• Andersson, E., , and Järvinen H. , 1999: Variational quality control. Quart. J. Roy. Meteor. Soc., 125 , 697722.

  • Andersson, E., , and Garcia-Mendez A. , 2002: Assessment of European wind profiler data, in an NWP context. ECMWF Tech. Memo. 372, 14 pp.

  • Benjamin, S. G., and Coauthors, 2004a: An hourly assimilation-forecast cycle: The RUC. Mon. Wea. Rev., 132 , 495518.

  • Benjamin, S. G., , Schwartz B. E. , , Stoke E. J. , , and Koch S. E. , 2004b: The value of wind profiler data in U.S. weather forecasting. Bull. Amer. Meteor. Soc., 85 , 18711886.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bluestein, H. B., , and Speheger D. A. , 1995: The dynamics of an upper-level trough in the baroclinic westerlies: Analysis based upon data from a wind profiler network. Mon. Wea. Rev., 123 , 23692383.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bouttier, F., 2001: The use of profiler data at ECMWF. ECMWF Research Department Tech. Memo. 331, 22 pp. [Available from ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom.].

    • Crossref
    • Export Citation
  • Buehner, M., 2002: Assimilation of ERS-2 scatterometer winds using the Canadian 3D-Var. Atmos.–Ocean, 40 , 361376.

  • Bussinger, S., , Adams M. E. , , Koch S. E. , , and Kaplan M. L. , 2001: Extraction of geopotential height and temperature structure from profiler and rawinsonde winds. Mon. Wea. Rev., 129 , 17291739.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chouinard, C., , and Halle J. , 1999: The impact of TOVS radiances in the CMC 3D-Var analysis system. Proc. Int. TOVS Study Conf. X, Boulder, CO, WMO, 92–98.

  • Chouinard, C., , Charette C. , , Halle J. , , Gauthier P. , , Morneau J. , , and Sarrazin R. , 2001: The Canadian 3D-Var analysis scheme on model vertical coordinate. Preprints, 14th Conf. on Numerical Weather Prediction, Fort Lauderdale, FL, Amer. Meteor. Soc., 14–18.

  • Côté, J., , Gravel S. , , Méthot A. , , Patoine A. , , Roch M. , , and Staniforth A. , 1998: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part I: Design considerations and formulation. Mon. Wea. Rev., 126 , 13731395.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Courtier, P., , Thepaut J-N. , , and Hollingsworth A. , 1994: A strategy for operational implementation of 4D-Var using an incremental approach. Quart. J. Roy. Meteor. Soc., 120 , 13671387.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Daley, R., 1991: Atmospheric Data Analysis. Cambridge Atmospheric and Space Science Series, Vol. 2, Cambridge University Press, 457 pp.

  • Deblonde, G., 2001: Variational retrievals using SSM/I and SSM/T-2 brightness temperatures in clear and cloudy situation. J. Atmos. Oceanic Technol., 18 , 559576.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gal-Chen, T., 1988: A theory of the retrievals of virtual temperature from remote measurements of horizontal winds and thermal radiation. Mon. Wea. Rev., 116 , 13021319.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gauthier, P., , Charette C. , , Fillion L. , , Koclas P. , , and Laroche S. , 1999a: Implementation of a 3D variational data assimilation system at the Canadian Meteorological Centre. Part I: The global analysis. Atmos.–Ocean, 37 , 103156.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gauthier, P., , Buehner M. , , and Fillion L. , 1999b: Background-error statistics modeling in a 3D variational data assimilation scheme: Estimation and impact on the analysis. Proc. ECMWF Workshop on Diagnosis of Data Assimilation Systems, Reading, United Kingdom, ECMWF, 131–145.

  • Hollingsworth, A., , and Lonnberg P. , 1986: The statistical structure of short-range forecast errors as determined from radiosonde data. Part I: The wind field. Tellus, 38A , 111136.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Järvinen, H., , and Undén P. , 1997: Observation screening and background quality control in the ECMWF 3D-Var data assimilation system. ECMWF Research Department Tech. Memo. 236, 33 pp. [Available from ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom.].

  • Koch, S. E., , Benjamin S. G. , , Schwartz B. E. , , and Szoke E. J. , 2004: The value of wind profiler data in U.S. weather forecasting. Preprints, 20th Conf. on Weather Analysis and Forecasting/16th Conf. on Numerical Weather Prediction, Seattle, WA, Amer. Meteor. Soc., CD-ROM, 6.1.

  • Kuo, Y-H., , and Anthes R. A. , 1985: Calculation of geopotential and temperature fields from an array of nearly continuous wind observation data. J. Atmos. Oceanic Technol., 2 , 2234.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Laroche, S., , Gauthier P. , , St-James J. , , and Morneau J. , 1999: Implementation of a 3D variational data assimilation system at the Canadian Meteorological Centre. Part II: The regional analysis. Atmos.–Ocean, 37 , 281307.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P., 1993: Comparison of wind-profiler and radiosonde measurements in the tropics. J. Atmos. Oceanic Technol., 10 , 122127.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Trexler, B. L., , and Koch S. E. , 2000: The life cycle of a mesocale gravity wave as observed by a network of Doppler wind profilers. Mon. Wea. Rev., 128 , 24232446.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wagneur, N., , and Garand L. , 2003: Operational assimilation of GOES imager water vapor channel at MSC. Proc. Int. TOVS Study Conf. XIII, Ste. Adele, QC, Canada, WMO, 654–659.

