This paper describes a method of absolutely calibrating and routinely monitoring the reflectivity calibration from a scanning weather radar using a vertically profiling radar that has been absolutely calibrated using a collocated surface disdrometer. The three instruments have different temporal and spatial resolutions, and the concept of upscaling is used to relate the small resolution volume disdrometer observations with the large resolution volume scanning radar observations. This study uses observations collected from a surface disdrometer, two profiling radars, and the National Weather Service (NWS) Weather Surveillance Radar-1988 Doppler (WSR-88D) scanning weather radar during the Texas–Florida Underflight-phase B (TEFLUN-B) ground validation field campaign held in central Florida during August and September 1998.
The statistics from the 2062 matched profiling and scanning radar observations during this 2-month period indicate that the WSR-88D radar had a reflectivity 0.7 dBZ higher than the disdrometer-calibrated profiler, the standard deviation was 2.4 dBZ, and the 95% confidence interval was 0.1 dBZ. This study implies that although there is large variability between individual matched observations, the precision of a series of observations is good, allowing meaningful comparisons useful for calibration and monitoring.
The development in the early 1990s of a fully operational national network of Weather Surveillance Radar-1988 Doppler (WSR-88D) radars, also known as the Next Generation Radar (NEXRAD) network (Heiss et al. 1990; Baer 1991; Crum et al. 1993, 1998), has paved the way for a major advancement in the ability to provide quantitative precipitation estimates (QPEs) over most of the continental United States (Serafin and Wilson 2000; Smith et al. 1996). While this capability has led to a major improvement in operational nowcasting and short-range forecasting, it is widely recognized that there is still room for substantially improving the retrieval of precipitation estimates from WSR-88D radars (Glitto and Choy 1997; Baeck and Smith 1998; Serafin and Wilson 2000). Among the principal factors that affect the quality of QPEs are radar calibration, the variability of the vertical profile of reflectivity (VPR), range-dependent sampling issues, and the empirical Z–R relationships used to relate radar reflectivity to rain rate. The Z–R relations have been found to vary substantially depending on the variability of drop size distributions (DSDs) within the evolution of the storm (Brandes et al. 1999) and from location to location (Austin 1987; Atlas et al. 1995). The use of polarimetric radar will improve QPEs (Brandes et al. 2002), yet polarimetric observations will not be operationally available for several years to come.
Clearly, given the current reliance on the Z–R relationship to retrieve precipitation estimates from observed reflectivity, it is essential that the scanning radar be well calibrated. The calibration of the individual WSR-88D radars has recently been shown to vary as much as a few dBZ in relation to the National Aeronautics and Space Administration (NASA) Tropical Rainfall Measuring Mission (TRMM) precipitation radar (PR) (Atlas 2002; Anagnostou et al. 2001; Bolen and Chandrasekar 2000). Furthermore, comparisons of reflectivities from adjacent WSR-88D radar reflectivities in the contiguous United States differ by as much as a few decibels (Smith et al. 1996). Such differences can lead to substantial systematic errors in precipitation estimates (Ulbrich and Lee 1999). Consistent calibration between adjacent network radars is also crucial for constructing radar mosaics used in real-time weather forecasting. In this paper, we focus attention on the radar calibration issue and develop a methodology for using profilers to calibrate scanning radars and to monitor their calibration. The use of profilers to calibrate scanning radars was first outlined in Gage et al. (2000), and this paper develops a methodology that produces statistics from 2 months of profiling and scanning radar observations.
While sophisticated approaches are currently used to calibrate each WSR-88D radar, the calibration problem remains to be fully solved (Atlas 2002). There are many techniques in use today for verification of WSR-88D radar calibrations, including routine engineering measurements of the various subsystems (Serafin and Wilson 2000), and a very different approach that combines the radar calibration and Z–R relationships through the comparison of radar-derived rain maps with dense networks of rain gauges (e.g., Stellman et al. 2001; Anagnostou et al. 1999; Smith et al. 1996). Along the lines of this paper, direct comparisons of reflectivities from the WSR-88D radars with surface Joss–Waldvogel disdrometers (JWDs) have been utilized by Martner (1977) and Ulbrich and Miller (2001). However, the volume and space–time mismatches between the surface disdrometers and scanning radars add large uncertainties to these comparisons. Further, sampling issues that lead to underestimation of DSD and rainfall by disdrometers when only a few raindrops are detected are discussed by Smith et al. (1993).
