional RMS values were combined into a single value ZMRMS = (ZRMS + MRMS)/2 and then combined with the rejection value (ZMRMS + NR)/2. This curve is plotted for both RII and ERA-40 and the optimum parameters were considered to be those where this curve was a minimum (parameter sets 7, 8, and 12). Parameter set 8 used in the subsequent processing is shown to minimize RMS difference and rejection rates while only producing slightly larger RMS difference than the stricter five-beam quality control. As a final step, a 3-h sliding median sampled hourly is applied to the data to remove remaining outliers.
The improvement in height coverage and quality of the meridional winds using the new CSA–cluster processing is shown in Fig. 3. The standard processing used by Schafer et al. (2003) and data publicly available through the National Oceanic and Atmospheric Administration (NOAA) provide height coverage to only about 10–12 km because of low signal detectability at upper levels. The new processing with spectral averaging and cluster processing extends the observations to 19 km. The winds in both standard and new processing methods are consistent with each other below about 12 km, while the upper-level winds retrieved using the new processing are consistent with known atmospheric processes. For example a 4–5-day period mixed Rossby gravity wave structure is shown from about 25 July to 6 August with upward phase propagation below 13 km and a reversal of phase propagation near 13 km, consistent with the observed structure of this wave (e.g., Wheeler et al. 2000). While spectral averaging was used to improve the height coverage of VHF winds from 1991 to 2002, only moments were available from 1986 to 1990 (spectra were not saved). Despite this, the profiler had excellent height coverage during this period as it was operating in a higher power mode with a longer pulse and longer dwell that, while requiring more time to sample each beam, did not appear to impact the overall horizontal wind temporal resolution as it operated in a three-beam mode.
3. Comparison of wind profiler and rawinsonde
During July and August 1994 NCAR ATD launched about 60 Vaisala RS80–15N rawinsondes at Christmas Island. Of these launches, only 41 reached an altitude of 15 km or more. Winds were determined using Omega tracking with 240-s smoothing. Launch time and the time interval between launch to a specified altitude were recorded with the winds, so the closest wind profiler observation in time and height was matched to each rawinsonde observation and difference statistics were calculated. Acheson (1974) investigated sources of error with omega wind finding prior to the Global Atmospheric Research Program (GARP) Atlantic Tropical Experiment (GATE) and found that the accuracy of Omega tracking can be variable, with errors of 0.5–2 m s−1 depending on atmospheric signal propagation and the location of ground stations.
The zonal and meridional winds for the 41 soundings that reached at least 15 km and the corresponding wind profiler winds are shown in Fig. 4. These correspond to the vertical lines marked in Fig. 3, giving an overview of the profiler observed winds during this period. June and July correspond to a transition period marking a weakening of the upper-level westerlies (above about 6 km). This is shown in both the profiler and rawinsonde winds with relatively low wind speeds above 12 km. The meridional winds during this period suggest a local Hadley circulation with low-level southerlies becoming northerlies above about 5 km.