  • Weber, B. L., , and Wuertz D. B. , 1990: Comparison of rawinsonde and wind profiler radar measurements. J. Atmos. Oceanic Technol., 7 , 157174.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weber, B. L., , Wuertz D. B. , , Welsh D. C. , , and McPeek R. , 1993: Quality controls for profiler measurements of winds and RASS temperatures. J. Atmos. Oceanic Technol., 10 , 452464.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xu, Q., , and Wei L. , 2001: Estimation of three-dimensional error covariances. Part II: Analysis of wind innovation vectors. Mon. Wea. Rev., 129 , 29392954.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • View in gallery

    The NOAA Profiler Network.

  • View in gallery

    Std dev of the observation wind module error as a function of height for profilers (asterisk) and radiosonde (lozenge) data.

  • View in gallery

    Vertical profile of the observation wind components error of wind profilers for the zonal (asterisk) and the meridional (triangle).

  • View in gallery

    Vertical profile of the analysis increment in response to the vertical correlation of the observation.

  • View in gallery

    Vertical observation error correlation as a function of vertical distance (m) between levels for the (left) low and (right) high modes. The asterisks represent an average over a 250-m interval.

  • View in gallery

    Flowchart of wind profiler data processing.

  • View in gallery

    Zonal wind rms (thick) and bias (thin) as a function of pressure for the full resolution (solid) and one out of three (dashed) profiler 3DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

  • View in gallery

    Zonal wind rms (thick) and bias (thin) as a function of pressure for the control (solid) and one out of three (dashed) profiler 3DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

  • View in gallery

    Time series of the rms of the 6-h departure (O–F) of the zonal wind at 500 hPa for the control (solid) and the profiler 3DVAR experiment (dashed) for the first 14 days of Jan 2004.

  • View in gallery

    Difference in rms of the 48-h forecast error of wind speed (m s−1) between the profiler 3DVAR experiment and the control at 250 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the profiler experiment (dark shading with solid lines) than in the control.

  • View in gallery

    Difference in rms of the 48-h forecast error of geopotential (dam) between the profiler 3DVAR experiment and the control at 700 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the profiler experiment (dark shading with solid lines) than in the control.

  • View in gallery

    Zonal wind rms (thick) and bias (thin) as a function of pressure for the full resolution (solid) and one out of three (dashed) profiler 4DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

  • View in gallery

    Difference in rms of the 48-h forecast error of wind speed (m s−1) between the full vertical resolution profiler 4DVAR experiment and the vertically filtered profiler experiment at 250 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error for the vertically filtered profiler experiment (dark shading with solid lines) than in the full vertical resolution profiler experiment.

  • View in gallery

    Difference in rms of the 48-h forecast error of geopotential (dam) between the full vertical resolution profiler 4DVAR experiment and the vertically filtered profiler experiment at 700 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the vertically filtered profiler experiment (dark shading with solid lines) than in the full vertical resolution experiment.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 77 77 13
PDF Downloads 78 78 10

Assimilation of Wind Profiler Data in the Canadian Meteorological Centre’s Analysis Systems

View More View Less
  • 1 Canadian Meteorological Centre, Meteorological Service of Canada, Dorval, Québec, Canada
  • | 2 Meteorological Research Branch, Meteorological Service of Canada, Dorval, Québec, Canada
© Get Permissions
Full access

Abstract

Real-time horizontal wind observations from the National Oceanic and Atmospheric Administration’s (NOAA’s) Profiler Network (NPN) are assessed in preparation for their assimilation in the Canadian Meteorological Centre (CMC) analysis systems. As a first step, radiosonde winds from 20 stations were compared to the central U.S. profiler stations over the 2001/02 winter season. It was found that profilers are at least as good as conventional radiosonde data. The 2001/02 winter season data were also used to examine the vertical correlation structure of the observation error for profilers. Using a statistical analysis of innovations, the observation error standard deviation of the wind components is estimated as 2.2 m s−1 and the vertical correlation length is approximately 500 m. These results suggest that the data are vertically correlated because they are available every 250 m. Therefore, a thinning process is proposed in which one out of three data are selected in the vertical for each station.

Since January 2004, a close monitoring of NPN profiler data revealed significant errors at some stations in the lower and upper troposphere. Consequently, a monthly blacklist of NPN profilers is built based on data from the previous month.

A data impact study with both the three-dimensional variational data assimilation (3DVAR) and four-dimensional variational data assimilation (4DVAR) analysis systems was conducted using data from the 2003/04 winter season in which the vertical thinning was tested. It was found that the vertical thinning improves slightly the 6-h forecast error, especially in the 4DVAR over the central United States in which 6 times more profilers are assimilated. The impact of the vertical thinning is found to be neutral in the 3DVAR. Also, the impact of profiler data is significant over the central U.S. domain compared to a control run with the only difference being the addition of profiler data. These results were sufficiently good to implement NPN profilers in both the CMC global and regional analysis systems with the thinning process in fall of 2004.

Corresponding author address: Stéphane Laroche, Data Assimilation and Satellite Meteorology, 2121 Route Trans-Canadienne, Dorval, Québec, H9P 1J3, Canada. Email: stephane.laroche@ec.gc.ca

Abstract

Real-time horizontal wind observations from the National Oceanic and Atmospheric Administration’s (NOAA’s) Profiler Network (NPN) are assessed in preparation for their assimilation in the Canadian Meteorological Centre (CMC) analysis systems. As a first step, radiosonde winds from 20 stations were compared to the central U.S. profiler stations over the 2001/02 winter season. It was found that profilers are at least as good as conventional radiosonde data. The 2001/02 winter season data were also used to examine the vertical correlation structure of the observation error for profilers. Using a statistical analysis of innovations, the observation error standard deviation of the wind components is estimated as 2.2 m s−1 and the vertical correlation length is approximately 500 m. These results suggest that the data are vertically correlated because they are available every 250 m. Therefore, a thinning process is proposed in which one out of three data are selected in the vertical for each station.

