Enhanced methane monitoring: a globally harmonized daily 0.1° XCH₄ through machine learning-based fusion of GOSAT, GOSAT-2, and TROPOMI

Accurate global monitoring of atmospheric methane (CH₄) is essential for tracking progress toward climate mitigation targets such as the Global Methane Pledge (GMP). Ground-based measurement networks are too sparse to provide sufficient spatial coverage, while satellite-derived retrievals are hindered by systematic biases and uncertainties, limiting their reliability for consistent global monitoring. We present the first global fusion of GOSAT, GOSAT-2, and TROPOMI to generate a globally consistent daily 0.1° land dataset for 2020–2023 for enhanced global column-averaged dry-air mole fraction of atmospheric methane (XCH₄) mapping. The framework employs a three-step machine-learning (ML) approach: (1) sensor-specific bias correction using TCCON observations, (2) cross-sensor harmonization to GOSAT-2, the sensor with the strongest post-correction TCCON agreement, and (3) priority-based fusion. Tree-based ensemble regressors were trained with satellite retrieval parameters to reduce systematic biases and inter-sensor discrepancies. Independent validation at three withheld TCCON stations demonstrates robust generalization of the Fused product (R² = 0.81, RMSE = 10.78 ppb), outperforming standard and operational bias-corrected satellite products and previously reported ML-based approaches. Regional assessments show that fusion substantially improves data availability and reduces systematic errors, delivering up to 9.5 % relative coverage gains compared to TROPOMI operational products in challenging regions (South Asia, Amazon Basin, Eastern Siberia). The Fused dataset reveals intensifying positive XCH₄ anomalies (+60 ppb) over South Asia, East Asia, and Central Africa during 2020–2023, linked to MODIS-derived agricultural and urban land classes as well as known oil and gas fields. The dataset provides a scalable resource for regional CH₄ emissions assessment and continuous monitoring, with the framework extendable to upcoming satellite missions (GOSAT-GW, CO2M) for long-term GMP progress tracking.

1 Introduction

Methane (CH₄) is a potent greenhouse gas (GHG), contributing approximately 0.6 °C of global warming since the pre-industrial era and accounting for nearly 30 % of the observed increase in global mean surface temperature (Naik et al., 2021; IEA, 2025). Despite lower atmospheric concentrations than CO₂, its global warming potential is 28 times higher over a 100-year horizon (Myhre et al., 2013; IPCC, 2014). Its high radiative efficiency, combined with a relatively short atmospheric lifetime due to oxidation by hydroxyl radicals (OH), makes CH₄ an especially effective target for near-term climate mitigation (Prather et al., 2012; Nisbet et al., 2020). Recent observations show that CH₄ levels have been rising at an unprecedented rate, with 2020–2022 marking the fastest growth since systematic monitoring began (Saunois et al., 2025). Major CH₄ sources include anthropogenic activities (livestock farming, rice cultivation, fossil-fuel exploitation, and waste management) and natural wetland emissions (Saunois et al., 2025). Reflecting the urgency of mitigation, the Global Methane Pledge (GMP), launched at the 26th Conference of the Parties (COP26), aims to reduce global CH₄ emissions by at least 30 % from 2020 levels by 2030 (IEA, 2022). Accurate quantification of the column-averaged dry-air mole fraction of atmospheric methane (XCH₄) is therefore essential for identifying emission sources and evaluating progress toward climate goals.