Using all available observations from the rawinsonde launches, difference statistics were calculated for selected height ranges and are shown in Table 2. About 300 or more comparison points were used in each range, when both profiler and rawinsonde observations were available. The largest difference standard deviation (although with a relatively small median bias) occurs from 14 to 19 km and this may be a result of the large spatial separation between the profiler and rawinsondes at this height and remaining velocity outliers that were not removed by the cluster algorithm. During this period the VHF profiler operated in five-beam mode while the UHF profiler operated in a three-beam mode. This provided an opportunity to test the reliability of the VHF-derived winds above 2 km. Three estimates can be derived for the zonal or meridional winds. For example zonal velocity estimates based on a 1-h median of the east beam spectra, the west beam spectra, and the combined coplanar east and west beam spectra. A condition was specified that if all estimates are within 5 m s−1 the CSA estimate is considered reliable otherwise it is rejected. This additional quality control produces no changes in the statistics from 2 to 10 km where the signal return is generally good. However there is a small improvement in the 10–14-km range and larger improvement in the 14–19-km range. The large change in mean zonal wind bias in the 14–19-km range compared to median bias suggests that some outliers are not removed by the cluster algorithm. However, five-beam quality control may be too restrictive as good wind estimates from CSA may not be reflected in estimates from the individual beams with lower signal detectability. In addition many wind profilers and the majority of the Christmas Island observations use a three-beam mode, which does not have the redundant measurements. Statistics using standard processing are also shown in Table 2. The CSA combined with cluster processing in general improved the RMS agreement between the profiler and rawinsonde at all heights compared to standard processing with the exception of the 14–19-km range. The largest improvements in quality are in the UHF measurements from 0 to 2 km and in the VHF measurements from 10 to 14 km, where there is a 62% increase in the number of recovered velocities (32% increase using five-beam quality control). While the zonal RMS differences in the 14–19-km range are larger for CSA with cluster quality control than standard processing, the median absolute error (MAE) shows that the scaled error using the standard processing is larger. In addition the MAE is the same for the zonal velocities and smaller when applying CSA and cluster quality control to the meridional winds. These results show that the cluster algorithm performs well below 10 km and is effective at upper levels at rejecting outliers representing no signal. CSA improves the number of good estimates compared to the standard processing when using either the five-beam validation or the cluster algorithm. The median profiler and rawinsonde wind profiles show larger differences near the surface and at about 3 km. These differences are reflected in the 0–2- and 2–7-km difference statistics and may be a result of greater vertical (1000-m pulse length) and temporal (1 h) smoothing of the wind profiler observed winds. Because there is good agreement with the rawinsondes in the 0–10-km range with statistics similar to those of other studies, the profiler winds are shown to be reliable in this height range. The larger RMS difference at heights above 10 km even when the five-beam quality control is applied suggests that differences may be partially attributed to the large horizontal spatial separation of the rawinsonde and profiler as well as smoothing of the profiler winds. The median difference when using the cluster algorithm is smaller than the mean difference and similar in magnitude to the five-beam quality control (both mean and median). This suggests that differences are also associated with remaining outliers. The differences in the zonal winds and meridional winds are within the 0.5–2 m s−1 errors quoted for Omega rawinsondes and about 0.9–1.3 m s−1 (difference standard deviation) using beam-to-beam comparisons with VHF profilers.
Jasperson (1982) studied the space–time variability of wind observations using balloons launched at different times at the same location in Minnesota and at 20-m to 21-km spacing at the same time utilizing a novel self-contained high-resolution tracking system (Gage and Jasperson 1974). Rawinsondes launched at the same location (within 30 min apart) showed differences of 1.7 m s−1 for launches at the same time but when separated by 20 km they showed differences of 2.4 m s−1. Examination of the Christmas Island rawinsonde data shows that above about 3 km the rawinsondes have typically drifted over 20 km from their launch site. Given these values and the accuracies determined by Acheson (1974) the rawinsonde and profiler winds compare quite well, with differences similar to those reported by May (1993).
Statistics are also presented comparing the UHF and VHF wind profiler winds from 1990 to 2002 near 2 km, where the UHF and VHF data overlap (Table 3). As expected, the difference between the profilers is smaller than the difference between the profiler and rawinsonde, since the profilers measure in a fixed location. The difference standard deviation between the two profilers is less than the difference between winds observed at consecutive range gates for the individual profilers, and similar to the coplanar beam measurements of Strauch et al. (1987). This indicates that the systems match each other reasonably well.