Since January 2004, a close monitoring of NPN profiler data revealed significant errors at some stations in the lower and upper troposphere. Consequently, a monthly blacklist of NPN profilers is built based on data from the previous month.

A data impact study with both the three-dimensional variational data assimilation (3DVAR) and four-dimensional variational data assimilation (4DVAR) analysis systems was conducted using data from the 2003/04 winter season in which the vertical thinning was tested. It was found that the vertical thinning improves slightly the 6-h forecast error, especially in the 4DVAR over the central United States in which 6 times more profilers are assimilated. The impact of the vertical thinning is found to be neutral in the 3DVAR. Also, the impact of profiler data is significant over the central U.S. domain compared to a control run with the only difference being the addition of profiler data. These results were sufficiently good to implement NPN profilers in both the CMC global and regional analysis systems with the thinning process in fall of 2004.

Corresponding author address: Stéphane Laroche, Data Assimilation and Satellite Meteorology, 2121 Route Trans-Canadienne, Dorval, Québec, H9P 1J3, Canada. Email: stephane.laroche@ec.gc.ca

1. Introduction

The National Oceanic and Atmospheric Administration’s (NOAA’s) Profiler Network (NPN) is composed of 35 UHF radars operating in full capability since 1992 (Fig. 1). The radars provide hourly average measurements of horizontal wind vectors with high vertical resolution, which are available in real time over the Global Telecommunications System (GTS). The 404-MHz wind profilers operate in two separate modes—low [500–9250 m above ground level (AGL)] or high (7500–16 250 m AGL). To sample these higher altitudes a longer pulse (increased power) is needed. Therefore, with the longer (shorter) pulse in the high (low) mode, a lesser (higher) resolution of 900 (300) m is attained. Winds measured by the profiler are an average within each resolution volume, centered every 250 m vertically.

These data were first used mainly to detect and study mesoscale weather structures. The relative high spatial and temporal resolution of the network over central United States makes it possible to resolve meteorological features, such as sharp baroclinic zones and gravity waves (Bluestein and Speheger 1995; Trexler and Koch 2000). Substantial efforts have also been made to retrieve the mass field from the wind profiler data to further diagnose the structure of these phenomena (e.g., Kuo and Anthes 1985; Gal-Chen 1988; Bussinger et al. 2001).

NPN winds have been found to be of a high quality and reliable for data assimilation both in mesoscale and global forecasting systems (Bouttier 2001; Koch et al. 2004; Benjamin et al. 2004b). These winds are already assimilated at the National Centers for Environmental Prediction (NCEP) and at the European Centre for Medium-Range Weather Forecasts (ECMWF). In the Rapid Update Cycle (RUC) data assimilation system (Benjamin et al. 2004a), the forecast impact has been found to be significant only at short range (3–6 h) over the central United States below 300 hPa where there are relatively few aircraft reports (Koch et al. 2004; Benjamin et al. 2004b). Bouttier (2001) showed the benefits of assimilating the NPN wind data, not only at the short range over North America but also over Europe at longer range. This was, however, possible through a close data monitoring, which revealed that observations below 700 and above 400 hPa can be representative of phenomena such as lee waves or convective systems that are not well captured in the short-range forecast used as background field. This study also pointed out that the vertical correlation of the observation error for wind profilers might not be negligible. Nevertheless, wind profiler data in the ECMWF forecasting system were implemented by neglecting the correlation of the observation error and by setting the observation error variances to be the same as for radiosonde reports. This is consistent with intercomparison studies of similar NPN wind profiler and radiosonde data, which show that both measurements are of the same quality. For instance, Weber and Wuertz (1990) found rms differences of the order of 2.5 m s−1 over Colorado, while May (1993) found differences as low as 1.5 m s−1 in the Tropics. These differences are within the observational error of both instruments.

The Canadian Meteorological Centre (CMC) operates real-time global and regional data assimilation systems for the production of its medium- and short-range forecasts. Both systems use the same three-dimensional variational data assimilation (3DVAR) analysis scheme (Gauthier et al. 1999a; Laroche et al. 1999), along with the Global Environmental Multiscale (GEM) model (Côté et al. 1998). Extension of the analysis scheme to four-dimensional variational data assimilation (4DVAR) was under way during the writing of this paper in view of its operational implementation by early 2005.

This paper presents an assessment of the value of the NPN winds in view of their assimilation in the global forecasting system. In particular, the impact of the vertical observation error correlation is assessed. The NPN data have also been introduced in the regional data assimilation system but the results for this system will not be presented in this paper.

The variational data assimilation scheme is briefly described in section 2. The vertical structure of the observation error is examined in section 3 in order to estimate the vertical correlation and observational error variances. The profiler data are also compared to radiosonde wind data over the central United States. The quality control check of the profiler data is presented in section 4. Results from preimplementation tests during the winter period of 2003/04 in the global forecasting system are presented in section 5. Some preliminary results with a 4DVAR system will also be presented. The conclusions are summarized in section 6.