Ground-based monitoring systems such as the Total Carbon Column Observing Network (TCCON) provide high-precision XCH₄ measurements and serve as key reference data for satellite validation and bias correction. Despite their accuracy, TCCON sites are limited in number and spatially sparse (Lorente et al., 2021), leaving large observational gaps across Africa, South America, and large parts of Asia. Satellite platforms help address these limitations by providing broad spatial coverage. The Greenhouse gases Observing SATellite (GOSAT), GOSAT-2, and the TROPOspheric Monitoring Instrument (TROPOMI) have enabled global-scale mapping of XCH₄ and have been widely used for emission estimation and inverse modeling (Qu et al., 2021; Worden et al., 2022; Nesser et al., 2024; Maasakkers et al., 2019; Janardanan et al., 2020). However, these instruments exhibit distinct trade-offs in spatial resolution, revisit frequency, and retrieval accuracy. GOSAT and GOSAT-2 provide high spectral resolution (0.06 nm) but limited spatial and temporal coverage with ∼ 10 km circular footprints and multi-day revisit intervals (Kuze et al., 2016; Suto et al., 2021). In contrast, TROPOMI offers near-daily global coverage at finer spatial resolution (5.5 km × 7 km) but suffers from larger retrieval uncertainties due to lower spectral sensitivity and greater vulnerability to clouds and aerosols (Jacob et al., 2022; Hasekamp et al., 2022; Umezawa et al., 2025). As a result, no single satellite can offer both the accuracy and coverage desirable for robust global-scale XCH₄ monitoring, making multi-sensor data fusion necessary. However, merging multiple sources first requires addressing the systematic biases inherent to each sensor to ensure inter-sensor consistency.

Previous bias correction studies have relied on statistical regression approaches that account for factors such as surface albedo, aerosol loading, and viewing geometry (Inoue et al., 2016), as well as small-area approximation methods that derive surface albedo-related biases directly from satellite observations (Lorente et al., 2021). However, such methods can struggle to capture nonlinear, sensor-specific bias structures. Machine learning (ML) approaches, including Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost), have demonstrated improved bias correction performance by learning complex relationships between retrieval errors and environmental or observational predictors (Balasus et al., 2023; Schneising et al., 2023; Li et al., 2024).

Despite these advances, several limitations remain. Some studies improve spatial coverage but retain substantial residual bias (Li et al., 2024), whereas others reduce systematic bias by applying global offsets to match TCCON means (Balasus et al., 2023; Parker et al., 2020), an approach that can ignore regional variations in sensor bias. Moreover, most prior work has focused on single-sensor enhancement or limited dual-sensor combinations (e.g., GOSAT + TROPOMI), leaving the potential of multi-sensor fusion underexplored (Fan et al., 2024; Balasus et al., 2023; Li et al., 2024). In contrast, XCO₂ research has increasingly adopted multi-platform fusion (e.g., OCO-2, OCO-3, GOSAT, and GOSAT-2) as a standard approach (Chen et al., 2024; Wang et al., 2022), while comprehensive global XCH₄ fusion products remain comparatively scarce. To the best of our knowledge, GOSAT-2 has not yet been incorporated into a global XCH₄ fusion product, despite its improved spectral performance and expanded coverage relative to GOSAT.

To address these gaps and enable consistent global XCH₄ monitoring, we develop an ML-based multi-sensor harmonization and fusion framework that (1) corrects sensor-specific biases using TCCON observations, (2) harmonizes all satellite retrievals to the most accurate sensor (GOSAT-2), and (3) fuses the harmonized retrievals to maximize daily spatial coverage while preserving accuracy. The resulting product provides global daily XCH₄ over land (60° S–80° N) at 0.1° resolution for 2020–2023, with improved accuracy and spatiotemporal coverage, compared to individual sensor retrievals. This enhanced dataset enables more reliable regional emission analysis, hotspot identification, and source characterization.

Figure 1 Study domain and TCCON sites (GGG2020 XCH₄) used during 2020–2023. Cyan circles denote sites used for model training/validation. Green stars indicate independent test sites held out from training and used only for evaluation of the Fused XCH₄ product. Red boxes (A–E) denote regions of interest used for regional analyses in Sect. 4.3. The background map shows the 2020–2023 mean SWIR surface albedo from TROPOMI.

Table 1 TCCON (GGG2020 version) sites used for evaluation of satellite XCH₄ and machine learning based bias correction.

^a Sites excluded from the GOSAT evaluation under the GOSAT colocation criteria. ^b Sites excluded from the GOSAT-2 evaluation under the GOSAT-2 colocation criteria. ^c Independent test sites not used for training/validation of the machine learning based bias-correction model.

2 Datasets

2.1 Ground XCH₄ observations

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.

Table 2 Key specifications of GOSAT, GOSAT-2, and Sentinel-5P/TROPOMI XCH₄ products used in this study.

^∗ The number of XCH₄ retrievals per day represents the average soundings calculated per day over the full study period from 1 January 2020 to 31 December 2023.
LST: Local Solar Time.