4. Comparison with reanalysis
While Schafer et al. (2003) used the RI, this study uses both RI and the updated reanalysis (RII) described in Kanamitsu et al. (2002) and ERA-40. The closest reanalysis grid point to Christmas Island is at 2.5°N, 157°W (Fig. 1). Reanalysis winds at this grid point were interpolated to the same heights as the wind profiler wind observations. Hourly wind profiler observations at 0000, 0600, 1200, and 1800 UTC were then matched to the reanalysis data. Missing data in the profilers were excluded from the reanalysis statistics, to replicate the profiler sampling as closely as possible. Because assimilation systems may reject observations for a number of reasons it is difficult to know precisely what profiler data may have been assimilated. However, the profiler observations were made available over the global telecommunications system (GTS) and using a rawinsonde data format (World Meteorological Organization station identifier 91492). The Integrated Global Radiosonde Archive (IGRA) database includes Christmas Island profiler observations (with gaps) from mid-1994 to the end of 1997 and from 2000 to mid-2001. The limited profiler data in the IGRA database suggests that the profiler can be considered a quasi-independent measure of the quality of the reanalysis over the central equatorial Pacific.
The monthly zonal profiler and RII winds show the annual cycle of expansion and contraction of the zonal circulation driven by convection over the western Pacific and Maritime Continent (Fig. 5; Gage et al. 1996a; Schafer et al. 2003). These results extend the results of Schafer et al. (2003) by including winds below 2 km (UHF in the boundary layer) in addition to the winds above 12 km recovered by the improved processing. The results indicate that the circulation was strongest in the periods from about 1988–90 and 1999–2001 with a weaker cycle in other years. The weaker circulation from 1990–95 was shown by Gage et al. (1996b) to correspond to a prolonged warm event. As discussed in McAfee et al. (1992) and shown in the profiler and reanalysis plots the zonal circulation is weakest in El Niño years 1987, 1992, and 1998 when convection is displaced eastward, weakening the Walker circulation over the central equatorial Pacific. The 17-yr climatology shows the mean seasonal cycle with easterlies below about 6 km and westerlies above from October to June, and deep easterlies from July to September. The peak in the westerlies occurs just below the tropopause, as marked in the profiler composite and determined from the local peak in VHF reflectivity. During July–September, this is also the height range of the weakest easterlies.
Meridional winds are more variable than the zonal winds, and there are marked discrepancies between the profiler and RII meridional winds (Fig. 6). In the profiler, there is an indication of a local Hadley-like circulation, with southerlies below about 2 km and upper-tropospheric northerlies centered on 13 km (Fig. 6b). The profiler climatology (Fig. 6b) also indicates a secondary core of upper-level northerlies centered near the freezing level at around 5-km height throughout most of the year. A thin layer of weak southerlies is also seen during June–October at around 8 km, and these connect to thicker layers of stronger midtropospheric southerlies during the rest of the year. In the RII climatology this midtropospheric circulation is substantially weaker, while the surface southerlies seen in the profiler are weaker or completely missing during parts of the year. A weaker midtropospheric circulation is also observed in the ERA-40 composite with similar differences in the surface winds when compared to profiler (not shown). These findings are supported by results of McNoldy et al. (2004), which showed that July climatologies of NCEP–NCAR reanalysis meridional winds near the surface were significantly weaker than QuikSCAT satellite–derived winds. While ERA-40 also showed weaker equatorial meridional winds, the differences were smaller with improvements attributed to the assimilation of ERS scatterometer data. A complex multilayered meridional wind structure was also observed during the Line Islands Experiment (March–April 1967; Madden and Zipser 1970). Rawinsonde observations from this experiment at Palmyra revealed a structure similar to that shown in the meridional wind March–April climatology with northerlies near the tropopause, near 8 km, and near the surface (Fig. 6). It was suggested that these regions of strong shear may be regions of turbulence and hence significant energy sinks. The structure of the meridional and zonal wind profiles in the RII climatology are consistent with those shown for the original RI in Schafer et al. (2003) for reanalysis RI composites, so both reanalysis versions appear to have similar deficiencies over the central equatorial Pacific.
Analysis of outgoing longwave radiation (OLR) has shown that the circulation during this time of year is directly associated with convection in the ITCZ to the north of Christmas Island. The dominance of southerly low-level flow and upper-tropospheric return flow during most months is presumably due to the fact that the climatological position of the ITCZ remains in the Northern Hemisphere throughout the year (e.g., Waliser and Gautier 1993).