In this study, we utilize a vertically pointing radar profiler to provide a statistical transfer standard between the disdrometer and the scanning radar. The surface disdrometer is used to calibrate the collocated profiler, and the calibrated profiler provides a reflectivity transfer standard at multiple altitudes that may be compared directly and routinely with the reflectivities observed by a nearby scanning radar observing over the disdrometer/profiler site.
The fundamental problem with comparing reflectivities from the surface disdrometer, the profiling radar, and the scanning radar is that all three instruments have different spatial and temporal resolutions. This would not be a problem if the precipitation was not varying in time and space. But since precipitation is variable at different temporal and spatial resolutions, the uncertainties of the QPE from each instrument will be a combination of the instrument measurement error and the spatiotemporal variability of the precipitation.
This paper presents a comparison of scanning radar observations from the WSR-88D radar in Melbourne, Florida (KMLB), with those from two profiling radars deployed in central Florida during the TRMM Texas–Florida Underflight-phase B (TEFLUN-B) ground validation field campaign in August and September 1998. The KMLB radar observations have been used by NASA for TRMM ground validation since the launch of the TRMM satellite in November 1997 and are the subject of several related studies (Anagnostou and Krajewski 1998, 1999a, b). The results discussed here confirm the study of Anagnostou et al. (2001), finding that the KMLB radar was well calibrated during the TEFLUN-B campaign. The statistics from this study also indicate that while the variability between the individually matched observations is large, the precision of rigorously simultaneous time series means is quite good and allows meaningful comparison. In other words, finding the mean of rigorously simultaneous reflectivity observations filters out the variability due to both instrument error and the natural variation of precipitation fields sufficiently that calibrations are possible.
This paper is organized as follows. The disdrometer, profiling radars, and scanning radar datasets are introduced in section 2. The profiling radar measurement error is determined in section 3. The profiling radar reflectivity is absolutely calibrated using the surface disdrometer observations in section 4. Sections 5 and 6 develop the methodology for using the profiler to calibrate and monitor the scanning radar reflectivity. The discussion of results and conclusions is in section 7.
This paper presents observations from four different instruments: one surface disdrometer, two vertically pointing profiling radars, and one WSR-88D scanning radar.
a. Disdrometer observations
The observations from a model RD-69 surface disdrometer manufactured by Distromet, Inc. (Switzerland), recorded the precipitation reaching the surface. This disdrometer is also commonly known as the Joss–Waldvogel disdrometer in honor of its inventors (Joss and Waldvogel 1969). The disdrometer essentially measures the momentum of the raindrops hitting the 50-cm2 sensor head and converts these individual impacts into raindrop diameter estimates from 0.3 to 5.3 mm in 127 nonuniform size intervals [see Sheppard and Joe (1994) and the work referenced in that study for more details about the disdrometer operation]. The counts in the raindrop diameter size bins are accumulated over 1-min intervals to yield minute-by-minute DSDs, which are converted into estimates of the surface precipitation radar reflectivity.
b. Profiler observations
Two vertically pointing profiling radars were deployed by the National Oceanic and Atmospheric Administration (NOAA) Aeronomy Laboratory in support of TEFLUN-B. One operated at 915 MHz and the other at 2835 MHz. The two stationary, dish antennas were placed next to each other, and the observations were synchronized to ensure that the two radar systems were observing approximately the same atmospheric volumes at the same time. Both radar systems had 5° beamwidths and had the same vertical (105 m) and temporal resolution (60 s). Figure 1 shows a photograph of the surface disdrometers and profiling radars during the TEFLUN-B campaign, and Table 1 lists the system parameters.
c. Scanning weather radar observations
The NWS operates a WSR-88D radar (KMLB) near Melbourne, Florida, which is on the east coast of Florida and south of Cape Canaveral. The KMLB reflectivity used in this study was extracted from the Level II Archive WSR-88D radar data publicly accessible via the NOAA National Climatic Data Center (NCDC) archive (www.ncdc.noaa.gov). Descriptions of the WSR-88D radar operating conditions and NEXRAD operations can be found in Crum et al. (1993), Crum and Alberty (1993), and Crum et al. (1998). The operational Level II Archive data product consists of reflectivity, mean radial Doppler velocity, and spectral width at each measurement cell defined by the coordinates of time, range gate, azimuth angle, and elevation angle. These data were screened to extract only the closest observation directly over the profiler site for each elevation angle without any interpolation. Table 2 lists the scanning radar system operating parameters.
