1. Introduction

Weather Surveillance Radar-1988 Doppler (WSR-88D) Next-Generation Weather Radars (NEXRAD) include 158 deployed operational Doppler devices throughout the United States and represent superior resolution and improved observation accuracy compared to past radar systems. The observations from these radars aid National Weather Service (NWS) forecasters in issuing warnings and watches to citizens about dangerous weather. The spatial and temporal densities of WSR-88D raw radial wind data represent a rich source of high-resolution observations that can be assimilated in operational numerical weather prediction (NWP) models. For more than a decade, the WSR-88D has played an important role in the improvement of short-range nowcasts (see, e.g., the review article by Wilson et al. 1998) as well as warnings for severe thunderstorms, tornadoes, and flash floods, but until recently operational models did not obtain the benefit of these high-resolution observations in real time. The National Oceanic and Atmospheric Administration’s (NOAA) National Centers for Environmental Prediction (NCEP) pointed out that high-resolution observations were needed in real time for high-resolution regional models coming online for the next millennium. The first time radar radial winds were assimilated into NCEP operational models was in a demonstration project at the 1996 Olympics in Atlanta, Georgia,¹ to aid the forecasters at the weather forecast office who had the responsibility for issuing weather warnings for Olympic events. For this demonstration, a commercial radar data feed was used from the NEXRAD Information Dissemination Service (NIDS), which delivered a truncated resolution and precision compared to the radar returns at each individual radar.²

That such data were not supplied at a central site, capable of operational model ingest, was more a problem with the level of information technology in the Advanced Weather and Interactive Processing System (AWIPS). The impact from the addition of NIDS data into the regional model was not clear from the experiments that were run at the time of the 1996 Atlanta Olympic demonstration. This provides a motivation to compare the impact of higher-resolution/precision radar returns with delivered observations available a decade ago. The technology available then—for example, bandwidth—limited the large amount of data the radar generated that could be transmitted. The computations needed to perform preprocessing tasks such as converting radar return locations to model coordinates were limited by CPU resources. A practical and cost-effective way to transmit the information content of the high-resolution radar observations is to use an averaging technique to reduce the dataset size at the radar and utilize the computers located at each radar station to do the necessary computations. This is an early use of a parallel system of many distributed [potentially 158 across the continental United States (CONUS)] processors acting on the same software program independently at the same time and delivering the results to a central site. The reduction in the data size from the averaging technique at each radar site was 1500 to 1 (Istok et al. 2003).

A characteristic of these radar observations is the significant degree of redundant information present in the radar returns. Redundant observations impose a burden on an operational assimilation system since each datum is processed with repetitive interpolations from the analysis grid to its location and back again. This effort is carried out for each datum regardless of the information that can be attributed to it in the overall assimilation. The time and storage expended on mutually redundant data could be better spent on improving other aspects of the assimilation (Purser et al. 2000). Therefore, it is desirable to maximize whatever data compression the ensemble of fresh observations allows, while minimizing any degradation of the information content. The term for a surrogate datum that replaces several partially redundant actual data is a “super-observation” or “super-ob.” In the technique developed by Purser et al. (2000), and used in Liu et al. (2005), the redundant information is optimally windowed by a Gram–Schmidt orthogonalization algorithm. One may schematically think of this as creating the optimal size grouping of radial winds to reduce redundancy given all the raw radar information. Since data transmission bandwidth limits the number and precision of raw radar radial wind returns, we endeavor to get as much of the information content as possible, choosing a simple approach, by averaging radar observations in preselected cells. In addition, such averaging reduces the noise (errors) in the raw radar returns.

Super-obs have been assimilated in the NCEP operational analysis system for subsets of the WSR-88D radar radial wind observations from the NIDS commercial data feed. NIDS provided a way for public users and NWS to receive NEXRAD data. The NIDS contract expired on 31 December 2000, and NWS assumed the responsibility of distributing the NEXRAD data for operations, the content of which remained the same. In terms of radial wind, the NIDS data transmission truncated the number of antenna tilts to 4 from a possible 16, decreased data to 4 bits (16 levels) compared to 16 bits (65 536 levels of amplitude) of precision, and truncated radial resolution to 1 km along a radial line of a possible 0.25 km that is available at the radar site.

The precision and information content of the NIDS radial wind is improved if data at each radar site is directly utilized at the resolution and precision of the WSR-88D radar. We construct a super-ob at each radar site, acting on the complete set of radar data in several volume scans, and then deliver the reduced dataset to a central site with higher precision as described in the reports of Alpert et al. (2003, 2004). The radar wind super-ob takes the results of radar scans and averages data points within a prescribed time and spatial three-dimensional volume before transmitting reports to a central collection of radial wind data. One could create super-obs using the reflectivity and spectrum width in addition to the radial wind super-ob; however, use of these quantities requires an independent forward model so this study confines itself to radial winds for which the forward model is well defined. We note that beginning in 2005, with rapid improvements in information technology, direct transfer of all raw radar (level 2) data to a central site is scheduled, so that models can take advantage of the complete set of radar data and antenna tilts. This data feed will include more complex quality control measures because the entire matrix of events will be available before averaging is done (Liu and Xu 2005; Zhang et al. 2005). The super-ob product reported here will remain useful as a benchmark and backup system to compare with the level-2 data that may be ready for operations within a year.

