Mesoscale Data Assimilation
Advanced data assimilation systems for community use
Application of WRF 3D-Var in polar, tropical and global domains
The WRF three-dimensional variational (3D-Var) data assimilation system is an integral part of many MMM projects. In FY04, scientists at MMM applied the WRF 3D-Var system for data assimilation efforts over polar, tropical and global domains. Dale Barker, Wei Huang and Syed Rivzi recently upgraded the data assimilation component of the MM5-based Antarctic Mesoscale Prediction System (AMPS), which was jointly develop by the MMM Division and Byrd Polar Research Center of the Ohio State University, to WRF 3D-Var for real-time forecasting over high southern-latitudes. The WRF 3D-Var system was enhanced to assimilate several types of satellite remote sensing observations. Rizvi implemented the capability of WRF 3D-Var to assimilate ATOVS retrieved temperature and moisture profiles from the NOAA satellites. Barker has begun assimilating retrieved motion vectors from the MODerate resolution Imaging Spectrometer (MODIS) instrument aboard NASA's Aqua satellite. (Figure 18) shows a typical observation distribution of retrieved MODIS atmospheric motion vectors over a 2-h period. This data-source was provided by the University of Wisconsin’s Space Science Engineering Center. Given the dearth of conventional polar wind observations, the assimilation of MODIS winds is expected to lead to significant improvements in the forecasting of key polar processes (e.g. katabatic winds, polar cyclogenesis) as well as for medium-range mid-latitude NWP.
The performance of variational data assimilation techniques in the tropics is a challenging area of research. Balance constraints (e.g. geostrophic) used in mid-latitudes do not apply at low latitudes. Also, the proper representation of nonlinear processes (e.g. convection) is key to successful tropical NWP. In variational systems, these issues translate to
Tropical variational data assimilation work is supported by international collaborations with both Taiwan (Civil Aeronautics Administration, and Central Weather Bureau) and India (National Center for Medium-Range Weather Forecasts). The MM5-based MMM/NCMRWF collaborative work to date has concentrated on the specification of forecast error statistics for the Indian domain (Figure 19), and the impact of particular observations on Indian NWP. For example, differences between Figures 19a and 19b illlustrate the impact of Special Sensor Microwave/Imager (SSM/I) data on 24hr 850hPa geopotential height forecasts. The improved definition of the tropical cyclone when using the SSM/I surface wind speed data is clearly shown,
The WRF 4D-Var capability of WRF-VAR (the merged WRF 3/4D-Var system - see elsewhere) will be applied as it becomes available to worldwide applications of WRF. The impact of improved forecast error covariances, and additional observations (e.g. radiances) in WRF-VAR will be also assessed. High-resolution applications of WRF 3/4D-Var over IHOP cases over continental U.S. are ongoing.
Doppler Radar data assimilation using WRF 3D-Var system
In collaboration with Seoul National University and Korea Meteorological Administration, Q. Xiao, Jianfeng Gu (long-term visitor from Shanghai Meteorological Bureau), D. Barker, Wei Haung, and Ying-Hwa (Bill) Kuo worked on the Dopper radar data assimilation using WRF 3D-Var system for the Northwestern Pacific Typhoon Rusa (2002). Figure 20 compares the 850-hPa wind distributions with and without the assimilation of Doppler radar data. The typhoon initialization with Doppler radial velocities is much more compact. The wind around the typhoon is increased, and the radius of the maximum wind is decreased. As Doppler radar observations have very high resolution, assimilation of these data can adjust the typhoon vortex to the right position with improved vortex structure. As a result, the typhoon forecasts using the initialization enhanced by radar data are also improved.
Tropical cyclone initialization
Q. Xiao, B. Kuo, Ying Zhang (RAP) and D. Barker developed a typhoon bogus data assimilation algorithm and implemented in the MM5 3D-Var system. Numerical experiments on Typhoon Rusa (2002) case showed that the scheme works well and can improve the prediction of typhoon track and intensity. With MM5 3D-Var bogus data assimilation scheme, the initial typhoon vortex is dynamically balanced, and possesses a well-defined warm core structure. Compared with traditional bogussing in the 3D-Var background fields, MM5 3D-Var bogus data assimilation scheme has the advantage of much reduced problem in precipitation spin-down/up in the subsequent typhoon forecast. Sensitivity experiments show that assimilation of only the bogus sea-level pressure data produces too strong typhoon intensity, while assimilation of the bogus wind data alone produces a storm that is much weaker than the observations. Assimilation of both bogus sea-level pressure and wind data obtains the best hurricane initialization and subsequent forecast for this case. The landfall location, time and intensity all compare favorably with the observation.