Figure 2 Spatial distribution of daily XCH₄ retrievals over land on 1 January 2020 from GOSAT, GOSAT-2, and TROPOMI. Red markers indicate native pixel locations of individual soundings and are enlarged for visualization purposes.

2.2 Satellite XCH₄ products

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.

2.2.1 GOSAT

GOSAT (also known as Ibuki), launched in 2009, carries the Thermal And Near infrared Sensor for carbon Observation Fourier Transform Spectrometer (TANSO-FTS) and the TANSO Cloud and Aerosol Imager (CAI) in a sun-synchronous orbit at 666 km (Kuze et al., 2009). TANSO-FTS measures high-resolution spectra in the SWIR and thermal infrared bands, which are used to retrieve column abundances of CO₂ and CH₄ (Yoshida et al., 2011). We used the NIES SWIR L2 CH₄ product (v03.05) from GOSAT Data Archive Service (GDAS: https://data2.gosat.nies.go.jp/, last access: 26 December 2025), which provides both bias-uncorrected (hereafter standard) and bias-corrected retrievals calibrated against TCCON GGG2014 (Inoue et al., 2016). The standard product was used as ML model input, while the bias-corrected product was used for comparison.

2.2.2 GOSAT-2

GOSAT-2, the successor to GOSAT, was launched in 2018, carries upgraded TANSO-FTS-2 and CAI-2 with an intelligent cloud-avoidance pointing strategy, enabling enhanced sensitivity and expanded spatial coverage (Suto et al., 2021; Imasu et al., 2023; Park et al., 2024). GOSAT-2 extends Band 3 spectral coverage to 2.33 µm (compared to 2.08 µm for GOSAT), capturing additional CH₄ absorption features. We used the SWIR Level-2 full-physics XCH₄ product (v02.00; hereafter standard) from the GOSAT-2 archive (https://data2.gosat.nies.go.jp/, last access: 26 December 2025). To ensure high data reliability, only retrievals with quality flag = 0 were retained for bias-correction modeling.

Figure 3 Overview of the machine learning (ML)-based bias correction (BC) and data fusion framework used to generate the Fused XCH₄ product.

2.2.3 TROPOMI

TROPOMI, the sole payload on ESA's Sentinel-5 Precursor (S5P) mission launched in 2017, is a nadir-viewing push-broom hyperspectral imaging spectrometer that measures radiance from the ultraviolet (UV) to the SWIR (Veefkind et al., 2012). CH₄ retrievals are derived using the RemoTeC algorithm (Hasekamp et al., 2022), and the sun-synchronous orbit with an approximately 2,600 km swath width enables near-daily global mapping of atmospheric CH₄. We used the S5P Level-2 XCH₄ product (S5P_L2__CH4____HiR) from NASA Goddard Earth Sciences Data and Information Services Center (GES DISC; https://disc.gsfc.nasa.gov/datasets/S5P_L2__CH4____HiR_2/summary?keywords=methane, last access: 22 June 2026). This dataset provides both uncorrected XCH₄ (hereafter standard) and the posteriori albedo-dependent bias corrected XCH₄ (Lorente et al., 2021), along with associated retrieval parameters. For consistency with GOSAT and GOSAT-2, we selected standard XCH₄ retrievals with a quality assurance flag qa_value = 1 as model input, while the official bias-corrected product was used for comparison.

3 Methodology

To address limitations of single-sensor CH₄ retrievals and inter-satellite inconsistencies, we developed an integrated ML framework that combines bias correction, cross-sensor harmonization, and multi-sensor fusion of XCH₄ retrievals from TROPOMI, GOSAT, and GOSAT-2. The framework produces a globally harmonized daily 0.1° land XCH₄ product for 2020–2023 through three hierarchical steps (Fig. 3):

• Step 1: Sensor-specific bias correction using TCCON as the calibration standard.

• Step 2: Inter-satellite harmonization by aligning all datasets to the most accurate sensor (bias-corrected GOSAT-2).

• Step 3: Multi-sensor fusion to maximize daily spatial coverage while preserving accuracy.

