TCCON is a global network of ground-based Fourier transform spectrometers (FTSs) that record direct solar spectra in the near-infrared (NIR) region (Wunch et al., 2011). TCCON serves as a benchmark dataset for validating satellite-derived XCH4 because its measurements are minimally affected by aerosol loading, airmass uncertainty, and surface-reflectance variability compared with SWIR satellite retrievals (Dils et al., 2014; Malina et al., 2022). We obtained the latest TCCON GGG2020 dataset from the TCCON data archive hosted by CaltechDATA (https://tccondata.org/, last access: 26 December 2025) spanning January 2020 to December 2023, covering the overlapping observation period of TROPOMI, GOSAT, and GOSAT-2. Stations used in this study are shown in Fig. 1 and listed in Table 1.

This study used XCH4 products from TROPOMI, GOSAT, and GOSAT-2. These satellite platforms have been widely used for regional and global GHG monitoring, emission estimation, and multi-sensor XCH4 data integration (Balasus et al., 2023; Fan et al., 2024; Bao et al., 2025; Wang et al., 2025; Liu et al., 2026; Lee et al., 2025: Choi and Jeong, 2026). All three missions retrieve XCH4 using full-physics algorithms that simultaneously estimate surface albedo, atmospheric scattering, and trace-gas concentrations from SWIR and NIR spectra (Butz et al., 2011). However, full-physics retrievals are sensitive to clouds, aerosols, and surface heterogeneity, increasing retrieval bias and reducing data availability in high-aerosol regions and complex terrain (Jacob et al., 2022; Umezawa et al., 2025). GOSAT and GOSAT-2 provide high-precision measurements with spectral resolution approximately four times finer than that of TROPOMI (Kuze et al., 2009, 2016; Suto et al., 2021; Jacob et al., 2022; Noël et al., 2021; Table 2), whereas TROPOMI offers daily global mapping, which increases data density despite higher susceptibility to atmospheric interference (Hu et al., 2018; Qu et al.,2021). These contrasting characteristics reflect an inherent trade-off between sampling density and retrieval accuracy. Table 2 summarizes mission specifications, and Fig. 2 illustrates daily sampling patterns for 1 January 2020.

Reliable colocation between satellite retrievals and reference observations is essential for bias correction and validation. The sparse global distribution of TCCON sites prevents perfectly coincident measurements (Zhou et al., 2016). We therefore constructed collocated training pairs using satellite-specific spatiotemporal and elevation constraints, adopting criteria from previous studies and official validation strategies (Balasus et al., 2023; Yoshida et al., 2023). Using the criteria in Table 3, we first established Step 1 between each satellite and TCCON. To ensure adequate temporal representation across 2020–2023 and sufficient training samples, TCCON site with fewer than 20 collocated observations (i.e., eureka01) was excluded. We then constructed Step 2 inter-satellite colocations, pairing GOSAT and TROPOMI observations with TCCON calibrated GOSAT-2 using the thresholds in Table 3.

While GOSAT and TROPOMI provide official bias-corrected products (Inoue et al., 2016; Lorente et al., 2021), GOSAT-2 does not. To treat all sensors consistently, we developed a unified ML framework using TCCON as the calibration standard. Previous studies improved TROPOMI XCH4 by aligning to GOSAT official (Fan et al., 2024) or GOSAT proxy data (Balasus et al., 2023; Li et al., 2024), which applies a single global offset to force mean bias to zero and is available only through 2021 (Parker et al., 2020). Our approach directly calibrates each sensor against TCCON, enabling spatially and temporally adaptive bias correction throughout 2020–2023 and data-driven harmonization reference selection.

The fusion step integrates harmonized satellite datasets into a single daily 0.1° gridded product. We adopted a hierarchical priority-based fusion strategy (Chen et al., 2024), ranking datasets by their validation performance against TCCON and inter-sensor consistency. Because GOSAT-2 was selected as the harmonization reference and exhibits the strongest independent agreement with TCCON, it was assigned the highest priority. Based on the Step-2 evaluation, the final ranking used for fusion was: GOSAT-2(BC)→TROPOMI(GOSAT-2-like)→GOSAT(GOSAT-2-like)

For each day and each 0.1° grid cell (x,y), the Fused product was defined as: 6XCH4,Fused(x,y)=XCH4,GOSAT-2(x,y),if availableXCH4,TROPOMI(x,y),else if availableXCH4,GOSAT(x,y),else if availableNaN,otherwise

This rule retains the highest-priority valid observation at each location and time, thereby improving both spatial coverage and precision of the Fused daily XCH4 product.

