Inverse Modeling of CO Surface Sources on a Global and Monthly Scale

1
Inverse Modeling of CO Surface Sources on a
Global and Monthly Scale

• Gabrielle Pétron
• gap_at_ucar.edu

National Center for Atmospheric Research Boulder,
Colorado, USA
2
March 2000 Total column of CO MOZART2 (top)
and MOPITT (bottom)
model
observations
3
Overview

• Define the object
• Formulate the problem and the hypotheses
• Method
• Results and Discussion
• Conclusions

4
Object CO emissions

• known nature of sources
• uncertain
• intensities
• location
• timing, seasonality, interannual variations
• splitting fossil fuel/biofuel....
• tools to study CO budget
• observations, emissions inventories, models

5
Observations

• few direct observations of emissions
• ?yet many observations of CO distribution in
  the troposphere
• NOAA/CMDL stations
• N/S gradient, seasonality in sources and
  sinks, seasonality in CO
• high precision
• -- remote regions, mostly NH,representativity?

6
CMDL
7
Observations

• few direct observations of emissions
• ?yet many observations of CO distribution in
  the troposphere
• NOAA/CMDL stations
• N/S gradient, seasonality in sources and
  sinks, seasonality in CO
• -- remote regions, mostly NH,representativity
• MAPS (space shuttle) 4 campaigns
• biomass burning impact
• IMG few days from 8/1996 to 6/1997

8
Remote Sensing of CO
spectroscopy around 4,7 ?m
April -October 1994

• space shuttle
• MAPS 4 flights
• (12-14/11/1981 5-13/10/1984
  9-19/04/1994 30/09-11/10/1994)
• ?max sensitivity in upper troposphere
• satellite
• - IMG a few days between August 1996 and June
  1997 total CO column (?sensitivity max around 6
  km)

Connors et al., 1999
4 days June 1997
Clerbaux et al., 2001
9
Observations

• few direct observations of emissions
• ?yet many observations of CO distribution in
  the troposphere
• NOAA/CMDL stations
• N/S gradient, seasonality in sources and
  sinks, seasonality in CO
• -- remote regions, mostly NH,representativity
• MAPS (space shuttle) 4 campaigns
• biomass burning impact
• MOPITT (satellite) since spring 2000
• CO good tracer of continental air pollution

10
MOPITT Data

• Since spring 2000, the MOPITT instrument
  monitors the tropospheric content of CO, covering
  the global surface of the Earth in a few days.

• Characteristics
• Global coverage in 3 days
• 20km x 20km pixel
• 7 levels
• Lower Precision/in situ
• Filtered for clouds
• CO mixing ratios are Level3 data
• We need to know the averaging Kernel what
  MOPITT sees!!

CO at 500 hPa
11
Emissions inventories

• Fossil fuel 300-600 TgCO/yr
• Biomass burning 300-900
• (forests, savannas, agric. waste burning, fuel
  wood use)
• Vegetation 50-200
• Oceans 6- 30
• Methane oxidation 400-1000
• HCNM oxidation 300-1000
• TOTAL Source 1400 3700 TgCO/yr
• Photochemical sink 1400-2600
• Surface deposition 150-500
• TOTAL Sink 1550 3100 TgCO/yr

LARGE UNCERTAINTIES
12
Chemistry-Transport model

• représentation numérique discrète (x,t) de la
  troposphère globale
• opérateur de chimie
• opérateur de transport
• advection, turbulence diffusion turbulente et
  convection
• conditions aux limites
• émissions, dépôt, échanges avec la stratosphère
• résolution ?x 200/500km , ?t 20min/1 h
• modèle avec champs climatologiques mensuels
  IMAGES (5ox5o)
• modèle avec champs analysés MOZART (2,8ox2,8o)

13
Models

• pretty good... yet...
• at best monthly emissions
• resolution 2.8ox2.8o
• off-line transport
• Only tool to track COs origins

day 2
day 65
14
Question....

• Can observations of CO distribution in the
  troposphere help better constrain CO sources?

15
then....

• HOW?

