• Alishouse, J. C., S. Snyder, J. Vongsathorn, and R. R. Ferraro, 1990:Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens.,28, 811–816.

    • Crossref
    • Export Citation
  • Black, T. L., 1994: The new NMC mesoscale Eta model: Description and forecast examples. Wea. Forecasting,9, 265–278.

    • Crossref
    • Export Citation
  • Buizza, R., and A. Montani, 1999: Targeting observations using singular vectors. J. Atmos. Sci.,56, 2965–2985.

    • Crossref
    • Export Citation
  • Derber, J. C., and W.-S. Wu, 1998: The use of TOVS cloud-cleared radiances in the NCEP SSI analysis system. Mon. Wea. Rev.,126, 2287–2299.

    • Crossref
    • Export Citation
  • Diak, G., 1987: Assimilation of scalar versus horizontal gradient information from the VAS into a mesoscale model. J. Climate Appl. Meteor.,26, 738–748.

    • Crossref
    • Export Citation
  • Laudon, J., and D. Lenoski, 1997: The SGI Origin: A ccNUMA Highly Scalable Server. Proc. 24th Annual Int. Symp. on Computer Architecture (ISCA ’97), Denver, CO, SGI, 241–251.

    • Crossref
    • Export Citation
  • Leslie, L. M., G. A. Mills, L. W. Logan, D. J. Gauntlett, G. A. Kelly, M. J. Manton, J. L. McGregor, and J. M. Sardie, 1985: A high resolution primitive equations NWP model for operations and research. Aust. Meteor. Mag.,33, 11–35.

  • Menzel, W. P., F. C. Holt, T. J. Schmit, R. M. Aune, A. J. Schreiner, G. S. Wade, and D. G. Gray, 1998: Application of GOES-8/9 soundings to weather forecasting and nowcasting. Bull. Amer. Meteor. Soc.,79, 2059–2077.

    • Crossref
    • Export Citation
  • Nieman, S. J., W. P. Menzel, C. M. Hayden, D. Gray, S. T. Wanzong, C. S. Velden, and J. Daniels, 1997: Fully automated cloud-drift winds in NESDIS operations. Bull. Amer. Meteor. Soc.,78, 1121–1133.

    • Crossref
    • Export Citation
  • NWS, cited 1999: EDAS. NWS Technical Procedures Bulletin. [Available online at http://www.nws.noaa.gov/om/tpb/3d-eta.htm.].

  • Parrish, D., J. Purser, E. Rogers, and Y. Lin, 1996: The regional 3D variational analysis for the Eta model. Preprints, 11th Conf. on Numerical Weather Prediction, Norfolk, VA, Amer. Meteor. Soc., 454–455.

  • Raymond, W. H., and R. M. Aune, 1998: Improved precipitation forecasts using parameterized precipitation drag in a hydrostatic forecast model. Mon. Wea. Rev.,126, 693–710.

    • Crossref
    • Export Citation
  • Reale, A. L., 1995: Departures between derived satellite soundings and numerical weather forecasts: Present and future. Tech. Proc. Eighth Int. TOVS Study Conf., Queenstown, New Zealand, WMO, 395–404.

  • Rogers, E., D. Parrish, Y. Lin, and G. DiMego, 1996: The NCEP Eta data assimilation system: Tests with regional 3-D variational analysis and cycling. Preprints, 11th Conf. on Numerical Weather Prediction, Norfolk, VA, Amer. Meteor. Soc., 105–106.

  • ——, and Coauthors, 1997: Changes to the NCEP Operational “Early” Eta Analysis/Forecast System. NOAA/NWS Tech. Procedures Bull. 447, 16 pp. [Available from Office of Meteorology, National Weather Service, 1325 East–West Highway, Silver Spring, MD 20910.].

  • Schmidlin, F. J., 1988: WMO international radiosonde comparison, phase II final report, 1985. Instruments and Observing Methods Rep. 29, WMO/TD No. 312, WMO, Geneva, Switzerland, 113 pp.

  • Smith, W. L., H. M. Woolf, C. M. Hayden, D. Q. Wark, and L. M. McMillin, 1979: The TIROS-N operational vertical sounder. Bull. Amer. Meteor. Soc.,60, 1177–1187.

  • Treadon, R., 1993: The NMC Eta model post processor: A documentation. Office Note 394, National Oceanic and Atmospheric Administration, 42 pp. [Available from NASA Goddard Space Flight Center, Wallops Flight Facility, Wallops Island, VA 23337.].

  • Velden, C. S., S. J. Nieman, W. P. Menzel, and S. T. Wanzong, 1997:Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc.,78, 173–195.

    • Crossref
    • Export Citation
  • View in gallery

    EDAS analyses from the control run valid 0000 UTC 6 Feb 1998. (a) Mean sea level pressure contoured every 4 hPa. (b) 500-hPa geopotential height contoured every 60 m. (c) Accumulated precipitation during the last 3 h of the assimilation cycle (2100–0000 UTC) contoured every mm

  • View in gallery

    Combined GOES-8 and -9 infrared window (11 μm) image for 2345 UTC 5 Feb 1998

  • View in gallery

    EDAS 0000 UTC 8 Feb 1998 analysis of (a) mean sea level pressure contoured every 4 hPa and (b) 500-hPa geopotential height contoured every 60 m. A 48-h forecast of the same fields from a control run initialized 0000 UTC 6 Feb 1998 is shown in (c) and (d)

  • View in gallery

    The 0000 UTC 6 Feb 1998 data locations for (a) RAOB2, (b) ACAR2, and (c) GOESC. For easier identification, the dots are larger for RAOB2 than the other two data types

  • View in gallery

    Sensitivity of the EDAS analysis for this case to each data type. Rms differences between the control run’s temperature and each experimental run’s temperature were computed over (a) the entire model domain and (b) an area extending only over CONUS. The statistics were computed at three levels: 850, 500, and 300 hPa

  • View in gallery

    Temperature sensitivity at 500 hPa in the EDAS due to (b) RAOB1, (c) TOVCD, and (d) ACAR1. These sensitivities are contoured every 0.3 K and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 3 K. For clarity, no zero line is plotted in (b), (c), and (d)

  • View in gallery

    The same as Fig. 5 except the statistics are for the u component of the wind

  • View in gallery

    The u-component sensitivity at 300 hPa in the EDAS due to (b) RAOB2, (c) GOESC, and (d) ACAR2. These sensitivities are contoured every 1 m s−1 and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 10 m s−1. For clarity, no zero line is plotted in (b), (c), and (d)

  • View in gallery

    The same as Fig. 5 except the statistics are for relative humidity

  • View in gallery

    Relative humidity sensitivity at 850 hPa in the EDAS due to (b) RAOB1, (c) SSMI1, and (d) GOESM. These sensitivities are contoured every 5% and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 20%. For clarity, no zero line is plotted in (b), (c), and (d)

  • View in gallery

    Stacked bar charts displaying the (a) temperature (K), (b) u component (m s−1), and (c) relative humidity (%) sensitivity of the Eta 48-h forecast for this case to each data type. The statistics are limited to a subsection of the entire model domain east of roughly 110°W