2. Description of the super-ob

The full-resolution WSR-88D base radial wind data provide sufficient data amounts for statistically significant averaging. The super-ob product is programmed at each WSR-88D site using the open systems Radar Product Generator (RPG) to control all aspects of the calculation. The open RPG is the system that operates between the Radar Data Acquisition (RDA) system and more sophisticated display devices, such as the National Weather Service’s AWIPS. The RDA collects data from the WSR-88D radar and forwards base data products to the open RPG. These base data products consist of reflectivity, radial wind, and spectrum width. The open RPG creates the special purpose products from the base data and forwards them to other systems for display or for further processing. Super-ob is one of the newly enabled, enhanced products under the open RPG. The new product is super-obs of radar radial winds and is known in this study as level-2.5 super-ob. These will be compared with the NIDS radar radial wind super-ob referred to as level-3 super-obs as well as a no-radar radial wind case in analysis and forecast experiments using the NCEP Eta Data Assimilation System (EDAS), which is a three-dimensional variational data assimilation (3DVAR) analysis system (Parrish et al. 1996). Forecasts are from the meso-Eta model, in its operational configuration with a resolution of 12 km (information on current and past operational analysis and meso-Eta model configurations can be found at http://www.emc.ncep.noaa.gov/mmb/mesoscale.html). The level-3 super-obs were placed in the July 2003 implementation of the Eta mesoscale regional model (Ferrier et al. 2003). The level-2.5 super-ob described here was implemented in the operational Eta Model renamed as the North American Mesoscale (NAM) model in the spring bundle of April 2005 (DiMego and Rogers 2005; Rogers et al. 2005). In this paper, we refer to the NAM as the meso-Eta.

3DVAR is ideally suited for assimilating Doppler radar wind data because the radial wind super-obs are a scalar having only one component and the vector wind cannot be uniquely determined. To determine both wind components (u, υ), additional information is needed. The variational procedure requires a forward model to project the model’s winds onto the radar observation locations, a linear process. In effect, the assimilation (inverse) process utilizes the model’s knowledge, through dynamics and physics from its forecast error statistics, to project the scalar radial wind observations onto the model wind vectors.

Adaptable parameters for super-ob are the time window, cell range size, cell azimuth size, maximum range, and minimum number of points. The values of these parameters, which define the super-ob averaging in time and space, are shown in Table 1. The default settings indicate that at each elevation angle, a wedge shaped volume of 6 azimuth degrees by 5 km along a radius, averaged over a time of 60 min, define each super-ob cell. The processing of the radial wind observations, at each radar, requires choosing a time-averaging period and a super-ob cell size. A large enough area and time period are needed to include enough observations to insure a good average and to minimize the dependency between adjacent super-obs. The projection of the radial wind field in the observation space can be calculated in the 3DVAR analysis from the orthogonalization of the observation forward model (Liu et al. 2005). Each of the super-obs contains no fewer than 50 points, and no cell extends past 100 km, as the radar beamwidth becomes too wide, and returns become less certain at larger distances from the radar. These adaptable parameter values are programmed through the RPG. The range of possible values, shown in Table 1, allows for the super-ob product to adjust to different analysis resolution requirements as they occur. All the WSR-88D sites will create a super-ob data product by this process and transmit it to a central site. The data transmission precision of the super-ob (time mean) radial wind is 0.01 m s⁻¹ as described in Stephenson (2002). Another improvement in the level-2.5 super-ob over the level-3 super-ob is the transmission of a maximum of 15 antenna tilt levels compared to four levels (0.5, 1.5, 2.4, 3.3), which are the first four tilt angles of what is available at each radar. The super-ob product is potentially a time average of many volume scans that depend on the number that can fit into the Table 1 time window interval value. This number varies with the radar volume coverage pattern (VCP) mode. The standard deviation about the mean within each cell is calculated from the time and spatial averaging and included in units of meters per second.

A number of tests have been made to the radar system within the Systems Engineering Center at the NOAA/NWS Office of Science and Technology. An example of the radial wind scalar field is shown on a vector plot in Fig. 1a and as a color-filled display of magnitudes in Fig. 1b. The data for this example is from KATX, Seattle, Washington, radar, at an antenna tilt of 0.4°, from 1800 UTC 28 May 2002 and represents a typical return. The characteristic sinusoidal pattern of radar radial returns is apparent, and there are large areas that do not have returns. The radial wind magnitude is seen to increase southwest and northeast at larger distances from the radar. The number of returns is variable, dependent on the radar VCP. The more precipitation that occurs, the more radial wind returns cover the domain and, therefore, more super-obs will be calculated.

