Improved Ensemble-Based Satellite-Derived XCO₂ and XCH₄ Climate Data Records Using EMMA v5.1

Abstract. Satellite observations of the column-averaged dry-air mole fractions of carbon dioxide (XCO₂) and methane (XCH₄) are essential for monitoring their global atmospheric distribution and variability, as well as for related climate applications. To generate long-term, consistent Climate Data Records (CDR), ensemble merging approaches such as the ensemble median algorithm (EMMA) have been used.

Here, we introduce a new version of the EMMA algorithm, v5.1, and the four merged datasets that were created with it, namely merged Level 2 (L2) and corresponding gridded Level 3 (L3) products for both XCO₂ and XCH₄. EMMA v5.1 keeps the core idea from previous EMMA versions of the ensemble median being a good choice for merging multiple satellite data products, while taking a new approach: for each gas and time period, one of the L2 input products is used as a reference, and the ensemble is used to compute an empirical spatio-temporal adjustment for this input dataset. This new design eliminates conceptual limitations of former EMMA versions, making the L2 dataset eligible for hotspots emission estimations in periods in which the reference product has sufficient spatial coverage. Additionally, new and updated input L2 products were included in the ensemble, and the L3 datasets have higher spatial resolution, i.e. monthly 1°x1°, compared to the previous monthly 5°x 5° resolution. The four merged datasets cover the time period from 2003 to 2024.

All four merged datasets were validated against ground-based Total Carbon Column Observing Network (TCCON) retrievals, showing good agreement. The XCO₂ L2 (L3) merged datasets have a spatio-temporal bias of 0.47 ppm (0.48 ppm). The estimated single-sounding (single-cell) random error is 1.51 ppm (1.82 ppm), very close to the average reported uncertainty of 1.66 ppm (1.65 ppm). The estimated linear drift is 0.00 ± 0.12 ppm y⁻¹ (0.00 ± 0.11 ppm y⁻¹).

The XCH₄ L2 (L3) merged datasets have a spatio-temporal bias of 7.37 ppb (4.31 ppb). The estimated single-sounding (single-cell) random error is 20.87 ppb (9.01 ppb), very close to the average reported uncertainty of 21.11 ppb (11.05 ppb). The estimated linear drift is 0.42 ± 1.14 ppb y⁻¹ (0.11 ± 0.62 ppb y⁻¹).