comparisons of nowcasting techniques for oceanic convection huaqing cai, cathy kessinger, nancy...
DESCRIPTION
CNO Based on TITAN (Dixon and Wiener, 1993) TITAN for Radar Data An Example of 1 Hr CNO-TITAN *1 hr nowcast of CDO valid at 1315 UTC on August 19, 2007 using TITAN technique is shown on the right; red lines on the right represent CDO = 2.5 verification. *Advantages of TITAN: computationally efficient; capability of addressing growth/decay. *Disadvantages of TITAN: polygons can only roughly represent storm shapes; tends to over- forecastingTRANSCRIPT
COMPARISONS OF NOWCASTING TECHNIQUES FOR OCEANIC
CONVECTION Huaqing Cai, Cathy Kessinger, Nancy Rehak, Daniel Megenhardt
and Matthias Steiner
National Center for Atmospheric ResearchBoulder, CO
14th Conference on Aviation, Range, and Aerospace MeteorologyAtlanta, GA
18-21 January, 2010
ACKNOWLEDGMENTS This study is supported by NASA ROSES and NASA ASAP program and in collaboration with NRL and MIT LL
Oceanic Diagnosis and Nowcasting SystemConvective Diagnosis Oceanic
(CDO) identifies convective cells
CDOInterest
CDOBinary
Product
Convective Nowcasting Oceanic (CNO-Titan) makes
1-hr and 2-hr nowcasts of storm location using an
object tracker (Titan)
CNO-Titan
Nowcast
CNO-Gridded produces gridded nowcasts that will
more closely resemble storm
structures
CNO-GriddedNowcast
CNO-RFRandomForest
Nowcast
CNO-RF utilizes environmental and model-based inputs
to better predict storm initiation and
decay[Cai et al. (2009)]
CTop CClass GCD
With
Growth/Decay
With
Growth/Decay
Without
Growth/Decay
CNO Based on TITAN (Dixon and Wiener, 1993)
TITAN for Radar Data An Example of 1 Hr CNO-TITAN
*1 hr nowcast of CDO valid at 1315 UTC on August 19, 2007 using TITAN technique is shown on the right; red lines on the right represent CDO = 2.5 verification.*Advantages of TITAN: computationally efficient; capability of addressing growth/decay.*Disadvantages of TITAN: polygons can only roughly represent storm shapes; tends to over-forecasting
CNO Based on Modified TITAN---- Gridded Forecast
An Example of 1 Hr CNO-Gridded Forecast
TITAN Motion Vectors at t0
Gridded 0 hr TITAN Motion Vectors
Temporal & Spatial Smoothing
15-60 min Motion Vectors
15-60 min Forecastsby Advecting OriginalSatellite Data at t0
Gridded 1 hr TITAN Motion Vectors
Gridded 2 hr TITAN Motion Vectors
Merged with GFS Winds Closest in Time
Temporal & Spatial Smoothing
Temporal & Spatial Smoothing
75-120 min Motion Vectors
135-180 min Motion Vectors
75-120 min Forecastsby Advecting 60 minNowcasts
135-180 min Forecastsby Advecting 120 minNowcasts
Merged with GFS Winds Closest in Time
Merged with GFS Winds Closest in Time
1-3 Hr CNO-Gridded Forecast Flow Chart
*Advantages of CNO-Gridded: realistic looking storms; low bias.*Disadvantages of CNO-Gridded: could be computationally expensive; no explicit growth/decay capability
CNO Based on Random Forest Statistical Analysis and Data Fusion
• The random forest technique produces an ensemble of decision trees from labeled training instances– during training, RF generates estimates of predictor importance– RF trees “vote” on classification of new data points, comprising
a nonlinear empirical model that provides both deterministic predictions and probabilistic information
Vote: 1=> 40 votes for “0”, 60 votes for “1”; consensus category “1”
Data pt.
Tree 1
Vote: 0
Data pt.
Tree 2
Vote: 0
Data pt.
Tree 3
Vote: 1
Data pt.
Tree 4
Vote: 0
Data pt.
Tree 100…
For details of RF technique in CoSPA R&D effort, referred to paper J10.4 by Ahijevych et al. on Thursday *Slide courtesy of John Williams and Dave Ahijevych
An Example of CNO-RF Forecast
Compared with CNO-TITAN ( 1 hr)
*1 hr forecasts valid at1315 UTC on August 19,2007 for both techniques; Red lines representCDO = 2.5 verification
*Advantages of random forest technique: more realisticlooking storms; taking into account of storm environment to address storm growth/decay.