The level-2.5 super-ob from radar returns close to the radar can have errors due to electronic gating deficiencies that are not screened out (the first two Doppler gates being “negative distance” and the radial wind values considered suspect). Therefore, the level-2.5 super-obs may average some of these data that should be screened out from its smallest radius. The level-2.5 super-obs near the radar, encompassing the first two innermost gates, will contain at most 10% suspect raw returns, and should not affect the results of the experiments. Figure 2 represents a one-dimensional representation of the super-ob returns shown in Figs. 1a,b as a function of azimuth and radius. Figure 2a shows the height variation plotted for the super-ob dataset according to the first observation, which is along 0°N, closest to the radar, extending out to larger radii from the radar, and then clockwise around in azimuth for the remaining returns in the report, in this case over 700 super-obs for a complete circle of azimuth. For example, the first line segment in Fig. 2a coincides approximately with 0 azimuth pointing local north (Figs. 1a,b) and radius of ∼2.5 km and the next point along that radius, 7.5 km farther out, as radial distance increases and continuing clockwise from north around with increasing azimuth of 6°. Each asterisk in Fig. 2a represents a super-ob point. As the radar beam extends to larger radial distances from the radar, the height of the returns increases as shown by the spikes in Fig. 2a until either there are no more returns along a particular radius or the radius reaches the default maximum of 100 km (Table 1). In the case of the super-ob radial wind shown in Figs. 1a,b, for the lowest elevation angle, the height ranges from about 250 m to 2 km as shown in Fig. 2a. The radial wind standard deviation is calculated within each super-ob, 6° of azimuth by 5 km radius, and transmitted in the super-ob record with a precision of 1 m s⁻¹ as shown in Fig. 2b and, for clarity, in a color-filled display in Fig. 1c.

Because this analysis was initially created to assimilate Doppler radar winds, all winds are treated as line-of-sight winds. A conventional wind observation becomes two line-of-sight observations, one along a north-pointing line (earth υ component) and one along an east-pointing line (earth u component). All line-of-sight winds, radar or conventional, are then assigned the angle that each observation line of sight makes with the Eta grid x axis. The forward model (which estimates the observed quantity from model variables) is just

View Expanded

where (u, υ) are the components of the real wind at the same height as the super-ob and θ is the azimuth. This equation reduces from the three-dimensional system (Parrish 2005) with the vertical motion set to zero and at constant elevation angle. Figure 2c shows each super-ob radial wind plotted along a particular radius and clockwise for each 6° of azimuth; therefore, a sinusoidal pattern of the radial wind results due to the radial wind definition (1). Each successive super-ob moves out along the radii with increasing beam height causing a range of super-ob wind magnitudes along each azimuth. The super-ob radial wind returns labeled “mean wind” in Fig. 2c range from a maximum of 13.1 m s⁻¹ (away from the radar) at point 649, to a minimum of −14.5 m s⁻¹ (toward the radar) at point 457. At any radial distance from the radar observing the returns in increasing azimuth will result in a characteristic sinusoidal pattern. The range of wind values superimposed on the sinusoid is from varying heights and location of the return as well as the result of natural variance in the winds in the radar vicinity due to synoptic conditions. However, a number of super-ob wind values in Fig. 2c extend beyond the normal range as defined by its neighbors. One can scan this dataset, and examine values outside the expected range of wind, and compare the standard deviation, height, and the location of a return in question.

A few radial wind speeds in Figs. 1a,b sometimes appear as outliers compared with their neighbors (Fig. 2c). For example, point 109 at (latitude, longitude) (48.85, −121.632) has a super-ob wind of 12.1 m s⁻¹ at a height of 1549 m. The standard deviation corresponding to this point on Fig. 2b is among the lower values (<2 m s⁻¹ also shown by gray areas in Fig. 1c) that may arise from a preponderance of ground clutter, in this case from the mountains around Seattle. Points 459 and 475 at (47.263, −123.238) and (47.579, −123.110) respectively (see SW quadrant of Figs. 1a,b), show radial wind magnitudes of −14.12 and −12.58 m s⁻¹ (Fig. 2c), which are outliers being larger than their neighbors. The returns for these super-obs are expected to have larger wind magnitudes with increased elevation (Fig. 2a) as they are farther away from the radar. An example of a larger standard deviation return is shown for point 590 at (48.039, −123.789) with a value of 12 m s⁻¹ and wind magnitude of −10.33 m s⁻¹. The height for point 590 is 1572 m, which makes it among the larger distances from the radar, and higher height, therefore a larger beamwidth causing uncertainty in returns. Figure 2b shows a moderate spike at point 590 in line with the increased height in Fig. 2a and the increased negative super-ob wind. In this case, the value is suspect because of the large standard deviation in the radar radial winds making up the super-ob. One may expect a larger standard deviation of radial wind values because of the widening of the radar beamwidth with greater distance from the radar as well as increased wind speeds at higher altitudes. When the radar beam grows wider with increasing distance from the radar, the vertical position of the return cannot be determined with precision. The standard deviation is useful in determining the quality of the returns. The quality control of the super-ob data in the analysis program may be able to utilize the reflectivity to determine locations of severe convection (and vertical motion), which could cause errors in the radial wind, and this is under consideration for assimilating level-2 raw returns.