Comparison of Variational and Ensemble Data Assimilation at High-Latitudes
Data assimilation at high latitudes is particularly challenging given the relative sparcity of conventional observations and unique physics present in these regions. Given this, and the fact that most modern data assimilation systems are designed with an emphasis on mid-latitudes, Antarctic mesoscale data assimilation is a demanding test of current assimilation strategies. In FY04, D. Barker performed an assessment of the relative performance and costs of MM5 3D-Var and ensemble square-root filter (EnSRF) algorithms in AMPS. His preliminary results indicate comparable performance of 3D-Var/EnSRF during a one month trial period in December 2002. For example, Figure 21 shows 12hr temperature forecast verification for EnSRF, 3D-Var and "NoObs" forecasts run from interpolated GFS data (the latter's superiority is due at least in part to the lack of radiance assimilation in the MM5 EnSRF and 3D-Var algorithms). In a paper submitted to the Monthly Weather Review, D. Barker highlights the detrimental impact of EnSRF ensemble sampling error on subtle multivariate forecast error covariances. D. Barker also studies the use of randomized samples of 3D-Var covariances to provide initial and lateral boundary condition perturbations for the EnSRF.
MM5/WRF Research Community Variational Data Assimilation System
With the team effort by D. Barker, W. Huang, Yong-Run Guo, Q. Xiao, and S. Rizvi, an upgraded version of the WRF 3D-Var (V2.0) data assimilation system was released to the research community in May 2004. The code, documentation and online tutorial are available from the WRF web site http://www.wrf-model.org/wg4. New capabilities available to the community in V2.0 include a vertical velocity analysis, radar radial velocity assimilation, preconditioned conjugate gradient minimization (to improve convergence), and an improved assimilation of surface observations via the use of boundary layer parameterizations in the interpolation from model to observation locations.
Assimilation of GPS radio occultation data
Assimilation of GPS RO Data for an Intense Antarctic Storm
Recently, Tae-Kwon Wee (COSMIC), B. Kuo and David Bromwich (Ohio State University) collaborated on a study to assimilate the GPS radio occultation (RO) data from CHAMP and SAC-C missions using an updated version of MM5 4D-Var assimilation system, and assessed their impact on short-range prediction of an intense cyclone that took place in December 2001 over the Ross Sea. The intensity of the December 2001 storm (which has a minimum pressure of 936 hpa) had not occurred near McMurdo Station over the 46 years of record keeping. Wee et al. performed parallel data assimilation experiments with and without the use of GPS RO data over an 8-day period for this storm. Continuous assimilations with a window size of 12-h were carried out over 48 hours (e.g., 4 cycles). For this case, various types of observations were collected and processed.
The CHAMP and SAC-C GPS RO data were obtained from the COSMIC Data Analysis and Archive Center (CDAAC) at UCAR. Other observations collected include TEMP (rawinsonde and pibal), SYNOP (surface, ship, drifting buoy, and PAOB), aircraft observations (AIREP, PIREP, and AIRCAR), ATOVS retrieved soundings from NESDIS (NOAA 15 and 16), MODIS/Terra retrieved soundings, satellite motion vectors, SSM/I retrievals (rainfall rate, total liquid water, precipitable water vapor, and surface wind speed) from DMSP 13-15, and QuikSCAT surface wind vector.
A significant positive impact of GPS RO data on short-range forecasts was found for experiments with 48 hours of continuous assimilation (Figure 22). The assimilation of GPS RO data over an extended period (e.g., 48 h) improved all parameters in all forecast ranges. Moreover, the amount of positive impact as a result of GPS RO data assimilation increased proportionally to the lengthened analysis period. With continuous cycling assimilation, the analysis errors were kept to a degree comparable to those of the NCEP operational analysis. A significant error reduction was noted over the interior of the Antarctic continent when the forecasts were verified against observed GPS RO data. This study demonstrates the importance of GPS RO data for weather prediction over the Antarctic.
