A STATISTICAL EVALUATION OF TAMDAR DATA IN

1
A STATISTICAL EVALUATION OF TAMDAR DATA
IN SHORT-RANGE MESOSCALE NUMERICAL MODELS Neil
A. Jacobs1 and Yubao Liu2 1AirDat, LLC,
Morrisville, NC 27560 2National Center for
Atmospheric Research, Boulder, CO 80307
2
Vertical Resolution Case Studies Hypothesis
Increasing the number of model ?-levels in the
lower to mid-troposphere will better utilize the
greater observation density provided by TAMDARs,
and result in a more accurate forecast.
3
Great Lakes late-season snow event (April 22-25,
2005)
Model Description RT-FDDA ? First-guess
field Little_R (MM5 Objective analysis) ? INTERPF
? MM5 Domains 1/2 (36-km/12-km) Grell CP, MRF
PBL, Mixed-phase (Reisner-1) microphysics 4
Simulations (96 hwas 14 h) TAMDAR (36
?-levels) Cntl No TAMDAR (36 ?-levelssame as
TAMDAR) TAMDAR (48 ?-levels 6 ? 1.5 km /
all 12 ? 5.5 km) Cntl No TAMDAR (48
?-levelssame as TAMDAR) Initialized 1100 UTC
22 April 2005
4
TAMDAR
Cntl
TAMDAR
Cntl
Stage IV (4 km)
1-h precipitation forecast (mm) and sea-level
pressure (mb), as well as the 1-h Stage-IV
analysis (mm), valid 1200 UTC 22 April 2005.
5
Cntl / Cntl Scatter plot of 1-h QPF totals
versus the 1-h Stage-IV analysis comparing
matching grid point magnitudes (above 5 mm
threshold) summed over each of the 14 forecast
hours. The Cntl is blue, and the Cntl is red.
TAMDAR / TAMDAR Scatter plot of 1-h QPF totals
versus the 1-h Stage-IV analysis comparing
matching grid point magnitudes (above 5 mm
threshold) summed over each of the 14 forecast
hours. The TAMDAR is blue, and the TAMDAR is
red.
6
Precipitation cell isolation who cares? A crude
method to quantify QPF performance
We can isolate cells based on magnitude, and
retain only the precipitation associated with
that cell
Laplacian edge detection
Can also detect noise, so a smoothing Gaussian
filter was testedLaplacian of Gaussian

• Both methods yield near-identical results
• All forecasts are regridded to 12-km
• An assumption is made that the closest cell
  (radial search) was the predicted cell.
• A weighted score was applied to the magnitude of
  the cell based on the linear distance
• (to the maximum) from truth (Stage-IV).
• For example, Stage-IV compared against itself
  would receive full weight. A score of 0
• would mean either no cell was detected, or the
  distance was gt 2(dm_cellds4_cell).

7
Total
5-mm
10-mm
15-mm
20-mm
EXAMPLE Raw Stage-IV 3-h accumulated
precipitation data (Total), and the postprocessed
(no minimum) isolated cells. The domain-2 data
are mapped on the x-y grid.
8
Comparison of 12-h (4x3-h) QPF between TAMDAR,
TAMDAR, Cntl, Cntl, and other various models
regridded to the smallest grid (12-km), as well
as the Stage-IV analysis (truth) for 20-mm
cells with 2-mm minimum bound.
The results presented here are consistent with
preliminary findings from similar studies
conducted at NCAR.
9

• Results suggest that the addition of TAMDAR data
  in conjunction with increased vertical resolution
  improves the forecast skill for certain output
  parameters.
• GFS, NAM, and RUC were included as reference
  models, and the improvement of the Cntl over
  these models is attributed to the 4DVAR ingestion
  technique of the RT-FDDA system.
• However, the TAMDAR run shows significant
  improvements of 18-22 over the TAMDAR, Cntl, and
  the Cntl for this case.
• This suggests that proper utilization of TAMDAR
  data plays a crucial role in forecast skill.