Figure 3 illustrates the location of each super-ob. The centroid of WSR-88D returns in space is calculated before super-ob averaging and made part of the transmitted header as the location in longitude and latitude of each super-ob value. This is the reason the super-ob longitude–latitude locations, shown by a “+” in Fig. 3, or the tails of the vectors in Fig. 1a, are not orientated in exact concentric circles. A time deviation is also reported as the valid time of a particular super-ob cell as the average of several scans occurring within each time window (defaults shown in Table 1).

Analyzed conventional winds at a height of 0.4 km interpolated from the NCEP Global Forecast System (GFS) Spectral Statistical Interpolation (SSI; Parrish and Derber 1992) valid 1800 UTC 28 May 2002, are shown in Fig. 1d. The wind barbs are displayed over the same domain and time as the super-ob radial wind observations in Fig. 1a. The operational SSI, used here as an independent (low resolution) analysis, uses all available remote and conventional observations to create the analysis at a resolution (spectral triangular truncation T254) of less than 1° of latitude and longitude. The SSI analysis is used as a proxy for observed winds in the radar area because the number of wind observations available around the radar is few. This is done with the knowledge that a global 1° × 1° analysis at a particular height level is not a definitive way to verify these high-resolution observations, but it serves as a coarse check on the validity of the super-ob returns. The super-ob radial winds located close to the radar are at or near a height of approximately 0.4 km as shown in Fig. 2a; therefore, one can compare Fig. 1d with Fig. 1a. Given the definition of radial wind from Eq. (1), there is one direction where the analyzed wind will be identical to the radial wind, and ideally, winds perpendicular to this will have zero radial wind. The direction of the analyzed wind north and west of the radar in Fig. 1d is southwesterly at ∼15 m s⁻¹. The radial wind along a line in the northwest quadrant in Fig. 1a lines up in the same direction as analyzed winds shown in the same quadrant in Fig. 1d. To the east of the radar, the radial wind magnitudes are far less, 0–5 m s⁻¹. This is because the radial wind in this region is perpendicular to the analyzed southerly wind (Fig. 1d), so according to the definition [Eq. (1)] the component in the radial direction is small. The analyzed winds in the southwest quadrant tend toward larger magnitudes as do the radial winds shown in Figs. 1a,b. Plots at 1- and 1.5-km levels of regular analyzed winds show that winds at higher elevations in this area have increased magnitudes even though the angle of the wind (southerly) and the radial wind are at large angles (not shown). The regular winds in the southeast quadrant are SSE with similar magnitudes to the radial wind. Overall the analyzed winds show changes that are reflected in the radial wind pattern as shown in Figs. 1a and 1d.

The number of radar returns, and therefore the corresponding super-obs, vary with the amount of precipitation, increasing with more precipitation. Convective regions can distort the returns, and these are filtered by observing the standard deviation of the returns, which is transmitted with each super-ob. The super-ob standard deviation is used as a proxy for the error in the radial wind measurement. The larger the standard deviation, the smaller the weight the analysis system will apply to a super-ob. Even in clear-air mode, super-obs are successfully created as the radars are sensitive to Doppler motions.

The quality control used in the EDAS for the level-2.5 and level-3 super-obs is identical. One quality control element of the super-ob ingest into the EDAS analysis is that the vertical component of the radar beamwidth is assumed to increase with radar range at a rate of 20 m km⁻¹. This is roughly 20% larger than the actual beamwidth and is done to account for the uncertainty in beam propagation. The VAD quality mark is used as a proxy for quality control of the radar winds since the VAD wind is calculated from a series of measurements of the raw radial winds. It is used mainly to detect bird contamination (Collins 2001). Winds within the maximum range, as set in Table 1 (presently set at 100 km), are used. The radial winds are not used if there is no VAD observation or if it fails a quality control (QC) check (Ferrier et al. 2003). This approach combines the QC algorithm for removing radial winds contaminated by birds with other checks used on the VAD winds. These QC algorithms were originally developed for the level-3 super-ob and should be considered preliminary (Parrish and Purser 1998). In addition, the radial winds are not used if 1) the beam envelope extends below the Eta Model terrain height, 2) the observation error is larger than 6 m s⁻¹, or 3) it fails the same gross checks used for conventional winds. Sophisticated QC algorithms developed by Zhang et al. (2005) and Liu and Xu (2005) are currently under development for NCEP operations with the full level-2 dataset.

The analysis contains an “inflate” parameter that magnifies the observation error relative to other observation types at each model level. The weights for any observation type are inversely proportional to the product of the observation errors at each model level, p, and a constant inflate parameter as discussed by Zapotocny et al. (2000):

View Expanded

The level-3 super-ob is assigned an inflate parameter of 3. Experiments will be shown for the level-2.5 super-ob, with inflate parameters set to 3 and 2, since this data has improved precision. The level-2.5 super-ob radial winds were operationally implemented with the same inflate parameter as the level-3 super-ob. In addition, using the standard deviation of the super-ob as the observation error estimate gives information for quality control in the analysis system for each observation event. For example, in strong convection, the super-ob standard deviation will be larger, which diminishes the weighting in the analysis system relative to other observations.

