Machine Learning-Based Swath Bias Correction for OCO-3 Snapshot Area Mapping Mode Observations: 2019 - 2025

We trained a random forest classifier on 1,723 manually labeled OCO-3 SAMs spanning August 2019 to February 2025 to distinguish scenes requiring bias correction from those with legitimate spatial gradients in XCO2. The model uses five key diagnostic features identified through systematic feature selection: inter-swath XCO2 discontinuity, scene homogeneity, aerosol optical depths, and spectral alignment corrections. A median-alignment algorithm then removes detected biases by adjusting swath-level offsets while conserving total column CO2. This data set is the result of this approach further detailed in a forth coming publication. All corrected NetCDF files retain a small subset of the original OCO-3 Level 2 Lite variables (sounding_id, latitude, longitude, time, Sounding/operation_mode, Sounding/orbit, Sounding/target_id, xco2, xco2_quality_flag) and metadata, with the addition of two new fields: xco2_swath_bc containing the bias-corrected XCO2 retrievals and swath_bias_corrected providing a binary flag indicating whether swath bias correction was applied (0 = no change, 1 = correction applied).