  • View in gallery

    The same as Fig. 6 except the statistics are for 48 h and displayed in the subsection

  • View in gallery

    The same as Fig. 8 except the statistics are for 48 h and displayed in the subsection

  • View in gallery

    The same as Fig. 10 except the statistics are for 48 h and displayed in the subsection

  • View in gallery

    Stacked bar charts displaying the temperature (K), u component (m s−1), and relative humidity (%) sensitivity of the Eta 0-, 12-, and 48-h forecast for RAOB1, ACAR1, SSMI1, GOESM, RAOB2, ACAR2, and GOESC. The 0-h results include the entire model domain, while the 12- and 48-h results use only the subsection east of 110°W

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 33 33 6
PDF Downloads 16 16 4

A Case Study of the Sensitivity of the Eta Data Assimilation System

View More View Less
  • * Cooperative Institute for Meteorological Satellite Studies, Madison, Wisconsin
  • | + National Environmental Satellite, Data, and Information Service, Madison, Wisconsin
  • | # Environmental Modeling Center, National Centers for Environmental Prediction, Washington, D.C.
  • | @ Cooperative Institute for Mesoscale Meteorological Studies, Norman, Oklahoma
© Get Permissions
Full access

Abstract

A case study is utilized to determine the sensitivity of the Eta Data Assimilation System (EDAS) to all operational observational data types used within it. The work described in this paper should be of interest to Eta Model users trying to identify the impact of each data type and could benefit other modelers trying to use EDAS analyses and forecasts as initial conditions for other models.

The case study chosen is one characterized by strong Atlantic and Pacific maritime cyclogenesis, and is shortly after the EDAS began using three-dimensional variational analysis. The control run of the EDAS utilizes all 34 of the operational data types. One of these data types is then denied for each of the subsequent experimental runs. Differences between the experimental and control runs are analyzed to demonstrate the sensitivity of the EDAS system to each data type for the analysis and subsequent 48-h forecasts. Results show the necessity of various nonconventional observation types, such as aircraft data, satellite precipitable water, and cloud drift winds. These data types are demonstrated to have a significant impact, especially observations in maritime regions.

Corresponding author address: W. Paul Menzel, UW–CIMSS, 1225 West Dayton St., Madison, WI 53706.

Email: paul.menzel@ssec.wisc.edu

Abstract

A case study is utilized to determine the sensitivity of the Eta Data Assimilation System (EDAS) to all operational observational data types used within it. The work described in this paper should be of interest to Eta Model users trying to identify the impact of each data type and could benefit other modelers trying to use EDAS analyses and forecasts as initial conditions for other models.

The case study chosen is one characterized by strong Atlantic and Pacific maritime cyclogenesis, and is shortly after the EDAS began using three-dimensional variational analysis. The control run of the EDAS utilizes all 34 of the operational data types. One of these data types is then denied for each of the subsequent experimental runs. Differences between the experimental and control runs are analyzed to demonstrate the sensitivity of the EDAS system to each data type for the analysis and subsequent 48-h forecasts. Results show the necessity of various nonconventional observation types, such as aircraft data, satellite precipitable water, and cloud drift winds. These data types are demonstrated to have a significant impact, especially observations in maritime regions.

Corresponding author address: W. Paul Menzel, UW–CIMSS, 1225 West Dayton St., Madison, WI 53706.

Email: paul.menzel@ssec.wisc.edu

1. Introduction

The Cooperative Institute for Meteorological Satellite Studies (CIMSS) has been conducting data assimilation studies since the middle 1980s. The CIMSS Regional Assimilation System (CRAS) (Raymond and Aune 1998) evolved from the Australian Meteorology Research Center’s regional model (Leslie et al. 1985). The CRAS has served as a valuable tool for initial work on the assimilation of satellite data into numerical forecast models (Diak 1987), and as a source of information for satellite data producers about the strengths and limitations of the data. However, results from the CRAS have limited operational utility for a number of reasons. First, results from one numerical weather prediction (NWP) system do not ensure similar results from another. Second, assimilation procedures do not necessarily migrate in a meaningful way from one system to another. Third, different initial observational data types in two systems can be a source of inconsistent results.

For the above reasons, CIMSS has started using operational NWP systems to diagnose the importance of satellite data. Specifically, CIMSS has acquired the complete Eta Data Assimilation/Forecast System (EDAS) (Black 1994) from the National Centers for Environmental Prediction (NCEP) (Rogers et al. 1996). The goals of the CIMSS EDAS efforts are to maintain a system that is consistent with the operational EDAS, and to use the system in coordination with NCEP for investigating the impact of current and planned satellite data sources on operational forecasts.

This paper describes the results of an early experiment utilizing the CIMSS EDAS. Before planning long-term weekly or seasonal impact studies on experimental satellite data types, it was necessary to gain some knowledge about the sensitivity of the EDAS to data used operationally within the assimilation system. The EDAS was run repeatedly on one case study with varied sets of input data. A 48-h control run utilized all NCEP operational data types within the EDAS at the valid time of this case (6 Feb 1998). Separate experiments were then run to 48 h in which one of the 34 operational data types was denied. Differencing the individual experiments with the control run yields a measure of the sensitivity of the analysis system and forecast to each data type. Sensitivity results are presented for nine standard meteorological fields; they are 300-, 500-, and 850-hPa u component, temperature, and relative humidity. The regions of study at analysis time (0 h) include both the entire model domain and a subsection represented by the continental United States (CONUS). However, as a result of contamination from the lateral boundary conditions, the 48-h sensitivity results are limited to a domain east of roughly 110°W.

It is important to note that this is a single case study and that the sensitivity results presented here may not be representative of other synoptic flow regimes, geographical locations, or other data assimilation systems. It is equally important to note that, since a corresponding verification of the 48-h forecast was not done, a large impact by a particular data type does not necessarily translate to a better forecast.

The paper is organized as follows. Section 2 describes specifics of the CIMSS implementation of the EDAS for this experiment. Section 3 presents a synoptic description of the case study. Section 4 discusses the analysis system and its relationship to data amount and quality. Section 5 presents the analysis and 48-h results for various data denials. The results are summarized in section 6.

2. Experimental design and implementation

In February 1998, three-dimensional variational analysis (3D-VAR) (Parrish et al. 1996; Rogers et al. 1996) became the data assimilation tool of the operational EDAS. Several other changes were also made to the EDAS at that time, including an increase in horizontal resolution from 48 to 32 km, an increase in the number of vertical levels from 38 to 45, and an increase in the number of soil moisture levels from 2 to 4. Procedures for utilizing EDAS forecasts as the first guess, thereby “fully cycling” the system, were also developed and implemented in operations in the summer of 1998. Details of the operational EDAS configuration used in this study can be found in the work of Rogers et al. (1997);the EDAS has since been updated (NWS 1999).