3. Results from Eta Model impact experiments

To test the impact of the level-2.5 super-ob on regional model forecasts, three experiments were performed with the NCEP meso-Eta regional model with a resolution of 12 km, 60 vertical levels, and a North American domain as in Rogers et al. (2001). Each experiment included a separate analysis from the EDAS cycling three-dimensional variational analysis system (Parrish et al. 1996; Wu et al. 2002) starting 12 h before the initial condition time of 1200 UTC 28 Oct 2003. This case study was chosen randomly for this date with a goal of showing the impact of the level-2.5 and level-3 super-obs for a typical forecast. All conventional and nonconventional observations that were available in the operations were assimilated in these experiments. In addition, the operational model run contains the assimilated level-3 super-ob over a 3-h window; therefore, the operational run will be the control run for experiments. Experiments are conducted to show the impact of assimilating the level-2.5 super-ob, which we designate as expt 1. Expt 2 is the operational run that includes the level-3 super-ob, and expt 3 has no radar radial wind data present, as indicated in Table 2.

The mean sea level pressure map and precipitation in Fig. 4 are 12-h forecasts that are representative of the synoptic situation from expt 1. We use a 12-h forecast for convenience to show generally the areas where precipitation is indicated, and the synoptic situation of the experiments. In the central United States there are two major low pressure systems, one extending to northwestern United States and a second over the Great Lakes. There is also an inverted trough over the eastern U.S. coast. This figure also shows the 12-h forecast of precipitation to indicate regions with more radar returns and a likely corresponding greater number of super-obs compared to regions with clear-air returns. A low pressure system and its associated precipitation having a northwest–southeast orientation is forecasted to be approaching the Great Plains. These precipitation patterns move to the south and east and provide areas where the radar responds in its precipitation mode, usually providing more than the normal number of super-obs through 36 h.

There are not sufficient verifying observations to compare the high-resolution super-ob data directly using collocated conventional observations like rawindsondes because of scale differences. To compare the impact of higher-precision level-2.5 super-obs with the lower-precision level-3 super-obs, a case study approach is employed to show the impact of the different super-ob-generating methods. The lower-precision level-3 super-obs may not have as much influence on the analysis due to ultimately larger deviations from the guess wind field compared to the other observations. The higher-precision level-2.5 super-obs should have less error and therefore have more influence in the analysis by virtue of being closer to other observations. We first show the vector wind difference and magnitudes between level-2.5 (expt 1) and level-3 (expt 2) super-obs, in Fig. 5a, and then the level 3 and no radar radial wind (expt 3) to see the impact of both super-obs in Fig. 5b. The level-2.5 and level-3 super-ob (Fig. 5a) differences are of small scale, the largest difference being generally greater than 1 m s⁻¹. These are located in areas along the inverted trough (East Coast), low pressure (from the Great Lakes through the Northwest), and an area in the Southwest mountains. The magnitude and aerial extent are larger in Fig. 5a than in Fig. 5b implying that the level-3 super-obs (expt 2) have little or no impact in the analysis compared with the level-2.5 super-obs. This is confirmed in the difference plot of level 2.5 minus no radial wind (not shown).

The analyses from each of the three experiments were used to initialize separate meso-Eta forecasts at 12-km resolution, resulting in a 48-h forecast for each experiment. Shown is the (36-h) forecasted wind differences between the level-2.5 super-ob and level-3 super-ob (expt 1 minus expt 2) in Fig. 5c and the (36 h) forecasted wind differences between the level-3 super-ob and no radar radial wind present (expt 2 minus expt 3) in Fig. 5d. We use 700 hPa as the lowest representative level above most mountains. The largest changes appear for the level-2.5 super-ob (expt 1 minus expt 2), which had wind magnitude differences ranging from −7 to 5 m s⁻¹ compared to the differences between level-3 super-ob and no radar integration (expt 2 minus expt 3), ranging from −2.5 to 3.5. Thus, by the 36-h forecast, the influence of the level-2.5 super-ob is double that of the level-3 super-ob. The shape of the differences appear as filament-like areas in the level-2.5 super-ob, while the level-3 super-ob differences are barely discernable and appear random. The main differences between the level-2.5 super-ob and the level-3 super-ob are the precision in the delivered wind values (0.05 m s⁻¹ compared to 5 m s⁻¹, respectively) and the extra antenna tilt angles (16 compared to 4, respectively). The largest differences seen in Fig. 5c appear in regions of precipitation as shown in Fig. 4 for 12 and 36 h (not shown).

The number of super-ob winds with wind magnitude error (as measured by the standard deviation for each super-ob cell) within a 1 m s⁻¹ range for the level-2.5 and level-3 super-obs is shown in Table 3. The simple quality control used for the level-2.5 and -3 super-obs has a cutoff of 2 m s⁻¹ standard deviation, so Table 3 does not show the results for errors less than 2 m s⁻¹. However, the number of super-obs with standard deviation less than 1 m s⁻¹ is typically small (Fig. 2b). Most of the level-2.5 super-obs appear within 2 m s⁻¹ error whereas the level-3 super-ob errors peak at 3 m s⁻¹ as shown in boldface in Table 3. Also note that the number of level-2.5 super-obs is typically 5 times that of the level-3 super-obs. (An operational real-time monitoring Web site containing a display of observation counts can be found at http://www.nco.ncep.noaa.gov/pmb/nwprod/realtime/.)