3. Uncertainties in profiler reflectivity estimates
It is very difficult to determine the error characteristics of instruments that measure precipitation parameters. This difficulty is largely due to the variability of precipitation in space and time. To quantify the precipitation error characteristics for improved QPE, it is informative to break the error characteristics down into factors related to the instrument technology and factors related to the precipitation processes. If Zobserved and Ztrue represent the observed and true reflectivity, the error characteristics can be written as (Bringi and Chandrasekar 2001)
The error characteristics can be quantified using the variance statistic
The representativeness error represents the precipitation spatiotemporal variation. It is very difficult to quantify the representativeness error because discretely sampling moving precipitation systems scales the horizontal structure of the precipitation (Fabry 1996). More work is needed to quantify and understand this error for QPE purposes.
The measurement error quantifies the limitations of the instrument and is more tractable. Here we quantify it by observing the same volume of precipitation with two similar radars. The two instruments have the same sampling volumes and dwell times so that the representativeness error is nearly identical for both instruments. The observed variance of the reflectivity differences is used to quantify the measurement error.
The 915- and 2835-MHz vertically pointing profiling radars were configured to observe the same atmospheric volumes with the same temporal resolution. Figure 2 shows the reflectivity difference between the 2835- and 915-MHz profiler observations for all 105-m resolution observations between 0.5 and 4.0 km during the TEFLUN-B campaign. To ensure that only precipitation observations were analyzed, all used observations had reflectivities greater than 10 dBZ and downward velocities greater than 4 m s−1. These two constraints limit the radar dataset to 74 585 observations and ensure that both radars are identifying the Rayleigh scattering component of the Doppler spectrum [See the work of Ralph (1995) and Williams et al. (2000) to identify hydrometeors in profiler observations.]
The statistical bias, standard deviation, and 95% confidence interval for the profiler reflectivities were estimated using
where n is the number of samples, Z2835 MHz and Z915 MHz are the observed reflectivities from both profilers, and t95% is the 95th percentile of the t distribution with n degrees of freedom (Anagnostou et al. 2001). Figure 2c shows the probability distribution function (PDF) of the 915-MHz reflectivity expressed in dBZ units [10log(mm6 m−3)]. Figure 2a shows the PDF of the reflectivity difference (2835 − 915 MHz). The shape of the PDF is important for interpreting the statistical calculations between the two datasets. The near-Gaussian-shaped PDFs indicate that the datasets are nearly normally distributed. The standard deviation and 95% confidence interval estimates have statistical significance only when the datasets are normally distributed. Since the observed reflectivity PDFs expressed in linear units (mm6 m−3) were not normally distributed (not shown), all statistical calculations in this study were performed with the reflectivity in dBZ units.
The total bias in reflectivity between the two profiling radars was zero because both profilers were absolutely calibrated using the surface disdrometer. Section 4 describes this calibration. The standard deviation of the reflectivity difference was 0.57 dBZ and is shown with straight lines in Fig. 2a. The 95% confidence interval about the mean difference for these 74 585 observations was 0.0034 dBZ. For each 1-dBZ interval of 915-MHz reflectivity the bias and 95% confidence interval were calculated and are shown with box symbols in Fig. 2b.
This analysis provides a quantitative estimate of the measurement error of both profilers. Assuming that both profilers have the same error characteristics, the measurement error variance for either profiler is estimated to be 0.16 dBZ2. This is calculated from var(Z2835 MHz − Z915 MHz) = 0.572 = 2σ2measurement error. Quantifying the measurement error has two implications. First, the measurement error of an individual reflectivity estimate can be quantified using the standard deviation statistic. For these profilers, for example, the uncertainty of an individual reflectivity estimate can be quantified as Zobserved ± σmeasurement error, or ±0.4 dBZ. And second, as is shown later in this paper, the observed variance of reflectivity differences between a profiler and a disdrometer and between a profiler and a scanning radar were much larger than σ2measurement error, verifying the contribution of the representativeness error.
