![Page 1: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/1.jpg)
Institut für
Physik der Atmosphäre
Institut für
Physik der Atmosphäre
Object-Oriented Best Member Selection in a
Regional Ensemble Forecasting System
Christian Keil and George Craig
Institut für Physik der AtmosphäreDLR Oberpfaffenhofen, Germany
![Page 2: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/2.jpg)
Institut für
Physik der Atmosphäre
Regional Ensemble Prediction System
COSMO-LEPS regional ensemble (ARPA SMR)
1. Identify ten clusters from ECMWF 51 member ensemble2. Use a representative member from each cluster to drive a
regional model (DWD Lokal Model)
To make use of forecast ensemble, need to weight members equally probable use cluster populations, or use most recent data (e.g. satellite imagery)
But, for local severe weather, phase errors may dominate, so use a nonlinear pattern recognition algorithm
![Page 3: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/3.jpg)
Institut für
Physik der Atmosphäre
Main components:1. COSMO-LEPS: based on ECMWF EPS providing initial
and boundary conditions and Lokal-Modell (LM Δx=7km)
2. LMSynSat: forward operator to compute synthetic satellite imagery in LM
3. Objective Pattern Recognition Algorithm using Pyramidal Image Matching
Regional Ensemble Prediction System
![Page 4: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/4.jpg)
Institut für
Physik der Atmosphäre
Clustering of 1 EPSs fc range +48..60h (2002070912-00)using 4 discriminating variables at 3 pressure levels(u,v,Φ,q at 500/700/850 hPa):
Clustering method -----> COMPLETE LINKAGE Selection mode --------> MINIMIZE INT/EXT RATIO Ensemble --------------> 1 Initial Date ----------> 2002 7 7 12 UTC Forecast range (hours) -> 48 - 60 Area Limits (N/S/W/E) --> 60.0 30.0 -10.0 30.0 Number of clusters ----> 10 Explained Variance(%) -> 42.8 Cluster ---------------> 1 2 3 4 5 6 7 8 9 10 Size ------------------> 6 8 10 6 6 4 4 5 1 1 Internal variance(%) --> 5.8 9.8 12.3 6.9 6.8 4.6 4.6 6.5 .0 .0 Radius -----------------> 12.3 13.8 13.8 13.3 13.3 13.4 13.4 14.2 .0 .0
CL 1: ( 5) 0 5 17 24 40 41 CL 2: ( 1) 1 4 9 11 18 32 33 49 CL 3: ( 31) 2 3 10 12 26 28 31 34 46 50 CL 4: ( 39) 6 16 22 29 39 42 CL 5: ( 43) 7 13 36 38 43 48
CL 6: ( 45) 8 25 27 45CL 7: ( 44) 14 35 37 44CL 8: ( 15) 15 20 21 30 47CL 9: ( 19) 19CL 10: ( 23) 23
COSMO-LEPS case-study: 9 July 2002
![Page 5: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/5.jpg)
Institut für
Physik der Atmosphäre
Generation of synthetic satellite images in LM: LMSynSat
• RTTOV-7 radiative transfer model (Saunders et al, 1999)
• Input: 3D fields: T,qv,qc,qi,qs,clc,ozone
surface fields: T_g, T_2m, qv_2m, fr_land
• Output: cloudy/clear-sky brightness temperatures for
Meteosat7 (IR and WV channels) and
Meteosat8 (eight channels)(Keil et al, 2005)
![Page 6: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/6.jpg)
Institut für
Physik der Atmosphäre
Lokal Modell: all 10 clusters
Meteosat 7 IR 16:00 UTC
Case Study with COSMO-LEPS: 9 July 2002
![Page 7: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/7.jpg)
Institut für
Physik der Atmosphäre
Pyramidal Image Matching
1. Project observed and simulated images to same grid
2. Coarse-grain both images by pixel averaging, then compute displacement vector field that minimizes the total squared error in brightness temperature; search area +/- 2 pixel elements
3. Repeat step 2 at successively finer scales
4. Displacement vector for every pixel results from the sum over all scales
![Page 8: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/8.jpg)
Institut für
Physik der Atmosphäre
Image Matching: BT< -20°C and coarse grain
Meteosat 7 IR
1 Pixelelement = 8x8 LM GP
![Page 9: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/9.jpg)
Institut für
Physik der Atmosphäre
Image Matching: BT< -20°C and coarse grain
Model Cluster 7Observed
![Page 10: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/10.jpg)
Institut für
Physik der Atmosphäre
Displacement Vectors
Image Matching: BT< -20°C and coarse grain
Model Cluster 7Observed
![Page 11: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/11.jpg)
Institut für
Physik der Atmosphäre
Image Matching: successively finer scales
![Page 12: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/12.jpg)
Institut für
Physik der Atmosphäre
Image Matching: successively finer scales
![Page 13: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/13.jpg)
Institut für
Physik der Atmosphäre
Displacement vectors and matched image
![Page 14: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/14.jpg)
Institut für
Physik der Atmosphäre
Rank 1 2 3 4 5 6 7 8 9 10 corr.
subjective 2 10 4 7 9 1 5 6 8 3 0.85new measure 7 2 9 4 10 1 8 6 3 5 0.77population 3 2 4 1 5 7 6 8 9 10 -0.34
• Magnitude of displacement vectors consistent with subjective ranking• Cluster population shows no correlation
Ranking using different Quality Measures
![Page 15: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/15.jpg)
Institut für
Physik der Atmosphäre
A new Quality Measure
FQI = 0.33 * [ nordispl + (1-LM/Sat)+ + (1-corr)]
good
bad
![Page 16: Object-Oriented Best Member Selection in a Regional Ensemble Forecasting System](https://reader035.vdocument.in/reader035/viewer/2022062720/56813509550346895d9c5bac/html5/thumbnails/16.jpg)
Institut für
Physik der Atmosphäre
Conclusions
1. Pyramidal image matching provides a plausible measure of forecast error (consistent with subjective rankings)
2. COSMO-LEPS cluster populations are a poor indicator of local skill
3. Persistence of skill for about 12 hours owing to change of weather regime in region
Future: adaptive forecasting system: stochastic physics, assimilation of MSG and radar data