The spatial and temporal generalization performance of the BC was quantitatively evaluated using step-specific validation strategies. In Step 1, spatial generalization was assessed using leave-one-site-out cross-validation (LOSOCV) to account for site-to-site differences in background XCH4 levels and spatial heterogeneity. Temporal generalization was evaluated using leave-one-month-out cross-validation (LOMOCV) and leave-one-year-out cross-validation (LOYOCV) to account for interannual variability and growth trends (Jang et al., 2024). In Step 2, leave-one-band-out cross-validation (LOBOCV) was applied across 11 latitudinal zones (60° S–80° N; Fig. S1 in the Supplement) to account for latitudinal gradients in CH4 emissions (Maasakkers et al. 2019; Hwang et al., 2026), with temporal validation using LOMOCV and LOYOCV. Model performance was evaluated using three standard metrics: coefficient of determination (R2), root mean square error (RMSE) and mean absolute error (MAE) (Eqs. S1–S3 in the Supplement).

We evaluated the spatiotemporal generalization performance of official satellite products and ML-based bias-correction against TCCON. In Fig. 4, LOSOCV results and TCCON observations were aggregated at each TCCON site to examine site-level mean bias, within-site variability, and station-to-station consistency. The standard products systematically deviated from the 1:1 line, with pronounced negative biases at low-XCH4 sites such as darwin01, lauder03, and sodankyla01 (Fig. 4a, d, and g; see Fig. 1 for site locations). These sites are located in the Southern Hemisphere or high latitudes, where challenging retrieval conditions (e.g., high solar zenith angle and high surface albedo) and limited TCCON sampling can amplify errors. In addition, operational bias corrections tuned around typical global XCH4 levels (≈ 1850–1900 ppb) can leave larger residual deviations at the low and high tails.

In contrast, the ML-based correction tightened site-level agreement for all sensors, reducing both RMSE and MAE. This correction used sensor-specific models selected based on CV performance (Table S5; XGBoost for TROPOMI, RF for GOSAT and GOSAT-2). For TROPOMI, the RMSE decreased from 14.26 ppb (standard) and 5.49 ppb (operational bias-corrected) to 3.79 ppb (ML), corresponding to reduction of 73 % and 31 %, respectively (Fig. 4a–c). For GOSAT, RMSE reductions of 54 % and 45 % were achieved relative to the operational standard and bias-corrected products, respectively, while GOSAT-2 showed a 41 % reduction relative to the operational standard (Fig. 4d–h). Site-wise error statistics (Figs. S2–S4) confirmed these improvements, with ML-based correction reducing both station-wise mean bias (μ) and variability (σ), indicating more uniform performance under spatially heterogeneous conditions. Given a global XCH4 growth rate of ∼ 8–14 ppbyr-1 during the study period (NOAA GML; https://gml.noaa.gov/ccgg/trends_ch4/, last access: 26 December 2025), these reductions are large enough to influence long-term trend interpretation.

The LOMOCV results further highlighted differences in temporal generalization (Fig. 5). Across sensors, the standard products showed negative biases and broad interquartile ranges (IQRs) throughout the year, with strong negative shifts from June to September. The seasonal aggregation confirmed that this negative bias was most pronounced in June–July–August (JJA). This seasonal pattern is consistent with known sensitivities of SWIR-based XCH4 retrievals to surface albedo and atmospheric scattering by aerosols and cirrus (Inoue et al., 2016; Hu et al., 2016; Lorente et al., 2021; Oshio et al., 2020). These factors can modify the effective light path and vary seasonally with vegetation, humidity, cloud conditions, and solar geometry. Operational bias correction shifted the median bias toward zero but still exhibited larger seasonal dispersion than the ML-based correction. The ML-based correction maintained seasonal medians near zero and reduced both the IQRs and non-outlier spread, with most IQRs remaining within ± 8 ppb, indicating more stable performance under seasonal variability.