4. Profiler calibration using disdrometer observations
The profilers were absolutely calibrated using the surface disdrometer. While the surface disdrometer measured the raindrops impacting the sensor head, the lowest 105-m radar pulse volume was centered 308 m above the ground. The mismatch in volume size and the displacement of the two sample volumes add to the reflectivity difference variance between the two measurements. To ensure that both instruments were simultaneously observing rain, the observations were screened to require that the profiler had reflectivities greater than 10 dBZ and mean downward Doppler velocities greater than 4 m s−1, while the surface disdrometer was required to have observed at least 25 raindrops in the minute sample. The reflectivity differences between the simultaneous minutes of the 2835-MHz profiler and surface disdrometer observations are shown in Fig. 3 using the format introduced in Fig. 2.
The PDF of profiler minus disdrometer reflectivity difference is shown in the left-hand panel of Fig. 3. The profiler reflectivity was calibrated so that the mean reflectivity difference was zero. The standard deviation of the reflectivity difference of 2.84 dBZ (shown as straight lines in Fig. 3a) is much larger than the profiler-to-profiler reflectivity difference standard deviation of 0.57 dBZ estimated in section 3. The increased reflectivity difference variance is due to the different technologies of the two instruments, the different sampling volumes, and the physical separation of the two sampling volumes. The only consequence, however, is that a somewhat larger number of observations are needed to achieve a desired standard error of the mean.
In examining Fig. 3b, it is interesting to note that the bias of the reflectivity difference varies as a function of the JWD reflectivity. The reflectivity difference becomes more negative as the JWD reflectivities increase. While the bias varies with JWD reflectivity, the 95% confidence intervals are insensitive to the JWD reflectivity below 40 dBZ. The bias variation is curious and indicates an unaccounted effect in either the profiler or disdrometer observations. The cause is uncertain. One possibility is that near the surface, convective rains with large reflectivities are also associated with downdrafts due to evaporative cooling (Srivastava 1987). Thus, while the air motion is zero at the surface, the raindrops in a downdraft would be falling faster than their terminal velocities when they impacted the disdrometer sensor. This would cause the disdrometer to overestimate both the raindrop size and the reflectivity. More analysis is needed to validate this hypothesis and to understand the trend in the bias.
5. Calibrating the scanning weather radar
Before the calibrated profiling radar can be used to calibrate and monitor the scanning radar reflectivity, the difference in spatial and temporal sampling of the two instruments needs to be reconciled. The scanning radar observations have larger radar pulse volumes and longer gaps of time between samples than the profiling radars. Even after the profiling radar observations are averaged in altitude to approximate the vertical extent of the scanning radar pulse, and after the radars are matched in time and height, there is still considerable mismatch between the two sampling volumes that will contribute to the calculated reflectivity difference variance.
a. Differences in sampling volumes: Temporal and spatial
During the TEFLUN-B campaign, the KMLB radar sequentially swept through all azimuth angles in approximately 1° increments at 14 elevation angles from 0.5° to 19° above the horizon until a complete volume scan was obtained. During the 5 min required to complete one volume scan, the radar antenna was constantly rotating at a rate of approximately 19° s−1. Thus, the radar dwelled over the profiler for only a fraction of a second during each elevation scan. A WSR-88D radar vertical profile of observations over the profiler site consists of a short dwell time (approximately 50 ms) at each elevation angle, followed 20 s later (after the WSR-88D scans 360° in azimuth) by an observation at the next higher elevation angle. The dwell at any particular elevation is repeated once every 5 min.
For each scanning radar observation over the profiler site, six profiler range gates are averaged to match the vertical extent of the scanning radar observation. At a range of 36 km, the vertical extent of the 0.95° WSR-88D radar beam is 630 m when rounded to a multiple of the profiling radar pulse length of 105 m. The 1-km-long scanning radar pulse length (constructed from averaging four 250-m consecutive radar pulses) implies that the WSR-88D radar sample volume is of the order 3 × 108 m3, which is three orders of magnitude larger than a single profiler sampling volume.
b. Simultaneous profiler and scanning radar observations
During the months of August and September 1998, the surface disdrometer, vertically pointing profilers, and scanning radar were in continuous operation. Figure 4 shows an example of the vertical and temporal resolution of the observed reflectivity over the profiler by the vertically pointing 2835-MHz profiling radar and by the Melbourne WSR-88D radar. Figure 4a shows the profiler reflectivity at the original 105-m vertical and 1-min temporal resolution. The 30 original profiler observations (5 in time and 6 in range) that matched the 5-min time and 630-m altitude resolution of the WSR-88D radar measurement cell were averaged in linear units and are shown in dBZ units in Fig. 4b. The variation of reflectivity within the WSR-88D radar measurement cell is quantified by estimating the standard deviation of the 30 original profiler observations expressed in log units, σ5min,630m, and is shown in Fig. 4d. The Melbourne WSR-88D radar reflectivities for this rain event are shown in Fig. 4c.