CIMSS has acquired a Silicon Graphics Origin 2000 workstation (Laudon and Lenoski 1997) as a platform for EDAS research. The CIMSS Origin workstation has 2.0 gigabytes of available memory and eight 195-MHz microprocessors, although only four processors are typically used for EDAS research. A full 12-h assimilation cycle and 48-h forecast for the EDAS on the 48-km native grid requires approximately 17 h of wall clock time. As such, the control simulation plus the 34 data denial simulations presented here took 595 h of wall clock time. Such constraints severely limit the ability to perform week-long or even multiple case studies at this resolution. However, since CIMSS EDAS research is not meant to be an operational forecast tool, this performance is acceptable. Due to other modeling tasks that must be completed in real time, NCEP does not currently have the computational resources to routinely execute this quantity of EDAS parallel runs.

The case study described in this paper provides a first look at the sensitivity of the EDAS to all 34 operational data types used in the assimilation and forecast cycle for this case. Sensitivity of the fields is evaluated at both 0 and 48 h. The root-mean-square (rms) sensitivity S is defined as
i1520-0434-15-5-603-e1
where C is the control assimilation containing all 34 data types, D is the assimilation containing all but the one denied data type, and N is the total number of grid points on the isobaric level being evaluated. Grid points underground are not included in the statistics. Finally, while the experiments were run at 48 km, all results were interpolated to and displayed on the 80-km Nested Grid Model super C grid (grid number 104) using the Eta postprocessor (Treadon 1993). The statistics and plots herein were also obtained from grid 104.

3. Case study description

The case study presented comprises data describing atmospheric conditions at 0000 UTC 6 February 1998. The mean sea level pressure field from the control run of the EDAS (Fig. 1a) shows large low pressure centers off both coasts of the United States. The West Coast low is analyzed with a central pressure of approximately 980 hPa, while the East Coast low is approximately 986 hPa. A much weaker 1008-hPa cyclone is situated near the Big Bend region of Texas. An anticyclonic ridge extending from the Northwest Territories to Texas covers a large portion of CONUS and Canada.

Examination of the accompanying 500-hPa geopotential height analysis (Fig. 1b) reveals that the West Coast cyclone is vertically tilted toward colder air aloft, implying favorable conditions for further development as the low moves northeast and onshore. The East Coast low is nearly occluded and hence appears less subject to future development. Elsewhere, a fairly strong ridge is situated over the north-central United States, while a shortwave trough extends northeast–southwest from the middle Mississippi Valley into Mexico.

Accumulated precipitation during the last 3 h of the assimilation cycle of the control run (Fig. 1c) shows a strong and elongated area of intense precipitation off the California coast, associated with the frontal structure of the West Coast storm. Precipitation associated with the East Coast storm is weaker and for the most part north and east of the circulation center. Another smaller but intense region of precipitation is situated over the northwestern Gulf of Mexico.

Inspection of a combined and remapped Geostationary Operational Environmental Satellite (GOES) 11-μm infrared satellite image (Fig. 2) suggests that the precipitation fields shown at the end of the model assimilation cycle agree well with the observed cloud patterns. A convective feature near the Texas and Louisiana coast is represented in both the precipitation analysis and the satellite image. Large-scale cloud features are also indicated both east and west of CONUS. As such, the EDAS is satisfactorily characterizing the most important features of this case.

Figure 3 presents the synoptic distribution for 0000 UTC 8 February 1998 from the NCEP EDAS analysis (Figs. 3a and 3b) and after a 48-h Eta control run initialized 0000 UTC 6 February 1998 (Figs. 3c and 3d). Inspection of the analysis mean sea level pressure distribution (Fig. 3a) indicates that the cyclone originally off the East Coast has exited the model domain. The cyclone initially in southwest Texas has tracked across the Gulf of Mexico to a position east of South Carolina and intensified to approximately 990 hPa. The cyclone originally west of Oregon has moved inland to Montana and weakened considerably. A new 984-hPa cyclone is now positioned west of Oregon on the southeastern side of a large cyclonic circulation covering the northern Pacific Ocean. An anticyclonic ridge still covers much of CONUS and Canada, with the 1030-hPa core centered over southeast Hudson Bay.

Figure 3b shows that at 500 hPa a very strong trough is positioned south and east of Florida. Strong westerlies associated with this feature extend as far south as Jamaica and the Dominican Republic. There is still a relatively strong ridge over the north-central United States and a new negatively tilted shortwave trough positioned west of Oregon. The strong flow found in the central Pacific Ocean has been a persistent feature throughout the forecast period and has become more zonal with time.

Figures 3c and 3d indicate that the 48-h Eta Model control forecast is very similar to the verification. While the position or intensity of an individual feature may be shifted or slightly too strong or weak, this particular 48-h forecast accurately captured the strong cyclogenesis east of South Carolina and west of Oregon. The anticyclonic ridge, both at the surface and aloft, is also well represented over central CONUS and southern Canada.

4. Analysis and data

The EDAS regional 3D-VAR system has evolved from the global analysis system used at NCEP (Derber and Wu 1998). The analysis system minimizes an objective function of the form
JJbJoJbal
The Jb term corresponds to the fit of a prior estimate of atmospheric conditions, or background. The Jo term represents the fit to all types of real observations. The explicit Jbal term represents a weak balance constraint between the mass and velocity variables. When all three terms are expanded, the equation takes the form
i1520-0434-15-5-603-e3
This function is minimized at observation locations to yield the analysis. The resultant series of analysis values is represented by x. The series of background values is given by xb. The background error covariance matrix is B. The observational error covariance matrix is R. The series of observations is given by yo. The forward model is represented by H, which is used to convert the NWP model variables to estimates of the observations. The forward model for the balance relationship, Hbal, is a finite-difference approximation to the thermal wind equation evaluated at each grid point. The magnitude of the error assigned to the balance relationship, Bbal, determines the amount of balance enforced during the analysis.

Many factors influence the sensitivity of the variational analysis to a specific data type. The amount of observational data is important. The number of reports for each operational data type at each of the five analysis times during the 12-h assimilation cycle is shown in Table 1; this includes all routine data types used operationally within the EDAS for this case. In Table 1, every observation at a given level is considered to be a separate report, accounting for the existence of well over 5000 rawinsonde reports of temperature at T-12 and T-0 over North America. At each assimilation time, all data within 1.5 h are included. Therefore, data types such as surface airways observations and GOES precipitable water, which are routinely produced hourly, often have multiple reports at a given location for a given assimilation time. Real-time constraints on forecast generation usually lead to somewhat fewer reports at T-0 than are found at T-12.

Most data types used in the EDAS contain either mass (temperature, moisture, and pressure/height) or motion (wind) information, but not both. However, some data sample both mass and motion information (e.g., rawinsondes and aircraft reports). Data types with multiple information will be further clarified in this paper by appending a 1 after the mass observations and a 2 after the wind observations. Table 1 and subsequent Tables 2–6 also indicate, under the “Type” column, whether a data type is sampled predominantly in a clear/cloudy and/or land/marine environment.

