Validation of OCO-2 error analysis using simulated retrievals

Characterization of errors and sensitivity in remotely sensed observations of greenhouse gases is necessary for their use in estimating regional-scale fluxes. We analyze 15 orbits of simulated OCO-2 with the Atmospheric Carbon Observations from Space (ACOS) retrieval, which utilizes an optimal estimation approach, to compare predicted versus actual errors in the retrieved CO2 state. We find that the non-linearity in the retrieval system results in XCO2 errors of ~0.9 ppm. The predicted measurement error (resulting from radiance measurement error), about 0.2 ppm, is accurate, and an upper bound on the smoothing error (resulting from imperfect sensitivity) is not more than 0.3 ppm greater than predicted. However, the predicted XCO2 interferent error (resulting from jointly retrieved parameters) is a factor of 4 larger than predicted. This results from some interferent parameter errors larger than predicted, as well as some interferent parameter errors more strongly correlated with XCO2 error than predicted. Variations in the magnitude of CO2 Jacobians at different retrieved states, which vary similarly for the upper and lower partial columns, could explain the higher interferent errors. A related finding is that the error correlation within the CO2 profiles is less negative than predicted, and that reducing the magnitude of the negative correlation between the upper and lower partial columns from −0.9 to −0.5 results in agreement between the predicted and actual XCO2 error. We additionally study the post-processing bias correction affects errors. The bias corrected results found in the operational OCO-2 Lite product consists of linear modification of XCO2 based on specific retrieved values, such as the CO2_grad_delta (a measure of the change in the profile shape versus the prior) and dP (the retrieved surface pressure minus the prior). We find similar linear relationships between XCO2 error and dP or CO2_grad_delta, but see a very complex pattern of errors throughout the entire state vector. Possibilities for mitigating biases are proposed, though additional study is needed.