Royal Meteorological Institute of Belgium, Uccle, Belgium
York University, Mathematics and Statistics, Toronto, Ontario, Canada
University of Science and Technology of China, Statistics and Finance, Hefei, Anhui, China
Abstract. Inversion models can use atmospheric concentration measurements to estimate surface fluxes. This study is an evaluation of the errors in a regional flux inversion model for different provinces of Canada, Alberta (AB), Saskatchewan (SK) and Ontario (ON). Using CarbonTracker model results as the target, the synthetic data experiment analyses examined the impacts of the errors from the Bayesian optimisation method, prior flux distribution and the atmospheric transport model, as well as their interactions. The scaling factors for different sub-regions were estimated by the Markov chain Monte Carlo (MCMC) simulation and cost function minimization (CFM) methods. The CFM method results are sensitive to the relative size of the assumed model-observation mismatch and prior flux error variances. Experiment results show that the estimation error increases with the number of sub-regions using the CFM method. For the region definitions that lead to realistic flux estimates, the numbers of sub-regions for the western region of AB/SK combined and the eastern region of ON are 11 and 4 respectively. The corresponding annual flux estimation errors for the western and eastern regions using the MCMC (CFM) method are -7 and -3 % (0 and 8 %) respectively, when there is only prior flux error. The estimation errors increase to 36 and 94 % (40 and 232 %) resulting from transport model error alone. When prior and transport model errors co-exist in the inversions, the estimation errors become 5 and 85 % (29 and 201 %). This result indicates that estimation errors are dominated by the transport model error and can in fact cancel each other and propagate to the flux estimates non-linearly.
In addition, it is possible for the posterior flux estimates having larger differences than the prior compared to the target fluxes, and the posterior uncertainty estimates could be unrealistically small that do not cover the target. The systematic evaluation of the different components of the inversion model can help in the understanding of the posterior estimates and percentage errors. Stable and realistic sub-regional and monthly flux estimates for western region of AB/SK can be obtained, but not for the eastern region of ON. This indicates that it is likely a real observation-based inversion for the annual provincial emissions will work for the western region whereas; improvements are needed with the current inversion setup before real inversion is performed for the eastern region.
How to cite. Chan, E., Chan, D., Ishizawa, M., Vogel, F., Brioude, J., Delcloo, A., Wu, Y., and Jin, B.: Investigation of error sources in regional inverse estimates of greenhouse gas emissions in Canada, Atmos. Chem. Phys. Discuss., 15, 22715–22779, https://doi.org/10.5194/acpd-15-22715-2015, 2015.
Received: 24 Jun 2015 – Discussion started: 26 Aug 2015
This paper uses synthetic data experiments to investigate the impacts of different error sources associated with prior flux, transport model and optimisation method on the atmospheric greenhouse gas inverse estimates. Results indicate that estimation errors are dominated by the transport model error and can propagate to the flux estimates non-linearly. It is necessary to obtain stable and realistic results in synthetic data experiments before a real observation-based inversion is performed.
This paper uses synthetic data experiments to investigate the impacts of different error sources...