The radar radial wind observations that are analyzed by the EDAS use knowledge of past forecasts to calculate error statistics for the dependant variables. The knowledge necessary for the analysis to project the radial wind onto the analyzed wind field comes through the adjoint of the observation operator or the adjoint of the forward model. The definition of radial wind is a scalar as shown in Eq. (1), and there is only one direction for which the regular wind will be identical to the radial wind. We cannot uniquely compose the radial winds into regular wind vectors. The EDAS analysis uses the information of how the regular winds are related to the radial winds to make an optimal decision of how the observed radial wind projects onto the regular winds and other variables.

4. Verification with the Eta32-km model

The 32-km NCEP EDAS analysis with Eta parallel forecast model containing only the level-3 super-ob (noted as ETAW, control) is compared with only level-2.5 super-ob (noted as ETAY, parallel) integration. The experiments were run twice daily at 0000 and 1200 UTC for 12 days between 8 June 2004 and 20 June 2004 for 24 integrations to 60-h forecasts. These experiments were run with the 32-km model to show the super-ob verification scores over a 2-week period to obtain verification scores within computer resource constraints. These results were the basis for deciding if level-2.5 super-obs should become part of the “implementation bundle.” Equitable threat score and bias for 24-h accumulated precipitation and RMS height errors are shown (only 1200 UTC cycles are used in 48-h forecasts and only 0000 UTC cycles are used in 60-h forecasts). The 48-h level-2.5 super-ob forecast precipitation threat scores (Figs. 6a,b) show a small improvement at 0.01- and 0.1-in. thresholds and at the 1.5-in. threshold over the control run. It is slightly inferior at the 0.75-in. threshold. Except for the 1.5-in. threshold, the number of events is significant exceeding 1000 for 1-in. categories and less. The precipitation bias score is unchanged. The number of events for larger thresholds is too small to be statistically significant. Extending the Eta Model precipitation forecasts to 60 h (Figs. 6c,d) shows that there is improvement of bias and equitable threat scores for all thresholds for the level-2.5 super-ob experiment. Precipitation is a sensitive parameter, dependent on complex physical interactions, so small threat score enhancements imply that the level-2.5 super-ob precision improvements are making an impact on the EDAS analysis and forecast system.

The RMS height error versus rawinsonde observations (raobs) over CONUS for the two experiments, 60-h Control Eta-32, and 60-h Parallel Eta-32, is shown in Fig. 7a as a function of pressure level. The level-2.5 super-ob shows smaller errors at all levels compared to the level-3 super-ob (Fig. 7a). Figure 7b shows the height bias errors as a function of pressure against raobs over the CONUS from Eta-32 60-h forecasts for the two experiments. A significant reduction in the height bias is seen at upper levels, but there is a small increase in the bias errors at low levels that is not significant. The upper-level improvement in geopotential height may arise from radial wind observations at larger tilt angles and the additional number of antenna tilts in each radar volume scan. This provides superior information content to the analysis system compared with the four tilts in level-3 super-obs. The RMS wind and temperature errors (not shown) show marginal improvement at midtropospheric and jet levels. At shorter forecast times, the low-level RMS wind and temperature error differences between the experiments are insignificant. For the cases presented, the presence of the level-2.5 super-obs provides a beneficial impact in precipitation threat scores as well as reduction in RMS and bias errors in geopotential height, particularly over upper levels.

5. Case study of EDAS analysis inflate parameter impact

The NCEP 3DVAR analysis of the EDAS performs a regional analysis adapted to the Eta Model in operations at the time of these experiments. Each observation type used in the analysis has a certain weight relative to the error of other data types. For example, the errors assigned to rawinsonde temperatures at various pressure levels are approximately 1 K. Each competing observation type is weighted according to its estimated error. The super-ob standard deviation is calculated at each radar cell and is provided in the transmission of the super-ob product. As indicated in section 2, the standard deviation in each super-ob cell is taken as the observation error estimate of the radial wind super-ob. The analysis weight of the super-ob is set as shown in Eq. (2). The inflate parameter from Eq. (2) (section 2) is assigned the value of 3 for the level-3 super-ob, which is the Eta control experiment. Since the level-2.5 super-ob is obtained from higher-resolution radar returns and not filtered (from 0.25 km to 1 km) as in the level-3 super-ob case, and maintains high-precision values during level-2.5 super-ob transmission compared to the level-3 super-ob (0.01 versus ∼5 m s⁻¹), the level-2.5 super-ob may deserve to be treated with higher confidence in the EDAS. In addition, the level-2.5 super-ob benefits from the full complement of up to 16 antenna tilt levels, which correspond with an increase in the information content of the level-2.5 super-ob set relative to other data types in the EDAS. This motivates a more extensive investigation to operationally change the level-2.5 super-ob influence through the inflate parameter.