We define the systematic retrieval bias for a given satellite observation as:

$$\begin{array}{}\text{(1)}& {\mathrm{\Delta }}_{\text{Bias}}={\mathrm{XCH}}_{\mathrm{4},\mathrm{Satellite}}-{\mathrm{XCH}}_{\mathrm{4},\mathrm{Reference}}\end{array}$$

where XCH_{4, Satellite} denotes the standard (uncorrected) satellite retrieval and XCH_{4, Reference} denotes the reference value (TCCON in Step 1; bias-corrected GOSAT-2 in Step 2). For BC, we used ensemble tree-based regressors (RF and XGBoost) due to their strong performance and stability in tabular regression with nonlinear interactions (Hwang and Song, 2023; Shwartz-Ziv and Armon, 2022). For GOSAT and GOSAT-2, the predictor set additionally included co-retrieved XCO₂-related variables because CH₄ and CO₂ are retrieved simultaneously from the same SWIR spectra using a shared full-physics forward model (Butz et al., 2010). Scattering-induced light-path errors often affect both gases in a correlated manner (Oshio et al., 2020). As TROPOMI does not retrieve CO₂, only XCH₄ retrieval parameters were selected based on established practice (Balasus et al., 2023). Using these parameters, mission-specific BC models were trained to predict bias (${\widehat{\mathrm{\Delta }}}_{\text{Bias}}$) and the predicted bias was subtracted from the original retrieval:

$$\begin{array}{}\text{(2)}& {\mathrm{XCH}}_{\mathrm{4},\mathrm{BC}}={\mathrm{XCH}}_{\mathrm{4},\mathrm{Satellite}}-{\widehat{\mathrm{\Delta }}}_{\text{Bias}}\end{array}$$

Table 3 Colocation criteria used for TCCON-based bias correction (Step 1) and inter-satellite harmonization (Step 2). All thresholds were chosen based on satellite-specific characteristics and standard validation practices.

3.1 Data preprocessing and colocation

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.

3.2 ML-based bias correction and harmonization

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.

3.2.1 Step 1: TCCON-based bias correction

For each satellite, we trained ML models using satellite–TCCON collocated pairs to predict the systematic bias between raw satellite retrievals and TCCON:

$$\begin{array}{}\text{(3)}& {\mathrm{\Delta }}_{\text{Bias},\text{step1}}={\mathrm{XCH}}_{\mathrm{4},\mathrm{Satellite}}-{\mathrm{XCH}}_{\mathrm{4},\mathrm{TCCON}}\end{array}$$

Satellite-specific retrieval parameters were used as predictors to model systematic error in the raw XCH₄ retrievals. Because each mission provides a large number of candidate retrieval parameters (> 100), we applied a three-stage feature selection procedure:

• 1.

  Correlation screening: Pearson and Spearman analyses to rank predictors based on linear and monotonic associations with Δ_Bias,step1.

• 2.

  Recursive feature elimination (RFE): RFE with cross-validation to identify a compact predictor subset.

• 3.

  Refinement: final adjustment using feature-importance rankings from preliminary models.

The final predictor lists are provided in Tables S1–S3 in the Supplement. The trained model predicts ${\widehat{\mathrm{\Delta }}}_{\text{Bias},\text{step1}}$, and the bias-corrected product is computed using Eq. (2).

3.2.2 Step 2: Inter-sensor harmonization

To select a harmonization reference, we evaluated the Step 1 bias-corrected outputs from all three sensors against common TCCON observations. GOSAT-2 exhibited the highest agreement with TCCON and was therefore selected as the harmonization reference (detailed results in Sect. 4.2). Using inter-satellite collocations, we developed ML models to reduce sensor-to-sensor discrepancies relative to calibrated GOSAT-2:

$$\begin{array}{}\text{(4)}& {\mathrm{\Delta }}_{\text{Bias},\text{step2}}={\mathrm{XCH}}_{\mathrm{4},\mathrm{Satellite}}-{\mathrm{XCH}}_{\mathrm{4},\mathrm{BC}\mathrm{GOSAT}\text{-}\mathrm{2}}\end{array}$$