A secondary core of northerly flow above the boundary layer was also observed over the easternmost equatorial Pacific by Zhang et al. (2004) and McGauley et al. (2004), who also pointed out the existence of such a flow in Christmas Island profiler data. The northerlies over the eastern Pacific appear to maximize at around 2–3 km (Zhang et al. 2004; McGauley et al. 2004), as opposed to 4–5 km at Christmas. The cause of the two-mode vertical structure of meridional wind and its potential relationship to deep and shallow convection is the subject of continuing research, but may be related to the vertical structure of diabatic heating in shallow and deep convection and the stratiform decks of the decaying systems.
The smallest difference between the profiler and reanalyses zonal winds occurs in the boundary layer (0–2 km) where only UHF profiler winds have been used (Table 4). This likely reflects the significant increase in observations available for assimilation near the surface including the TAO array (e.g., McPhaden et al. 1998). Differences in this lowest layer may in part be errors of representativeness. For example, despite the fact that Christmas Island is relatively small and flat, there is a diurnal signal representing a sea breeze (not shown). There is significant improvement in the agreement between zonal reanalysis and profiler winds in the 10–14-km height range with a reduction in difference standard deviation of about 2 m s−1 from the previous results of Schafer et al. (2003). The smallest VHF differences occur from 2 to 14 km where the signal detectability is typically good while larger differences occur from 14 km to the lower stratosphere. While the RMS and difference standard deviations appear small for the meridional winds, the MAE shows that these differences are almost equal to the median magnitude of the wind speed. The meridional wind climatologies indicate a significant difference in the distinct layer of upper-level northerlies from December to May just below the tropopause, with profiler winds showing consistency with results from the Line Island experiment. ERA-40 meridional composites show a closer match to the profiler upper-level southerlies during December–February, but the vertical structure during March–May is similar in both reanalyses.
The difference in the meridional winds in particular at low levels suggest differences in reanalysis and profiler estimates of the backing or veering of the wind in the boundary layer. This is a critical issue because of the important uses of reanalyses for large-scale dynamical and atmosphere–ocean coupling studies. Errors in estimating the boundary layer winds will have implications for models that frequently use reanalysis-derived wind stress to force ocean models. Williams et al. (1992) investigated the backing of the wind near the surface in profiler and the ECMWF analysis using one month of UHF profiler data at Christmas Island. They found that while the profiler winds showed backing in the boundary layer, the ECMWF winds veered. It was suggested that these differences may be due to inaccurate modeling of the position of the intertropical convergence zone (ITCZ) or inaccurate representation of cold-air advection associated with cooler sea surface temperatures to the south of Christmas Island. This previous study that only used 1 month of data has been extended and the new results are presented as a hodograph for the RII and ERA-40 reananalyses and profiler observations from 1990 to 2002 (Fig. 7). In contrast to the Williams et al. (1992) study, all wind datasets show backing in the lowest 2 km. However, the profiler, RII, and ERA-40 winds show backing angles of −13° and −3° and −6°, respectively, suggesting that the reanalyses underestimate the backing of the winds in the boundary layer. The results become more variable if individual months are compared (not shown). For example observations and reanalysis using only March 1991 show backing profiler winds and veering reanalysis winds in agreement with the differences shown by Williams et al. (1992) for ECMWF analysis in March 1990. Seasonal variations of these differences between profiler and reanalysis winds will be examined in future studies. The profiler shows better agreement with ERA-40 at all height levels. In the boundary layer, this may be a result of the ERA-40 assimilating ERS-1 and ERS-2 scatterometer winds, which are not assimilated by RI or RII consistent with the results of McNoldy et al. (2004). It is also shown that the differences between RII and ERA-40 are similar in magnitude to the difference between the profiler and reanalyses.
