�ݺ�ߣ

Validation of different nowcasting
models based on the Meteosat
Second Generation satellite data
M. de Rosa1, M. Picchiani1,2, M. Sist1,2, F. Del Frate2
1 GEO-K srl, via del Politecnico, Rome,Italy
2 “Tor Vergata” University of Rome, Department of
Civil Engineering and Computer Science, via del
Politecnico, Rome, Italy

Outline
• Nowcasting
• The StormTrack model
• The validation
• The benchmark
• Validation at South Africa (case studies)
• Validation at Europe (Summer 2015)
• Summary
• Future tasks
2015-09-23 Eumetsat Conference 2015 - Toulouse

Nowcasting
• Very short term weather prediction (within
few hours) over a certain area
• Prediction of extreme weather events like
thunderstorms, floods, hurricanes, tornadoes
• Near real time computation time: in parallel
with observations (weather stations,
soundings, satellite images, weather radar)
• Very useful to the outdoor activities, air traffic
control, agrometeorology

The actors
StormTrack
The benchmark
The ground truth
The validation tool

The StormTrack model
• MSG as unique data source
• (Early) Detection of the convective objects
• Tracking of the detected objects
• Cells lifecycle monitoring
• Temporal and spatial extrapolation of the
detected objects
• High computation efficiency and reliability
• Easy to use

• Cell Detector (CDT)
• Cell Tracker (CTK)
• MSG Nowcasting Engine (MNE)
• CTWriter
STK

JSON Public APIs

The algorithm: CDT
• Use the 5,6 and 9
channels
• BTD6,9 cloud base
detection (early
detection)
• BTD5,9: cloud top
detection (Kolios and
Feidas, 2010)
• Connected components
• Convex approximation
• Object definition
(properties)

The Validation
• MET is a set of verification tools developed by the Developmental
Testbed Center (DTC) for use by the numerical weather prediction
community to help them assess and evaluate the performance of
numerical weather predictions.
• The primary goal of MET development is to provide a state-of-the-
art verification package to the NWP community. By “state-of-the-
art” it means that MET will incorporate newly developed and
advanced verification methodologies, including new methods for
diagnostic and spatial verification and new techniques provided by
the verification and modeling communities.
• Several tools are part of the MET package and the MODE (object
oriented validation) tool has been chosen for the validation of the
StormTrack algorithm.
Don’t reinvent
the wheel
Write new
routines?
New validation
techniques?

The Validation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD FAR

The benchmark
• RDT, Rapid Development Thunderstorms, has been (and is)
developed by Meteo-France in the framework of the
EUMETSAT SAF in support to Nowcasting.
• Using mainly geostationary satellite data, it provides
information on clouds related to significant convective
systems, from meso-alpha scale (200 to 2000 km) down to
smaller scales (few pixels).
• The RDT algorithm includes three steps:
– The detection of cloud systems
– The tracking of cloud systems
– The discrimination of convective cloud objects
• Only objects flagged as convective have been taken into
account
RDT: the state-of-the-art
in storms detection

Validation at South Africa
• 6 case studies during summer
2014 (12 UTC – 18 UTC)
• Lightning network: accuracy
500m, CG strikes, 19
detectors
• Strikes 5 mins before/after
MSG slot time
• RDT setup: MSG HRIT, NWP,
lightning data (thanks to
SAWS)
• StormTrack setup: MSG HRIT

South Africa case study: 2014/12/16

Event statistics
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
0
50
100
150
200
250
300
350
400
Strikes Density (strikes/px)
Strikes
Density

StormTrack vs RDT
PODs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD STK POD RDT
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
0
50
100
150
200
250
300
350
400
Higher lightning activity

StormTrack vs RDT
FARs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FAR STK FAR RDT
0.00
1.00
2.00
3.00
4.00
5.00
0
50
100
150
200
250
300
350
400
Higher lightning activity

Event statistics
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
18.00
0
500
1000
1500
2000
2500
Strikes
Density

StormTrack vs RDT
PODs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD STK POD RDT

StormTrack vs RDT
FARs
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
18.00
0
500
1000
1500
2000
2500
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FAR STK FAR RDT

Event statistics
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
0
500
1000
1500
2000
2500
3000
3500
Density
Strikes

StormTrack vs RDT
PODs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
POD STK POD RDT
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
0
500
1000
1500
2000
2500
3000
3500