Figure 4 depicts data locations at 0000 UTC 6 February 1998 for three data types. Rawinsonde winds (RAOB2), Aircraft Communications Addressing and Reporting System (ACARS) winds (ACAR2), and GOES high-density infrared cloud drift winds (GOESC) are displayed in Figs. 4a–c, respectively. The RAOB2 distribution is fairly typical of the twice-daily North American network. The ACAR2 distribution displayed here is also fairly typical, with many more flight-level observations over CONUS than any other region. However, a different time window could encompass more international flights. Finally, at this particular observation time, there is a wealth of GOESC information in both the Pacific and Atlantic Oceans. In general, raob’s have a larger number of observations at 0000 and 1200 UTC since they extend throughout the column, while ACARS and GOES have a more dense horizontal and temporal coverage but are single-level data. It is important to note that the location of many of these data types can change greatly on a daily or even hourly basis.

Given that 173 440 temperatures are reported cumulatively during the five analysis times from the Television Infrared Observation System (TIROS) Operational Vertical Sounder (TOVS) (Reale 1995) cloudy (TOVCD) retrievals, while only 14 477 rawinsonde temperatures are reported, one might expect a larger sensitivity to TOVS data than rawinsonde data within the 3D-VAR component of the EDAS. However, as will be shown later, the reverse is true. It is generally accepted that a data type with 10 reports has less chance to influence the analysis than a data type with 10 000 reports. It is less apparent that a 10 000 report data type may have the same amount of influence on the analysis as a data type with 100 000 reports. This is partially due to steps the 3D-VAR system used for this case study takes to reduce the volume of data, whereby the datum representing the median of all observational values in a given grid cell is chosen, and all others are ignored. [The procedure to reduce the volume of data within a grid cell has been amended in more recent versions of the EDAS (NWS 1999).]

Recent work has concentrated on the location of observations at critical times or the targeted observations approach (Buizza and Montani 1999). In such a case it is possible for 10 targeted observations to influence the analysis more significantly than a group of 10 000 observations with high observational error in a region where the forecast error is small. As such, the data’s spatial distribution can be more important than the number of observations. However, since this study deals with a predetermined set of available observations and times, the targeted observations approach is not pursued further.

A difference in EDAS characterization of the observational error estimates accounts for the difference in sensitivity to rawinsonde and TOVS temperatures. To characterize each data type, the operational EDAS assigns profiles of error estimates at each model level. The weights are inversely proportional to the errors assigned to an observation. Tables 2–6, respectively, show the observation errors assigned for this case at five pressure levels for temperature, velocity, precipitable water, specific humidity, and pressure/height. Note that errors assigned to rawinsonde temperatures are approximately 1 K, while errors assigned to TOVS cloudy temperature retrievals are about 7 K (Table 2). This accounts for the relatively small sensitivity to TOVS data presented in section 5, even though they represent a huge volume of data. It should be noted that NCEP has changed several of the observation error estimates since the time of this study. Specifically, the precipitable water weights presented in Table 4 have been increased to eight.

Another factor influencing the sensitivity of the analysis to a given data type is the accuracy of the forward model (H). The forward model incorporates both the calculations involved in converting model variables to observation estimates and the calculations involved in interpolating model variables from model grid points to observation locations and vice versa. A conversion is often necessary when incorporating satellite data types. For example, retrievals of multilayer (GOESM and GOESL) or vertically integrated (SSMI1) satellite precipitable water used in this study must be related to the model’s discrete profile of specific humidity. Future versions of 3D-VAR will also need to convert model quantities into GOES radiances so that the radiances may be assimilated directly.

Sensitivity of the variational analysis to data is also dependent on the accuracy assigned to the model forecast that is used as a background during the analysis. Error statistics compiled from one hundred, 80-km, 17 level, 24–36-h forecasts of the Eta Model are used to estimate the quality of the background. The use of 24–36-h forecasts is only meant to serve as a guide for obtaining an estimate of the horizontal and vertical variation of error variances, and vertical variation of the correlation lengths. The actual amplitude of the error is scaled linearly with time to give reasonable estimates of the errors for the 3-h forecasts needed during the 12-h assimilation cycle.

The background errors for height, temperature, and wind are a function of latitude and vertical position only, and do not evolve with time. The error variance is the product of a vertical table with two simple functions of latitude, which introduce some degree of realistic variation in the horizontal. The correlation lengths are only a function of vertical position. Tables 7 and 8 give approximate values for error amplitudes and correlation lengths, respectively. In Table 7 the units are degrees Kelvin for temperature, meters for height, and meters per second for wind. In Table 8 the units for the correlation lengths are kilometers. The wind has both longitudinal (along wind) and transverse (across wind) correlation lengths. The effective correlation functions are generated from recursive filters and are approximately equivalent to exp{−x2/(2*L2)}, for a separation distance x and correlation length L. The error variance for moisture is fixed at 80% of the background saturation specific humidity.

If a data type does not affect the analysis, there are three possible explanations. First is that the data were either too few in number or received too little weight. Second is that the environment sampled by the observation was already successfully depicted in the EDAS. Third is that the data may not have passed an assimilation quality control check. It is beyond the scope of this study to separate these possibilities.

Finally, although not discussed in detail, it is important to note that research on the error characteristics of data types and forward models has historically been minimal, especially for satellite retrievals. This probably accounts for the conservative assignment of satellite observation errors used operationally. It is generally agreed that forecast accuracy can be improved by better definition of observation and forward model errors.

5. Results

The EDAS was run 35 times for this case study. The control simulation utilized all of the operational data types used in the EDAS. One of the 34 data types was then denied for each subsequent experiment. Each run consisted of the complete 12-h assimilation cycle and a 48-h forecast. Data were assimilated via 3D-VAR at 1200 UTC 5 February (T-12), 1500 UTC 5 February (T-9), 1800 UTC 5 February (T-6), 2100 UTC 5 February (T-3), and 0000 UTC 6 February (T-0), using 3-h Eta Model forecasts between each assimilation step. The result of the 12-h assimilation cycle is the complete analysis before the Eta Model’s 48-h forecast cycle begins. Differences of the various experimental 12-h assimilation runs with the control 12-h assimilation run as determined by (1) yield a measure of the sensitivity of the EDAS to each individual data type for this case. Similarly, differences in the 48-h forecasts indicate the sensitivity of the Eta Model to each data type for this case.

Rms differences were computed for various quantities at three levels in the atmosphere (850, 500, and 300 hPa). These statistics were computed at analysis time over both the entire model domain and CONUS. However, during the forecast, statistics were limited to a region east of roughly 110°W. The need for using this subsection of the entire model domain during the forecast evaluations was determined largely by the persistent strong zonal flow over the Pacific Ocean described in section 3. Since the individual data types could not be denied from the lateral boundary conditions common to all simulations, the strong zonal flow rapidly propagated information of the individual data types from the boundary conditions eastward into the model domain.

a. The analysis results

The analysis results are summarized in Figs. 5–10. The bar graphs (Figs. 5, 7, and 9) are stacked bar graphs, so the contribution from each of the three levels is shown, and the total sensitivity is readily seen by the cumulative height of each bar. Contrasting the overall statistics (top chart) with those computed only over CONUS (bottom chart) gives an appreciation for the continental and maritime contributions of the various data types. Furthermore, in the colored geographical plots, positive sensitivities are shown in red.