Development of nonlocal observation operator for GPS radio occultation
The radio occultation (RO) sounding technique that uses signals transmitted by the Global Positioning System (GPS) has evolved as an important global observing technology. The Constellation Observing System for Meteorology and Climate (COSMIC) will launch six microsatellites in late 2005, and will provide approximately 2,500 GPS radio occultation (RO) soundings per day, distributed uniformly around the globe. The COSMIC GPS RO data have the potential to make a significant contribution to operational numerical weather prediction. However, the GPS RO soundings are non-traditional measurements. Effective and efficient assimilation of GPS RO data into operational models remains a challenge. In FY04, B. Kuo and Wei Wang collaborated with Sergey Sokolovskiy on the development of a new observation operator, known as the “nonlocal excess phase” operator. This new observation operator takes into account the effects of horizontal inhomegeneity of atmospheric refractivity. It is considerably more accurate than the “local refractivity” operator (which does not account for horizontal inhomegeneity), and much more computationally efficient than the sophisticated bending observation operator (which is nearly two order of magnitudes more expensive). This observation operator was evaluated in observing system simulation experiments using WRF 4-km simulation of Hurricane Isabel, as well as against actual CHAMP GPS RO observation. It was proven to be an accurate and effective observation operator. The results were presented in two papers submitted to the Mon. Wea. Rev. The development of this observation operator greatly facilitates the use of COSMIC data in operational forecasting.
Research with ensemble Kalman filters
Assimilation of Doppler radar observations using ensemble Kalman filters
Mesoscale data assimilation is hampered by two facts. First, observations which are plentiful (e.g. Doppler radar measurements of wind and reflectivity) involve only a subset of atmospheric variables, while observing platforms such as radiosondes that measure all relevant variables are sparse and resolve mesoscale motions poorly. Second, the balances between variables, such as geostrophy, that pertain at large scales in the atmosphere are questionable at the mesoscale; these balances are an important component of more traditional assimilation schemes such as 3DVar. To overcome these difficulties, there has been substantial effort within MMM and the Data Assimilation Initiative to explore the potential of the ensemble Kalman filter (EnKF) for mesoscale assimilation. Following the results of Snyder and Zhang (2003), who showed using simulated observations from a single radar that the EnKF could produce accurate analyses of convective storms, Fuqing Zhang (Texas A&M), C. Snyder and Juanzhen Sun have explored how the assimilation is influenced by changes in the available observations or in the quality of the initial estimate of the storm (Zhang et al. 2004). Alain Caya (MMM and Data Assimilation Initiative [DAI]), C. Snyder and J. Sun have also compared the EnKF to a four-dimensional variational assimilation scheme (again in the context of simulated observations) and find that the EnKF performs somewhat better given several radar scans but somewhat worse than the variational scheme when only 2--3 scans are available.
In addition, a team led by David Dowell (presently Cooperative Institute for Mesoscale Meteorology at Oklahoma University, formerly ASP and MMM postdoc), and including F. Zhang, C. Snyder, A. Crook and Louis Wicker (NOAA/NSSL), has explored the assimilation of real Doppler-radar observations of the 1981 Arcadia, Oklahoma with the EnKF. They find that the EnKF, using observations from a single radar, performs at least comparably to traditional dual-Doppler retrieval techniques. Together, these results are the first for the EnKF outside of global atmospheric models and hold substantial promise for the application of the EnKF to meso- and convective scales.
The foregoing experiments with the EnKF all assume that the larger-scale environment in which the storm develops is known perfectly. In the context of a simulated squall line, William Skamarock and C. Snyder have examined the effects of uncertainty in the environmental sounding. They show that the EnKF errors are significantly degraded for errors that are comparable to likely uncertainty in soundings (several m/s), and they have begun exploring the possibility of estimating the environmental sounding based on the radar observations.
Development and application of EnKF system based on WRF
This past year progress was made in developing an EnKF system for mesoscale data assimilation based on WRF. This is a leading-edge effort with few counterparts, and which complements other research with the EnKF at global scales. Lateral boundary conditions are an important source of uncertainty in limited-area forecasts and this uncertainty must be accounted for in the EnKF. To this end, Ryan Torn, Greg Hakim (both University of Washington), and C. Snyder have begun exploring various treatments of the lateral boundary conditions for short-range limited-area ensemble forecasts. They have demonstrated that the preferred treatment of the lateral boundaries, in which an EnKF on a larger domain provides an ensemble of boundary conditions, often yields only marginal benefits over various simple, ad hoc treatments, such as drawing boundary perturbations from a climatological time series or generating them using the background covariance model from the WRF three-dimensional variational assimilation scheme.