Figure 6 further illustrates condition-dependent biases by relating the XCH4 ratio (TCCON/satellite) to key retrieval variables. For TROPOMI, the standard product shows a strong negative dependence on SWIR surface albedo (R = -0.49; Fig. 6a), consistent with known sensitivity to scattering-related light-path errors under low-albedo conditions (Lorente et al., 2021; Balasus et al., 2023; Li et al., 2024). Bias correction substantially reduces this dependence (Fig. 6b), yielding near-zero correlation (R = -0.02) and indicating that albedo-driven bias is largely mitigated. For the GOSAT series, operational bias correction is implemented as a multiple linear regression using explanatory variables (e.g., ΔPs, XH2O, airmass, and ΔT) (Inoue et al., 2016; Yoshida et al., 2023). Consistently, ΔPs dependence is weak for GOSAT (R=0.22; Fig. 6c) and more evident for GOSAT-2 (R=0.35; Fig. 6e) in the standard products, and it is reduced after correction (Fig. 6d and f). Overall, the ML-based correction reduces not only mean bias but also bias sensitivity to key retrieval conditions across sensors.

Multi-sensor fusion requires minimizing inter-sensor inconsistencies by defining a common reference scale. Table 4 compares the ML-based bias-corrected products from all three sensors evaluated against TCCON using common collocated samples under LOSOCV, ensuring that the reference selection is based on independent out-of-site performance. GOSAT-2 achieved the highest accuracy (R2 = 0.67; RMSE = 10.67 ppb) and was therefore selected as harmonization reference.

To improve cross-sensor consistency, the GOSAT and TROPOMI products were calibrated to the ML-based bias-corrected GOSAT-2 scale (i.e., generating “GOSAT-2-like” XCH4). Across the evaluated algorithms, XGBoost yielded the best harmonization performance (Table S6). Figure 7 shows distributional shifts before and after harmonization: after harmonization, bias distributions for GOSAT and TROPOMI moved toward near-zero mean bias relative to GOSAT-2 and their spread was reduced. This inter-sensor harmonization further assessed the major retrieval-parameter dependencies identified in the TCCON-based analysis over broader parameter ranges (Fig. 8). After ML-based harmonization, the SWIR surface albedo dependency of TROPOMI and the ΔPs dependency of GOSAT were nearly removed, indicating improved consistency with the ML-based bias-corrected GOSAT-2 scale.

After harmonization, TROPOMI exhibited slightly higher agreement with GOSAT-2 (Table S6) than GOSAT and provided approximately 6 times more collocated observations (183 550 for TROPOMI vs. 32 244 for GOSAT). Accordingly, the fusion priority was set to GOSAT-2 → TROPOMI → GOSAT to balance reference consistency and sampling density and to enhance robustness of the Fused XCH4 product. Finally, we generated a daily 0.1° Fused XCH4 product by integrating the harmonized observations from the three satellites.

Figure 9 evaluates the Fused XCH4 product at three independent TCCON sites (xianghe01, edwards01, and garmisch01) and includes a comparison with the Blended TROPOMI + GOSAT XCH4 product of Balasus et al. (2023). These sites were excluded from the TCCON-based bias correction in Step 1 to assess generalization under independent conditions. The three stations represent distinct environments: a peri-urban site influenced by strong anthropogenic emissions (Xianghe, China; mean = 1903.16 ppb), an arid high-desert site with high surface albedo (Edwards, USA; mean = 1880.37 ppb), and a mountainous site (Garmisch, Germany; mean = 1872.98 ppb). Across all sites, the Fused XCH4 showed strong agreement with TCCON (R2 = 0.81, RMSE = 10.78 ppb; Fig. 9a). The time series further indicates that the Fused product reproduces temporal variability, while reducing inter-sensor offsets relative to the individual satellite products (Fig. 9b). For 2022, the Fused product achieved higher accuracy than the Blended XCH4 product at the same locations, (R2 = 0.74, RMSE = 10.38 ppb vs. R2 = 0.67, RMSE = 11.82 ppb; Fig. 9c), which may reflect the use of a harmonized multi-sensor framework that incorporates the higher-accuracy GOSAT-2 product.