*As a relatively new, novel technique for nowcasting, its potential needs to be fully explored
CNO
Hurricane DeanA
B
C
D
Hurricane DeanA
B
C
D
CNO-RF
CNO-TITAN
Statistical Evaluation of the Three Nowcasting Techniques
CSI BIAS
•5 days of data from Aug 19-23, 2007 over the Gulf of Mexico domain are used to calculate the statistics with a grid size of ~ 5 km and CDO threshold of 2.5•All three techniques show skill over persistence•RF and gridded forecast perform best at 1 hr lead time•TITAN is the best at 2-3 hr lead time•Gridded forecast is the best for 4-6 hr lead time
Relative Skill Comparisons of Three Nowcasting Techniques versus Persistence
•Gridded and RF nowcasts ~10 % better than persistence at 1 hr lead time•TITAN is the best for 2 and 3 hr lead time (~20-30% improvement)•Gridded nowcasts the best for 4,5 and 6 hr lead time (~ 15-25% improvement)•Overall, gridded technique seems to be the best performer
Examples of 1 hr Gridded Forecast over the Gulf of Mexico Domain
*White lines are CDO=2.5 verification, satellite data available every 30 min
A B
C
DIssue time: 1215 UTC 2009/09/05Valid time: 1315 UTC 2009/09/05
1 HR
Examples of 3 hr Gridded Forecast over the Gulf of Mexico Domain
*White lines are CDO=2.5 verification, satellite data available every 30 min
A B
C
DIssue time: 1215 UTC 2009/09/05Valid time: 1515 UTC 2009/09/05
3 HR
Examples of 6 hr Gridded Forecast over the Gulf of Mexico Domain
*White lines are CDO=2.5 verification, satellite data available every 30 min
A B
C
DIssue time: 1215 UTC 2009/09/05Valid time: 1815 UTC 2009/09/05
6 HR
Examples of 3 hr CDO Forecasts Based on CNO-Gridded Technique in the West Pacific Domain
*White lines are CDO=1.5 verification; Cloud class is not used in CDO*Satellite data are available every 3 hrs
A
B
C
D
Issue time: 2100 UTC 2009/12/28Valid time: 0000 UTC 2009/12/29
3 HR
Examples of 3 hr CDO Forecasts Based on CNO-Gridded Technique in the West Pacific Domain
*White lines are CDO=1.5 verification; Cloud class is not used in CDO*Satellite data are available every 3 hrs
A
B
C
D
Issue time: 0000 UTC 2009/12/29Valid time: 0300 UTC 2009/12/29
3 HR
Examples of 3 hr CDO Forecasts Based on CNO-Gridded Technique in the West Pacific Domain
*White lines are CDO=1.5 verification; Cloud class is not used in CDO *Satellite data are available every 3 hrs
A
B
C
D
Issue time: 0300 UTC 2009/12/29Valid time: 0600 UTC 2009/12/29
3 HR
Summary Statistics of CNO-Gridded Forecasts
•30 days of data from Sep 1-30, 2009 over the Gulf of Mexico domain are used to calculate the statistics with a grid size of ~ 5 km and CDO threshold of 2.5•The results showed here could serve as benchmark performance of extrapolation-based nowcasting techniques for oceanic convection•Similar verification for model forecasts need to be done so that a comparison of convective forecasting skills between model and extrapolation can be obtained
The black squares are
statistics from Aug 19-22,
2007
What are the GFS model scores for oceanic
convection???
Summary and Future Work• Three nowcasting techniques (CNO-TITAN, CNO-Gridded and CNO-
Random Forest) for oceanic convection forecasting in the 1-6 hr time frame are implemented, tested and compared in the Gulf of Mexico domain
• At 1 hr lead time, CNO-Gridded and CNO-RF got the best performance scores
• At 2-3 hr forecast lead time, CNO-TITAN outperforms the other techniques
• After 3 hr, CNO-Gridded outperforms the other techniques• Based on the overall performance statistics, CNO-Gridded forecasts for 1-
8 hr are implemented in the Gulf of Mexico and Pacific domain in the realtime Oceanic Diagnosis and Nowcasting System
• The summary statistical performance of CNO-Gridded extrapolating technique could serve as benchmark for future blending work of GFS model and extrapolating for oceanic convection in 1-8 hr time frame
Thanks for Your Attention!
Questions and Comments?