Uncertainty-Aware Machine Learning Bias Correction and Filtering for OCO-2: 1

Abstract

The Orbiting Carbon Observatory-2 (OCO-2) makes space-based radiance measurements of reflected sunlight. Using a physics-based retrieval algorithm, these measurements are inverted to estimate column-averaged atmospheric carbon dioxide dry-air mole fractions (XCO2). However, biases are present in the retrieved XCO2 due to sensor calibration errors and discrepancies between the physics-based retrieval and nature. We propose a Random Forest (RF), a non-linear, interpretable machine learning (ML) technique, to correct these biases. The approach is rigorously validated, comes with quantified uncertainties, and is derived independent of carbon flux models. Compared to the operational approach, our method reduces unphysical variability over land and ocean and shows closer agreement with independent ground-based observations from the Total Carbon Column Observing Network. The RF-bias correction is suitable for integration into the operational processing pipeline for the next version of OCO-2 products, pending additional testing and validation. It is inherently generalizable to other existing and planned greenhouse gas monitoring missions. This paper (Part 1) describes the RF bias correction, while a second paper (Part 2) describes the development of a data filtering strategy specifically designed for a subset of retrievals exhibiting irreducible errors that remain inadequately corrected by the ML bias correction.