Testing the OCO-2 Retrieval Algorithm with 3D Radiative Transfer Simulations

The Orbiting Carbon Observatory-2 (OCO-2) Level 2 (L2) retrieval algorithm uses an optimal estimation approach to infer the column-averaged carbon dioxide concentration (XCO₂) from high spectral resolution radiance measurements. In order to implement a practical algorithm, a highly optimized 1-D radiative transfer model must be used in the retrieval. Even with a highly optimized 1-D model, the computational expense is significant, which limits the application of the retrieval to a small fraction of the total collected data. A key aspect of the full L2 algorithm is the prescreening step, which identifies the subset of the collected data likely to produce accurate XCO₂ retrievals. Clouds are the most likely source of retrieval errors, implying that the prescreening is primarily a cloud screening method.

Previous tests of the OCO-2 L2 algorithm with synthetic spectra from a 1-D radiative transfer model showed that low clouds can be difficult to detect in prescreening. Quality control filters can be applied to the retrieved state information to screen out cloud contaminated retrievals, which mitigate the resulting errors in the retrieved XCO₂.

In this study, we perform new tests of the L2 algorithm using simulated OCO-2 observations from a 3-D radiative transfer model, SHDOM. We focus on an idealized scene containing a single, spatially unresolved boundary layer cloud. These synthetic observations are processed with the L2 prescreening and retrieval algorithms. Our results show that in many cases, existing prescreening methods are effective at detecting the unresolved cloud. The prescreening is less effective in cases when the observation is primarily affected by the cloud shadow. Quality control filters applied to the retrieved states remain effective at mitigating the cloud contamination. In cases where the observation is not screened or filtered by quality control, the errors in the retrieved XCO₂ are relatively low. However, other retrieved state variables, such as aerosol optical depth and surface albedo, show larger errors.