Improving Satellite XCO₂ Measurements Accuracy: A Bayesian Bias Correction Approach Considering Spatiotemporal Bias Characteristics

Measurements of column-averaged dry air mole fraction of CO2 (XCO2) from satellite contain systematic errors and regional scale biases are often induced by the limitations of the retrieval algorithm. Although the global linear bias correction (BC) model has successfully reduced some of these systematic errors, the unified correction formula and filter have proven too restrictive for certain regions, resulting in high remaining systematic errors. We propose a BC method through Bayesian optimal estimation considering spatiotemporal bias characteristics (ST-BCs) to address this issue. The prior distribution is provided by the global flux model, and the corrected samples are calculated by: 1) training a nonlinear model of satellite correction parameters and truth proxy based on XGBoost machine learning; 2) adjusting spatial characteristics bias based on spatial correlation of systematic errors; and 3) using the total carbon column observation network (TCCON) long-term measurement trend to correct temporal characteristics bias. The posterior correction value is iteratively calculated through Bayesian optimal estimation. We conducted experiments using OCO-2 V10 retrieval in the study areas in Europe, Asia, and North America, where TCCON coverage is extensive. The results demonstrated that the method effectively reduces regional bias and also significantly diminishes the correlation between state parameters and bias. The application of ST-BC enhances the correction accuracy of ACOS V10 products from 1.07 to 0.88 ppm. Particularly, the method reduces bias by 0.78 ppm for complex summer measurements. This study improves the accuracy of satellite retrieval XCO2 and will advance our understanding of the global and regional carbon sources and sinks.