Even with the 630-m vertical resolution, the time–height displays of reflectivity from the profiler and scanning radar can be used to diagnose precipitation regimes fairly well. The large reflectivities at low altitudes and the extension of reflectivities to high altitudes (∼10 km) near 0530 UTC are indicators of convective rain. The stratiform rain is identified by the radar bright band that is well defined in the 105-m profiler observations but not so well distinguished in the 630-m observations from 0630 to 0730 UTC.
While Fig. 4 shows the time–altitude comparison between the profiler and scanning radar reflectivity, Fig. 5 shows reflectivity profiles obtained during two different volume scans. A convective (stratiform) profile at 0531 (0636) UTC in Fig. 5a (Fig. 5b) shows the scanning radar reflectivity with filled circles and the five 1-min profiler reflectivity profiles with thin lines. The convective rain profile (Fig. 5a) has more variation in profiler reflectivity than the stratiform rain profile (Fig. 5b) and has reflectivities greater than 10 dBZ extending above 9 km in altitude. The stratiform rain profile has a well-defined radar bright band resolved by the profiling radar but is not resolved by the scanning radar because of the unique combination of tilt angles and range from the scanning radar. The bright band is resolved by the scanning radar in other precipitation events when the bright band is at a lower altitude.
The increased reflectivity variance for the convective rain profile compared with the stratiform rain profile is consistent with our understanding of the natural variations in the precipitation processes (see Houze 1993). Obviously, then, the best comparisons between two instruments that have close but different sampling volumes would occur in stratiform situations, when the precipitation processes have the least amount of spatiotemporal variation. The profiler reflectivity standard deviation, σ5min,630m, over each 5-min and 630-m time–altitude window that corresponds with the WSR-88D radar resolution quantifies the temporal and vertical variation of the reflectivity. Indeed, the σ5min,630m is a fair indicator of precipitation regime. For example, it is larger during the convective region of the storm (0515–0545 UTC) and smaller during the stratiform regime (0600–0715 UTC). Note, however, that σ5min,630m increases near the end of the event from 0730 to 0745 UTC due to gradients near the edges of the storm. In these observations, the radar bright band is clearly identified by large σ5min,630m values due to the change in reflectivity over the 630-m vertical altitude used in calculating σ5min,630m.
c. Profiling and scanning radar reflectivity differences—21 August 1998 case study
The Melbourne WSR-88D radar reflectivities are tested for bias using the calibrated profiler reflectivities from the one rain event of 21 August 1998 shown in Fig. 4. Using only the six lowest tilt angles that correspond to altitudes between 0.4 and 3.4 km to avoid the altitudes that include the melting layer, 179 individual profiler and scanning radar reflectivity differences were obtained and are shown in Fig. 6 versus profiler reflectivity and versus σ5min,630m.
For this single rain event, the observed reflectivity bias between the two radars is 0.44 dBZ, the standard deviation is 2.58 dBZ, and the 95% confidence interval is 0.32 dBZ. Thus, while the calculations indicate large variability in individual profiler to scanning radar comparisons, the precision of the series is good. These calculations are consistent with other profiling to scanning radar comparisons (May et al. 2002). The profiler reflectivity PDF shown in Fig. 6c indicates a peak number of observations with reflectivities near 30 dBZ. The spread in the reflectivity difference near 30 dBZ is approximately plus and minus the standard deviation while there is more variation in reflectivity difference at smaller reflectivities. Figure 6d shows the observed reflectivity differences as a function of σ5min,630m, and Fig. 6e shows the PDF of σ5min,630m. Note that the spread in the reflectivity difference is greater for σ5min,630m ∼ 2 dBZ, suggesting that σ5min,630m can be used to identify observations with smaller reflectivity variance.