StormTrack vs RDT
FARs
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
FAR STK FAR RDT
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
0
500
1000
1500
2000
2500
3000
3500

Summary: South Africa case studies
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
12,00 13,00 14,00 15,00 16,00 17,00
Hours UTC
StormTrack vs RDT
Mean StormTrack POD
Mean RDT POD
Mean StormTrack FAR
Mean RDT FAR
Lower StormTrack FAR (0.2 vs 0.3)
Higher RDT POD in the morning
Higher StormTrack POD in the afternoon

Validation at South Africa:
Summarising
• Mean StormTrack Accuracy (POD): 0.5
– 0.41 in the morning
– 0.61 in the afternoon
• Mean RDT Accuracy (POD): 0.54
• Mean StormTrack FAR: 0.2
– O.2 in the afternoon
• Mean RDT FAR: 0.32

Validation at Europe
• Mid June 2015 – Mid Sep
2015 (00 UTC-21 UTC)
• ATDNet lightning data
(sampled every 5 mins)
• Strikes 5 mins before, 10
mins after MSG slot time
• RDT setup: MSG HRIT,
NWP, lightning data
(thanks to AEMET)
• StormTrack setup: MSG
HRIT

Credits Jocken Kerkmann (EUMETSAT)

• Over 5000 samples collected (Mid June 2015 –
Mid Sept 2015) for StormTrack
• Over 3000 samples collected (Mid June 2015 –
Mid Aug 2015) for RDT
• No filters on ground data: investigating about
the model(s) sensitivity
Density
Strikes
Density
Strikes
Density
Strikes
Density
Strikes

Validation at Europe:
Scores vs Strikes
StormTrack
RDT
Discriminant: 250 strikes, POD>FAR
Discriminant: 130 strikes

Scores vs Matched objects area
StormTrack
RDT
Discriminant:750 pixels
Discriminant: 360 pixels

Scores vs Strikes density
StormTrack
RDT
Discriminant: 3.5 strikes/px
Discriminant: 1.9 strikes/px2015-09-23 Eumetsat Conference 2015 - Toulouse

Hourly Scores
RDT
Higher STK
POD
Lower STK
FAR2015-09-23 Eumetsat Conference 2015 - Toulouse

Summarising
• Mean StormTrack Accuracy (POD): 0.6
• Mean RDT Accuracy (POD): 0.7
• Mean StormTrack FAR: 0.54
• Mean RDT FAR: 0.49
• Sensitivity:
– Strikes: STK 250, RDT 130
– Matched area: STK 750, RDT 360 (px)
– Strikes density: STK 3.5, RDT 1.9 (strikes/px)
STK suitable on extreme events
RDT suitable on different conditions

Our nowcasting equipment
DISH: 110 cm
LNB Inverto
Reception station: 8 GB Ram
Quad core 3 GHz TBS 6983 DVB S2 card
Win 7
BDADataEx DVBS2 card sw
STK HW:I7 Quad core 3.2 GHz 8 GB
FD Processing time: 7 mins
Linux CentOS 6.0

Summary
• Validation over South Africa (6 case studies) and Europe
(Summer 2015)
• MET framework for validation (Object Oriented using
MODE tool)
• StormTrack setup: MSG HRIT
• RDT (benchmark) setup: MSG HRIT, NWP, lightning data
• Good results in the afternoon on South Africa case studies
• Over Europe better RDT POD but comparable FAR. BTW
good STK POD on average (no ground data or NWP)
• STK flexible and “light” (less than 2 GB RAM needed on FD)
• Issues on cold base clouds

Future tasks
• Validation over Europe will go on (divide the
region into sub regions)
• Automatic trajectories extrapolation ready but
not yet operational (15/30 mins ahead)
• RSS integration
• Neural nets to improve POD/reduce FAR (pruning
using other MSG channels)
• Public APIs development to share data (first
version online)
• App and IoT (Internet of Things) integration

Acknowledgements
• Estelle de Coning (SAWS) and SAWS for the RDT and the lightning
data
• Cecilia Marcos (AEMET), Ana Sánchez Piqué (AEMET) and AEMET
for the hints about MET, the RDT and the lightning data
• Italian Air force Meteorological Office for the support

Thanks for your attention
michele.derosa@geo-k.it
Questions?

�ݺ�ߣ

Validation of different nowcasting models based on the Meteosat Second Generation satellite data

More Related Content

Validation of different nowcasting models based on the Meteosat Second Generation satellite data