This paper describes new processing methods that have been applied to a relatively long archive of UHF and VHF wind profiler observations at Christmas Island. These UHF and VHF profilers have operated almost continuously for 13 and 17 yr, respectively, in a data-sparse area of the central equatorial Pacific. This is the first time that Christmas Island winds have been compared to rawinsonde observations and the first time that the full merged dataset of UHF and VHF winds from 500 m to 19 km ASL has been compared to RI, R2, and ERA-40. New processing involving coplanar spectral averaging, a cluster algorithm, and improved detection of sea clutter and other contamination sources have improved the signal detectability, spatially smoothed the wind field, and significantly improved the height coverage of the profiler data. While CSA improves the quality of observations and number of good observations, the quality control step is critical for identifying these good observations and rejecting outliers associated with failed signal detection. The final gridded dataset consists of hourly winds with 480-m vertical resolution in the 2–19-km height range (covering a period from 1986 to 2002), and 200-m vertical resolution in the 500-m–2-km height range (covering a period from 1991 to 2002).
The processing improved the agreement between the upper-level reanalysis and profiler winds. However, relatively large differences between the reanalysis and profiler meridional winds remain with difference standard deviations having equivalent size to the magnitude of the meridional winds. While there are similarities in the mean structure of the meridional winds, the strength and timing of the circulations appear to be underestimated in the reanalysis. The most significant differences, however, are in the lowest level where profiler and reanalysis winds often have opposite direction. Mean seasonal zonal wind patterns are, however, clearly represented in the profiler and reanalysis zonal winds, but there are differences in the strength of the mean winds in particular in the layer of maximum westerlies of the zonal circulation where profiler observed winds are stronger than the reanalysis.
A short rawinsonde campaign in July and August 1994 conducted by NCAR ATD provided an opportunity to evaluate the new processing. This was considered particularly valuable for the observation levels that had not previously been compared against reanalysis or other instruments (i.e., the lowest layer using UHF profiler observations and the upper VHF observation heights above 12 km). The newly processed winds agree well with observations from a short rawinsonde campaign. However, there were larger differences at heights above 14 km. Use of redundant wind information obtained using five-beam observations suggests that much of this difference is associated with the large horizontal spatial separation between the profiler and rawinsonde at these heights and the smoothing of the wind profiler observations, with a smaller contribution from outliers that were not removed by the cluster quality control algorithm. The difference standard deviations are similar to those found in other studies and have a magnitude similar to that expected as a result of spatial wind variability and the accuracy of rawinsonde omega winds versus profiler observations.
The profiler winds compare well with rawinsonde measurements and the zonal winds compare well to the reanalysis giving confidence in the newly processed winds. The new processing techniques have produced a wind dataset that can be used to investigate time scales from circulations in individual thunderstorm cells to long time-scale climatological changes (e.g., those due to the Southern Oscillation). This work has also highlighted a number of features in the profiler data that appear to be underestimated in the reanalyses. One important example is the complex vertical structure of the meridional flow that may be directly related to the nature of convection to the north of Christmas Island, rather than a single overturning cell having a simple two-layered flow structure. Profiler observations can be used to validate the various reanalyses products, but more importantly they are critical for studying the boundary layer and troposphere because of their much higher temporal resolution compared to reanalysis or rawinsonde. Continuing work will look at the structure and evolution of convective systems over multiple time scales and how this may dynamically force the climatological structure of the meridional flow.
Support for this work was provided by NSF Grant ATM 0116178. NCEP–NCAR reanalysis data were provided by the CIRES Climate diagnostics Center, Boulder, Colorado. Rawinsonde data were provided by NCAR ATD. Unprocessed profiler data were provided by the NOAA/Earth System Research Laboratory (ESRL). The authors thank David Carter and Paul Johnston for their contributions to the archival and quality control of the profiler data, helpful discussions on profiler data, and the maintenance of the profiler systems. The Christmas Island profiler was supported by the NOAA/Office of Global Programs.
Corresponding author address: Robert Schafer, CDM Optics, 4001 Discovery Dr., Suite 130, Boulder, CO 80303-7816. Email: Robert.Schafer@cdm-optics.com