Exploiting Satellite Observations of Tropospheric Trace Gases Ross N. Hoffman, Thomas Nehrkorn, Mark Cerniglia Atmospheric and Environmental Research, Inc. Lexington, MA Dylan B. A. Jones Department of Physics University of Toronto, Toronto, ON Colette L. Heald, Daniel J. Jacob, Robert M. Yantosca Division of Engineering and Applied Sciences And Department of Earth and Planetary Sciences Harvard University, Cambridge, MA Oliver Wild Frontier Research Systems for Global Change Yokohama, Japan Isabelle Bey Swiss Federal Institute of Technology Lausanne, Switzerland
Inversion Methodology The state x that minimizes the cost function J ( x ) is the state with the maximum probability Assumes that statistics are Gaussian and unbiased Biases will result in an non-optimal estimate of the state Improve model parameters by minimizing the mismatch between the observations and the forward model simulation, weighted by the uncertainty of the observations and model simulation
Characterizing Model Error 2. Model Intercomparison Compare GEOS-CHEM and FRSGC/UCI models Estimate model error by comparing the modelled CO fields, calculated using the same CO emission inventory (Street et al. in Asia, Logan et al. in rest of the world) FRSGC/UCI model ECMWF meteorological fields spectral T63 (1.8 º x1.8 º ) Vertical resolution: 30 eta levels GEOS-CHEM model GMAO meteorological fields Horizontal resolution: 2 º x 2.5 º Vertical resolution: 48 sigma levels 3. Compare GEOS-CHEM with MOPITT observations Apply MOPITT averaging kernels to modelled CO fields and estimate the relative error based on the statistics of the differences between the modelled and observed CO (binned on the GEOS 2 º x 2.5 º grid) for the TRACE-P period 1. The NMC method (Parrish and Derber, 1992): Assume that the differences between forecasts of CO for the same time are representative of the model error structure Compare pairs of 48-hr and 24-hr forecasts of CO
Model Errors in Column CO (for Feb-April 2001) Model errors based on UCI-GEOS differences are twice as large as NMC errors, but the patterns are consistent (correlation coefficient r = 0.71) NMC method GEOS - FRSGC/UCI differences percent Error covariance based on 89 pairs of forecasts
NMC Model Error Correlation Altitude = 8 km Altitude = 1.5 km Correlations are: not isotropic influenced by local meteorology
Mean (March) CO column differences between GEOS and FRSGC/UCI Systematic errors in the CO columns are small (less than 10%) Largest biases are associated with convection in Africa and south Asia percent
Mean (March) Convective Mass Fluxes GEOS-CHEM FRSGC/UCI Mean difference (FRSGC/UCI - GEOS) Vertically integrated mass flux (up to 200 hPa) Biases in CO columns in Africa reflect differences in central African convection kg m -1 s -1 More widespread convection in central Africa in GEOS-CHEM
Comparison between NMC and MOPITT derived errors Spatial variability has been filtered by spectrally truncating fields to 24 wavenumbers Large-scale error patterns from NMC method are consistent with MOPITT derived errors percent NMC model error in column CO (with MOPITT averaging kernels) Model error in column CO based on model-MOPITT differences
MOPITT CO: Asian Monsoon Plume (350 hPa) June 2000 Sept July Aug. Oct. Nov. Spatial and temporal variability reflect changes in CO emissions, convective lifting, and advective outflow [Kar et al., 2004]
GEOS-CHEM CO: Asian Monsoon Plume June July August Sept. Oct. Nov. Challenge: segregating the error components associated with the emissions, observations, and transport
Summary The largest errors in the simulation of CO are in regions of strong convection Errors in convection introduce significant biases in the models The three approaches for characterizing model error provide different but complementary information Future Work Extend the NMC study to other seasons using GMAO forecasts of CO and CO 2 Examine the seasonal and interannual dependence of the MOPITT derived model errors Compare GEOS-CHEM with GEM-AQ (York University)