2. Variational data assimilation system

The variational analysis scheme developed at the CMC is based on the incremental formulation proposed by Courtier et al. (1994). It is a 6-h intermittent data assimilation cycle in which the analysis increment δx with respect to the background state xb (a 6-h forecast) is obtained by minimizing the following objective function:
i1520-0426-22-8-1181-e1
where 𝗚 is a general linear approximation of the operator G, which maps the model state into the observation space; 𝗕 and 𝗥 are the covariance matrices of background and observation errors, respectively, and d′ is the innovation defined as
i1520-0426-22-8-1181-e2
where d represents the observations. This very general formulation is valid for both 3DVAR and 4DVAR. In 3DVAR, G is the observation operator (usually referred to as H), while in 4DVAR, G also includes the prediction model, which propagates the background state to the time of the observations. At the minimum, the solution for the analysis increment corresponds to
i1520-0426-22-8-1181-e3
Global and regional analyses are performed 4 times a day at synoptic times (0000, 0600, 1200, 1800 UTC). In the global forecasting system, the horizontal resolution of the GEM model is 0.9°, whereas the analysis increment is calculated at a lower resolution of 1.5°.

The main advantage of a variational scheme is its ability to assimilate various types of observations. Observations that are currently assimilated are listed in Table 1. Since the implementation of 3DVAR at the CMC in 1997, radiances from the Television Infrared Observation Satellite (TIROS) Operational Vertical Sounders (TOVS) (Chouinard and Halle 1999), temperature from radiosondes and aircrafts (Chouinard et al. 2001), and radiances for the Geostationary Operational Environment Satellite (GOES) (Wagneur and Garand 2003) have been introduced. Other types of observations, such as the vertically integrated water content from the Special Sensor Microwave Imager (SSM/I) (Deblonde 2001) and surface wind from scatterometers (Buehner 2002), have also been examined and will be assimilated in the near future. In the current state of the CMC 3DVAR system, NPN data are used as vertical profiles of horizontal wind components, and those that are valid at synoptic times are assimilated. In the 4DVAR context, all hourly reports over the 6-h assimilation window are considered. A complete description of the background term and its associated error statistic 𝗕 can be found in Gauthier et al. (1999b). Finally, the data quality control, which is an important part of the variational data assimilation, is described in section 4.

3. Observation error structure

The hourly averaged wind profiler data, gathered through the GTS, have already gone through a quality control check, which basically relies upon the continuity and consistency of data over time and over height (Weber et al. 1993). Although this a priori check eliminates most erroneous measurements, some data may still be contaminated by undetected and representativeness errors. It is, therefore, important to estimate the observation error structure that includes both measurement and representativeness errors, and to perform additional quality control that is more specific to our data assimilation and forecasting systems.

a. Innovation method

The innovation method (Hollingsworth and Lonnberg 1986; Daley 1991; Xu and Wei 2001) is used to estimate observation errors. First, longitudinal and transverse wind components (l, t) of the innovation are computed for each observation station pair. Then, their associated covariances are calculated for each pair of observation stations and binned for each distance interval of 100 km over the range of 0–2000 km. The observation and forecast errors can be partitioned under the assumption that the observational error is horizontally uncorrelated and that the forecast error is horizontally homogeneous. In this case, the vertical observation error covariance matrix can be estimated from
i1520-0426-22-8-1181-e4
i1520-0426-22-8-1181-e5
where 𝗖ob(m, n) is the observation error covariance matrix, 𝗖(r, m, n) is the total innovation covariance matrix, the angle brackets denote the ensemble mean over the time period, subscripts i and j refer to two stations, subscripts m and n refer to two vertical levels, and r is the distance between stations i and j. First 𝗖(r → 0, m, n) is evaluated with 〈lim, ljn〉 + 〈tim, tjn〉 at the zero-distance limit of (5). Then 𝗖ob(m, n) is obtained by subtracting 𝗖(r → 0, m, n) from the average innovation covariance matrix calculated for each station. See Xu and Wei (2001) for more details. With this method, both observation error variance and vertical correlation can be estimated. The error variance as a function of height is simply the diagonal of 𝗖ob(m, n), while the error correlation is obtained from
i1520-0426-22-8-1181-e6
where σ is the observation error standard deviation.

b. Profiler and radiosonde observation errors

As a first step, we compared profiler data with independent data, such as radiosonde winds for the 2001/02 winter season. Observation error statistics for both radiosondes and profilers were calculated with the innovation method. We used 20 radiosonde stations nearest to the 30 profiler stations over central United States. Figure 2 shows the vertical profile of the standard deviation of the observation wind module error for both profiler and radiosonde data over the central United States as a function of height. Near the surface, the radiosonde observation error is smaller than the profiler observation error. This may be a result of the higher representativeness error in the planetary boundary layer, as suggested by Bouttier (2001). In the midtroposphere, both observation errors compare well, and above 8500 m the profiler observation error is smaller than the radiosonde observation error. Also, the profiler observation error variations are smoother in the vertical, unlike radiosondes whose observation errors vary a little more with height, especially in the upper troposphere. Because of fewer radiosonde stations used in the error estimations and the larger distance separating the stations compared to the distance between profiler stations, the observation statistics in the upper troposphere are a little more irregular. The vertical smoothness of the profiler observation errors seen in Fig. 2 suggests that they are correlated with height. Overall, profiler data are of as good a quality as radiosonde data, which are in agreement with previous studies (Weber and Wuertz 1990; May 1993).