10
Hurricane Katrina (August 29, 2005)
Model Description RT-FDDA ? First-guess
field Little_R (MM5 Objective analysis) ? INTERPF
? MM5 Domains 1/2 (36-km/12-km) Grell CP, MRF
PBL, Mixed-phase (Reisner-1) microphysics 4
Simulations (96 hwas 14 h) TAMDAR (36
?-levels) Cntl No TAMDAR (36 ?-levelssame as
TAMDAR) TAMDAR (48 ?-levels 6 ? 1.5 km /
all 12 ? 5.5 km) Cntl No TAMDAR (48
?-levelssame as TAMDAR) Initialized 2300 UTC
29 August 2005
11
TAMDAR
Cntl
Stage-IV
Sea-level pressure (mb) and 1-h precip.
(in) Valid 0600 UTC 30 AUG 2005 (7-h Fcst)
12
Cntl
TAMDAR
Stage-IV
Sea-level pressure (mb) and 1-h precip.
(in) Valid 0900 UTC 30 AUG 2005 (10-h Fcst)
13
Cntl
TAMDAR
Stage-IV
Sea-level pressure (mb) and 1-h precip.
(in) Valid 1200 UTC 30 AUG 2005 (13-h Fcst)
14
Cntl
TAMDAR
Stage-IV
Sea-level pressure (mb) and 1-h precip.
(in) Valid 1500 UTC 30 AUG 2005 (16-h Fcst)
15
Cntl
TAMDAR
Stage-IV
Sea-level pressure (mb) and 1-h precip.
(in) Valid 1800 UTC 30 AUG 2005 (19-h Fcst)
Precipitation bands TAMDAR is 4 mb deeper
16
850-hPa Relative Humidity () Valid 1800 UTC 30
AUG 2005 (19-h Fcst)
TAMDAR
Cntl
17
850-hPa Relative Humidity () Valid 0600 UTC 30
AUG 2005 (7-h Fcst)
Cntl
TAMDAR
18
850-hPa RH Analysis Difference TAMDAR minus
Cntl 2300 UTC 29 AUG 2005
Regions of RH responsible for future band
formation
From outer 36-km grid
850-hPa T Analysis Difference TAMDAR minus
Cntl 2300 UTC 29 AUG 2005
19
RAOB verification of RH band (case 2) is tough
because of the space-time void. TAMDAR was meant
to fill this void, but verification against
itself is a last choice. Grell CP scheme trigger
function is dependent on saturation (or near
saturation) of moisture fields. Minor
differences in magnitude that exist near the CP
schemes trigger threshold can tip the scales in
a huge way hours later, which can be good or bad.
Thus, proper assimilation of accurate data is
key!
20
Why was the cyclone in TAMDAR 4 mb deeper when
increased RH was the only difference seen in the
analysis? An increase in lower-tropospheric PV
seen in the TAMDAR run appears to be linked to
latent heat release from the precipitation bands
around the cyclone (e.g., Bretherton
1966). Preliminary findings suggest that the
majority of geopotential height difference can be
attributed to this additional PV.
21
Precipitation Forecast Comparison 46 cases (22
March 2005- 4 June 2005) AIRDAT (RT-FDDA-MM5 -
TAMDAR) AIRNOT (RT-FDDA-MM5 - no TAMDAR) RUC,
NAM, GFS Stage-IV "truth" Originally 49 cases,
but 3 cases in May omitted based on
initialization errors.
22
(No Transcript)
23
(No Transcript)
24
NCAR verification 0.2 to 15-mm threshold
Object-oriented verification technique
Developed by Barbara Brown et al. (NCAR), and
presented at previous GLFE. All 49 cases3 May
outliers not removed GFS not shown
25

• Improvement of QPF accuracy for short-range
  severe precipitation
• 8-11 - Control RT-FDDA-MM5 w/out TAMDAR
  (apples-to-apples)
• 34-55 - RUC w/out TAMDAR (apples-to-oranges)
• 71-84 - NAM (apples-to-squash)
• gt90 - GFS (apples-to-spaghetti)
• A conservative estimate of potential improvement
  because
• 4-month study utilized only 36 ?-levels
• Stage-II/IV bias adjustment typically on low
  side (Smith and Krajewski 1991)
• Weighting/optimization of ingestion/parameteriza
  tions are still being refined
• AirDat and NCAR findings are consistent despite
  different techniques

26
Cold-Start-MM5 Sensitivity Tests
168 simulations on a CONUS 36-km grid 7 winter
events (12 combinations) Table 144-h average of
3-h error Error Forecast - ASOS
Blackadar good in winter despite 5-layer
LSM Snow cover / lack of veg. ? LSMs
influence KF -gt Grell may ? feedback in
warm-start 12/12 combinations ? error with
TAMDAR Automated Surface Observing
System (NWS/FAA/DOD)
27
2-m temp. error averaged for all 7 cases using
KF/Blackadar Cntl (No TAMDAR) Exp (TAMDAR)

• This is a trend seen in all 7 cases!
• not just an artifact of one outlier.
• ?
• Better QPF gt more accurate snow cover,
• albedo, and/or surface radiation gt long-
• range surface temp. impact ?
• Better forecasted feedback from
• downstream blocking ?
• Weird Hovmoller teleconnection ?
• Lucky-7 ?

Objective was to obtain CPU speed benchmark for
new 3GHz dual-core
28

• Upcoming studies
• Sampling Rate Impact Study
• 3 Parallel Simulations on Cold-Start MM5
• 36-km CONUS ? 12-km GLFE
• 48 ?-levels
• CNTL No TAMDAR data
• EXP1 TAMDAR data at old original sample
  rate
• EXP2 TAMDAR data at new increased sample
  rate
• Variable Sampling Rate ?
• Based on forecasted dynamics
• Weighting Studies
• Independent testing of ascent/descent
• and independent testing of RH, T, winds
• Testing of radius and magnitude w.r.t.
  seasonal and diurnal variations
• QPF Verification Round 2

29
Acknowledgments Barbara Brown, Randy
Bullock, and Wei Yu (NCARs object-based QPF
verification) Stan Benjamin and William Moninger
(NOAA/ERL/FSL/GSD) NCAR (OSSE computer support,
etc.) NASA Aeronautics Research Offices
Aviation Safety Program FAA Aviation Weather
Research Program AirDat, LLC
30
(No Transcript)