1. Spring AGU, Montreal May 20, 2004
Comparative inverse analysis of MOPITT and aircraft observations to estimate Asian sources of carbon monoxide
Colette Heald, Daniel Jacob, Dylan Jones, Paul Palmer, Jennifer Logan, David Streets, Glen Sachse, John Gille
Spring AGU, Montreal
May 20, 2004

2. How well does MOPITT constrain regional sources?
Do satellite and aircraft observations provide consistent/complementary/redundant constraints on pollution sources?
How well does MOPITT constrain regional sources?
EMISSIONS
MOPITT
TRACE-P
With the recent advent of satellite observations of the troposphere, it is important that we understand how these observations fit into the traditional framework of in situ observations
We try to address this in the context of how satellite and aircraft observations constrain pollution sources during the TP campaign. Do they provide consistent/complementary or redundant information?
I’m also going to try to provide some insight on how well we can constrain regional sources in general with MOPITT data.

3. APPROACH: INVERSE MODEL (February-April 2001)
Forward Model
(GEOS-CHEM)
2°x2.5° resolution
Anthropogenic CO
[Streets et al., 2003]
and Logan & Yevich
TRACE-P
Aircraft CO
EMISSIONS
OBSERVATIONS
This is the framework for what I’m going to talk about today.
Current emission inventories are very uncertain, but represent our best bottom-up information on sources.
We expect that growth in East Asia in the next decades will be one of the major drivers for atmospheric change. And so we’re of course very interested in improving our estimates of those emission sources in particular.
The observations from the TRACE-P aircraft campaign in spring 2001 are really ideally suited to characterizing Asian outflow, as this mission coincided with both the peak in biomass burning activity in Asia as well as the time of maximum outflow from the continent.
At the same time MOPITT observations of carbon monoxide became available and I will be focusing on obs from this region during the TRACE-P mission.
We can use a forward model to link emissions to observations. For this purpose I use the GEOS-CHEM global 3D CTM at 2x2.5 resolution
In order to exploit the constraints that these observations can provide on CO emissions we can use a standard Bayesian inverse model. One of the advantages of this formal approach is that it provides an estimate of the a posteriori errors – although as I will discuss, these errors may not represent a true uncertainty for our solution.
An important point is that the inverse solution depends critically on the error characterization, both on the bottom up (or a priori) emissions and on the error on our observations and in our model, which we call collectively the observational error
Inverse Model
S are error covariance matrices, K is the Jacobian matrix of the forward model
MOPITT CO
(daily v.3 column)
Biomass Burning CO
[Heald et al., 2003a]

4. QUANTIFYING OBSERVATIONAL ERROR (SS)
Aircraft: Residual Relative Error (RRE) approach [Palmer et al., 2003]:
RRE
Mean bias
Altitude [km]
SS = (y*RRE)2
(measured-model) /measured
MOPITT: mathematically derived result:
Observational Error
in “MOPITT space”
Variance of the
MOPITT/Model differences
= RRE
We also include the spatial covariance of the observational error [Jones et al., 2003]
Here is a quick overview of how we characterize these errors. I’m happy to take questions on the details.
We base our approach on the RRE by Palmer et al. This plot illustrates this in the case of the aircraft – where the relative difference is plotted as a function of altitude. We assume that the bias (in red here) is primarily a result of sources but that the variability about this mean bias (in blue) reflects the observational error.
We mathematically derived a very similar result for the satellite observations, which essentially shows that the observational error can also be approximated as the RRE. The plot here shows the geographical variability of this error for the MOPITT column observations which varies from 5-20% in our regions of interest.

5. WHAT STATE VECTOR SHOULD BE USED?
Can all these regions be
constrained independently
by our observations?
Correlation of the a posteriori errors suggest that biomass burning and anthropogenic sector emissions are correlated within any given region
We reduce the state vector to 11 elements
The observations we are using are clearly chosen to quantify regional sources in Asia and we can arbitrarily specify our state vector (the vector of emissions we are trying to solve for). We specified the following regions, and separate anthropogenic and biomass burning emissions in the regions shown in blue. But the question is can all these regions be constrained independently by our observations?
When we examine the a posteriori errors we find sources within a region are correlated, and we therefore merge the original 16 element state vector to 11 elements.
We can get further insight into the information provided by these observations by calculating what is called the “effective rank” of the system – which is the number of pieces of information that can be retrieved above the noise in the system. For this 11 element state vector, we calculate a rank of 10 for MOPITT and 5 for the aircraft. This clearly demonstrates that there is more information in the MOPITT observations for constraining Asian sources.
Effective rank of the system = SVD of
Further Insight…
MOPITT rank = 10
Aircraft rank = 5
There is less information in aircraft observations to constrain Asian sources