The estimated standard deviations of the observation wind component errors of profilers are plotted in Fig. 3 as functions of vertical levels. The values for both the zonal and meridional winds vary slightly with height. Based on Fig. 3, the standard deviation of the observation error in the analysis scheme is set to 2.2 m s−1 at all levels. This observation error is slightly smaller than that for the pilot reports, which is 2.5 m s−1 in our analysis system. This result is also in good agreement with other studies on the quality of wind profiler data.

c. Vertical observation error correlation

As shown in the previous section, the vertical observation error correlation for the wind profiler is not negligible. This problem becomes even more important in the 4DVAR context in which hourly wind data are assimilated. For a 6-h assimilation time window, the volume of data may be 6 times greater. Technically, it is not an easy task to deal with observation error correlation in an operational data assimilation context. Two avenues are possible to avoid accounting for error correlation—thinning the observations in the vertical or inflating the observation error variances, as suggested by Bouttier (2001). The objective of both approaches is the same: to obtain the proper weight of the new source of observations in the analysis.

Figure 4 shows the impact of hypothetical observations for which their errors are correlated in the vertical (60 levels). We suppose here that the innovation is equal to 1 across the vertical domain (dotted line); the background and observation errors statistics are Gaussian with characteristic scale lengths of 3 and 1.25 grid points, respectively, with error variances of 1 for both. All the units are nondimensional. This experimental setup was designed to replicate, in a simplified manner, the assimilation of the profiler data in our variational analysis scheme. The thick solid curve is the analysis increment obtained from (3) when the observation error correlation is properly taken into account, while the thin solid curve is the result obtained when the observation error correlation is neglected. When the error correlation is neglected, the analysis increment is shifted toward the innovation value, meaning that the weight of the observation is too strong. The dashed curve is the analysis increment obtained by thinning the observations (1 out of 3), while the dashed–dot curve is obtained by inflating the observation error variance by 225%. In both cases, the observation error correlation is omitted. As expected, these two solutions are similar and provide the right weight to the observation. Because fewer observations are assimilated as a result of the thinning process, this approach is cheaper than the inflation approach. We have, thus, retained the thinning process in the treatment of the profiler data.

Figure 5 shows the vertical correlation statistics for the observation error as a function of distance between the vertical levels. The correlation structures are calculated with (6) for the two separate modes (i.e., low and high). The hypothesis of homogeneity and isotropy in the vertical is applied for each mode. The characteristic length for the vertical observation error correlation is 500 m for the low mode and 513 m for the high mode. This result is somewhat surprising, because a longer pulse is used in the high mode, suggesting that the length scale may be longer in the upper troposphere. Because the wind measurements are the result of an average within a resolution volume of 300 (900) m in the low (high) mode, the correlation lengths and Fig. 5 suggest that the wind profiler data reported at every 250 m vertically are, indeed, correlated.

From these results, we adopted the following thinning procedure: the first two levels above ground level are rejected and the third level AGL is used, followed by the rejection of the next two levels, etc. Of the 63 levels that are available for assimilation, approximately 20 observations are actually used in the analysis. The vertical thinning process reduces the time required for the assimilation and allows the use of a diagonal error covariance matrix.

4. Quality control

Like all the of the data types at CMC, wind profiler data undergo various quality control checks. A monthly monitoring of wind data compares wind measurements with short-range (6 h) forecasts at each NPN station. After several months of monitoring, we noticed significant errors at certain levels, especially near the surface and in the upper troposphere. As pointed out in Bouttier (2001), this is thought to be a combination of data reliability problems and a lack of realism in model physics and insufficient model resolution, especially if there is deep convection or orographic drag in the vicinity. These deficiencies led us to consider the use of a monthly blacklist for suspect station levels. Also, the monitoring of profiler data clearly indicates problems at specific levels for some stations, hence, a blacklist composed of levels per station. Based on procedures for the exchange of monitoring results by the World Meteorological Organization, two criteria are established for the production of the monthly blacklist. The gross error limits that are used for the observed minus background fields of wind speed are listed in Table 2. If the number of observation minus background fields cases at a certain height exceeds the limit by more than 50% during the previous month, this level appears on the blacklist. When at least 10 observations per level from the previous month are available, we calculate an rms vector departure from the background field (gross errors are not considered). If the rms vector departure exceeds 10 m s−1, the level for that particular station appears on the blacklist. An example of a blacklist for February 2004, produced with profiler measurements of January 2004, is given in Table 3. Profiler levels on the blacklist are not assimilated in the analysis scheme.

The profiler data are submitted to the following quality control in the data assimilation cycle. During the background quality control (Järvinen and Undén 1997), the model counterparts for profiler observations are calculated through the observation operator. The square of the background departure is considered as being suspect when it exceeds its expected variance by more than a predefined limit set to 8. This limit is defined as the ratio between the variance of the background departure, and the sum of the variance of the background error and the variance of the observational error. The background quality control rejects the observations with obvious gross errors.

Finally, a variational quality control (Andersson and Järvinen 1999) is performed during the 3DVAR. It operates as part of the variational problem and is based on Bayesian probability theory. The cost function is modified during the iterative minimization process in order to take into account the non-Gaussian nature of the gross errors. This has the effect of reducing the analysis weight that is given to data with large departures from the current or preliminary analysis.

The operational data processing of wind profilers is described in Fig. 6. First, the blacklisting process withholds from the analysis all data levels of the stations that appear on the blacklist. The remaining data are submitted to a background quality control that eliminates observations with obvious gross errors. The NPN profilers that are valid at the analysis time are chosen as part of the temporal and station selection process. Then, the vertical thinning process, which selects about 20 of the 63 levels, is applied to the profilers. Finally, profilers are assimilated in the analysis scheme and the variational quality control is activated during the minimization process.

