On model-data fusion approaches for wet deposition in North America

Alain Robichaud1, Amanda S. Cole2, Mike D. Moran3, A. Lupu4, M. Beauchemin5 and V. Fortin6

Environment and Climate Change Canada’s ADAGIO project (Atmospheric Deposition Analysis Generated by optimal Interpolation from Observations) generates maps of wet, dry and total annual deposition of oxidized and reduced nitrogen (N) and sulphur (S) in Canada and the United States by fusion of observed and modeled data. Optimal interpolation methods are used to integrate seasonally-averaged surface concentrations of gaseous, particulate, and precipitation species into concentration fields predicted by Environment and Climate Change Canada’s in-line regional air quality model GEM-MACH. The result, a product called objective analysis (OA), corresponds to an interpolation of the difference between the modeled and measured values at network observation sites. In model-data fusion, the weights attributed to model and observations are usually derived from error statistics from both model and observation. However, there is no real consensus on how to choose the best method to derive error statistics (weights for model and observations) for data-fusion products. Most methods have been successfully applied in the free troposphere and stratosphere but the extrapolation to the surface is not straightforward. This is because at the surface, model-data fusion deals with observations near emission sources and the presence of turbulent flows highly dependent on the topography or land cover.  Consequently, near the surface, the representativeness error is likely to be more significant compared to the free troposphere. Therefore, modification of existing techniques is required. A new approach is presented here for model-data fusion based on a modification of the Hollingsworth and Lonnberg method. It consists of three steps: 1) performing sensitivity tests using 5 independent ensembles (each containing 80% of the data) to identify the optimum correlation length Lc (i.e. having the lowest analysis error), 2) with these initial error statistics, running the solver for the objective analysis by scaling both the model and observation error by the value of a diagnostic parameter (χ2 : xhi squared) obtained by the first pass in the OA solver. The second pass would then correspond to the tuned weights or error statistics (model and observation errors), 3) independent validation using the remaining 20% of the data. The method is robust, converges very rapidly (no more than a few iterations) and performs well when verifying against independent observations. In this presentation, the method will be described along with examples of wet deposition and concentration for sulfate, nitrate and ammonium. The resulting trends in these fields from 2010 through 2016 will also be shown.

 

 

1Environment and Climate Change Canada, alain.robichaud@canada.ca
2Env. and Climate Change Canada, amanda.cole@canada.ca
3idem, mike.moran@canada.ca
4idem, alex.lupu@canada.ca
5idem, marc.beauchemin@canada.ca
6idem, vincent.fortin@canada.ca