To explore the variability quantitatively, four different thresholds of σ5min,630m are determined from the mean plus multiples of the standard deviation of σ5min,630m [e.g., + n std(σ5min,630m), where n = 0, 1, 2, 3]. For each threshold, all observations associated with σ5min,630m less than the threshold are used to estimate the rain event reflectivity difference standard deviation, and the results are plotted with the dashed line using the right-hand axis in Fig. 6e. While the method of estimating the threshold could be refined, this simple threshold still provides an objective method of removing outliers from the rain event population based on a physical measurement.
d. Profiling and scanning radar reflectivity differences—2 months of observations
During the 2-month period of the TEFLUN-B experiment, 21 rain events passed over the profiler site and were observed by the surface disdrometer, profiling radars, and the Melbourne WSR-88D radar. These 21 rain events are summarized in Table 3, which indicates the storm duration and the statistics of simultaneously matched profiling and scanning radar observations. Section 5c shows that large values of the σ5min,630m statistic are associated with outliers. Therefore, all of the rain events were screened to remove all observations with σ5min,630m greater than the mean plus one standard deviation of σ5min,630m. Although this screening only removes about 10% of the observations in each rain event, it successfully removes the largest outliers. The thresholds for each rain event are also listed in Table 3.
The statistics from all 21 rain events are listed near the bottom of Table 3 and are shown in Fig. 7 using the multipanel format introduced in Fig. 6. The reflectivity difference PDF plotted in Fig. 7a shows that the distribution is slightly asymmetric with a bias value of 0.48 dBZ. The reflectivity difference statistics are shown as a function of every 2-dBZ interval of profiler reflectivity in Fig. 7b using the box symbols defined in Figs. 2 and 3. Note that the bias appears independent of profiler reflectivity until about 40 dBZ. After 40 dBZ, the number of observations decreases substantially and the observations are predominately from convective situations, suggesting that the time–volume mismatch is playing an important role.
The reflectivity difference plotted as a function of σ5min,630m in Fig. 7d clearly shows the tendency of the spread to increase as the convective situation becomes more variable. It also shows a tendency for the WSR-88D profiler bias to become more negative as σ5min,630m increases. This tendency may or may not be real, as the number of observations for large σ5min,630m is quite small (see Fig. 7e).
Figures 7d and 7e show that as σ5min,630m increases, the 95% confidence interval broadens and the bias becomes more negative. While each storm has been filtered by the storm σ5min,630m threshold, the thresholds listed in Table 3 indicate that some storms are more variable than others. By examining each rain event it was determined that the events with large σ5min,630m thresholds were convective events that contained short-duration (on the order of 30 min or less) convective cells. Seven of these convective rain events were subjectively screened out, and the remaining 14 rain events were combined to produce the ensemble statistics listed in Table 3 and shown in Fig. 8. Since the largest σ5min,630m threshold of the selected 14 rain events is 5.80, the horizontal axis of Figs. 8d and 8e only extends to 6 dBZ. The reflectivity difference bias, 95% confidence interval of the bias, and the standard deviation for the 14-rain-event dataset are 0.66, 0.09, and 2.36 dB, respectively.
6. Monitoring the scanning radar reflectivity calibration
The analysis presented in section 5 shows how the profiler and scanning radar reflectivities can be combined to calibrate the scanning radar. This calibration procedure can be performed for each rain event (as described in section 5c), or the individual rain events can be combined to produce an ensemble of observations (as described in section 5d). Deviations in the storm-to-storm statistics from the ensemble statistics or trends in the storm-to-storm statistics may indicate that the performance of the scanning or profiling radar is changing with time. If this change in performance is detected by the scanning radar personnel, then they could take appropriate actions in interpreting the precipitation products. Also, they can contact the engineering staff, who would determine the cause of the problem and make the necessary corrections. The benefit of this instrument monitoring methodology is that the WSR-88D radar performance can be tracked from storm to storm. This will provide more information to the radar operators and the confidence that the radar is operating correctly between the periodically scheduled maintenance.
The storm bias and 95% confidence interval for the 21 rain events simultaneously sampled by the profiling and scanning radars are shown in Fig. 9a. The ensemble bias and 95% confidence interval is shown on the right-hand side of Fig. 9a with the total bias shown as a straight line throughout the 2-month period. The bias and 95% confidence interval statistics are shown in Fig. 9b for the 14 rain events that are used to generate the ensemble statistics shown in Fig. 8. The seven rain events identified as having large σ5min,630m thresholds are also the outlier storms shown in Fig. 9a. No trend was evident in the storm-to-storm statistics or a sudden change in the storm statistics. This indicates that both the scanning and profiling radars were stable throughout this 2-month campaign.