5. Data impact study

A data impact study has been carried out from 11 December 2003 to 2 February 2004 with the global data assimilation system. Data from the 33 stations were used every 6 h, centered at the analysis validity time. The 3DVAR and 4DVAR experimental assimilation cycles were run with 10-day forecasts that ran at 0000 and 1200 UTC each day. In both cases, two experiments were conducted—the assimilation of profiler data with and without vertical thinning. These experiments were compared with each other in order to evaluate the impact of the vertical thinning. In the case of the 3DVAR, the best experiment was then compared to a control run with the only difference being the addition of the profiler data. The results are evaluated in terms of radiosonde (O) minus analysis (A) and radiosonde minus background field (F) departure statistics for the 2-month period (123 cases). The departure statistics are evaluated over the central United States (28°–40°N, 85°–111°W). As for the 10-day forecasts, they are compared to its respective analysis up to 48 h.

In the case of the 4DVAR, the results presented here are considered as preliminary because our 4DVAR scheme was still under preparation during the writing of this paper. In the 4DVAR context, profiler data were assimilated hourly leading to 6 times more data than in the 3DVAR. All of the other data types were the same as the ones used in the 3DVAR experiments.

a. 3DVAR results

The impacts of the NPN profiler data are first examined in the assimilation cycle. Vertical profiles of departure statistics for both experiments using profiler data are first compared over the central United States in Fig. 7. The analysis and background fields are compared to radiosonde data. The analysis departure is smaller for the vertically thinned profiler experiment. The assimilation cycle with less profiler data compares better to radiosonde data in the rms statistics, as expected, because less weight is given to the profiler data. Vertical thinning of profiler data reduces the competition with the radiosonde data giving more weight (or a better fit) to these data. The background departure statistics are very similar for both experiments with and without the vertical thinning. A slight change in the bias is noticeable, but the rms differences are practically identical. Based on these statistics, there is no need to assimilate the profiler data at its full vertical resolution. The need for vertical thinning will also be examined in the 4DVAR context. The remaining comparisons will be done with the one-out-of-three data selection experiment.

Vertical profiles of departure statistics for the control versus profiler experiment are shown in Fig. 8 over the central United States. The addition of profiler data increases the O–A departure over 200 hPa, which indicates that the profiler data are not always in agreement with the radiosonde data, and provides more competing data in the analysis. The profiler experiment has a smaller O–F departure than the control experiment, especially in the rms statistics. Short-range forecasts of other variables (geopotential, temperature, and humidity, not shown) also benefit from the assimilation of profiler data because of the multivariate effect of the analysis scheme. This improvement is, however, limited to the central United States where most of the NPN profiler data are available. Time series of the background departure (O–F) of the zonal wind at 500 hPa for the profiler experiment and control are displayed in Fig. 9 for the first 14 days of January 2004. The variations on a day-to-day basis clearly indicate a positive impact of assimilating profiler wind data most of the time. The positive impact is sometimes bigger depending on the weather regime over the central United States, as is also reported by Koch et al. (2004).

The rms difference in the 48-h forecast error of the wind speed at 250 hPa (profiler minus control) is shown in Fig. 10. Contour solid lines indicate lower errors in the experiment using profiler data. The assimilation of NPN profilers has a positive impact over almost all of North America. This positive impact can also be seen in the geopotential at 700 hPa (Fig. 11). However, the forecast impact is significant only at the short range and does not extend beyond 48 h.

b. 4DVAR results

Departure statistics of O–A and O–F against radiosonde data are presented in Fig. 12 over the central United States. The analysis departure with less profiler data compares better with the radiosonde data. As for the 6-h forecast, the filtered data experiment compares slightly better to radiosonde data. The vertical selection of the profiler data has a larger impact in the 4DVAR than in the 3DVAR because 6 times more data are assimilated in the 4DVAR (every hour instead of one every 6 h in the 3DVAR).

Difference in the rms of the forecast error between the full vertical resolution profiler experiment and the vertically filtered profiler experiment is evaluated for the same period as for the 3DVAR experiments. The rms difference in the 48-h forecast error of the wind speed at 250 hPa (the full vertical resolution profiler minus the vertically filtered profiler) is shown in Fig. 13. Here, the contour solid lines indicate lower errors in the vertically filtered profiler experiment. The thinning process allows for the 48-h wind forecast to be more consistent with its own analysis over most of North America. The rms difference of the geopotential at 700 hPa is also shown in Fig. 14. The positive impact of the vertically filtered profiler data is more concentrated over the North American east coast, where the downstream flow may be affected by forecasts originating from the profiler domain.

6. Conclusions

Data from the NOAA Profiler Network have been assessed in the CMC global analysis system in view of their operational implementation. An intercomparison study with wind data from the 20 nearest radiosonde stations for the 2001/02 winter season points to wind speed observation errors of the order of 3 m s−1. The observation error for the wind profiler is slightly higher than that for radiosondes in the lower troposphere, but becomes lower above 8500 m. This suggests that NPN profilers seem to be at least as good as conventional radiosonde data. Innovations were also used to determine the observation error of the wind components that is estimated as 2.2 m s−1. From the 2001/02 winter season data study, and from the close monitoring of the wind profiler data since January 2004, a monthly blacklist is established in which certain levels from some stations are discarded. The background and variational quality controls reject the few outliers that may remain.

The vertical correlation of the observation error was estimated by using the innovation method. A characteristic correlation length scale of the order of 500 m was found both in the low and upper troposphere. Because the data are available every 250 m in the vertical, the observation errors of the adjacent level are, thus, strongly correlated. To avoid explicitly accounting for vertical correlation in our analysis scheme, it was decided to filter the data in the vertical, which is a cheaper option than inflating the observation error variance. As a result, only one observation out of three in the vertical is selected.