In addition, A. Caya, together with Jeffrey Anderson (CGD), Joshua Hacker (RAP) and C. Snyder, has continued the development of an EnKF system for WRF based on the Data Assimilation Research Testbed (DART, developed by DAI). Observation operators for surface fields, radar reflectivity and radial velocity have been implemented, along with the capability to use the EnKF on multiple nested domains, which will be required for assimilations that span the meso- and convective scales. This is one of the first assimilation systems that handles nested domains together rather than performing analyses separately for each domain. A. Caya and C. Snyder have continued their assimilation work using the WRF/DART system. In simulated-observation experiments, they have found that the EnKF is an effective scheme on a continental-U.S. domain. They have begun initial tests of the system with real observations, including radiosondes, aircraft observations and a subset of satellite winds.
Design of optimal global observing system
Developing optimal observing strategies requires evaluating the costs and benefits of different observations. Although the meteorological community has discussed the question of optimal investment in observations for more than three decades, it still lacks a practical, systematic framework for analyzing the issue. In FY04, R. Morss, Kathleen Miller (ESIG), and Maxine Vasil (graduate student, University of Colorado) developed an economic approach to analyzing the optimal observing system. In a publication currently in press in Monthly Weather Review, they present the approach, then combine meteorological, cost, and benefit information to demonstrate the approach for a simplified version of the real observing and prediction system. They also identify gaps in existing knowledge that must be addressed before the framework can be implemented to analyze the real observing system.
Choices about implementation of publicly funded observing systems require
understanding meteorological aspects of observing systems, but are made
in the public policy arena. To help meteorological knowledge and meteorologists
contribute more effectively to public policy decisions on observing systems,
R. Morss analyzed meteorological observing system design from a public
policy perspective. The results are presented in a manuscript, in press
in BAMS, which illustrates the importance of careful problem definition
and builds towards an appropriate multidisciplinary problem definition
for observing system design.
Microphysical parameter retrieval and improvement of warm rain microphysical parameterization scheme
Juanzhen Sun, Ed Brandes (RAP), and Guifu Zhang (RAP) conducted a study on Microphysical parameter retrieval and improvement of warm rain microphysical parameterization scheme. This project was supported by the NCAR opportunity fund. They used a four-dimensional variational scheme, a simple cloud model, and single Doppler radar observations to optimally determine both the initial conditions of the numerical model and some of the warm rain microphysical parameters. Experiments were performed using both simulated data and real data. Effort was also made to implement a new microphysical scheme that was based on constrained Gamma drop size distribution. This new scheme was compared with the existing Marshal-Palmer scheme. They have found that both the optimal determination of the microphysical parameters and the constrained Gamma microphysical scheme showed potential improvement in the prediction of thunderstorms.
Assimilation of multiple Doppler radar observations to improve 0-12 hour QPF
J. Sun, Q. Xiao, A. Crook, C. Snyder, J. Miller, D. Barker, and B. Kuo conducted the research on the assimilation of multiple Doppler radar observations into mesoscale models. They worked on a squall line case observed during IHOP. Data from 12 NEXRAD radars for a period of 6 hours were quality-controlled and preprocessed. Simulations using WRF initialized without radar observations showed the model was unable to predict the initiation and evolution of the squall line. OSSE experiments are being performed using the WRF 3D-Var to answer the question whether a 3D-Var system is able to retrieve the unobserved wind component through continuous cycling. Real data experiments using WRF 3D-Var, EnKF, and a 4D-Var system have just begun.
Assimilation of Doppler radar for an Oklahoma supercell
A. Crook and David Dowell (CIMMS/University of Oklahoma) have continued
their data assimilation studies of the Arcadia Oklahoma supercell. The
focus in FY04 has been on the retrieval of unobserved parameters and
fields. Using the microphysical retrieval code of Sun and Zhang discussed
above it is shown that making the precipitation fallspeed a control parameter
(rather than a constant) allows for a much better fit to the observed
radar data. It is shown that the primary mechanism that determines the
fallspeed is the improvement in the fit to the temporal evolution of
the reflectivity data. It is also shown that the optimal fallspeed is
less than the default value which is probably due to the lack of ice
physics in the underlying cloud model. In related work, a simple cloud
model (without dynamics) and its adjoint have been developed to examine
how the total water field is adjusted in regions of precipitation. Given
observations of the rainwater field at two different times, it is possible
to retrieve the total water field via the adjoint of the microphysical
Next Topic: Prediction and Predictability