Recent progress of NCEP OSSE project

1
Assessing/Reducing/Representing Uncertainty in
observations    
Yucheng Song Michiko Matsutani Jack
Woolen Zoltan Toth
Acknowledgements Sharan Majumdar Steve
Lord Russ Treadon Mark Iredell
IMSG/EMC/NOAA/NWS/NCEP
3rd NCEP/NWS Ensemble User Workshop (1030-1050)
Laurel, MD
2
Collaborators

• Sharanya Majumdar Univ. of Miami/CIMAS
• Craig Bishop Naval Research Lab
• Christopher Velden Univ. of Wisc./CIMSS
• Milija Zupanski Colo. State Univ./CIRA
• Thomas Hamill NOAA/Climate Diagnostics Lab
• Istvan Szunyogh Univ. of Maryland
• Robert Atlas NASA/GSFC
• David Emmitt, Simpson Weather Associates,
  Charlottesville, VA

3
OUTLINE

• Uncertainties related to observing system
• - Instrument errors
• - Data Processing errors (Brads example)
• - Representativeness errors (NWP)
• Reducing uncertainties originating from
  observations
• - More/better quality observations
• Adaptive observation method
• Collection
• Processing
• OSSE studies
• Winter storm reconnaissance
• Summary

4
Uncertainties related to observation system (1)

• Instrument errors
• Random errors
• Systematic bias errors
• (accuracy and precision calibration issues)
• OFCM's site on accuracy and precision

High accuracy but bad precision
High precision but bad accuracy
Processing errors (Brads example) - needs
sophisticated quality control
An example showing Track-Check Problems with
African AMDAR Data Courtesy of Dr. Bradley
Ballish
5
Uncertainties related to observing system (2)

• Representativeness errors?
• Why bother?
• Numerical Weather Prediction models all have
  spatial resolution
• e.g. NCEP T382L64 Operational GFS
  model (1152x576 or 768x384)
• Ideally, observing system should provide
  information on all model variables at each
  initial time, representative on the model scale
  on the model grid
• Limited Spatial /temporal/variable coverage
  (observation Gaps) NWP first guess fills the
  gaps
• Parameterization schemes are based on the model
  resolution (SAS, RAS)
• Small scale variability not accounted by NWP
  but will be sampled by the observation
• Importance of assessing these initial errors
  quality of analysis and forecast depends on good
  error estimates

Area closed by dashed line is the half grid-point
box value on the grid point represents the
average of the box

6
Estimation for Representativeness Error
340km twice the model grid length at 170km
Jet region may add more observation errors
Power spectrum of model resolvable scale vs.
un-resolvable scales used in estimation
From Lorenc 1992
7
An example of observing errors

• Example, ADPUPA (220) Rawinsondes from NCEP GDAS

0.85000E03 0.10000E10 0.10000E10 0.15000E01
0.10000E10 0.10000E10 0.80000E03 0.10000E10
0.10000E10 0.16000E01 0.10000E10 0.10000E10
0.75000E03 0.10000E10 0.10000E10 0.16000E01
0.10000E10 0.10000E10 0.70000E03 0.10000E10
0.10000E10 0.16000E01 0.10000E10 0.10000E10
0.65000E03 0.10000E10 0.10000E10 0.18000E01
0.10000E10 0.10000E10 0.60000E03 0.10000E10
0.10000E10 0.19000E01 0.10000E10 0.10000E10
0.55000E03 0.10000E10 0.10000E10 0.20000E01
0.10000E10 0.10000E10 0.50000E03 0.10000E10
0.10000E10 0.21000E01 0.10000E10 0.10000E10
0.45000E03 0.10000E10 0.10000E10 0.23000E01
0.10000E10 0.10000E10 0.40000E03 0.10000E10
0.10000E10 0.26000E01 0.10000E10 0.10000E10
0.35000E03 0.10000E10 0.10000E10 0.28000E01
0.10000E10 0.10000E10 0.30000E03 0.10000E10
0.10000E10 0.30000E01 0.10000E10 0.10000E10
0.25000E03 0.10000E10 0.10000E10 0.32000E01
0.10000E10 0.10000E10 0.20000E03 0.10000E10
0.10000E10 0.27000E01 0.10000E10
0.10000E10 P T
q Wind
Psfc TPW
8
Reducing uncertainties related to observation
system for NWP