where XCH_{4, Satellite} is the standard retrieval from the target sensor (GOSAT or TROPOMI) and ${\mathrm{XCH}}_{\mathrm{4},\mathrm{BC}\mathrm{GOSAT}-\mathrm{2}}$ is the Step-1 bias-corrected GOSAT-2 retrieval. Feature selection followed the same three-stage procedure as in Sect. 3.2.1 (correlation screening, RFE, and importance refinement). Final predictor lists are provided in Tables S3 and S4. We trained ML models to predict ${\widehat{\mathrm{\Delta }}}_{\text{Bias},\text{step2}}$ and generated GOSAT-2-like harmonized products by subtracting the predicted bias:

$$\begin{array}{}\text{(5)}& {\mathrm{XCH}}_{\mathrm{4},\mathrm{Harmonized}}={\mathrm{XCH}}_{\mathrm{4},\mathrm{Satellite}}-{\widehat{\mathrm{\Delta }}}_{\text{Bias},\text{step2}}\end{array}$$

3.3 Step 3: Data Fusion

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:

$$\begin{array}{c}{\displaystyle }\text{GOSAT-2}\left(\text{BC}\right)\to \text{TROPOMI}\left(\text{GOSAT-2-like}\right)\to \\ {\displaystyle }\text{GOSAT}\left(\text{GOSAT-2-like}\right)\end{array}$$

For each day and each 0.1° grid cell (x,y), the Fused product was defined as:

$$\begin{array}{}\text{(6)}& \begin{array}{rl}& {\mathrm{XCH}}_{\mathrm{4},\mathrm{Fused}}(x,y)=\\ & \left\{\begin{array}{ll}{\mathrm{XCH}}_{\mathrm{4},\mathrm{GOSAT}\text{-}\mathrm{2}}(x,y),& \text{if\hspace{0.17em}\hspace{0.17em}available}\\ {\mathrm{XCH}}_{\mathrm{4},\mathrm{TROPOMI}}(x,y),& \text{else\hspace{0.17em}\hspace{0.17em}if\hspace{0.17em}\hspace{0.17em}available}\\ {\mathrm{XCH}}_{\mathrm{4},\mathrm{GOSAT}}(x,y),& \text{else\hspace{0.17em}\hspace{0.17em}if\hspace{0.17em}\hspace{0.17em}available}\\ \mathrm{NaN},& \mathrm{otherwise}\end{array}\right.\end{array}\end{array}$$

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.

3.4 Model evaluation

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).

Figure 4 Site-wise scatterplots of column-averaged dry-air methane (XCH₄) between Total Carbon Column Observing Network (TCCON) and satellite products; (a–c) TROPOMI XCH₄, (d–f) GOSAT XCH₄, (g–h) GOSAT-2 XCH₄ for the operational standard and operational bias-corrected products, and machine learning (ML) based bias-corrected result (under leave-one-site-out cross-validation), respectively. Each marker represents the mean XCH₄ at one TCCON site over the study period (2020–2023), with horizontal and vertical error bars indicating one standard deviation of TCCON and satellite XCH₄, respectively. Red lines denote least-squares linear regression fits.

4 Results and discussion

4.1 Overall validation against TCCON

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 ppb yr⁻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.

Figure 5 Monthly and seasonally aggregated XCH₄ bias distributions under leave-one-month-out cross-validation (LOMOCV). Box-and-whisker plots show bias distributions (Satellite − TCCON) for each calendar month and their seasonal aggregations for (a) TROPOMI, (b) GOSAT, and (c) GOSAT-2. The compared datasets include the uncorrected operational standard products, operational bias-corrected products, and ML-based bias-corrected results. Boxes indicate the interquartile range (IQR) with the median shown as a horizontal white line. The shaded band marks the ± 8 ppb range around zero bias.

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 Dependence of XCH₄ ratios ($\text{TCCON}/\text{Satellite}$) on key retrieval variables and its reduction after bias correction. (a, b) ($\text{TCCON}/\text{TROPOMI}$) XCH₄ ratio as a function of SWIR surface albedo for the standard product (a) and after bias correction (b). (c–f) Same as (a), (b) but for GOSAT and GOSAT-2 as a function of surface pressure difference (ΔPs): standard product in (c), (e) and bias-corrected results in (d), (f). Solid lines denote the fitted least-squares regressions, and the dashed horizontal line indicates the unbiased reference level (ratio = 1). The Pearson correlation coefficient (R) and regression equation are reported in each panel.