The standard deviation values (Fig. 2b) range from 0 to 15 in units of meters per second, and with an inflate parameter of 3, can produce weight reciprocals that range from 0 (handled appropriately) to 45 [Eq. (2)]. A gross toss (removal) of a super-ob is applied for standard deviation values greater than 6 m s⁻¹ or weight ≥21. Since the level-2.5 super-ob is constructed at each radar with 0.25-km radius and transmitted at higher precision to a central site, a reduction of the inflate parameter should cause the analysis to draw closer (by giving more weight) to the level-2.5 super-ob radial wind observations. Two experiments were run to study the influence of the inflate parameter for the level-2.5 super-ob and a comparison was made with the operational control (level-3 super-ob with inflate parameter set to 3). The two experiments consist of consecutive initial conditions (29 and 30 September 2004, chosen randomly) integrated for 48 h each. One cannot draw definitive statistical conclusions from only two experiments in terms of statistics as shown in the last section; however, a comparison of precipitation patterns between the experiments can determine whether the impact from the changes are significant. Precipitation patterns will be used to highlight the differences between the experiments.

The relevant synoptic situation at 1200 UTC 1 October 2004 corresponding to these experiments (not shown) has two light to moderate precipitation bands oriented northeast to southwest through Nebraska and South Dakota as well as on the Oklahoma–Texas state line (Fig. 10). Figures 8a,b show an enlarged excerpt of the total Eta domain from 30° to 45°N and 110° to 90°W. The 12-km Eta, 24-h forecast of accumulated precipitation, valid at 1200 UTC 1 October 2004, is shown in Fig. 8a when the inflate parameter is set to 2 and using the level-2.5 super-ob. Figure 8b shows the operational control, with the level-3 super-ob, and inflate parameter set to 3. Setting the inflate parameter to 2 with the level-2.5 super-ob has enhanced precipitation for moderate amounts, for example, the maxima in Figs. 8a,b. The enhanced precipitation amounts seen with the inflate parameter set to 2 are small in horizontal scale and extend south of the Oklahoma–Texas line as well as in the trisection of the state boundaries of Nebraska–Colorado–Wyoming. The precipitation difference (Fig. 8a minus Fig. 8b) shown in Fig. 9a indicates the most significant precipitation differences over north Texas and Oklahoma, with smaller maxima in the Texas panhandle, and in the tristate region (Nebraska–Colorado–Wyoming). These patterns are isolated, indicating that they are from precipitation amount increases rather than from phase displacement of the banded patterns. For example, differences in the precipitation banded pattern in South Dakota (centered at 44°N, 100°W) arise from forecast phase error as shown by the dipole pattern in the difference plot. As forecast times lengthen, the larger scale phase error of the model increases and will overcome the small-scale effects we intend to study. The larger precipitation difference values in the tristate region (41°N, 104°W) are isolated and not in a dipole pattern like the differences found in the Oklahoma–Texas area at 48-h forecast (Fig. 9b). The level-2.5 super-ob experiment enhanced the precipitation amounts in both regions.

The verification for this experiment is the observed 24-h accumulated 10-km River Forecast Center (RFC) mosaic precipitation observations from rain gauges (Fulton et al. 1998) and is shown in Fig. 10. The patterns from the observed 10-km rain gauges are far more detailed than the 12-km Eta forecasts. This is expected as the rain gauges are analyzed from extremely high resolution data while the 12-km Eta model has various filters and other nonmeteorological forcing, and is hence smoother than the rain gauge observations. There are two prominent precipitation bands at the Oklahoma–Texas line and a northern pattern from Nebraska–Colorado–Wyoming tristate area extending northeast through South Dakota. The main features are captured by both experiments, but the observed maximum precipitation is about double (see the 1″ and 0.5″ contour at the Oklahoma–Texas line in Fig. 10) what was forecasted (Figs. 8a,b), and the 2″ contours in the observed Texas panhandle and near the tristate area (in Fig. 10) intersection are too small scale for the Eta Model to resolve. The level-2.5 super-ob experiment forecast precipitation enhances the precipitation amount over the operational level-3 super-ob, which is evident in the difference field of Fig. 9a, significantly, for both above-mentioned precipitation areas by about 0.25″ to 0.5″. A forecast experiment with the only difference being a change of the inflate parameter from 2 to 3 using the level-2.5 super-ob was run, and the impact to the precipitation forecasts is shown in Fig. 11. This shows the impact of the analysis system with more influence from the level-2.5 super-ob compared to other conventional observation types. The precipitation differences are positive throughout most of the domain, indicating that more influence from the super-ob observations produce precipitation in the direction of what is observed. This has to be taken in the context of a single case study in view of the nonlinear nature of convective schemes.

A 48-h forecast started from initial conditions 24 h earlier so that the verifying times are the same was run with the 24-h accumulated precipitation difference shown in Fig. 9b. This run is identical to the previous one in which the inflate parameter was set to 2 and the results compared with the operational run with level-3 super-ob and inflate parameter set to 3. The level-2.5 super-ob has precipitation amounts (inflate 2) greater than the operational level-3 super-ob in most regions, particularly near the Oklahoma–Texas border (Fig. 9b). This is in the correct direction compared to the high-resolution rain gauge observations shown in Fig. 10. Experiments with the inflate parameter set to 3 with the level-2.5 super-ob and with these initial conditions (not shown) at 48 h fall in between that shown in Fig. 9b. This is a similar result to that indicated by Fig. 11 showing further improvement by giving more weight to the level-2.5 super-ob. By 48 h, the large-scale phase errors in the precipitation forecasts are growing, as indicated by the dipole structure in the difference plots (Fig. 9b compared with Fig. 9a).

