Radar Data Assimilation with WRFDA

Juanzhen Sun

August 5–August 7, 2015, Boulder, CO

OUTLINE

• Introduction: motivation, milestones, and current

capabilities

• WRFDA-radar procedure and methods

• Configure and run WRFDA-radar

• Applications and ongoing research

 Why radar data assimilation
Objectives of radar data
assimilation

• Improve short-term prediction

of high impact weather
Nowcast • For the very short-term range,
(extrapolation-based) an ambitious goal is to
forecast the timing and
location of storms at
county/city scale

NWP models • Improved understanding of

mesoscale processes
contributing to the formation
of convective weather
(may require advanced DA
techniques)
Courtesy of Isztar Zawadzki (2012)
 Radar data: the good and the bad
Good
•High spatial and temporal resolutions
at the convective-scale Radial velocities from 20
•Observes wind (radial velocity) as well WSR-88D radars
as microphysics (reflectivity)
•Accurate observations
•Observations are mostly in the lower
atmosphere

Not so good
•Indirect observations – need observation
operators
•Incomplete coverage – limited range and
limited detection ability in clear air
•Only radial velocity and reflectivity
•Nontrivial for QC
•Locally available
 WRFDA-radar development milestones

2005: radial velocity data assimilation with WRFDA 3DVar

(Xiao et al. 2005)
2007: reflectivity data assimilation using a partition scheme to
obtain microphysics
(Xiao et al. 2007)
2013: indirect assimilation of reflectivity using qr and qc as
control variables
(Wang et al. 2013)
2014: adjoint of Kessler scheme for 4DVar radar data
assimilation
(Wang et al. 2014)
2015: new momentum control variables (u/v) for radar data
assimilation
(Sun et al. 2015)
 Current capabilities

3DVar
- Assimilate both radial velocity and reflectivity
- Direct and Indirect assimilation of reflectivity
- “Warm start” by assimilating estimated humidity within cloud
- Options for choice of momentum control variables
- Operational capability since 2012
- Hourly cycle tested recently in a real time demonstration in CO

4DVar
- Use WRF tangent linear model as constraint with multiple outer loops
- Can be run with multi-incremental option
- Adjoint of physics schemes: modified Kessler microphysics, large-
scale condensation, a simple cumulus scheme, and diffusion scheme
- 4DVar framework is fully consistent with 3DVar
 4DVar physics options

WRFTL & AD WRFNL

mp_physics_ad=98 mp_physics can be set to any

Large scale condensation options for WRF
mp_physics
Mp_physics_ad=99 It can also be set to 98 or 99,
Modified Kessler scheme same as WRFTL & AD

cu_physics = 0:
no cumulus scheme
cu_physics cu_physics=98: Same as the left column
Simplified CU scheme
Any other numbers will be
Defauted to 98
OUTLINE
• Introduction: motivation, milestones, and current
capabilities

• Radar DA methods in WRFDA

• Configure and run WRFDA-radar

• Applications and ongoing research

 Radar DA methods (options)

• Radial velocity or reflectivity or both

use_radarobs=true,
use_radar_rv=true,
• 3DVar (default) or 4DVar
var4d=true
• Two methods for reflectivity DA
use_radar_rf = true
use_radar_rhv = true
• Microphysics control variables
cloud_cv_options = 3 BE of regular variables are from gen_be; cloud variables hard coded
• Cloud analysis scheme (assimilate estimated water vapor)
use_radar_rqv=true
• u/v momentum control variables
cv_options = 5 5: standard psi/chi CV 7: new u/v CV
• Choice of data: radar only, GTS+radar, or GTS then radar
 Observation operators
vr - (u,v,w,qr) Relation:

x − xr y − yr z − zr
vr = u+ v+ (w − VT (qr ))
r r r

Z-qr Relation (assume Marshal-Palmer DSD)

Z = 43.1+ 17.5 log10 ( ρ qr ) Sun and Crook (1997)

Can also use empirical relations obtained by fitting with disdrometer data

Z-qs and Z-qh follow Gao and Stensrud (2012)

 Two methods for reflectivity DA
1. Direct assimilation of reflectivity (Xiao et al. 2007)
- Requires an observation operator to link the reflectivity
with microphysics
- No cloud control variables
- Vertical velocity is diagnosed using the Richardson
equation
- Microphysics are diagnosed using a warm rain partition
scheme