Figure 6 further illustrates condition-dependent biases by relating the XCH₄ ratio ($\text{TCCON}/\text{satellite}$) 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.

4.2 Inter-satellite harmonization

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.

Table 4 Accuracy comparison of machine learning (ML) based bias-corrected XCH₄ from three satellite sensors (TROPOMI, GOSAT, GOSAT-2), evaluated against TCCON using common collocated samples (N=309) under leave-one-site-out cross-validation. Boldface indicates the satellite with the best performance for each error metric.

Figure 7 Inter-satellite harmonization to the ML bias-corrected GOSAT-2 ML BC scale. Distributions of XCH₄ bias relative to the ML-based bias-corrected GOSAT-2 product (a: TROPOMI − GOSAT-2; b: GOSAT − GOSAT-2) before and after cross-satellite harmonization. Solid vertical lines denote the mean bias for each distribution, and the legend reports the corresponding mean and standard deviation (SD).

Figure 8 Retrieval-parameter dependency of harmonized XCH₄ relative to the final target. Scatter density plots show the ratio of the final target (i.e.,ML-based bias-corrected GOSAT-2 XCH₄) to satellite XCH₄, as a function of key retrieval parameters using satellite co-location samples. The upper panels show the dependency of TROPOMI XCH₄ on SWIR surface albedo, and the lower panels show the dependency of GOSAT XCH₄ on the surface pressure difference, ΔPs. For each sensor, the left panels show the standard products, while the right panels compare the operational bias-corrected products and the ML-based harmonized results.

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.

Figure 9 Independent validation of the Fused XCH₄ at three TCCON test stations and comparison with a Blended product (Balasus et al., 2023). (a) Scatter plot of Fused XCH₄ versus TCCON XCH₄ at three independent sites (Xianghe01, Edwards01, and Garmisch01); colors indicate individual stations. The red line denotes the fitted regression. (b) Time series of station-mean XCH₄ averaged across the three test stations (equal-weight mean), comparing TCCON, Fused, and the standard products from GOSAT, GOSAT-2, and TROPOMI. (c) Comparison against TCCON for 2022 at the same test stations, contrasting the Fused XCH₄ (blue) with the Blended XCH₄ product based on TROPOMI and GOSAT (orange). The solid-colored lines denote fitted regressions, respectively.

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.

4.3 Coverage improvements from multi-satellite fusion

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.

Figure 10 Regional assessment of multi-sensor XCH₄ fusion benefits in 2023 across five regions representing distinct retrieval challenges. (a) Coverage enhancement relative to TROPOMI showing absolute gain in percentage points (bars, left axis) and relative improvement (line, right axis). (b) Annually dominant sensor per pixel based on valid retrieval frequency in 2023 for each ROI (GOSAT: blue, GOSAT-2: orange, TROPOMI: grey, missing data: white). Orange and blue pixels indicate locations where TROPOMI data are unavailable but coverage is provided by GOSAT-2 or GOSAT.

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.

Figure 11 Spatiotemporal characteristics and growth of Fused XCH₄ during 2020–2023. (a) Monthly latitudinal distribution of Fused XCH₄ concentrations at 0.1° spatial resolution from 2020 to 2023. (b) Comparison of monthly global mean methane time series from the Fused XCH₄ product (orange), NOAA marine surface CH₄ measurements (blue), and TCCON ground-based XCH₄ observations (green). Values in the legend indicate mean annual growth rates over 2020–2023 with associated uncertainties. (c) Monthly hemispheric year-over-year differences (solid lines) and annual mean growth rates (markers) for the Northern Hemisphere (NH; 0–80° N, red), Southern Hemisphere (SH; 0–60° S, blue), and high-emission zone (0–40° N, orange). Error bars represent the standard deviation of monthly growth rates within each year. All trends shown in panels (b) and (c) were statistically significant (p < 0.001).