6. Summary and conclusions

A radial wind product at all NEXRAD radars was added to NCEP operations to reduce the degree of redundant information in the radar returns, increase the precision of the radial wind values delivered to a central site, and decrease the computational burden and bandwidth needed to assimilate high-resolution radar radial winds into operational NWP models. This was done by simply averaging many observations into one at each radar site and transmitting the result to a central site at higher precision with navigation data. We have used the term super-ob for the averaging, and distinguish between lower precision level-3 super-ob and a new higher precision level-2.5 super-ob product. The level-3 super-ob has 15 levels (4 bits) of radial wind to span ±50 m s⁻¹ possible of radar wind values that are filtered from the raw radar data to 1 m s⁻¹ precision and four antenna tilts over 1-km radial distance. The level-2.5 super-ob has the advantage of calculating the super-ob averaging at the radar with 0.25-km radial resolution, 9–16 antenna tilts for each radar scan, and a capable precision of transmitting the results at ±0.002 m s⁻¹ to a central site. In addition, the super-ob transmission includes the standard deviation within the averaging area, which is used as error information for quality control in the operational analysis system.

A case study with the Eta-12km model comparing the level-3 with the level-2.5 super-obs showed that there was almost no impact to the analyzed wind fields (700 hPa shown) for the level-3 super-ob. Two weeks of twice-daily runs with the Eta-32km version showed improvement from the level-2.5 super-ob for upper-level height and wind fields and precipitation threat scores. These 32-km model experiments were identical, except for the control run including the level-3 super-ob and the experimental run including the level-2.5 super-ob. There is some indication that the lack of impact for the level-3 super-obs is that the precision is not sufficient to provide winds that project closely enough onto the analyzed wind field compared to other sources. This will happen because the observation error is larger due to lower precision of the transmitted radar observations. Also, the larger number of antenna tilts in the level-2.5 super-ob product will provide more observations to radial winds compared to the level-3 super-ob. Improving the super-ob precision by having access to the raw radar data appropriately filtering the redundant data, which allows the communications bandwidth to transmit higher-precision values and more information from the extra tilts in each radar volume scan, provides superior information content to the analysis system.

High-resolution precipitation data from rain gauges were used to verify a case study comparing the level-3 and level-2.5 super-ob forecasts. Definitive conclusions cannot be drawn from a single case study but details in the verifying precipitation field that correspond to model forecasts indicate predictive capability. The details of the precipitation patterns show improved predictability for both the orientation and amount for the level-2.5 super-ob when only the level-3 super-obs are assimilated over regions where heavier rain events appear. In addition, when the analysis was forced to take into account the level-2.5 super-ob and draw more for these observations, the 24-h forecast details in precipitation amount and location improved. How well the analysis system will project the radial wind information onto the analyzed wind is determined by the error in the observations. The standard deviation of the super-ob area averaging is used to calculate observational error estimates. We obtain an error estimate for each super-ob by assuming that the standard deviation is inversely proportional to the weights. We studied the result of more or less reliable super-ob observations by comparing a parameter to inflate the errors of the super-ob relative to other observation types. The results showed that a lower error estimate for the level-2.5 super-ob enhanced (improved) precipitation amount compared to the verifying independent rain gauge observations.

The indication that model forecasts might utilize more influence from the level-2.5 super-ob by lowering the inflate parameter, giving these observations more weight compared to other observation types, can be tested by many experiments in an operational environment. The improvements in quality control afforded from having the raw radar data in one place as well as the ability to act on the entire raw radar dataset will undoubtedly improve the accuracy of high-resolution models. Efforts are under way at NCEP to utilize a comprehensive quality control of radial winds as well as reflectivity with the level-2 raw radar data for use in the North American and Weather Research and Forecast models (http://www.wrf-model.org/wrfadmin/publications.php).

Since the level-2.5 super-ob is obtained from higher-resolution radar returns and not filtered (from 0.25 km to 1 km) as in the level-3 super-ob, and maintained high-precision values during level-2.5 super-ob transmission compared to the level-3 super-ob (0.01 versus 9 m s⁻¹), the level-2.5 super-ob may deserve to be treated with higher confidence in the EDAS. After April 2005, the level-2.5 super-ob was implemented in EDAS/Eta operations, but if the level-2.5 super-ob is not present, then the NIDS super-obs are used. However, the inflate parameter settings continue to be set at 3. This study is the first step to motivate an investigation to operationally change the level-2.5 super-ob influence through the inflate parameter. With the use and delivery of all the raw radar to a central site, which is imminent, the super-ob 2.5 product can serve as a benchmark to test improvements in quality control and other radar enhancements.