2. Indirect assimilation of reflectivity (Wang et al. 2013)

- Diagnose microphysics (qr, qs, qg) and humidity from
reflectivity
- Assimilate the diagnosed quantities
- Cloud control variables and vertical velocity control
variable
 Cost Function
Indirect method with cloud control variables

J = J b + J o + J vr + J qr + J qv
For radar DA

• Control variables :
u/v (or ψ/χu), T (or Tu), Ps (or Psu), RHs, qc, qr, and w

• 3DVar critically depends on a cloud analysis scheme

that assimilates estimated in-cloud humidity

• A modified Kessler scheme along with its adjoint produces

analyses of microphysics in 4DVar
 Multi-incremental 4DVar
Following Courtier et al. 1994, WRFDA uses
a multi-incremental formulation, which means

•The forward prediction model within the 4DVar

window is an approximation of the nonlinear model

•The control variables are increments from the

forward model trajectory

•The formulation requires the update of analysis

increment in an inner loop but also the update of
the nonlinear model in an outer loop

•It makes the cost function better conditioned and allows different
spatial resolutions for the inner and outer loops – multi-incremental

Details in Huang et al. (2010) and Wang et al (2013)

Cost function reduction with different
lengths of windows

# of outerloops
 OUTLINE

• Introduction: motivation, milestones, and current

capabilities

• Radar DA methods in WRFDA

• Configure and run WRFDA-radar

• Applications and ongoing research

 Basic steps for radar DA
Step 1: prepare radar data in the correct format and write the data into ob.radar
(ob01.radar, ob02.radar, … for 4DVar)
- Use your own QC software
- (ϕ,r,θ) => (x,y,θ) => Lat/lon profiles => merge the radars into one file

Step 2: produce 1-3 month WRF forecasts (12h & 24h) over the study domain, and then
compute BE using the WRFDA utility gen_be

Step 3: modify the namelist.input to make radar DA choices

Step 4: conduct single observation tests to tune the length scale and variance for your
specific domain

- 200km for GTS data and 30km for radar data are commonly used

Step 5: configure WRFDA: to invoke the “CLOUD_CV” option, do the following in the
configure script prior to compiling
setenv CLOUD_CV 1 - for csh
export CLOUD_CV = 1 - for both ksh and bash

STEP 6: link ob.radar and other other observation files and first guess
 Data format
write(301,'(a14,i3)') 'Total number =', nrad
write(301,'(a)')
'#-----------------------------#'
write(301,'(a)') ' '
do irad = 1, nrad ! nrad: total # of radar
!--- Write header
write(301,'(a5,2x,a12,2(f8.3,2x),f8.1,2x,a19,2i6)') 'RADAR', &
radar_name, rlonr(irad), rlatr(irad), raltr(irad)*1000., &
trim(radar_date), np, imdv_nz(irad) write(301,'(a)') &
'#---------------------------------------------------------#'
write(301,*)
!----Write data
do i = 1,np ! np: # of total horizontal data points
write(301,'(a12,3x,a19,2x,2(f12.3,2x),f8.1,2x,i6)') 'FM-128 RADAR', &
trim(radar_date), plat(i), plon(i), raltr(irad)*1000, count_nz(i)
do m = 1,count_nz(i) ! count_nz(i): # of vertical elevitions for each radar
write(301,'(3x,f12.1,2(f12.3,i4,f12.3,2x))') hgt(i,m), &
rv_data(i,m), rv_qc(i,m), rv_err(i,m), &
rf_data(i,m), rf_qc(i,m), rf_err(i,m)
enddo
enddo
enddo
 Example of data format
date
Name of radar Lat&lon of radar Altitude of radar
Total # of sounding

# of elevations

3 elevations

altitude
Lat & lon of data
RV

QC index
RF

Err variance
 Tuning BES parameters
To change BES variance and length scale, do the following
in your namelist.input: U increment before tuning
To decrease the weight of the
background
VAR_SCALING1=2.0
VAR_SCALING2=2.0
VAR_SCALING3=2.0
VAR_SCALING4=2.0
VAR_SCALING5=2.0
To decrease the length scale U increment after tuning
LEN_SCALING1=0.5
LEN_SCALING2=0.5
LEN_SCALING3=0.5
LEN_SCALING4=0.5
LEN_SCALING5=0.5
 A note on u/v momentum CV

