Data Compression to Improve Inverse Modeling with Large Satellite Datasets：An Example from OCO-2

The number of greenhouse gas observing satellites has greatly expanded in the past five years, and these new satellite datasets could provide an unprecedented constraint on global CO₂ and CH₄ fluxes. However, a continuing challenge for satellite-based inverse modeling is the sheer size of many satellite datasets. The large number of observations often makes it prohibitive to run trajectory models (e.g., STILT or FLEXPART), save the Jacobian or sensitivity matrix, or implement some inverse modeling calculations (e.g., calculate posterior uncertainties). Many satellite datasets of atmospheric gases are very large, but the information content of any single total column observation can be modest due to column averaging and due to retrieval errors. What's more, the information content of each retrieval often has partial redundancy with neighboring retrievals. These facts indicate that many satellite trace gas datasets may be compressible to much smaller size without losing information about the underlying fluxes or emissions. In this study, we develop a satellite-based data compression approach based on geostatistical principles. We subsequently tune and evaluate the data compression approach using both synthetic and real CO₂ observations from NASA's Orbiting Carbon Observatory-2 (OCO-2). We compare CO₂ sources and sinks estimated from the original and compressed satellite datasets using a geostatistical inverse modeling approach. We can compress the satellite dataset to an average of one observation per 50 km without compromising the flux estimate, but we find that compressing the data further has a noticeable, detrimental impact on the estimated fluxes. We also find that data compression can efficiently reduce the size and computational burden of inverse modeling with OCO-2. Furthermore, the approach developed here could be applied broadly to a range of inverse problems that use very large trace gas datasets.