1) Temperature sensitivity

Figures 5 and 6 present temperature sensitivity of the EDAS to the 34 data types. Overall, Fig. 5a shows that EDAS temperature analyses are most sensitive to conventional rawinsonde observations (Schmidlin 1988) of temperature, specific humidity, and pressure/height (RAOB1), which show a total sensitivity of 1.2 K. Rawinsonde wind observations (RAOB2), TOVS cloudy retrievals (TOVCD), ACARS temperatures (ACAR1), ACARS winds (ACAR2), Special Sensor Microwave/Imager (SSM/I) precipitable water (SSMI1), and GOES cloud drift winds (GOESC) are also significant, with total sensitivities ranging from 0.4 to 0.6 K. Over CONUS, Fig. 5b reveals significant cumulative contributions of greater than 0.5 K from RAOB1, RAOB2, ACAR1, and ACAR2. It is important to note that satellite temperature retrievals are produced, but not used over land in the operational EDAS.

Figure 6 shows geographical plots of the temperature sensitivity at 500 hPa due to rawinsonde (RAOB1), TOVS cloudy (TOVCD), and aircraft (ACAR1) data. Inspection of Fig. 6b reveals that the Alaskan, Canadian, and Mexican rawinsonde stations have the largest influence on the analysis, while Fig. 6d indicates that ACARS temperature data have little influence outside of CONUS at 500 hPa. Several factors probably contribute to these sensitivity patterns. First, there are relatively few raob and ACARS reports over oceanic regions (Figs. 4a and 4b). Second, outside of CONUS aircraft flights are typically at pressures less than 300 hPa; therefore, it takes longer to impact the fields at 500 hPa. Consequently, the small areas of temperature sensitivity seen east of Florida in Fig. 6d are predominantly a result of the advection of information that was assimilated over land during the 12-h assimilation cycle.

Modifications to the 500-hPa EDAS temperature analysis over oceans are mainly due to the assimilation of satellite data types. GOES high-density infrared cloud drift winds (GOESC) (Nieman et al. 1997), TOVS-retrieved temperatures (TOVCD) (Smith et al. 1979), GOES-retrieved multilayer precipitable water (GOESM) (Menzel et al. 1998), and Special Sensor Microwave/Imager (SSM/I) vertically integrated precipitable water (SSMI1) (Alishouse et al. 1990) all have a significant contribution over the entire model domain (Fig. 5a), but are much smaller over CONUS (Fig. 5b). Cumulative contributions from these four satellite data types range from 0.2 K for GOESM to 0.55 K for TOVCD over the entire domain to generally less than 0.2 K over CONUS. The utilization of TOVS cloudy retrievals leads to a modification of EDAS temperatures greater than 0.6 K at 500 hPa over the northern Pacific Ocean and southern Hudson Bay (Fig. 6c). The precipitable water datasets (SSMI1 and GOESM) may impact the temperature analysis within 3D-VAR through the balance term in (2). However, a larger domain-wide impact is realized through such effects as the thermal wind balance and lateral diffusion as the 3-h Eta forecasts proceed within the 12-h analysis time window. Eventually each data type included in the analysis impacts all grid points in the entire model domain.

Other data types also have a significant impact on EDAS temperatures. Conventional surface measurements of temperature and specific humidity (SFCM1, SFCL1) show approximately half the impact at 850 hPa as rawinsonde information, with each showing sensitivity near 0.2 K (Fig. 5a). Conventional aircraft reports (AIRE1) also have a measurable contribution, especially over oceanic regions at 300 hPa (0.3 K). Profiler (PROF) and velocity azimuth display (VAD) radar (RADAR) winds are shown to have a small 0.15 K signal overall (Fig. 5a), but the CONUS statistics reveal that they are twice as significant in the region they are measured (Fig. 5b). Data from one reconnaissance flight (RECO1) from San Francisco into the West Coast cyclone did have a small influence on the temperature analysis; its impact is approximately 0.2 K for both the entire domain and CONUS. Finally, since they are few in number or at the surface, the inclusion of most other data sources, such as Geostationary Meteorological Satellite (GMS) and Meteosat winds (GMSHI, GMSLO, METHI, METLO) and surface buoys (SFCBY), caused little change to the temperature analyses at 850, 500, or 300 hPa for this case (Fig. 5a). These data types are included only for completeness.

2) Wind sensitivity

Figures 7 and 8 show u-component wind sensitivity of the EDAS system to the various data types. The overall EDAS wind field is altered most by ACARS winds (ACAR2), both sets of radiosonde measurements (RAOB1 and RAOB2), GOES cloud drift winds (GOESC), SSM/I-retrieved precipitable water (SSMI1), conventional aircraft reports (AIRE2), and GOES water vapor motion winds (GOESW) (Fig. 7a). As with temperature sensitivity, the conventional data have the most influence on the analysis over land, while satellite data show most of their influence in oceanic regions. Similar to satellite temperature information, satellite wind information is produced over land but not used by the EDAS. The geographical distribution of changes made to the 300-hPa u component of the wind within the analysis by including RAOB2, GOESC, and ACAR2 wind measurements is shown in Fig. 8.

For this case, the EDAS u-component analysis is very sensitive to raob winds (RAOB2) (Fig. 8b) near radiosonde locations in Canada, Mexico, the southern Gulf of Mexico, and to a lesser extent on the west coast of the United States. The lack of sensitivity throughout most of the United States is probably due to the relatively high accuracy of the forecast background. The forecast should be very accurate over CONUS due to a wealth of other observations, and the comparatively easy forecast scenario of a high pressure system over the majority of the region. The west coast of the United States offers a different situation. In this region a large cyclonic circulation is upstream, over a region largely devoid of conventional observations. As such, by the end of the 12-h assimilation cycle, there is some significant information in the West Coast radiosonde data.

Some of the highest sensitivities to the inclusion of ACAR2 are also found predominately on the west coast of North America, presumably for the same reasons (Fig. 8d). The large volume of ACARS wind reports at all five assimilation times (Table 1) leads to a much more continuous field of observations over CONUS than is possible with rawinsondes. Larger differences in the wind analysis over CONUS due to ACARS winds than rawinsonde winds are the result. There is a smaller but possibly significant sensitivity to ACARS wind information for this case along flight paths from the United States to international destinations (see Fig. 8d, south of Alaska). However, the density of ACARS data over the oceans is much less than that over CONUS for this particular time window near 0000 UTC (see Fig. 4b). Furthermore, most oceanic ACARS data are taken at flight-level pressures less than 300 hPa. This accounts for the reduced sensitivity in oceanic regions compared to CONUS, which include much more ascent/descent information. Other time windows or days could conceivably have significantly more transoceanic flight information.

GOES high-density cloud drift winds (GOESC) are abundant in oceanic regions throughout the troposphere (Fig. 4c). They are also produced over land but are not utilized by the NCEP operational models. Thus, the EDAS 300-hPa wind analysis shows little sensitivity to cloud drift winds over CONUS, but widespread sensitivity over the Pacific and Atlantic Oceans (Fig. 8c). The oceanic cloud drift wind impact at 300 hPa is slightly less than the ACARS data impact over oceanic regions (cf. Figs. 8c and 8d). Finally, the entire model domain sensitivity of the 850-hPa wind analysis to GOESC is larger than any other data type at that level (Fig. 7a). This is probably due to the lack of large numbers of any other wind data type in the lower atmosphere near the two coastal cyclones. With large circulations like these, low-level cloud drift wind information is abundant and, at least for the Pacific circulation, propagates inland as the integration proceeds.