4.4 Global XCH₄ mapping and spatiotemporal analysis

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 ppb yr⁻¹ (Fig. 11b). This agrees well with TCCON (13.56 ± 1.25 ppb yr⁻¹) and NOAA (14.77 ± 0.64 ppb yr⁻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 ppb yr⁻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 ppb yr⁻¹ in the NH, 11.4 ± 3.9 ppb yr⁻¹ in the SH, and 6.4 ± 3.0 ppb yr⁻¹ in the high-emission zone by 2023, consistent with recent inverse modeling estimates (Pendergrass et al., 2025).

Figure 12 Annual XCH₄ anomaly maps for (a) 2020, (b) 2021, (c) 2022, and (d) 2023 relative to the four-year mean. Parentheses show global mean anomalies (ppb). Positive (red) and negative (blue) shading indicate regions with XCH₄ above and below the four-year baseline, respectively.

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.

Figure 13 Regional XCH₄ emission hotspots and corresponding land-cover characteristics for 2023. The central map displays the global distribution of Fused XCH₄ concentrations with six representative high-emission regions marked (a–f). Each regional panel pair shows XCH₄ concentrations (left) and MODIS MCD12Q1 land-cover classifications (right) for the Niger Delta (a), Ganges–Brahmaputra Delta (b), Indo–Gangetic Plain (c), Yangtze River Delta (d), Congo Basin (e), and Barnett Shale region (f). Land-cover legend follows the MODIS MCD12Q1 LC_Type1 classification.

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.

Figure 14 Seasonal distributions of Fused XCH₄ concentrations and latitudinal gradients in 2023. Left panels (a, c, e, g) show seasonal mean XCH₄ concentrations at 0.1° resolution for March–April–May (MAM), June–July–August (JJA), September–October–November (SON), and December–January–February (DJF). Right panels (b, d, f, h) display corresponding zonal mean profiles (solid purple line) with shaded regions representing spatial variability (± 1σ standard deviation) across each latitude band. Latitudes with data coverage below 2 % were excluded from zonal mean calculations to ensure statistical robustness. Regional definitions and seasonal statistics are detailed in Table S7.

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).

4.5 Limitations

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.

5 Conclusion

This study presents the first global multi-sensor fusion framework that integrates XCH₄ retrievals from GOSAT, GOSAT-2, and TROPOMI. Using a three-step ML strategy – (1) TCCON-calibrated bias correction, (2) inter-sensor harmonization, and (3) fusion – we produced a globally harmonized daily 0.1° land XCH₄ dataset for 2020–2023. Independent evaluation at withheld TCCON sites demonstrates strong generalization (R² = 0.81, RMSE = 10.78 ppb), outperforming the standard and operationally bias-corrected satellite products.

A key outcome is that GOSAT-2 provides an effective anchor for cross-sensor harmonization, offering high agreement with TCCON while maintaining broader sampling than GOSAT. The bias-correction and harmonization analyses also revealed and mitigated sensor-specific residual bias structures, including the SWIR surface albedo dependency in TROPOMI and the ΔPs dependency in GOSAT. These results indicate that the proposed framework not only improves the final Fused product but also provides diagnostic insight into retrieval-condition-dependent limitations of the individual satellite products.

Regional assessments show that fusion improves data availability in retrieval-limited environments, increasing coverage by ∼ 1.8 % to ∼ 9.6 % depending on the region compared to the TROPOMI operational product. Although the overall global coverage gain is modest, the Fused product provides complementary XCH₄ information in regions or conditions where TROPOMI retrievals are limited.

The Fused product reproduces key spatiotemporal characteristics of atmospheric XCH₄, including interannual hemispheric variability and persistent emission hotspots in tropical and subtropical regions associated with wetlands, agriculture, and fossil-fuel infrastructure. The dataset can support downstream applications such as XCH₄ gap filling, multi-sensor intercomparison and harmonization, regional methane assessment, hotspot identification, and the development of advanced fusion strategies. The sensor-specific corrected and harmonized products can also serve as reference-scale benchmarks for future XCH₄ bias-correction studies. The proposed framework is extensible to upcoming missions (GOSAT-GW, CO2M), enabling enhanced monitoring. Continued advances in multi-sensor harmonization and uncertainty characterization will be crucial for supporting progress toward the GMP target of a 30 % CH₄ emission reduction by 2030.