16
Overview

• Define the object
• Formulate the problem and the hypotheses
• Method
• Results and Discussion
• Conclusions

17
Known Limitations

• no optimization of CO production
• transport model is not perfect
• observations are not perfect
• is the problem linear ?
• if one double the emissions,
  do we get double CO?? NO!
• is problem weakly non-linear ? test
• dimensions!!!!!

18
DIMENSIONS

• Observations
• CMDL 40 stations, weekly data over gt 10 years,
  uncertainty quite low
• MOPITT tons of data, need filtering for clouds,
  errors not well described,
• best for comparisons with models
• Monthly averages (then No (MOPITT) gt5000)
• Emissions in model!
• 700 to 2700 land grid cells

19
Remark

• If some things are better/well known in the model
  they should not be part of the optimization
  process.
• example the location of a source is know, then
  it should be fixed

20
More questions...

• Do we know the errors on the observations ?
  biases ?
• Shall we use a set of priori emissions ?
• Yes
• Problem is weakly non linear unique solution
• Inverse Pb is under-determined
• What kind of error do we put on prior?

21
Why do we need errors???

• Let x be your unknown
• Let y1 and y2 be two observations of x
• Let s1 and s2 be E(y1 -x)2 and E(y2 - x)2
  Lets suppose E(y1 -x)E(y2 -x)0
• The Best Unbiased Linear Estimate of x is

22
Overview

• Define the object
• Formulate the problem and the hypotheses
• Method
• Results and Discussion
• Conclusions

23
Formulation of the problem

• We are going to optimize
• p monthly source processes vector x
• using
• n monthly averaged observations vector z

source-regions, Xsurface station
tagging
24
Formulation of the problem

• We are going to optimize
• p monthly source processes vector x
• using
• n monthly averaged observations vector z

We need to solve the following system
xxbxb zh(x)e E(xb)0, E(e)0, E(xbxbT), E(e
eT)R
25
Linearization

• h(x)h(x, u) where u(x) ....
• where u is for example the chemical production of
  CO, or the OH concentration
• we linearize around xb
• h(x)h(xb)H(x-xb)d dox-xb
• The system becomes
• xxbxb
• zh(xb)H(x-xb)ed

error due to linearization
26
H observation matrix

• h(x)h(xb)H(x-xb)d

The coefficients of H are calculated using the
model with fixed OH concentrations.
Hij normalized impact of source i to observation j
27
Solution
sum of information
pxp
nxn
28
Hypotheses

• Transport model is perfect
• Statistics of the observations known (mean and
  cov matrix), ? R is diagonal (no correlation)
• Statistics of the a priori sources known,
  ?B is diagonal (no
  correlation)
• All Errors are gaussian and independent
• CO sink not optimized.
• Chemistry weakly non linear.
• a posteriori sources close ENOUGH to a priori
  sources ? no big change in
  OH
• ? Use linearized version of the model

29
Overview

• Define the object
• Formulate the problem and the hypotheses
• Method
• Results and Discussion
• Conclusions

30
CMDL / IMAGES inversion

• representative of the 1990-1996 period
• 33 monthly source-processes
• 39 monthly averaged observations

31
a priori emissions
a posteriori emissions
Asian annual emissions are multiplied by 2

32

• Timing of biomass burning in Southern Hemisphere
  changed
• emissions peak in September

a posteriori emissions (open symbols)
tagged contributions of various sources to total
CO at Ascension Island
33
a priori / a posteriori uncertainties
impact of changing the a priori sources
uncertainties increasing the uncertainties
gives more weight to the observations
34
CO budget

• Hyp
• 50 uncertainties on a priori fluxes,
• min 10 errors on observations

35
Improvement at stations
Observed and simulated CO mixing ratios (ppbv)
black dots observations. bold lines with no
symbol standard deviation of the observations.
open diamonds CO simulated using the a priori
emissions open squares CO simulated using the
a posteriori emissions.
36
Obs
IMAGES
New sources
Monthly Average CO (ppbv)
37
Validation/ Tests

• Test1 Inversion with pseudo-data
• Test2 improvement of modeled CO at the
  stations
• 30.47 ppbv with a priori sources
• 17.56 ppbv with a posteriori sources
• Test3 CH4 and MCF lifetimes
• 8.74 yr (8.42 w/ a priori) / Krol et al.
• 4.41 yr (4.27 w/ a priori) / Krol et al.