6. (IN)CONSISTENCY OF AM VS. PM MOPITT DATA
= “best case”
Based on this “optimized” state vector, we performed a number of sensitivity tests.
I’m comparing here the results obtained when using the am MOPITT cross-over data to the pm cross-over data.
The plot shows for each element of the state vector (or region) the magnitude of emission. The bottom-up a priori is marked with a cross and the grey bar is the associated bottom-up uncertainty. The dots show the estimates from the inversions.
We find substantial differences in the results when we use ‘am’ vs. ‘pm’ data – and these differences far exceed the a posteriori uncertainty on these sources, which are in all cases smaller than the circles on this plot
Since only the daytime MOPITT observations have been validated, they represent our “best case” and I will only use ‘am’ data in everything that follows.
Night cross-over retrieval not validated – we only consider am overpass in our results

7. SENSITIVITY TO ERROR SPECIFICATION
This plot shows the sensitivity of the solution to the error specification. The general consistency of these results suggests that the inversion solution is relatively robust.
I compared my best case, which includes geographical variation in the correlation of the errors, with a case where error correlations were quantified with a single length scale (green) and when no correlation was implemented (blue). We find that the solution is sensitive to the correlation in the errors
We find the solution is relatively insensitive to the a priori error specification (yellow and purple), but more so to the observational error (which we test by setting the observational error to a constant percentage 7% or 20% (light blue and black) of the observations.
Solution is sensitive to the magnitude of the observational error and
its spatial correlation

8. SENSITIVITY TO TEMPORAL AVERAGING OF MOPITT DATA
I also found that the solution was sensitive to temporal averaging of the data.
Here I contrast the daily data to weekly averages and to the 7-week mean. By averaging the data, we loose synoptic information on the outflow and indeed we find that the effective rank of the system drops from 10 to 9 to 6 as we temporally average the data.
Optimistic b/c reduced errors by sqrt(N)
Information is lost when the MOPITT observations are temporally averaged

9. MOPITT INVERSION: RANGE OF SOLUTIONS
a priori (grey bar = uncertainty)
“best case” inversion
I a posteriori error on “best case”
I range of inverse solutions
This figure summarizes my analysis with the MOPITT data. The a priori is now shown with the black circle and the best case is shown in red. You can see here the small a posteriori uncertainties associated to the best case solution.
In addition to this best case I am showing a range of solutions (all using am only data) as defined by my sensitivity tests. We feel that this type of “ensemble analysis” provides an estimate of the true uncertainty on the solution. We see that this range does however represent a reduction in uncertainty from the a priori.
In terms of the sources in Asia, we see that regions dominated by anthropogenic emissions, such as China are underestimated whereas regions dominated by biomass burning emissions are overestimated.
Regions dominated by anthropogenic emissions are underestimated,
Regions dominated by biomass burning emissions are overestimated.
The range of solutions provides a better estimate of uncertainty than a posteriori error.

10. COMPARING AIRCRAFT AND SATELLITE RESULTS
Aircraft not well-suited
to sampling all of SE
Asian outflow
I compare here the inverse solution obtained with MOPITT observations to that obtained with aircraft. I am showing this on a reduced state vector, because if you will remember I earlier indicated that the aircraft has less information and it is unlikely to be able to separate all of the regions in my 11 element state vector.
The a priori is shown in black here. If we compare the red (aircraft) with the green (MOPITT) we see that they suggest remarkably similar emission magnitudes, with the exception of SE Asia, where we feel that the outflow is not fully characterized by the aircraft downwind (MOPITT has the advantage of obs over the continent).
MOPITT validation during TRACE-P suggests that there was a 6% bias, if we remove that bias, we get the result in blue which marginally improves the agreement with the aircraft
And finally, if we reduce the MOPITT observations to the outflow region sampled by the aircraft over the Pacific, we can see that we generally further improve the agreement with the aircraft. However by doing so we loose substantial information (effective rank drops to 6) therefore the results shown in green which include continental data likely provide a better characterization.
Aircraft and MOPITT generally consistent, but differ on quantitative partitioning

11. CONCLUSIONS
Aircraft and MOPITT provide a consistent characterization of sources in Asia.
MOPITT satellite observations provide more information towards constraining emission sources in Asia than aircraft (due more to geographical coverage of data than to data density)
BUT, aircraft observations are essential to the science-based validation of satellite instruments, so that we can use the satellite data accurately. In addition aircraft data can provide correlative flux information
The range of inverse solutions exceeds the a posteriori uncertainty.
This range of solutions is a reduction from the a priori uncertainty
(We have improved our understanding of Asian sources)
In Asia anthropogenic emissions appear to be underestimated, while biomass burning emissions appear to be overestimated
These are the main conclusions from what I’ve presented here. I can leave them up and am happy to take questions.