1. The SEVIRI Precipitating Clouds Product of the Nowcasting SAF: First results
15 October 2004
IPWG-2, Monterey
Anke Thoss
Swedish Meteorological and Hydrological Institute
Ralf Bennartz
University of Wisconsin

2. Contents
Introduction
Algorithm
Examples
Performance
Plans

3. Problem overview: NWCSAF approach:
Except for strong convection, VIS/IR features are not
strongly correlated with precipitation.
 likelihood estimates in intensity classes
more appropriate than rain rate retrieval
NWCSAF approach:
2 complementary products for Nowcasting purposes
Precipitating Clouds (PC) product gives likelihood of precipitation in coarse intensity classes
2. Convective Rain Rate (CRR) product estimates rain rate for strongly convective situations

4. three classes of precipitation intensity from
PC product:
three classes of precipitation intensity from
co-located radar data
Rain rate
Class 0: Precipitation-free mm/h
Class 1: Light/moderate precipitation mm/h
Class 2: Intensive precipitation mm/h

5. Data sets for algorithm development
Colocated sets of:
AVHRR
NWP Tsurface (HIRLAM)
radar reflectivities (dBZ),
gauge adjusted, of the
BALTRAD Radar Data Centre
BRDC (Michelsson et.al. 2000)
No quantitative tuning to MSG performed for version 1.0 which is presented here!

6. Input: NWCSAF Cloud type product NWP surface temperature (ECMWF)
MSG channels : 0.6 m, 1.6 m, 3.9 m, 11 m and 12 m

7. Algorithm development:
Based on Cloud type output
Correlation of spectral features with precipitation investigated
Special attention to cloud microphysics (day/night algorithms)
Precipitation Index PI constructed as linear combination of spectral features
Algorithms cloud type specific

8. Correlation of Spectral features with rain
Correlation with class, all potentially raining cloudtypes T
Tsurf - T
T11-T R R
ln(R0.6/R3.7)
R0.6/R
3.7m day algorithm, all 0.35
1.6m day algorithm, all 0.44
night algorithm, all

9. Probability distribution, all raining Cloudtypes
Night algorithm
3.7 Day algorithm
1.6 Day algorithm

10. Precipitation Index Example AVHRR 3.7 day algorithm, all cloud types:
PI= (Tsurf-T11)+5.99(ln(R0.7/R3.7))-3.93(T11-T12)
Example AVHRR 1.6 day algorithm, all cloud types:
PI = *abs(4.45-R0.6 /R1.6)+0.495*R (T11-T12)
+0*Tsurf+0*T11
MSG day algorithm:
Blend of 3.7µm day algorithm (applied to 3.9 µm channel)
and 1.6 µm algorithm with equal weight, some additional
features introduced for later use in quantitative tuning (a8-a10):
PI=a0 +a1*Tsurf +a2*T11+a3*ln(R0.6/R3.9)+a4*(T11-T12)
+a5*abs(a6-R0.6/R1.6)+a7*R0.6
+ a8*R1.6+a9*R3.9+a10*(R1.6/R3.9)
MSG night algorithm still identical to PPS

11. Cloud type dependence Algarithm 0 All precipitating cloud types
Reported 30min. rain frequency at Hungarian gauges
March-June 2004
Algorithm1
Medium level clouds
14.9%
5027 colocations
Algorithm2
High and very high opaque clouds
31.4%
4126 colocations
Algorithm3
Medium to thick cirrus
5.3%
5999 colocations
Thick cirrus most rain
Algorithm4
Cirrus over lower cloud
No Precipitation
All cloudfree classes,
low and very low clouds,
thin cirrus, fractional cloud
0.1% for cloudfree (of 9255)
0.9% for nonprecipitating cloud types (of 11459)

12. Cloud type and total precipitation likelihood (day), March 2004, 12UTC
100% -
70%
60%
50%
40%
30%
20%
10% 0%
Cloud type and total precipitation likelihood (day), March 2004, 12UTC

13. Night algorithm, courtesy of M
Night algorithm, courtesy of M. Putsay, Hungarian Meteorological Service

14. Day algorithm, , 1045

15. 05:30
06:30
07:30
Upper:PC1, lower:PC2,

16. 30 min. sampling
10 min. sampling

17. high+ very high opaque
medium level
Cirrus moderate-thick
Ci over lower cloud