38
Major Results CMDL /IMAGES

• Large increase in Asian Sources (//Kasibhatla et
  al., 2002)
• Change in timing of biomass burning in Southern
  Africa (//Galanter et al. 2000)
• Uncertainties decrease the most for Northern
  Hemisphere sources
• Not enough data to constrain Southern Hemisphere
  emissions
• Optimization of Chemical Production of CO not
  done!

39
What we need to get better results

• Use assimilated meteorology
  (to reduce bias due to errors in model
  transport)
• Include a test on OH for various lat. bands
• More observations (esp. in SH)
• More isotopes measurements
• d13C and d18O (strong signature of source type)
• Optimize methane and NOx sources as well

40
MOPITT /MOZART inversion

• Same inverse technique used
• MOPITT binned on MOZART grid
• use data at 700 hPa
• inversion done for each month of observations
• Ne 3000 obs /month
• 15 continental regions / 4 oceanic biogenic
• chemical production of CO not optimized
• 100 relative error on observations and on prior
  emissions

41
Chemistry Transport MODEL

• MOZART 3D-global
• surface to 2hPa
• 31 vertical levels
• Horizontal resolution 2.8º x 2.8º
• 56 chemical species
• Dynamical Fields NCEP, ECMWF, DAO
• timestep 20 min

42
Method Iterative Inversion
Analysis at month k1

• Observations MOPITT CO at 700 hPa
• 65oN-65oS, ½ of the data, April 2000-March 200
• Modeled CO
• projected on MOPITT 700 hPa level (aver. kernel)
• Diagonal covariance matrices
• Observation error 100
• A priori emissions inventory
• 50 error on technological CO sources
• 100 error on other CO sources

dimensions emissions U 106
observations z 3000
43
MOPITT /MOZART inversion

• April 2000 March 2001

44
Inversion results
Annual CO emissions from fossil fuel, biofuel
combustion and biomass burning (TgCO/yr)
45
Monthly global CO surface sourcesApril 2000-
March 2001
a priori

• shift in biomass burning maximum (august?
  sept)
• biofuel use emissions maximum in winter
  (x 2 / summer time)
• fossil fuel emissions maximum in winter
  (30 / summer time)

a posteriori
a posteriori
46
CO Biomass Burning EmissionsTgCO/yrApril
2000-March 2001
a priori
a posteriori
47
Validation
using CMDL CO dataset
The agreement between the modeled CO and the
observations improves at 26 stations (out of 33)
when using the optimized sources.
48
(No Transcript)
49
Overview

• Define the object
• Formulate the problem and the hypotheses
• Method
• Results and Discussion
• Conclusions

50
Conclusions (1) CO budget

• First time inversion of CO monthly sources
  using satellite data.
• Large underestimation in current inventories
• of Asian emissions (fossil fuel and biofuel)
• of biomass burning emissions in Africa
• New estimates of CO emissions from biomass
  burning (no hypothesis on EF, biomass burned...)

51
References

• http//twister.caps.ou.edu/OBAN2002/Assim_concepts
  .pdf
• Bouttier and Courtier notes on
  Data Assimilation concepts and methods
• articles by Enting, Talagrand, Tarantola

52
Conclusions (2) perspectives

• Test
• Impact of the uncertainties assigned to
  observations and to a priori emissions
• Impact of meteorological fields used
• Impact of BL ventilation parameterization
• Future work
• Better description of errors, biases of satellite
  data
• Optimization of chemistry, validation of OH
  distribution
• Multi-year inversion
• Integrate information from various platforms
• Implement 4D variational assimilation
  (development of the adjoint model of MOZART in
  progress)

53
Formalism

• System to be solved xr xb xb
• z - yb H(xr - xb ) ?
• ? measurement error and model error eoem
• hyp E(xb)E(?)E(?.xbT)0 E(xbxbT)Pb
  E(?.?T)R

unknown xr
54
March 2000 Total column of CO MOZART2 (top)
and MOPITT (bottom)
model
observations
55
July 2000 Total column of CO MOZART2 (top) and
MOPITT (bottom)
model
observations
56
(No Transcript)