Other sources of wind information also contribute noticeably to the EDAS wind analysis. PROF and RADAR wind measurements each have a significant influence of 0.4 and 0.55 m s−1, respectively, at 850 hPa and 0.40 and 0.45 m s−1, respectively, at 500 hPa over CONUS (Fig. 7b), even though this is a regime dominated by weak high pressure. Winds from conventional aircraft reports such as AIRE2 also play a significant role (0.75 m s−1), mainly over the oceans at upper levels (Fig. 7a). For this particular case, ACARS winds (ACAR2) greatly outnumber AIREP winds (AIRE2) (Table 1), and ACARS winds have a lower estimated error than AIREP winds (Table 3). This leads to a much more substantial sensitivity in the EDAS 300-hPa wind field over the entire domain (Fig. 7a) from ACAR2 (1.1 m s−1) than AIRE2 (0.75 m s−1). The 0.35 m s−1 ACAR2 versus AIRE2 “difference” is deemed significant, since many other data types have a 300-hPa sensitivity of less than 0.2 m s−1. The vertical and horizontal spatial distributions of these data types is also extremely important, especially when examining a single case study. This also could contribute to the large difference.

While only the u component of the wind was presented here, the υ component was also studied. For the most part, the υ-component sensitivity at analysis time was nearly identical to the u component presented. In fact, many of the rms differences were within several hundredths of a meter per second.

3) Moisture sensitivity

The 0-h EDAS moisture sensitivity is presented in Figs. 9 and 10. The analysis of moisture within the EDAS for this case is dominated by the inclusion of rawinsonde temperature, specific humidity, and pressure and height data (RAOB1) (Figs. 9a and 9b). The RAOB1 impact is largely confined to North America and regions downstream of North America (Fig. 10b). Both SSM/I (SSMI1) (Fig. 10c) and GOES marine precipitable water (PW) data (GOESM) (Fig. 10d) make smaller but still important (several regions greater than 15%) contributions over the oceans. Modifications to the EDAS 850-hPa relative humidity analysis due to the utilization of GOESM are greater than 15% in both the Pacific Ocean and Caribbean Sea (Fig. 10d). SSM/I retrievals are typically attempted in all regions except those where heavy rain is falling, while GOES retrievals are concentrated in areas of clear air. Unlike sounder temperature and wind information, GOES sounder moisture information is used over land (GOESL) by the EDAS, but the resulting contribution to the analysis is negligible everywhere except over the larger Caribbean islands and the Mexican Gulf Coast (not shown).

RAOB2, ACAR2, and GOESC also have a significant influence on the analyzed relative humidity over both the entire domain (Fig. 9a) and CONUS (Fig. 9b). Their respective individual cumulative sensitivities are 5.5%, 5.0%, and 4.8% over the entire domain and 7.5%, 8.1%, and 2.6% over CONUS. These larger contributions are due in part to the strengthening and weakening of the synoptic-scale flow and, therefore, the different advection of moisture during the assimilation cycle. Small changes in moisture can be significant, more so than comparable changes in temperature and wind. Subtle phase shifts in moisture can cause large changes in relative humidity values, due to the tighter horizontal and vertical gradients moisture exhibits relative to other state fields. Temperature increases/decreases or a different vertical motion field due to wind changes can also cause large changes in relative humidity during the assimilation cycle. Both of these changes alter the relative humidity distribution during the assimilation, even though they are not directly related to moisture.

b. The 48-h forecast results

Figures 11a and 12 present the EDAS temperature sensitivity results at 48 h. Even though the 48-h stacked bar charts only include the domain east of roughly 110°W, a comparison of Fig. 11a (48 h) and Fig. 5 (analysis) is revealing for many reasons. Nearly all sensitivities have grown by 48 h. For example, surface reconnaissance observations (SFCR), which numbered only 16 reports during the entire 12-h assimilation and had a negligible influence on the entire domain analysis (Fig. 5a), show a cumulative sensitivity of just over 0.2 K at 48 h in the subsection (Fig. 11a). Furthermore, while several data types had a cumulative sensitivity of less than 0.1 K initially, no data type shows a cumulative sensitivity of less than 0.2 K at 48 h. Since the assimilation aspect of the EDAS has finished at analysis time, it is likely that this baseline of sensitivity has grown from 0.1 K initially to 0.2 K at 48 h due to growth of perturbations rather than observational causes. It also appears that a significant level of cumulative sensitivity is not reached until approximately 0.3 K at 48 h (Fig. 11a).

Some data types with a relatively large initial temperature impact, such as SFCM1 and AIRE2, have stayed nearly the same or decreased somewhat by 48 h. The impact of other data types, such as RAOB2, have increased substantially relative to the entire domain analysis value. Four data types, GMSHI, GMSLO, METHI, and METLO, have negligible initial and 48-h impacts. Since these data types only reach the western and eastern edges of the EDAS domain, respectively, it is generally assumed that they are much more important in the global system, which provides boundary conditions for the EDAS.

Figure 12 displays the 48-h sensitivity of the same three data types (RAOB1, TOVCD, and ACAR1) shown in Fig. 6 for the analysis. RAOB1 differences, which were initially large over North America, have advected east and north by 48 h. Largest sensitivity differences are now located over northeastern Canada and the western Atlantic Ocean. Similarly for the TOVCD data type (Fig. 12c), large sensitivity differences initially over Hudson Bay (Fig. 6c) have more than doubled by 48 h and are now found northeast of Newfoundland (Fig. 12c).

Figures 11b and 13 show the 48-h 300-hPa u-component sensitivity results for the subsection. While ACAR2 showed the greatest initial cumulative sensitivity over CONUS (Fig. 7b), RAOB1 has nearly doubled by 48 h and has the greatest influence of any data type at all levels in the subsection. Initial GOESC sensitivities (Fig. 7a) show a larger signal at 850 hPa than 300 hPa. This is due to the wealth of good tracers provided by extensive areas of well-defined stratocumulus clouds off both coasts. By 48 h, GOESC exhibits nearly equal u-component sensitivities at 850, 500, and 300 hPa in the subsection (Fig. 11b). Finally, while GOES imager cloud-top water vapor winds (GOESW) have a cumulative impact of just greater than 1.5 m s−1 at 48 h (Fig. 11b), they are not as significant as GOESC. GOESW winds are important for diagnosing and forecasting tropical cyclones (Velden et al. 1997), but since their heights are typically 600–250 hPa they cannot capture the extensive, organized low-level cloud motions in a well-developed extratropical cyclone as well as GOESC winds. For this reason, it is likely that GOESC observations prove to have a slightly larger impact than GOESW for this case.