18. Day 20% No Rain MSG Rain (30 min) 84.1% 15.9% 24.0% 76.0% Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(30 min)
84.1%
15.9%
24.0%
76.0%
Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(10 min)
82.9%
17.1%
21.5%
78.5%
20%
likelihood
threshold
N=36466
Rain:
7.1% (30min)
4.9% (10min)
20%
POD= 0.76
FAR= 0.73
PODF= 0.16
HK= 0.60
BIAS= 2.85
ACC= 0.84
30%
POD= 0.58
FAR= 0.65
PODF= 0.08
HK= 0.50
BIAS= 1.66
ACC= 0.89
POD= 0.78
FAR= 0.81
PODF= 0.17
HK= 0.61
BIAS= 4.13
ACC= 0.83
POD= 0.62
FAR= 0.74
PODF= 0.09
HK= 0.52
BIAS= 2.42

19. Day 20% No Rain MSG Rain (30 min) 78.2% 14.7% 1.7% 5.4% Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(30 min)
78.2%
14.7%
1.7%
5.4%
Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(10 min)
78.9%
16.3%
1.0%
3.8%
20%
likelihood
threshold
N=36466
Rain:
7.1% (30min)
4.9% (10min)
Percent of total number 
Percent of gauge class 
Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(30 min)
84.1%
15.9%
24.0%
76.0%
Day 20%
Hungary,gauges
march-june 2004
No Rain
MSG
Rain
(10 min)
82.9%
17.1%
21.5%
78.5%

20. MSG PC Product validation with
surface observations
Dataset: 15 May – 18 June :00 UT:
MSG data and
Collocated surface observations of present weather (only ww classes indicating clearly rain or no rain considered)
PC product without use of cloud type (only a NN based cloud mask)

21. Validation of MSG PC product
Day, 45 N – 55 N, Total data points : (4.6 % raining)
Likelihood of precipitation agrees well with synop

22. Validation of MSG PC product
Night, 45 N – 55 N, Total data points : (4.6 % raining)
Likelihood of precipitation agrees well with synop

23. Validation of MSG PC product
Day, 30 N – 45 N, Total data points : 7218 (2.5% raining)
Likelihood of precipitation is over-estimated by the PC product

24. Validation of MSG PC product
Night, 30 N – 45 N, Total data points : 7218 (2.5 % raining)
Likelihood of rain is over-estimated by the PC product

25. Day 30N 45N No Rain MSG Rain Synop ww 84.2% 15.8% 9.8% 90.2% Night
81.3%
18.7%
Rain Synop ww
11.4%
88.6%
20%
likelihood
threshold
N=12123
4.6%
raining
20%
POD= 0.90
FAR= 0.78
PODF= 0.16
HK= 0.74
BIAS= 4.12
ACC= 0.84
POD= 0.88
HK= 0.72
BIAS= 4.17

26. Day 30N 45N No Rain MSG Rain Synop ww 87.4% 12.6% 7.8% 92.2% Night
85.5%
14.5%
9.8%
91.1%
20%
likelihood
threshold
N=7218
2.5%
raining
20%
POD= 0.92
FAR= 0.84
PODF= 0.13
HK= 0.79
BIAS= 5.78
ACC= 0.86
POD= 0.91
FAR= 0.86
PODF= 0.14
HK= 0.77
BIAS= 6.51

27. Score summary for MSG hardclustering threshold 20%
PC Product
POD
FAR HK
BIAS
Details
AMSU LAND
0.89
0.83
0.47
Against BALTRAD radar
AMSU skill to resolve intensity not considered here
AMSU SEA
0.88
0.75
0.57
MSG day 45-55N
0.90
0.78
0.74
4.12
Alg.0 (no cloud type), May/June
45-55N against Synop WW
MSG night 45-55
0.72
4.17
MSG day 30-45N
0.92
0.84
0.79
5.78
30-45N against Synop WW
MSG night 30-45
0.91
0.86
0.77
6.51
MSG day 30min.
0.76
0.73
0.60
2.85
Cloud type dependant,
March-June 2004
against Hungarian gauges
MSG day 10min
0.81
0.61
4.13

28. Open questions
Why does verification against SYNOP WW look better than for gauge comparison (POD)?
(parallax adjustment, alg0 better than alg1-alg4, May/June
easier, all difficult ww excluded …)
Timescale / horizontal scale (real effect or convenient Bias correction?)
How can false alarms be reduced further?

29. Algorithm Performance – Summary
 Night algorithm seems OK for strong convection, but overestimates precipitation (extent and intensity) for frontal situations
 Day algorithm better in general, but has no skill to class precipitation intensity  recommended to display total precipitation likelihood
 Discontinueties between day and night algorithm
 Precipitation likelihood fairly correct between 45-55N
 South of 45N precipitation likelihood overestimated

30. What is next? Status: ongoing
 tuning against European synop, covering a years cycle
Status: ongoing
while tuning, try to decrease discontinuaty
between day and night algorithm, especially for PC2
need more gauge data for PC2 tuning
 later: investigate usefulness of additional channels