verification methods for high resolution ensemble forecasts
DESCRIPTION
Evaluation and Diagnostic. Verification Methods for High Resolution Ensemble Forecasts. Barbara Brown 1 , Tara Jensen 1 , Michelle Harrold 1 , Tressa Fowler 1 , Randy Bullock 1 , Eric Gilleland 1 , and Brian Etherton 2 - PowerPoint PPT PresentationTRANSCRIPT
Barbara Brown1, Tara Jensen1, Michelle Harrold1, Tressa Fowler1, Randy Bullock1, Eric Gilleland1, and Brian Etherton2
1NCAR/RAL - Joint Numerical Testbed Program2NOAA/ESRL - Global Systems Division
AndDevelopmental Testbed Center
Verification Methods for High Resolution Ensemble Forecasts
Warn-On-Forecast Workshop: 8-9 February 2012
Evaluation and Diagnostic
Large variability in space and time Difficult to identify meaningful
impactsExtreme, high impact, weather
eventsSmall regions of importance
Difficult to identify impacts of forecast “improvements” across whole domain
Verification scores Desire for a single score... But CSI
alone does not give a lot of information about performance or improvements
Relationships and dependencies among scores
Double penalty for displaced high-res forecasts
Challenges for objective evaluation of convective scale short-term prediction...
•ObservationsUncertainties in qualityNeed to be on time and spatial scaled that support evaluation
Traditional Verification Scores
CSI is a nonlinear function of POD and FARCSI depends on base rate (event frequency) and Bias
Ex: Relationships among scores
FAR
POD
CSI
Very different combinations of FAR and POD lead to the same CSI value
1CSI 1 1 1POD 1 FAR
PODBias1 FAR
Freq Bias
Results from HMT-West 2010-2011 seasonCourtesy of Ed Tollerud, DTC/HMT Collab
9km - Ensemble Mean – 6h Precip
6HR>0.1 in.
6HR>0.5 in.
6HR>1.0 in.
6HR>2.0 in.
Performance DiagramAll on same plotPOD1-FAR
(aka Success Ratio)CSIFreq Bias
Best
Success Ratio (1-FAR)
Freq
Bia
s
CSI0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Here we see:Decreasing skill with higher thresholds even with multiple metricsRoebber (WAF, 2009)
Wilson (presentation, 2008)
Dots representDifferent leads
Looking at variability in space and time
14 May 2009 Init: 00 UTC MODE Objects Thresh: 30dBZ
No RadarAssim.
Objects ForecastField
ObservedField
RadarAssim.
SolidFCSTOBJ
LineOBS OBJ
What if you had this: Instead of this:
Objects are 30% too big (MODE area ratio=1.3) FBIAS = 1.3 for grid
Shifted west 40 km(MODE centroid distance = 10 gs) POD = 0.35
Moving too slow(MODE-TD angle diff = 15%)
FAR = 0.72
Peak Rain 1/2” too much(MODE diff in 90th percentile of intensities = 0.5”)
CSI = 0.16
Comparing objects can give more descriptive assessment of the forecast
QPE field Probability Field
Traditional MetricsBrier Score: 0.07Area Under ROC: 0.62
Spatial MetricsCentroid Distance:Obj1) 200 kmObj2) 88kmArea Ratio:Obj1) 0.69Obj2) 0.65
1
2 Median Of MaxInterest: 0.77
Obj PODY: 0.72Obj FAR: 0.32
Okay Forecast with
Displacement Error?
Or Bad Forecastbecause Too
SharpOr
Underdispersive?
QPE_06 >1 in. 50% Prob (APCP_06> 1 in.)
Perfe
ct Reli
abilit
y
OverFcst
UnderFcst
Reflectivity > 30 dBZ MODE Objects (01-06hr)
Looks like there is definitely a timing error
What would a spaghetti plot of MODE objects look like?
Hatched - Observed OBJ Solid - Fcst OBJ
Centroid of object
MODE Time Domain - (MODE – TD)
West
East
Color gives W-Emovement
Tim
e in
crea
sing
Preparing for implementation in Model Evaluation Tools (MET) in next release (Spring 2012)
Applied to model forecasts from 2010 HWT Spring Experiment
Example Attribute of MODE-TD
Conv. Rad. = 10 grid squaresConv Thresh. = APCP_01 > 3 mm
HWT SE 2010 DataMay 17 – June 16, 2010 (18 forecasts/days)
Number of objects:Obs 143, ARPS 161ARW 172, NMM 238
E-W speed (m/s)
Duration (h)
90th percentile intensity
MODE-TD Attributes
Continue with traditional scores (and methods of display) and not gain a true sense of why the ensembles are performing well and poorly- Or -
Adopt performance diagrams, reliability diagrams, rank histograms, and spatial methods for diagnostics and understanding
Take aways: To evaluate high resolution ensembles, you could…
Model Evaluation Tools (MET) developed by the Developmental Testbed Center (DTC)
R statistics package, including R-spatial developed by Eric Gilleland at NCAR/RAL Joint Numerical Testbed (JNT) and R-verification
Where to get these methods?
Recommended websites:http://www.dtcenter.org/met/users
http://www.r-project.org/ http://www.cawcr.gov.au/projects/verification/
http://www.ral.ucar.edu/projects/icp/http://www.dtcenter.org/eval/hwt/2011
The DTC and JNT Program can also serve as a resource for Evaluation Techniques
Extra MaterialToolsIntercomparisonsOutreachCollaboration
www.dtcenter.org/met/usersMETv3.1AvailableNow
METv4.0With GRIB2 support and MODE-TD to be released in late spring 2012
Object Oriented Method: MODEHow it works
OBSENS FCSTRadius=5Thresh>0.25”
Radius=5Thresh>0.25””
Merging
Matching
No false alarms
Misses
Merging
Matched Object 1
Matched Object 2
Unmatched Object
1. Smooth Field 2. Threshold 3. Restore Intensities
4. Merge within 5. Match between 6. Calculate Attributesand Categorical Stats
Example from HMT-West 2010-2011 seasonCourtesy of DTC/HMT Collaboration
Volume / intersection / union / symmetric difference
Axis anglesSpatial orientationAverage speedCentriods – space and time
No way to combine calculation of centriod of space and time because of their differing units
Fuzzy logic handles them by weighting each independently
Metrics to be available in MODE-TD
“Completely updated chapter on the Verification of Spatial Forecasts taking account of the wealth of new research in the area “
Authors: Brown, Ebert, Gilleland
Jolliffe and Stephenson, 2nd Ed
Spatial verification methods in “R”
Implemented by Eric Gilleland
Includes all major spatial methods
R spatial verification package
Example: Neighborhood methods
ICPInternational effort
focused on comparison of capabilities of the different spatial verification methods
Central U.S. precipitation forecasts and observations from HWT 2005
Many publications in WAF special collection
Preparing now for ICP2
Intercomparison project (ICP) and other collaborations
Web site: http://www.ral.ucar.edu/projects/icp/
Other collaborations• Collaborations with
WMO/WWRP Working Groups on Mesoscale and Nowcasting Research– Documents on verification of
mesoscale forecasts and cloud forecasts
• Workshops and tutorials
International collaborationGoals
Further testing of spatial and other methods from ICP1
Application to complex terrainVariables: Precipitation, WindForecasts (Model output), Obs analyses and
observations from 2007 COPS and MAP / D-PhaseVERA analyses (Include ensemble observation
analyses)2 or more sets of model output from MAP D-
PHASE
ICP2