A nonlinear data-driven approach to bias correction of XCO₂ for NASA's OCO-2 ACOS version 10

Abstract

Measurements of column-averaged dry air mole fraction of CO2 (termed XCO2) from the Orbiting Carbon Observatory-2 (OCO-2) contain systematic errors and regional-scale biases, often induced by forward model error or nonlinearity in the retrieval. Operationally, these biases are corrected for by a multiple linear regression model fit to co-retrieved variables that are highly correlated with XCO2 error. The operational bias correction is fit in tandem with a hand-tuned quality filter which limits error variance and reduces the regime of interaction between state variables and error to one that is largely linear. While the operational correction and filter are successful in reducing biases in retrievals, they do not allow for throughput or correction of data in which biases become nonlinear in predictors or features. In this paper, we demonstrate a clear improvement in the reduction in error variance over the operational correction by using a set of nonlinear machine learning models, one for land and one for ocean soundings. We further illustrate how the operational quality filter can be relaxed when used in conjunction with a nonlinear bias correction, which allows for an increase in sounding throughput by 14 % while maintaining the residual error in the operational correction. The method can readily be applied to future Atmospheric CO2 Observations from Space (ACOS) algorithm updates, to OCO-2's companion instrument OCO-3, and to other retrieved atmospheric state variables of interest.