• More observations in space/time/variables
• - Costly and more design work should be
  done first (OSSE, OSE)
• - DA overwhelmed by voluminous data
• Adaptive techniques
• Cost effective in areas of major interests
• Adaptive collection
• Reducing errors in initial conditions for
  forecasting crucial weather events
• Adaptive processing
• More detailed processing in targeted areas
• Reduce both random instrument errors and
  representative errors

9
Uniform Rawinsondes observation Experiment in OSSE
N. Hem. Forecast Skill Upper Tropospheric Wind
Time averaged from Feb13-Feb28 12 hourly
sampling Difference in AC for synoptic scale is
presented
Fibonacci Grid used in the OSSE (25mb vertical
interval )
OSSE can test impacts of analysis and forecast
due to equal area coverage observing network
10
DWL targeting example in OSSE
The task is to decide where to turn on the DWL
from Satellite for 10 of the time.
There has to be a match between satellite tracks
and max error level regions
11
Position of jet stream and adaptive sampling
region
10 Upper Level Adaptive sampling (based on the
difference of first guess and NR, three 3mins of
segments are chosen the other 81 mins
discarded)
The maximum sampling region located in the jet
core
12
Animation of adaptive data sampling
( full period every 6 hr)
13
Summary of Targeting Tests Results
A simple Adaptive sampling scheme shows
impressive results

• 10 DWL without targeting does not produce much
  impact
• A perfect 10 adaptive targeted DWL had a 3 day
  forecast skill similar to the 50 DWL experiment
• Target regions correspond well with the Northern
  Hemisphere jet stream

V 200hPa
By selecting a target efficiently, the data
impact could increase to the equivalent of as
much as 5 times the data in Northern Hemisphere
Winter.
Synoptic Scale AC score improvement
14
Winter Storm Reconnaissance Program

• Objective
• Improve Forecasts of Significant Winter
  Weather Events Through Targeted Observations in
  Data Sparse Northeast Pacific Ocean
• Approach
• 1) Collected Only Prior to Significant Winter
  Weather Events of Interest
• 2) Collected in Areas that Influence the
  Forecast the Most
• Results
• 60 80 of Targeted Numerical Weather
  Predictions Improve Significantly Due to Winter
  Storm Reconnaissance Program (Operational in
  January 2001)

15
ETKF-based targeting strategy
The ETKF spotted the target area
16
WSR overall statistics (2004-2006)
252219 66 OVERALL POSITIVE CASES.
010 1 OVERALL NEUTRAL CASES. 1078
25 OVERALL NEGATIVE CASES. 71.7
improved 27.1 degraded
OVERALL EFFECT
17
Summary

• Representativeness errors must be correctly
  assessed in NWP
• Adaptive observations can be used to reduce
  uncertainties related to observing system for NWP
  
• OSSE system can be used to study the adaptive
  data collection and processing techniques
• WSR provides a real example of how adaptive
  collection in sensitive regions can help improve
  forecast of major weather events

18
Background
19
Basic 3D VAR problem
20
Estimation for Jet region
The B region represents the Jet estimation So we
have a rough estimate of representativeness error
in the range of 2.43.8m/s
Jet region may add more observation errors to the
model
21
Data selection Cases (200mb Feb13 - Mar 6 average
)
50 Upper Level regular sampling
100 Upper Level
10 Upper Level
10 Upper Level tropics
(To be repeated with more uniform distribution)
22
10 Upper Level NH band
10 Upper Level NH Ocean
10 Upper Level Adaptive sampling (based on the
difference of first guess and NR, three 3mins of
segments are chosen the other 81 mins
discarded)
23
Forecast Impact(Jan 22-23,2005)
500mb Z
SLP and 1000mb Z
24
Forecast verification(Jan 22-23,2005)
SLP
250mb Height