meteocast: a real time nowcasting system

Geostationary MultispectralImagery Using Neural Models For

Meteorological Applications

Dr. Michele de Rosa1,2,Prof. Frank S. Marzano1,Dr. Antonio Eleuteri4, Dr. Giancarlo Rivolta3

1 Sapienza University of Rome, via Eudossiana, 18 - 00184 Rome – Italy2 T.R.S. S.p.A., via della Bufalotta, 378 – 00139 Rome – Italy3 Logica UK at the European Space Agency (ESA) - ESRIN (EOP-GTR), Po-Box 64 - 00044 Frascati (RM) - Italy4 Royal Liverpool & Broadgreen University Hospitals NHS Trust (RLBUHT) and the University of Liverpool, Prescot Street, L7 8XP, Liverpool, UK

RMets Conference 2011 - Exeter, 2011/06/28

RMets Conference 2011, Exeter UK

Summary

� Introduction� The problem� The starting point� The model� The case studies� The rainfall estimation� The future� The near future

Introduction

� Precipitation is a key factor in regulating equilibrium and life on Earth. It is a crucial geophysical parameter and one of the main actors in the global water cycle.

� 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.

Introduction: the global hydrological cycle

Introduction: The rainfall remote sensing

� The remote sensing provides an indirect measurements of rainfall.

� It is done through measurements of the radiative properties of the hydrometeors (i.e. inferring cloud/rain structures by measuring their radiative properties), both in a passive way (i.e. measuring the radiation spontaneously emitted by the hydrometeors and sensed by a radiometer) as in an active way (i.e. inferring the rain/cloud structure by measuring the reflected portion of the radiation emitted by a radar towards the precipitating cloud).

� Visible to IR estimates of rainfall are only indirect because they try to infer the underlying cloud structure from the top-of-the-cloud appearance

Introduction: The MSG (Meteosat Second Generation)�

� Geostationary satellite� First mission 1977 (Meteosat 1)�� 12 Channels (3 Vis,8 IR,1 HRis)�� Vis and IR resolution 3712x3712� HRis resolution 11136x5568� 15 minutes observation period� About 3 x 3 Km of resolution

Introduction: The MSG channels

Channel Spectral Band (µm) Main application

λ cen λ min λ max1 VIS0.6 0,635 0,56 0,71 Surface, clouds, wind fields 2 VIS0.8 0,81 0,74 0,88 Surface, clouds, wind fields3 NIR1.6 1,64 1,5 1,78 Surface, cloud phase4 IR3.9 3,9 3,48 4,36 Surface, clouds, wind fields5 WV6.2 6,25 5,35 7,15 Water vapor, high level clouds, atmospheric instability6 WV7.3 7,35 6,85 7,85 Water vapor, atmospheric instability7 IR8.7 8,7 8,3 9,1 Surface, clouds, atmospheric instability8 IR9.7 9,66 9,38 9,94 Ozone9 IR10.8 10,8 9,8 11,8 Surface, clouds, wind fields, atmospheric instability10 IR12.0 12 11 13 Surface, clouds, atmospheric instability11 IR13.4 13,4 12,4 14,4 Cirrus cloud height, atmospheric instability12 HRV Surface, clouds

Spectral Band (µm)

Broadband (about 0.4 – 1.1

The problem

� Develop a model based on the MSG frames to make nowcasts (from 30 MINs to 60 MINs) about the rainfield.� The model would predict the MSG IR

channels in order to predict the rainfield.� The model would be flexible, accurate and

quick.

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 IR channel nowcast� Model for IR-MW mapping (Neural net)�

The model: the multichannel approach

� MSG's images application� IR channels (4,5,6,7,8,9,10,11) nowcast (30 min-1

Hr)� Rain estimation from MW and IR sources, using

the IR channels nowcast� Bayesian approach to train the model� GLM nowcast model� Model for IR-to-Rain Rate mapping

The model: The multichannel 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 nowcast the next frame� The Dynamically Averaging Network (DAN)

Ensemble

The model: the Bayesian approach

� The bayesian framework was developed by David J. C. Mackay in the context of the neural networks.� The framework implements the Occam Razor

in order to penalize complex models vs. simple models.� The framework applies the evidence

approach to penalize the complex models.� The framework is general.

The model: the Ensemble of models

� An ensemble is a composition of different models.� In general, the ensemble is used to average

between different models.� The output of an ensemble minimizes the

average error with respect to the ensemble’s components (see Bishop C. M.).

The model: the different kinds of

Ensemble

� The GEM (General Ensemble Model)

iiiGEM xff

� The BEM (Basic Ensemble Model)

iiBEM xf

The model: the DAN Ensemble

� Let y be the output of a neural net and let be the probability associated to y.

� Let the DAN (Dynamically Averaging Networks) ensemble be defined as:

iiiDAN ywf

where:

1)(and

5.0≥yp

otherwise

Certainty

The model: the probability computation

� If is the estimated error bar related to the prediction y of the pixel (i,j) of the ensemble’s component n then:

yErrnji

yErryp

so that:

1)(0 , ≤≤ ypnji and ∑

nji yp

1, 1)(

The model: the multichannel approach layout

The case-studies

� The area analyzed is East longitude ranging from 7°to 18 °and North Latitude ranging from 36.5°to 48 °� 2006-07-24 � 2006-08-13� 2006-09-14� 2007-03-20 (for generalization test)� Each frame consists of 275x344 pixels

The case studies

The case studies: the performance indexes

( ) ( ) ( )[ ]∑ − kibkiest

bpoints

kεt,PTt,PT

1� BIAS (K)

( ) ( ) ( )[ ] 21

−∑ kibki

pointskε

t,PTt,PTN

=ts� RMSE (K)

( ) ( )[ ] ( ) ( )[ ]

( ) ( )[ ] ( ) ( )[ ] 21

−−

∑∑

kbkibkest

bkiest

kbkibkest

bkiest

tTt,PTtTt,PT

tTt,PTtTt,PT=tr� Correlation

index (%)

The case studies: the benchmarks

� The Persistence � The Steady State Displacement

ttt FF =∆+

vFF ttt

r+=∆+

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� Each bayesian GLM consists of 726 inputs

(nc=5, embed=6), 1 output.� 9 components and 27 GLMs

The case studies:

30 mins ahead mean values

0.20.40.60.8

11.21.4

DAN SSD Persistence

Ch. 10Ch. 11Ch. 4Ch. 5Ch. 6Ch. 7Ch. 8Ch. 9

The case studies :

DAN SSD Persistence

The case studies:

DAN SSD Persistence

Correlation

The case studies:

DAN SSD Persistence

The case studies :

DAN SSD Persistence

The case studies:

DAN SSD Persistence

Correlation

The case studies: