Uncertainty-Aware Machine Learning Bias Correction and Filtering for OCO-2. 2

Quality filtering of satellite XCO₂ retrievals is essential for every downstream science application, yet the community still grapples with the trade-off between retaining data availability and suppressing biases. For the Orbiting Carbon Observatory-2 (OCO-2) record in particular, the long-standing binary quality flag targets the single use case of global carbon flux inversion and is often too restrictive for small spatial scale analysis like quantifying emissions from coal-fired power plants. To address the need for flexible quality filtering we introduce a data-driven, ternary (three-state) quality flag constructed from the agreement of three independent sub-filters: two Random Forest classifiers trained on distinct “truth proxy” data sets and a third sub-filter based on the uncertainty estimate from a non-linear machine learning bias correction and the operational uncertainty product. We utilized a Bayesian multi-objective optimization to tune the sub-filters, balancing the competing goals of maximizing data throughput with minimizing error variance and retrieval uncertainty. The proposed ternary quality flag shows an improved reduction in root mean square error (RMSE) (22% for land and 53% for ocean) over the operational flag. This reduction is due in part to an improved ability to remove observations affected by 3D cloud biases. The flexible ternary flag can optionally increase data availability by 21% over land and 18% over ocean but is still competitive with the RMSE of the operational product. The proposed filter addresses the diverse needs of the science community and is generalizable to greenhouse gas monitoring missions such as GOSAT, CO2M and Carbon-I.