contact  |  support  |  links
level I inversion results
A “control” or “base” case inversion has been performed with the Level I TransCom model output submissions. We employ a Bayesian synthesis inversion formalism, specifying prior estimates of both the fluxes and their uncertainty, and optimizing with respect to atmospheric observations which are also uncertain.
The basic elements for this inversion are below.
All the ingredients and the basic inversion output are here.
Here are the basic output files generated by the IDL inverse code:
outsum.allpre.dat: model mean posterior fluxes for all 22 basis function regions and regional groupings. Also contains “within” and “between” uncertainties. All presubtraction fields are included.
outmod.allpre.dat: individual model posterior fluxes and uncertainties.
outgrp.allpre.dat: individual model results for the regional groupings.
The inversion is forced with 1992-1996 mean CO2 concentration data calculated from GLOBALVIEW-CO2 [2000]. This dataset encompasses 141 data records from 18 measuring programs including both flask and continuous measurements. Not all sites were operational during the full time period of interest but GLOBALVIEW uses a data interpolation/extrapolation scheme to fill data gaps and provide complete data records at all available observing sites. Since the CO2 records contain both a trend and a seasonal cycle, we choose here to use this interpolated data in calculating the 1992-1996 mean so that these values are not biased by missing poprtions of the data record. However we reject any site where the interpolated data is greater than 30% of the total data. Using this criterion, 76 sites are used in the inversion.
The data uncertainty needs to encompass a number a factors including measurement precision, network intercalibration, the ability of the model to represent the observation and interannual variability. We have found that the inversion is sensitive to the choice of data uncertainty used. While the choice made here for the control inversion is not intended to be optimal, it is designed to incorporate a range of the factors that impact the inversion through the data uncertainty specification. Previous inversions have used either constant uncertainties at all sites (with values ranging from 0.3-0.85 ppmv) or uncertainties which scale according to the variability at a site. Here we base the uncertainties on the residual standard deviation (rsd) given in GLOBALVIEW. This is a measure of the variability of individual flask samples around the smooth fitted curve from which the pseudo-weekly GLOBALVIEW data are generated. The rsd values are consequently higher for sites that are close to local, heterogeneous sources where datasets tend to be noisier. This, then, provides a measure of the uncertainty in modeling a particular location since we expect the modeling to be less reliable when close to sources.
The rsd values are modified in a number of ways to create the data uncertainties used here. We describe the process and its justification as a series of steps that are applied to each site.
1. The GLOBALVIEW annual rsd values are averaged over 1992-1996.
2. When the inversion is performed, a data mismatch is generated for each site. The methodology assumes that these mismatches, normalized by the initial data uncertainty for that site, will be normally distributed with mean 0 and standard deviation of 1. Initial tests produced standard deviations that were too small, indicating data uncertainties that were too large. We consequently reduce the rsd values by dividing by the square root of 5multiplied by the proportion of ‘real’ data contributed to the 1992-1996 mean. This effectively gives extra weight to those sites with more complete data records.
3. This would give a number of sites with data uncertainties less than 0.3 ppmv. This is unrealistic, both due to intercalibration issues between measuring programs and modeling constraints associated with approximating site locations with grid box average concentrations. We therefore set all data uncertainties to a minimum of 0.3 ppmv.
4. Finally, the observational network is spatially heteorogeneous. In the control case we present here, there are some locations with multiple data records and large parts of the globe with no measurements. We make some account for this heteorogeneity by giving less weight (i.e. larger data uncertainty) to data records that are co-located or close to each other. We do this by multiplying the data uncertainty for each site by the square root of the number of data records which are located within 4º latitude and longitude and 1000 m altitude of each other.
The prior flux estimates for the terrestrial basis functions represent an average of recent inventory studies. The prior flux for the North American boreal region: 0.0 Gt C/year; for the North American temperate region: -0.2 Gt C/year; for the American tropics region: 0.55 Gt C/year; for the South American temperate region: 0.0 Gt C/year (default value); for the Northern Africa region: 0.15 Gt C/year; for the Southern Africa region: 0.15 Gt C/year; for the Eurasian boreal region: -0.4 Gt C/year; for the Eurasian temperate region: 0.3 Gt C/year (a portion of what consider Tropical Asia); for the Tropical Asia region: 0.8 Gt C/year; for the Australia region: 0.0 Gt C/year; for the Europe region: -0.1 Gt C/year. The prior flux uncertainties for the terrestrial basis function regions were equivalent to the growing season net flux (GSNF, defined as the sum of carbon uptake for all months in which this was a positive number) as provided by the CASA model of net ecosystem production.
Oceanic prior fluxes are set at zero for each oceanic region. Ocean source uncertainties are based on density of pCO2 measurements and the area of each region.
The inversion code used for the Level I inversion was developed by Peter Rayner with further work by Rachel Law and Kevin Gurney. It is written in IDL. In addition to the observational data and the prior flux/uncertainties, this code requires a control file which contains paths to the model response functions, direction on which pre-subtracted field to include, and information on regional aggregation.
site design by Micek Design
All Images on this Website are the property of the stock.xchng, all rights reserved.
Copyright © 2011, Arizona State University all rights reserved.