ozone from gome data using neural network technique

Neural Network based Ozone ProfileRetrieval Using Combined UV/VI

and IR Satellite Data

Anton K. Kaifel, Jasmine Kaptur (ZSW),

Martin Felder (LRZ)

Center for Solar Energy & Hydrogen Research (ZSW), Stuttgart, Germany

Leibniz Computing Center (LRZ), Munich, Germany

ITSC-15, Oct. 2006, Maratea, Italy

Outline

1. Start with TOVS

2. NNORSY and Optimal Estimation (OE)

3. Approach for NNORSY

4. Local error estimation

5. Results of NNORSY-GOME/TOVS

6. Summary and future

Neural Network based Ozone Profile Retrieval Using CombinedUV/VIS and IR Satellite Data


Monthly mean values of56 WOUDC ground stationsalso used for TOMS validation

NNORSY-TOVS L3 (09/79 –12/02)

NNORSY-TOVS Total Ozone

global TOVS TOMSRMS: 9.1 D.U. 8.0 D.U.Bias: 2.0 D.U. -2.8 D.U.

NNORSY-NOAA-TOVS zonal mean 1987 - 2000

Do

bso

n U

nit

s


Op

tim

al

Est

imati

on

x

O3 profile

p(x|y)p(y|x) p(x)

p(y)=

x

"a priori"

y

GOMEspectrum

y

x

… …

k(x,y)

collocated measurements

yx RTM

OptimalEstimation

z.B. FURM, OPERA, ...

NNORSY

~1 min/profile

~1 ms/profile (!)

NN

OR

SY

Co

mp

ari

son

of

retr

ieval

ap

pro

ach

es


Overview NNORSY

operationalreal-time

processing

GOME profiles

Operation

trainingdata

neuralnetworktraining

randomdata

mixing

testdata

eval.data

networkparam.s

qualitycheck

Training

GOMETOVS N14

Level 1data

WOUDCsondes

collocatedspectra/ profiles

collocation

SHADOZsondes

SAGE, POAM, HALOE

profiles

collocation

collocation

Preprocessing


ozonesondes

test data set: 01/1996 – 07/2001, 12000 collocations

POAM III

SAGE II

HALOE

NNORSY: Collocation Example


training data

evaluationdata

y

t

y

t

hiddenlayer

weights w

o(yi)

outputvector

ti

targetvector

"learning process":minimize E=Σi(o(yi)-ti)2

Quality check by cross-validation

test data

y

t

yi

otherinput

param.s

+

inputvector

NNORSY (Neural Network Ozone Retrieval System)

stoppingpoint

Training and Test for NNORSY


NNORSY (Neural Network Ozone Retrieval System)

GOMEspectral

data

geolocation

meteorol. info

climatologicalpredictors

yi

Input data for neural network

NNORSY Input Parameters

270 – 325 nm (O3 Hartley & Huggins)380 – 385 nm (atmosph. window)598 – 603 nm (O3 Chappuis)758 – 772 nm (O2 → cloud info)SZA, SAAsensor scan angleLOS type (east, nadir, west)GEOS4 temperature profileLatitudeSeason and sensor age

HIRSMSU


TOVSTb

Example for NNORSY Ozone Retrieval

y

t

training data 1

y

t

training data 1

o

NNORSYretrieval

|o-t|

training data 2

error networktraining

Estimation of Local Retrieval Error

NNORSYerror estimation


NNORSY NN configurations: training/test


GM: GOMETV: TOVSTp: temperature profile

NNORSY NN configurations: profiles


GM: GOMETV: TOVSTp: temperature profile

NNORSY-GOME Single Ozone Profile I


NNORSY-GOME Single Ozone Profile II


Correlative Analysis: Artic


Correlative Analysis: 400 S / New Zealand


NNORSY-Climatology (TLLO): Time Series


NNORSY-Climatology (TllO): Single Profiles for Antarctic


NNORSY: Summary and Current

► 10 year NNORSY-GOME ozone profile data set

► comparison/validation with sonde/lidar/satellite and data assimilation

► Using NNORSY for setup of new ozone profile retrieval

MetOp RAO Meeting, May 2006, Frascati, Italy

► NNORSY-OMI for ozone profile retrieval

► NNORSY application to SCIAMACHY nadir

► Combining GOME with TOVS: small improvements in troposphere

► New dynamic ozone profile climatology available

NNORSY-MetOp: Future

► Total ozone column

► ATOVS, IASI, GOME, GOME+IASI

► Ozone profiles

► IASI, GOME, GOME+IASI


ozone from gome data using neural network technique

Documents