The Fused XCH4 product achieved annual global land coverage of 87.1 %, 86.9 %, 86.8 %, and 88.1 % for 2020–2023, corresponding to improvements of 0.45–0.74 percentage points over TROPOMI alone (Fig. S5a). Although global annual averages show modest gains, multi-sensor fusion yields substantial regional enhancements where individual sensors face systematic retrieval limitations.

To quantify fusion benefits under challenging retrieval conditions for TROPOMI, we examined five case-study regions (ROIs; see Fig. 1) in 2023 representing diverse environments and emission regimes: South Asia (ROI-A) and the Amazon Basin (ROI-B) as cloudy tropical regions with aerosol impacts; Eastern Siberia (ROI-C) as high-latitude wetlands with seasonal snow/ice constraints; the Pacific Northwest (ROI-D) as complex coastal terrain; and Central Europe (ROI-E) as Alpine topography (Fig. 10). Compared with TROPOMI, coverage improvements ranged from ∼ 1.8 % to ∼ 9.6 %, with the largest gain in South Asia (Fig. 10a). Source attribution maps illustrate how the GOSAT series spatially complements TROPOMI coverage in 2023 (Fig. 10b): GOSAT-2 provides the majority of additional observations, particularly filling monsoon-related gaps in South Asia and cloudy tropical regions, while GOSAT supplements coverage over Alpine terrain, where complex topography challenges both TROPOMI and GOSAT-2 retrievals.

Fusion also improves temporal resilience during periods of single-mission data loss. Missing VIIRS cloud-screening information in July–August 2022 and August 2023 (Borsdorff et al., 2024) led to substantial gaps in TROPOMI XCH4, which were partially filled by overlapping GOSAT and GOSAT-2 observations (Fig. S5b). This multi-sensor complementarity highlights the importance of global-scale fusion for comprehensive CH4 monitoring.

The Fused dataset reveals distinct spatiotemporal patterns across 2020–2023. Time–latitude diagrams (Fig. 11a) show a persistent interhemispheric gradient of 30–40 ppb between the Northern Hemisphere (NH) and Southern Hemisphere (SH). Elevated XCH4 is consistently observed in the northern subtropical belt (0–40° N), where mean concentrations increased from 1872.72 ppb to 1906.98 ppb over the study period (+34.26 ppb), exceeding the global mean increase. While the highest concentrations remained in the north, the Fused product captured a progressive southward expansion in XCH4 increases beginning in 2021, consistent with recent findings (Umezawa et al., 2025).

To quantify global growth rates, we compared the monthly global average of the Fused XCH4 dataset with independent observations from TCCON ground-based XCH4 measurements and NOAA marine surface CH4 records (https://gml.noaa.gov/ccgg/trends_ch4/, last access: 26 December 2025). Mean annual growth rates were calculated from the monthly time series over 2020–2023, while uncertainties represent the variability of monthly growth rates within each year. The Fused product captures a sustained increase from approximately 1850 ppb in early 2020 to nearly 1900 ppb by late 2023, corresponding to a mean global growth rate of 12.28 ± 1.00 ppbyr-1 (Fig. 11b). This agrees well with TCCON (13.56 ± 1.25 ppbyr-1) and NOAA (14.77 ± 0.64 ppbyr-1), supporting the ability of the Fused dataset to represent large-scale atmospheric CH4 variability and interannual trends.

Interannual growth rates exhibited strong regional variability (Fig. 11c). The NH (0–80° N), SH (60° S–0), and the high-emission zone (0–40° N) all peaked in 2021 at 14.3 ± 2.2, 15.4 ± 2.7, and 15.5 ± 2.1 ppbyr-1, respectively, consistent with record global XCH4 increases reported for that period (Saunois et al., 2025). Following this peak, growth rates generally declined during 2022–2023, reaching approximately 6.0 ± 1.8 ppbyr-1 in the NH, 11.4 ± 3.9 ppbyr-1 in the SH, and 6.4 ± 3.0 ppbyr-1 in the high-emission zone by 2023, consistent with recent inverse modeling estimates (Pendergrass et al., 2025).