A data impact study was carried out for the 2003/04 winter test period. Assimilation experiments with and without the vertical thinning were conducted in the global 3DVAR and 4DVAR systems. The assimilation of the wind profiler is especially beneficial for 6-h forecasts in the vicinity of the network. In the 3DVAR analysis system, these forecast errors for winds over the central United States are very similar with and without the vertical thinning, suggesting that there is no need to assimilate the profiler data at their full vertical resolution. The impact of the vertical correlation is, therefore, not significant in the 3DVAR context. However, in the 4DVAR analysis system, in which the profiler data volume is 6 times greater, the 6-h forecast error is slightly improved when the vertical thinning is applied. Moreover, the 48-h forecasts that are initialized with the vertically filtered profiler data are improved, compared to the assimilation of the full vertical resolution profiler data. This suggests that the vertically correlated profiler measurements do, indeed, require a vertical filtering process before their assimilation, especially in the 4DVAR. However, this benefit diminishes for a longer time range and does not extend to the 48-h range over North America. This is probably because of the small coverage of the network over the central United States and Alaska. It is expected that a wider network of wind profilers would be greatly beneficial for longer forecast ranges, as suggested by Koch et al. (2004).

A similar positive impact in the regional system was obtained (not shown here). These results were considered to be sufficiently good to justify the operational implementation in both the global and regional forecasting system in the fall of 2004.

In the near future, the quality control for profiler data will be improved by taking into account the quality control flags from the Profiler Hub, which captures bird migration problems. In the 4DVAR context, the temporal error correlation of the profiler data should also be examined. Finally, profilers from the European (Andersson and Garcia-Mendez 2002) and Japanese networks will be monitored and considered for operational implementation.

Acknowledgments

The authors wish to thank José Garcia, Lorraine Veillette, and Hamid Benhocine for the decoding and processing of the NPN profilers. The monitoring of the NPN profiler data was possible thanks to Réal Sarrazin. We are grateful to Dr. Louis Garand and Cortina Jone for their comments on the manuscript. We finally thank Dr. Stanley Benjamin for his constructive comments and suggestions.

REFERENCES

  • Andersson, E., , and Järvinen H. , 1999: Variational quality control. Quart. J. Roy. Meteor. Soc., 125 , 697722.

  • Andersson, E., , and Garcia-Mendez A. , 2002: Assessment of European wind profiler data, in an NWP context. ECMWF Tech. Memo. 372, 14 pp.

  • Benjamin, S. G., and Coauthors, 2004a: An hourly assimilation-forecast cycle: The RUC. Mon. Wea. Rev., 132 , 495518.

  • Benjamin, S. G., , Schwartz B. E. , , Stoke E. J. , , and Koch S. E. , 2004b: The value of wind profiler data in U.S. weather forecasting. Bull. Amer. Meteor. Soc., 85 , 18711886.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bluestein, H. B., , and Speheger D. A. , 1995: The dynamics of an upper-level trough in the baroclinic westerlies: Analysis based upon data from a wind profiler network. Mon. Wea. Rev., 123 , 23692383.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bouttier, F., 2001: The use of profiler data at ECMWF. ECMWF Research Department Tech. Memo. 331, 22 pp. [Available from ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom.].

    • Crossref
    • Export Citation
  • Buehner, M., 2002: Assimilation of ERS-2 scatterometer winds using the Canadian 3D-Var. Atmos.–Ocean, 40 , 361376.

  • Bussinger, S., , Adams M. E. , , Koch S. E. , , and Kaplan M. L. , 2001: Extraction of geopotential height and temperature structure from profiler and rawinsonde winds. Mon. Wea. Rev., 129 , 17291739.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Chouinard, C., , and Halle J. , 1999: The impact of TOVS radiances in the CMC 3D-Var analysis system. Proc. Int. TOVS Study Conf. X, Boulder, CO, WMO, 92–98.

  • Chouinard, C., , Charette C. , , Halle J. , , Gauthier P. , , Morneau J. , , and Sarrazin R. , 2001: The Canadian 3D-Var analysis scheme on model vertical coordinate. Preprints, 14th Conf. on Numerical Weather Prediction, Fort Lauderdale, FL, Amer. Meteor. Soc., 14–18.

  • Côté, J., , Gravel S. , , Méthot A. , , Patoine A. , , Roch M. , , and Staniforth A. , 1998: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part I: Design considerations and formulation. Mon. Wea. Rev., 126 , 13731395.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Courtier, P., , Thepaut J-N. , , and Hollingsworth A. , 1994: A strategy for operational implementation of 4D-Var using an incremental approach. Quart. J. Roy. Meteor. Soc., 120 , 13671387.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Daley, R., 1991: Atmospheric Data Analysis. Cambridge Atmospheric and Space Science Series, Vol. 2, Cambridge University Press, 457 pp.

  • Deblonde, G., 2001: Variational retrievals using SSM/I and SSM/T-2 brightness temperatures in clear and cloudy situation. J. Atmos. Oceanic Technol., 18 , 559576.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gal-Chen, T., 1988: A theory of the retrievals of virtual temperature from remote measurements of horizontal winds and thermal radiation. Mon. Wea. Rev., 116 , 13021319.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gauthier, P., , Charette C. , , Fillion L. , , Koclas P. , , and Laroche S. , 1999a: Implementation of a 3D variational data assimilation system at the Canadian Meteorological Centre. Part I: The global analysis. Atmos.–Ocean, 37 , 103156.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gauthier, P., , Buehner M. , , and Fillion L. , 1999b: Background-error statistics modeling in a 3D variational data assimilation scheme: Estimation and impact on the analysis. Proc. ECMWF Workshop on Diagnosis of Data Assimilation Systems, Reading, United Kingdom, ECMWF, 131–145.