Investigation and comparison of relative humidity sensitivities at 0 h (Figs. 9 and 10) and 48 h (Figs. 11c and 14) are also interesting. At 0 h, RAOB1 displayed the largest sensitivity over the entire domain and was on average six times larger than most other data types over CONUS (Fig. 9b). However, by 48 h its importance relative to ACAR1, SSMI1, GOESM, RAOB2, and ACAR2 is significantly reduced in the subsection. As the forecast evolves, the local maxima and minima initially located over CONUS (Fig. 10b) are advected east and north out of the subsection (Fig. 14b). Other significant but less dominant data types at 48 h include GOESL, TOVCD, SFCL1, RADAR, GOESC, GOESW, and GOESP. All other data types prove to be less significant, with cumulative rms relative humidity sensitivity values at or below 5% (Fig. 11c).

Figure 15 summarizes seven of the most important data types for this case at 0, 12, and 48 h; they are RAOB1, ACAR1, SSMI1, GOESM, RAOB2, ACAR2, and GOESC. This figure isolates a handful of the 34 overall data types so that the sensitivity from each observing system can be more readily viewed over the course of the integration. (Note that the 0-h results cover the entire model domain, while the 12- and 48-h results are for the subsection east of 110°W. The 12-h subsection results are presented to provide a clean comparison with the 48-h subsection results.)

In general, the 0-h temperature sensitivities for each data type are fairly uniform in the vertical (Fig. 15a), meaning the 850-, 500-, and 300-hPa sensitivities were of equal size by data type. However, by 48 h the largest temperature sensitivities exist predominately in the middle and lower troposphere for all data types (Fig. 15g). To a large extent just the opposite is true for the u component, where sensitivities are largest in the middle and upper troposphere at 48 h. The temperature and u-component results just discussed are for the most part an expected result, since the gradients of these two fields are largest in those portions of the atmosphere. The same conclusion does not hold for relative humidity, which shows less preference for largest rms differences at either high or low altitudes by data type or time. Cumulative relative humidity sensitivities also grow the least during the 48-h simulation, at least with respect to each data types y-axis scale. The temperature and u-component cumulative sensitivities show more overall growth during the forecast.

Of the seven data types shown in Fig. 15, one might expect that the two precipitable water (PW) data types (SSMI1 and GOESM) would have the largest impact on the relative humidity sensitivities; such is not the case. The cumulative relative humidity sensitivities for SSMI1 and GOESM remain fairly constant from 0 to 48 h (Figs. 15c, 15f, and 15i), while the cumulative sensitivities for the same two data types grow in most instances by at least a factor of 2 for the u component and temperature. The u-component and temperature sensitivities also grow within the subsection from 12 to 48 h. The additional PW data supplied by these two data types is presumably allowing differential heating associated with precipitation and radiation to modify the synoptic-scale flow during the simulation.

6. Summary

This case study reports on the sensitivity of the EDAS analysis and subsequent 48-h Eta Model forecast to all operational data types used within it at 0000 UTC 6 February 1998, a day with active cyclogenesis in both the Atlantic and Pacific Oceans. This case comes shortly after the EDAS began using 3D-VAR for the analysis and presents results where various nonconventional observation data types have a significant influence on the EDAS analysis and forecast, particularly data from oceanic regions.

Many of the 34 data types provide little impact on the forecast in a relative sense. Seven of the data types were shown to be among the most statistically important in terms of overall impact. They were rawinsonde temperature and specific humidity, rawinsonde winds, aircraft temperature, aircraft winds, GOES sounder multilayer marine precipitable water, GOES imager high-density infrared marine cloud drift winds, and SSM/I vertically integrated precipitable water. It is important to note that this is a single case study and that the results presented here may not extend to other synoptic flow regimes, geographical locations, or data assimilation systems. This study also does not address the issue of whether or not a large impact by a particular data type translates to a better forecast. More conclusive remarks about positive or negative forecast impact from a given data type must await expansion to case studies involving many synoptic regimes and seasons.

A further understanding of how to utilize these data, and future data types like them, is necessary if further improvements in forecast accuracy are to be realized. It is generally accepted that there is the greatest potential for forecast improvement on the west coast of the United States. This paper presents an approach for investigating the impact of various data types within the EDAS system. Future work will concentrate on longer-term weekly and seasonal studies involving several of the most significant data types identified here. Better characterization of observational errors, the forward model, and background errors also need to be investigated.

Acknowledgments

The authors wish to thank A. J. Mahajan, Raj Panda, and Kumaran Kalyanasundaram for their help in porting and optimizing the code on an SGI Origin 2000. Thanks are also due Ralph Petersen and William Raymond for their expert advice and interpretations of results and Robert Aune, Allen Lenzen, and Todd Schaack for visualization and verification tools. Special thanks are due Brad Hoggat and the anonymous reviewers for their enlightening scientific input. This research was supported under NOAA Grant NA67EC0100.

REFERENCES

  • Alishouse, J. C., S. Snyder, J. Vongsathorn, and R. R. Ferraro, 1990:Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens.,28, 811–816.

    • Crossref
    • Export Citation
  • Black, T. L., 1994: The new NMC mesoscale Eta model: Description and forecast examples. Wea. Forecasting,9, 265–278.

    • Crossref
    • Export Citation
  • Buizza, R., and A. Montani, 1999: Targeting observations using singular vectors. J. Atmos. Sci.,56, 2965–2985.

    • Crossref
    • Export Citation
  • Derber, J. C., and W.-S. Wu, 1998: The use of TOVS cloud-cleared radiances in the NCEP SSI analysis system. Mon. Wea. Rev.,126, 2287–2299.

    • Crossref
    • Export Citation
  • Diak, G., 1987: Assimilation of scalar versus horizontal gradient information from the VAS into a mesoscale model. J. Climate Appl. Meteor.,26, 738–748.

    • Crossref
    • Export Citation
  • Laudon, J., and D. Lenoski, 1997: The SGI Origin: A ccNUMA Highly Scalable Server. Proc. 24th Annual Int. Symp. on Computer Architecture (ISCA ’97), Denver, CO, SGI, 241–251.

    • Crossref
    • Export Citation
  • Leslie, L. M., G. A. Mills, L. W. Logan, D. J. Gauntlett, G. A. Kelly, M. J. Manton, J. L. McGregor, and J. M. Sardie, 1985: A high resolution primitive equations NWP model for operations and research. Aust. Meteor. Mag.,33, 11–35.

  • Menzel, W. P., F. C. Holt, T. J. Schmit, R. M. Aune, A. J. Schreiner, G. S. Wade, and D. G. Gray, 1998: Application of GOES-8/9 soundings to weather forecasting and nowcasting. Bull. Amer. Meteor. Soc.,79, 2059–2077.

    • Crossref
    • Export Citation
  • Nieman, S. J., W. P. Menzel, C. M. Hayden, D. Gray, S. T. Wanzong, C. S. Velden, and J. Daniels, 1997: Fully automated cloud-drift winds in NESDIS operations. Bull. Amer. Meteor. Soc.,78, 1121–1133.

    • Crossref
    • Export Citation
  • NWS, cited 1999: EDAS. NWS Technical Procedures Bulletin. [Available online at http://www.nws.noaa.gov/om/tpb/3d-eta.htm.].