• The new control variable (CV) option CV7 which uses

u/v instead of psi/chi as momentum control variables
is added in WRFDA3.7

• CV7 requires the computation of BES of u and v

• In the current version, correlation between variables

is not considered

• But will be studied and included in a future release

• Sun et al. (2015, MWR under review)

 OUTLINE

• Introduction: motivation, milestones, and current

capabilities

• Radar DA methods in WRFDA

• Configure and run WRFDA-radar

• Applications and ongoing research

 Scientific issues of radar DA
• Impact of radar data assimilation on convective forecasting

• Validity of the linearization of the observation

operator for reflectivity
(Wang et al. 2013)

• Validity of the tangent linear model

approximation for highly convective weather
(Wang et al. 2014)

• Comparison of 4DVar and 3DVar

(Sun and Wang 2014)

• Impact of the choice of momentum control variables

(Sun et al 2015, under review)

• Controlling noise in high-resolution analysis with multi-scale balance

(Vendrasco et al. 2015, under review)
 Operational performance of WRFDA-radar at BMB

Precipitation verification for June & July 2012

CSI  scores  for  1mm/hr  2012060100-­‐073121  
0.3  

0.25  
With radar
0.2  

0.15  
Without radar
0.1  

0.05  

0  
1   2   3   4   5   6   7   8   9   10  11  12   13   14  15  16  17   18   19  20  21  22   23   24  
• WRFDA-radar has been
BJRUCv1.0   BJRUCv2.0   Hour operational at Beijing
Meteorological Bureau
CSI  scores  for  10mm/hr  2012060100-­‐073121   since 2012
0.12  

0.1  
• The local forecasters has
0.08  
become dependent on
0.06  
the system to forecast
0.04  
localized convective weather
0.02  

0  
1   2   3   4   5   6   7   8   9   10  11  12   13   14  15  16  17   18   19  20  21  22   23   24  

BJRUCv1.0   BJRUCv2.0  
Hour Courtesy of S. Fan
 WRFDA realtime demonstration in CO Front
Range with hourly update cycles

Precipitation verification for 7 July – 15 August

Solid lines: no radar Dashed lines: with radar
t = 3h t = 12h

0mm
1mm
2mm
5mm

Courtesy of Barb Brown

 WRFDA realtime demonstration in CO Front
Range with hourly update cycles
3h accumulated rainfall in northeastern Colorado
during STEP Hydromet realtime experiment at t = 3h
OBS 3DVar without radar 3DVar with radar
7 August 2014

16 July 2014
Hourly4DVar radar DA for
Precipitation Typhoon Fanapi
forecasts
Hourly  precipita.on  at  t  =  6h  

OBS   GTS3DVar   RDR4DVar   Surface analysis

• 4DVar successfully predicts the two low pressure centers

with similar magnitudes as in the surface analysis
Sensitivity of physics in the TL and AD models
Precipitation forecasts at t = 2h for a Meiyu case occurred in Taiwan

Large-scale Kessler scheme

condensation

118oE 120oE 118oE 120oE 122oE

122oE
 Ongoing research
• Improve 3DVar hourly cycle, especially the issue of
overprediction in the first few hours
- Improve the cloud analysis scheme using PW from GPS
network
- Develop a scheme that maintains total water conservation

• How to configure 4DVar for radar data assimilation and short-

term convective forecasting?
- physics options
- cycling strategy
- multi-incremental method

• Reduce noise in the analysis using constraints in the cost

function
- Bias in the background; the imbalance issue in the analysis
 Future research
• Improve flow-dependent BE (error of the day) using the ensemble
method
- Application of WRFDA hybrid-3DVar to radar data
assimilation
• Improve climatology BE
- The effect of geographic inhomogeneity (Wang et al. 2015)
- The effect of diurnal cycle
• Improve 4DVar for convective-scale DA
- How to do multi-incremental DA for the convective-scale?
- The short window vs. long window dilemma
> multi-step approach?
> 3DVar/4DVar hybrid?
- En-4DVar
- Consider uncertainties in physics
• Polarimetric radar data assimilation