  • Hollingsworth, A., , and Lonnberg P. , 1986: The statistical structure of short-range forecast errors as determined from radiosonde data. Part I: The wind field. Tellus, 38A , 111136.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Järvinen, H., , and Undén P. , 1997: Observation screening and background quality control in the ECMWF 3D-Var data assimilation system. ECMWF Research Department Tech. Memo. 236, 33 pp. [Available from ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom.].

  • Koch, S. E., , Benjamin S. G. , , Schwartz B. E. , , and Szoke E. J. , 2004: The value of wind profiler data in U.S. weather forecasting. Preprints, 20th Conf. on Weather Analysis and Forecasting/16th Conf. on Numerical Weather Prediction, Seattle, WA, Amer. Meteor. Soc., CD-ROM, 6.1.

  • Kuo, Y-H., , and Anthes R. A. , 1985: Calculation of geopotential and temperature fields from an array of nearly continuous wind observation data. J. Atmos. Oceanic Technol., 2 , 2234.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Laroche, S., , Gauthier P. , , St-James J. , , and Morneau J. , 1999: Implementation of a 3D variational data assimilation system at the Canadian Meteorological Centre. Part II: The regional analysis. Atmos.–Ocean, 37 , 281307.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P., 1993: Comparison of wind-profiler and radiosonde measurements in the tropics. J. Atmos. Oceanic Technol., 10 , 122127.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Trexler, B. L., , and Koch S. E. , 2000: The life cycle of a mesocale gravity wave as observed by a network of Doppler wind profilers. Mon. Wea. Rev., 128 , 24232446.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wagneur, N., , and Garand L. , 2003: Operational assimilation of GOES imager water vapor channel at MSC. Proc. Int. TOVS Study Conf. XIII, Ste. Adele, QC, Canada, WMO, 654–659.

  • Weber, B. L., , and Wuertz D. B. , 1990: Comparison of rawinsonde and wind profiler radar measurements. J. Atmos. Oceanic Technol., 7 , 157174.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weber, B. L., , Wuertz D. B. , , Welsh D. C. , , and McPeek R. , 1993: Quality controls for profiler measurements of winds and RASS temperatures. J. Atmos. Oceanic Technol., 10 , 452464.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xu, Q., , and Wei L. , 2001: Estimation of three-dimensional error covariances. Part II: Analysis of wind innovation vectors. Mon. Wea. Rev., 129 , 29392954.

    • Crossref
    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

The NOAA Profiler Network.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 2.
Fig. 2.

Std dev of the observation wind module error as a function of height for profilers (asterisk) and radiosonde (lozenge) data.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 3.
Fig. 3.

Vertical profile of the observation wind components error of wind profilers for the zonal (asterisk) and the meridional (triangle).

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 4.
Fig. 4.

Vertical profile of the analysis increment in response to the vertical correlation of the observation.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 5.
Fig. 5.

Vertical observation error correlation as a function of vertical distance (m) between levels for the (left) low and (right) high modes. The asterisks represent an average over a 250-m interval.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 6.
Fig. 6.

Flowchart of wind profiler data processing.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 7.
Fig. 7.

Zonal wind rms (thick) and bias (thin) as a function of pressure for the full resolution (solid) and one out of three (dashed) profiler 3DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 8.
Fig. 8.

Zonal wind rms (thick) and bias (thin) as a function of pressure for the control (solid) and one out of three (dashed) profiler 3DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 9.
Fig. 9.

Time series of the rms of the 6-h departure (O–F) of the zonal wind at 500 hPa for the control (solid) and the profiler 3DVAR experiment (dashed) for the first 14 days of Jan 2004.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 10.
Fig. 10.

Difference in rms of the 48-h forecast error of wind speed (m s−1) between the profiler 3DVAR experiment and the control at 250 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the profiler experiment (dark shading with solid lines) than in the control.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 11.
Fig. 11.

Difference in rms of the 48-h forecast error of geopotential (dam) between the profiler 3DVAR experiment and the control at 700 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the profiler experiment (dark shading with solid lines) than in the control.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 12.
Fig. 12.

Zonal wind rms (thick) and bias (thin) as a function of pressure for the full resolution (solid) and one out of three (dashed) profiler 4DVAR experiments over the central United States. (left) The radiosonde minus analysis departures, and (right) the radiosonde minus 6-h departures.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 13.
Fig. 13.

Difference in rms of the 48-h forecast error of wind speed (m s−1) between the full vertical resolution profiler 4DVAR experiment and the vertically filtered profiler experiment at 250 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error for the vertically filtered profiler experiment (dark shading with solid lines) than in the full vertical resolution profiler experiment.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Fig. 14.
Fig. 14.

Difference in rms of the 48-h forecast error of geopotential (dam) between the full vertical resolution profiler 4DVAR experiment and the vertically filtered profiler experiment at 700 hPa over North America for 11 Dec 2003–31 Jan 2004. Negative contours indicate smaller error in the vertically filtered profiler experiment (dark shading with solid lines) than in the full vertical resolution experiment.

Citation: Journal of Atmospheric and Oceanic Technology 22, 8; 10.1175/JTECH1765.1

Table 1.

Observations currently assimilated in the CMC analysis system.

Table 1.
Table 2.

Gross error limits for wind profilers as a function of height.

Table 2.
Table 3.

Blacklist of the NOAA profilers for Feb 2004.

Table 3.
Save