Data Reduction For Inverse Modeling: An Adaptive Approach V1.0

The number of greenhouse gas (GHG) observing satellites has greatly expanded in recent years, and these new datasets provide an unprecedented constraint on global GHG sources and sinks. However, a continuing challenge for inverse models that are used to estimate these sources and sinks is the sheer number of satellite observations, sometimes in the millions per day. These massive datasets often make it prohibitive to implement inverse modeling calculations and/or assimilate the observations using many types of atmospheric models. Although these satellite datasets are very large, the information content of any single observation is often modest and non-exclusive due to redundancy with neighboring observations and due to measurement noise. In this study, we develop an adaptive approach to reduce the size of satellite datasets using geostatistics. A guiding principle is to reduce the data more in regions with little variability in the observations and less in regions with high variability. We subsequently tune and evaluate the approach using synthetic and real data case studies for North America from NASA's Orbiting Carbon Observatory-2 (OCO-2) satellite. The proposed approach to data reduction yields more accurate CO2 flux estimates than the commonly used method of binning and averaging the satellite data. We further develop a metric for choosing a level of data reduction; we can reduce the satellite dataset to an average of one observation per similar to 80-140 km for the specific case studies here without substantially compromising the flux estimate, but we find that reducing the data further quickly degrades the accuracy of the estimated fluxes. Overall, the approach developed here could be applied to a range of inverse problems that use very large trace gas datasets.