  • Parrish, D., J. Purser, E. Rogers, and Y. Lin, 1996: The regional 3D variational analysis for the Eta model. Preprints, 11th Conf. on Numerical Weather Prediction, Norfolk, VA, Amer. Meteor. Soc., 454–455.

  • Raymond, W. H., and R. M. Aune, 1998: Improved precipitation forecasts using parameterized precipitation drag in a hydrostatic forecast model. Mon. Wea. Rev.,126, 693–710.

    • Crossref
    • Export Citation
  • Reale, A. L., 1995: Departures between derived satellite soundings and numerical weather forecasts: Present and future. Tech. Proc. Eighth Int. TOVS Study Conf., Queenstown, New Zealand, WMO, 395–404.

  • Rogers, E., D. Parrish, Y. Lin, and G. DiMego, 1996: The NCEP Eta data assimilation system: Tests with regional 3-D variational analysis and cycling. Preprints, 11th Conf. on Numerical Weather Prediction, Norfolk, VA, Amer. Meteor. Soc., 105–106.

  • ——, and Coauthors, 1997: Changes to the NCEP Operational “Early” Eta Analysis/Forecast System. NOAA/NWS Tech. Procedures Bull. 447, 16 pp. [Available from Office of Meteorology, National Weather Service, 1325 East–West Highway, Silver Spring, MD 20910.].

  • Schmidlin, F. J., 1988: WMO international radiosonde comparison, phase II final report, 1985. Instruments and Observing Methods Rep. 29, WMO/TD No. 312, WMO, Geneva, Switzerland, 113 pp.

  • Smith, W. L., H. M. Woolf, C. M. Hayden, D. Q. Wark, and L. M. McMillin, 1979: The TIROS-N operational vertical sounder. Bull. Amer. Meteor. Soc.,60, 1177–1187.

  • Treadon, R., 1993: The NMC Eta model post processor: A documentation. Office Note 394, National Oceanic and Atmospheric Administration, 42 pp. [Available from NASA Goddard Space Flight Center, Wallops Flight Facility, Wallops Island, VA 23337.].

  • Velden, C. S., S. J. Nieman, W. P. Menzel, and S. T. Wanzong, 1997:Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc.,78, 173–195.

    • Crossref
    • Export Citation

Fig. 1.
Fig. 1.

EDAS analyses from the control run valid 0000 UTC 6 Feb 1998. (a) Mean sea level pressure contoured every 4 hPa. (b) 500-hPa geopotential height contoured every 60 m. (c) Accumulated precipitation during the last 3 h of the assimilation cycle (2100–0000 UTC) contoured every mm

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 2.
Fig. 2.

Combined GOES-8 and -9 infrared window (11 μm) image for 2345 UTC 5 Feb 1998

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 3.
Fig. 3.

EDAS 0000 UTC 8 Feb 1998 analysis of (a) mean sea level pressure contoured every 4 hPa and (b) 500-hPa geopotential height contoured every 60 m. A 48-h forecast of the same fields from a control run initialized 0000 UTC 6 Feb 1998 is shown in (c) and (d)

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 4.
Fig. 4.

The 0000 UTC 6 Feb 1998 data locations for (a) RAOB2, (b) ACAR2, and (c) GOESC. For easier identification, the dots are larger for RAOB2 than the other two data types

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 5.
Fig. 5.

Sensitivity of the EDAS analysis for this case to each data type. Rms differences between the control run’s temperature and each experimental run’s temperature were computed over (a) the entire model domain and (b) an area extending only over CONUS. The statistics were computed at three levels: 850, 500, and 300 hPa

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 6.
Fig. 6.

Temperature sensitivity at 500 hPa in the EDAS due to (b) RAOB1, (c) TOVCD, and (d) ACAR1. These sensitivities are contoured every 0.3 K and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 3 K. For clarity, no zero line is plotted in (b), (c), and (d)

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 7.
Fig. 7.

The same as Fig. 5 except the statistics are for the u component of the wind

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 8.
Fig. 8.

The u-component sensitivity at 300 hPa in the EDAS due to (b) RAOB2, (c) GOESC, and (d) ACAR2. These sensitivities are contoured every 1 m s−1 and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 10 m s−1. For clarity, no zero line is plotted in (b), (c), and (d)

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 9.
Fig. 9.

The same as Fig. 5 except the statistics are for relative humidity

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 10.
Fig. 10.

Relative humidity sensitivity at 850 hPa in the EDAS due to (b) RAOB1, (c) SSMI1, and (d) GOESM. These sensitivities are contoured every 5% and represent the difference of the experiment that denied the dataset in question minus the control with all of the data, which is shown in (a) and contoured every 20%. For clarity, no zero line is plotted in (b), (c), and (d)

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 11.
Fig. 11.

Stacked bar charts displaying the (a) temperature (K), (b) u component (m s−1), and (c) relative humidity (%) sensitivity of the Eta 48-h forecast for this case to each data type. The statistics are limited to a subsection of the entire model domain east of roughly 110°W

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 12.
Fig. 12.

The same as Fig. 6 except the statistics are for 48 h and displayed in the subsection

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 13.
Fig. 13.

The same as Fig. 8 except the statistics are for 48 h and displayed in the subsection

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 14.
Fig. 14.

The same as Fig. 10 except the statistics are for 48 h and displayed in the subsection

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Fig. 15.
Fig. 15.

Stacked bar charts displaying the temperature (K), u component (m s−1), and relative humidity (%) sensitivity of the Eta 0-, 12-, and 48-h forecast for RAOB1, ACAR1, SSMI1, GOESM, RAOB2, ACAR2, and GOESC. The 0-h results include the entire model domain, while the 12- and 48-h results use only the subsection east of 110°W

Citation: Weather and Forecasting 15, 5; 10.1175/1520-0434(2000)015<0603:ACSOTS>2.0.CO;2

Table 1.

Total number of reports for each operational data type at each of the five assimilation steps during the 12-h assimilation cycle. Multiple levels of any data type are considered separate reports. Also shown is the abbreviation for each data type used in subsequent figures and tables

Table 1.
Table 2.

Errors assigned to temperature observations (K) in the EDAS at five pressure levels. Data type is shown at left

Table 2.
Table 3.

Errors assigned to velocity observations (m s−1) in the EDAS at five pressure levels. Data type is shown at left

Table 3.
Table 4.

Errors assigned to precipitable water observations (mm) in the EDAS at five pressure levels. Data type is shown at left

Table 4.
Table 5.

Errors assigned to specific humidity observations (%) in the EDAS at five pressure levels. Data type is shown at left

Table 5.
Table 6.

Errors assigned to pressure/height observations (hPa) in the EDAS at five pressure levels. The data type is shown at left, and a missing value of (—) implies that no impact is possible from this particular data type at that level

Table 6.
Table 7.

Background error amplitudes by isobaric level (hPa) for temperature (K), height (m), and wind (m s−1)

Table 7.
Table 8.

Correlation lengths (km) by isobaric level (hPa) for temperature, specific humidity, longitudinal wind, and transverse wind

Table 8.
Save