Uncertainty-aware Machine Learning Bias Correction and Filtering for OCO-2 | 2014-2024

Overview

This is a dataset to explore the effect of applying a new bias correction and quality filtering approach to increase the accuracy of atmospheric CO2 measurements derived from the Orbiting Carbon Observatory-2 (OCO-2) satellite. This is not an official OCO-2 data product.

Data Description

The dataset contains OCO-2 retrieved XCO2 B11.2 that have been corrected and filtered with a new machine learning approach from 2014 to the end of 2024. There is one file for each year. The following variables are contained in the files:

sounding_id, xco2_ML*, xco2, xco2_x2019, xco2_quality_flag_ML*, xco2_quality_flag, bias_correction_uncert_ML*, xco2_uncertainty, latitude, longitude, time, land_water_indicator, operation_mode

* new variables that contain the new bias coorected xco2, quality flag, and uncertainty and are not contained in the official OCO-2 Lite Files.

xco2_ML: XCO2 Machine Learning corrected XCO2 on x2019 scale

xco2_quality_flag_ML: XCO2 ternary quality flag:  0 = best quality data, 1 = good quality data for increasing sounding throughput if needed, 2 = poor quality data (do not use)

bias_correction_uncert_ML: XCO2 bias correction uncertainty

For the full set of variables contained in the LiteFiles and description of each variable please refere to the data user guide of the official OCO-2 lite files: https://disc.gsfc.nasa.gov/datasets/OCO2_L2_Lite_FP_11.2r/summary?keywords=oco2

Current Bias Correction Approach

The operational bias correction method uses a multiple linear regression-like approach to adjust errors in XCO2 relative to elements of the state vector derived from the ACOS retrieval. Adjustments are currently made manually, guided by various success metrics like TCCON observation agreement, retrieval variability reduction, and flux model coherence.

New Bias Correction Approach

1. Computational Optimization: Replaces manual tuning with computational methods, enhancing transparency, traceability, and reproducibility.2. Non-linear Error Modeling: Allows more flexibility in modeling retrieval errors, reducing biases, particularly in previously unusable data.3. Independent Flux Inversion Models: Excludes flux inversion models from bias correction development, maintaining OCO-2 measurement independence.4. Quantified Correction Uncertainties: Provides uncertainty quantification for each bias correction at the per sounding level.

Data Usage

If you find anything unexpected in the data please report your findings to [e-mail] and [e-mail] so we can resolve any issues.

Additional Resources and Citation

Two preprints are currently available that describe the approach in detail and should be cited if the data is used. We will update the citations as soon as the papers are published:

https://doi.org/10.22541/essoar.174164198.80749970/v1https://doi.org/10.22541/essoar.174164203.37422284/v1

Copyright statement: © 2023 California Institute of Technology. Government sponsorship acknowledged.