It is very difficult to quantify the uncertainties of precipitation estimates because of the variability of the precipitation in time and space. These uncertainties can be described as instrumental measurement error and as representativeness error. The measurement error is associated with the precision of the instrument, and the representativeness error is associated with incomplete sampling of the variable precipitation field. In this paper we present the measurement error from profiling radars using two vertically pointing profiling radars to observe the same volume of the atmosphere. As was demonstrated in Gage et al. (2004), this methodology enables us to quantify the measurement error in a variable precipitation field. Using 74 585 simultaneous precipitation observations, the reflectivity difference standard deviation was 0.57 dBZ, resulting in a single observation precision of 0.4 dBZ. These results are consistent with the estimated measurement error for profilers used in TEFLUN-B (Gage et al. 2004) and with the estimated error of less than 0.5 dBZ reported by May et al. (2001) using a well-calibrated C-band polarimetric research radar to calibrate a 920-MHz profiler. The 95% confidence interval of the bias was 0.0034 dBZ for the 74 585 observations.
During the TEFUN-B campaign, a surface disdrometer was used to absolutely calibrate the vertically pointing profiling radar. There were 2462 simultaneous profiler and disdrometer observations during this 2-month campaign. The reflectivity difference standard deviation of 2.84 dBZ was substantially larger than the standard deviation between two profilers. This is partly due to the different sampling volume sizes and the 308-m separation between the two sampling volumes. Another factor is that the precision of a 1-min disdrometer observation is on the order of 2 dBZ (Gage et al. 2004). The large number of simultaneous observations enables the 95% confidence interval about the bias to be ±0.09 dBZ.
After the profiler was absolutely calibrated by the surface disdrometer, the profiler reflectivity was used as a transfer standard and compared with the simultaneous WSR-88D radar reflectivity measurements made over the profiler site. Since the scanning radar required 5 min to complete a volume scan and the vertical extent of the radar beam over the profiler was 630 m, the 1-min and 105-m vertical resolution profiler observations were averaged to match the WSR-88D radar resolution. A novel feature of this study was that the reflectivity variation during the 5-min and 630-m resolution window was quantified using the standard deviation of the profiler reflectivity, σ5min,630m. The direct correlation of σ5min,630m with the reflectivity difference variance led to the σ5min,630m statistic being used to detect and remove outliers associated with convective rain.
During the 2-month TEFLUN-B campaign, there were 21 rain events that had disdrometer, profiler, and scanning radar observations. The 2062 simultaneous profiler and scanning radar observations below the bright band were analyzed as an ensemble and for each rain event. Several of the storms had statistics that deviated from the ensemble statistics. After evaluating the meteorological conditions of these outlier storms, it was determined that they were short-duration (about 30 min over the profiler) convective cells exhibiting large reflectivity variability quantified by the σ5min,630m statistic. Ignoring these storms, the ensemble statistics suggested that the scanning radar was 0.66 dBZ higher than the profiler, the reflectivity difference standard deviation was 2.36 dBZ, and the 95% confidence interval was 0.09 dBZ. While the individual scanning radar to profiler comparisons show a large variation, the confidence interval about the mean bias is very small. There was very little variation in the bias statistic for each rainstorm, indicating that both radar systems were stable in their calibration during the TEFLUN-B campaign.
One advantage of comparing the reflectivity between two radar systems is that radar operators can determine if the calibration between the radars has changed with time. This could have an impact on short-term forecasts that rely on accurate calibrations to determine regions of potential hail and/or heavy rainfall and associated flash flooding. The comparison would provide a long-term record of the stability of both radar systems. This information could be implemented in a postcalibration processing to improve radar rainfall estimation. Long-term accurate estimates of areal rainfall are important for hydrologists to monitor water resources in their regions.
Since the location of the profiling and scanning radars usually do not change once they are deployed, the storm-to-storm and ensemble statistics can be easily automated and made graphically available. These products would become another tool that can be used to monitor the health and stability of the scanning radar. It needs to be stressed that the monitoring of the scanning radar does not require absolutely calibrating the profiling radar. The statistics can be calculated using the current calibration of both radar systems, and the change in statistics will be an indicator of a change in the performance of either radar system. The monitoring can be done in real time or during postcalibration analysis, depending on the resources and networking capabilities between the profiling and scanning radars.
This work has been partially supported by the NASA TRMM Project Office.
Corresponding author address: Christopher R. Williams, CIRES–NOAA/Aeronomy Lab, Mail Stop R/AL 3, 325 Broadway, Boulder, CO 80305. Email: Christopher.R.Williams@noaa.gov