To highlight emission hotspots relative to the global baseline, we computed annual anomalies by subtracting the 4-year mean (1869.95 ppb) from each year's XCH4 field (Fig. 12). By 2023, positive anomalies exceeding +60 ppb were apparent over South Asia, East Asia, and Central Africa, with maxima over the Indo-Gangetic Plain and the Ganges–Brahmaputra delta (Fig. 12d). The shift in global mean anomaly from -16.74 ppb (2020) to +14.16 ppb (2023) indicates increasing XCH4 accumulation over the study period.

To interpret drivers of elevated concentrations, we examined high-anomaly regions in 2023 using MODIS MCD12Q1 land-cover information (Fig. 13). Hotspots showed distinct land-cover signatures: the Niger Delta (Fig. 13a) exhibited strong anomalies over a mix of croplands, evergreen broadleaf forests, coastal wetlands, and urban areas associated with petroleum infrastructure (Olukaejire et al., 2024); the Ganges–Brahmaputra Delta (Fig. 13b) showed the largest anomaly over rice paddies and permanent wetlands; the Indo-Gangetic Plain (Fig. 13c) and Yangtze River Delta (Fig. 13d) displayed persistent enhancements over irrigated croplands and expanding metropolitan areas; the Congo Basin (Fig. 13e) showed elevated concentrations over tropical peatlands and evergreen forests; and the southern United States (Fig. 13f) showed enhancements over the Barnett Shale region in Texas, consistent with documented oil-and-gas emissions (Harriss et al., 2015; Lan et al., 2015). The spatial correspondence between enhanced XCH4 and wetland/cropland/urban land classes demonstrates the Fused product's ability to resolve regional emission patterns.

Finally, we examined seasonal patterns of global XCH4 in 2023. To preserve localized XCH4 enhancements, zonal-mean profiles were calculated at native 0.1° latitude resolution (Fig. 14), and seasonal statistics were summarized for major latitude zones (Table S7). The NH (0–80° N) exhibited the largest seasonal amplitude (32.14 ppb), peaking in winter (DJF; 1915.36 ppb) and reaching a minimum in spring (MAM; 1883.22 ppb). Within the NH, the high-emission zone (0–40° N) remained consistently elevated throughout the year, ranging from approximately 1901 ppb in JJA to approximately 1918 ppb in SON/DJF, with a moderate seasonal amplitude of 16.50 ppb. In contrast, the SH (60° S–0) showed a smaller amplitude (6.77 ppb), varying between 1850.66 ppb in MAM and 1857.43 ppb in JJA. The pronounced NH seasonality reflects strong source variability from wetland and agricultural emissions, whereas the smoother SH cycle is more strongly governed by the OH sink than by local emissions (East et al., 2024).

While the framework substantially improves retrieval accuracy across sensors, several limitations should be noted. First, the ML-based bias correction relies on TCCON as the primary reference dataset. Although TCCON provides high-precision ground-based XCH4 observations, its spatial distribution is limited and does not fully represent the diverse retrieval conditions encountered globally, particularly high-surface-albedo conditions. We partly addressed this limitation through LOSOCV strategy and additional satellite match-up analyses over broader surface albedo ranges, but independent validation under underrepresented retrieval conditions remains necessary. Second, the Fused dataset uses a fixed sensor-priority strategy (GOSAT-2 → TROPOMI → GOSAT) based on overall accuracy and sampling density, which may introduce regional bias because local observing conditions can favor a non-prioritized sensor. Third, although the fusion framework improves robustness, global annual coverage gains are modest (0.45–0.74 percentage points over TROPOMI). Fourth, ML-based correction in this study relies primarily on retrieval parameters to model systematic biases; incorporating external environmental covariates (e.g., temperature, humidity, and wind) may further reduce regional biases and improve generalization (Qu et al., 2025). Finally, the 0.1° resolution supports regional-scale analyses but remains insufficient for resolving fine-scale emission sources detectable by emerging high-resolution missions (e.g., EMIT or GHGSat). Future work could address these limitations via adaptive sensor weighting, uncertainty quantification, and integration of additional environmental covariates.