meteocast: a nowcasting model to predict extreme meteorological events

Post on 29-May-2015

101 Views

Category:

Education

3 Downloads

Preview:

Click to see full reader

DESCRIPTION

Presentation about the MeteoCAST very short term forecast model at the Eumetsat Conference 2013.

TRANSCRIPT

2013/09/16Eumetsat Conference 2013 –

Vienna

MeteoCAST: A Neural Ensemble Nowcasting Model based on Geostationary Multispectral Imagery for Hydro-Meteorological Applications

Dr. Michele de Rosa1,2, Prof. Frank S. Marzano1

1. “Sapienza” University of Rome, via Eudossiana, 18 - 00184 Rome – Italy2. GEO-K srl, via del Politecnico, 1 – 00133 Rome - Italy

2013/09/16Eumetsat Conference 2013 –

Vienna

Outline

Introduction The goal The starting point The model The case studies The rainfall estimation The present The future

2013/09/16Eumetsat Conference 2013 –

Vienna

Introduction

A relevant part of environmental risk can be ascribed to meteorological severe events with high precipitation rate.

Heavy precipitation associated to severe weather may cause serious damages in terms of economic losses and, in extreme cases, of human life losses.

Managing the environmental risk due to precipitation is strictly linked to monitoring and understanding the storms that produces hazards such as flash floods.

2013/09/16Eumetsat Conference 2013 –

Vienna

The goal

Develop a model based on the MSG frames to nowcast (from 30 Mins to 60 Mins) the rain field.

The model should predict the MSG IR channels in order to predict the rain field.

The model should be flexible, accurate and quick.

2013/09/16Eumetsat Conference 2013 –

Vienna

The starting point

The NeuCAST (Marzano et al.) Meteosat 7's images application IR channel (10.8 μm) nowcast (30 mins) Rain estimation from MW and IR sources, using

the nowcasting of the IR channels Model for IR-RR mapping (Neural net)

2013/09/16Eumetsat Conference 2013 –

Vienna

The model: the multi-channels approach MeteoCAST: Meteorological Combined

Algorithm for Storm Tracking Application on MSG images IR channels (4,5,6,7,8,9,10,11) nowcasting

from 30 mins to 60 mins Bayesian approach to train the model GLM nowcasting model Model for IR-to-Rain Rate mapping

2013/09/16Eumetsat Conference 2013 –

Vienna

The model: the multi-channels model tools Cao’s method to find the optimal temporal

window PCA (Principal Component Analysis) to

reduce the number of information sources: the 8 IR channels are replaced by a linear combination of them.

Bayesian model to make nowcasting about the next MSG image

The Dynamically Averaging Network (DAN) Ensemble

2013/09/16Eumetsat Conference 2013 –

Vienna

The model: the multi channels approach layout

2013/09/16Eumetsat Conference 2013 –

Vienna

The case-studies

The area of interest ranges from longitude 7° E to 18° E and from latitude 36.5° N to 48° N

Training 2006-07-24, 2006-08-13, 2006-09-14

Validation 2007-03-20

Test 2013-05-03

Each frame consists of 275x344 pixels

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: Ensemble setup 3 GLMs for each case-study: one GLM for the

lower correlation frame, one for the higher correlation frame and one for the median correlation frame (like the worst, best and mean case in computer science).

3 PCA channels 9 components and 27 GLMs Each bayesian GLM consists of 726 inputs

(nc=5, embed=6), 1 output.Pixel to project ahead

(i,j)

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: the benchmarks The Persistence

The Steady State Displacement (SSD)

ttt FF

vFF ttt

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: the performance indexes

BIAS

RMSE

Correlation index

kibkiestb

pointskε t,PTt,PT

N=tm

1

21

1 2

kibki

estb

pointskε t,PTt,PT

N=ts

21

)(

22

kbkibkestbki

estb

kbkibkestbki

estb

tTt,PTtTt,PT

tTt,PTtTt,PT=tr

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: training set 60 mins ahead mean performance

0

0.5

1

1.5

2

MeteoCAST SSD Persistence

BIAS

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: training set 60 mins ahead mean performance

0

5

10

15

20

MeteoCAST SSD Persistence

RMSE

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: training set 60 mins ahead mean performance

0

20

40

60

80

100

MeteoCAST SSD Persistence

Correlation

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

%

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2007/03/20 13:30 UTC 60 mins ahead

-0.5

0

0.5

1

1.5

2

MeteoCAST SSD Persistence

BIAS

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2007/03/20 13:30 UTC 60 mins ahead

02468

10121416

MeteoCAST SSD Persistence

RMSE

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2007/03/20 13:30 UTC 60 mins ahead

0

20

40

60

80

100

MeteoCAST SSD Persistence

Correlation

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

%

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: tornado over Modena

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: tornado over Modena from MSG

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2013/05/03 14:00 UTC 60 mins ahead

0

0.5

1

1.5

2

2.5

MeteoCAST SSD Persistence

BIAS

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2013/05/03 14:00 UTC 60 mins ahead

02468

10121416

MeteoCAST SSD Persistence

RMSE

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

K

2013/09/16Eumetsat Conference 2013 –

Vienna

The case studies: 2013/05/03 14:00 UTC 60 mins ahead

01020304050607080

MeteoCAST SSD Persistence

Correlation

Ch. 10

Ch. 11

Ch. 4

Ch. 5

Ch. 6

Ch. 7

Ch. 8

Ch. 9

%

2013/09/16Eumetsat Conference 2013 –

Vienna

The rainfall estimation

Use the produced synthetic images in a waterfall manner

Some intermediate products are generated: CM LST RR

Integration with the PGE01 and PGE05 products of the NWCSAF (for calibration purposes)

2013/09/16Eumetsat Conference 2013 –

Vienna

The rainfall estimation: the model layout

DEMMSGBTs

GLMCloud Mask

LSTEstimator

RR Classifier

RREstimator

First level

Second level

Third level

2013/09/16Eumetsat Conference 2013 –

Vienna

The rainfall estimation : tornado over Modena

Performance Indexes 60 Min

BIAS 2.17 mm/h

RMSE 10.49 mm/h

Correlation 53.00 %

2013/09/16Eumetsat Conference 2013 –

Vienna

The rainfall estimation: a static case 2010/01/26 10:15 UTC - 60 Mins ahead.

Performance Indexes 60 Min

BIAS 1.33 mm/h

RMSE 9.05 mm/h

Correlation 68.47 %

2013/09/16Eumetsat Conference 2013 –

Vienna

The present

The www.mondometeo.org website publishes the near real time outputs of the MeteoCAST model

The KMZ service The Augmented Reality service The Twitter service Covered countries: Italy, Swiss, Austria (almost

all covered) and Brazil (Sao Paulo region).

2013/09/16Eumetsat Conference 2013 –

Vienna

The future

CellTrack integration (attend the talk of Davide Melfi tomorrow morning)

RSS integration in order to improve the performance on heavy dynamic events

Integration with the NWCSAF v.2013 Synthetic images (extended to VIS) as input to

the NWCSAF Coverage of other countries: Africa and South

America Extension to other satellites: GOES and MTSAT

2013/09/16Eumetsat Conference 2013 –

Vienna

Acknowledgements

Thanks to the Italian Air Force Meteorological Office

for the support

2013/09/16Eumetsat Conference 2013 –

Vienna

Thank you for your attention

michele.derosa@geok.itmic_der@yahoo.it

top related