We use the blended TROPOMI+GOSAT XCH₄ observation product (Balasus et al., 2023) which corrects the operational TROPOMI data of Lorente et al. (2021) with a machine-learning (ML) algorithm trained on collocated GOSAT data. GOSAT is more precise and less subject to surface and aerosol artifacts because it uses a CO₂ proxy retrieval method for the 1.6 µm absorption band (Parker et al., 2020), whereas TROPOMI uses a full-physics retrieval method for the 2.3 µm absorption band. But GOSAT data are ∼ 200 times sparser than TROPOMI, and the blended TROPOMI+GOSAT product thus combines the density of TROPOMI observations with the quality of GOSAT observations. We refer to this blended product as TROPOMI in the text.

TROPOMI observations are limited to land and cloud-free scenes, and are fairly distributed across seasons over the domain of Fig. 1 (Fig. S1 in the Supplement). The number of observations ingested each year varies from 7.6 to 10.4 million. We average observations over the 0.25° × 0.3125° GEOS-Chem grid for each orbit to create super-observations (Eskes et al., 2003), where the reduction in retrieval error from averaging accounts for error correlations as described by Chen et al. (2023).

Smoothed TROPOMI concentrations applied as boundary conditions along the edges of Fig. 1 are produced on a 2.0° × 2.5° grid following Estrada et al. (2025) by sampling XCH₄ and its vertical distributions from a GEOS-Chem simulation with prior emission estimates, and correcting it with 15 d averages of TROPOMI observations over 8° × 10° domains. When such averages are not available, as for oceans, we apply a zonal mean correction for that latitudinal band. Using smoothed TROPOMI concentrations in this manner as boundary conditions ensures consistency with the observations used in the inversion. We estimate a 5–15 ppb error standard deviation on boundary conditions based on comparison with TCCON ground-based XCH₄ observations (Maasakkers et al., 2019).

We apply the IMI (version 2.1; https://carboninversion.com, last access: 4 December 2025) to infer emissions by minimizing the Bayesian cost function with normal error statistics (Brasseur and Jacob, 2017): 1Jx=x-xATSA-1x-xA+γy-KxTSO-1y-Kx, where x is the state vector of emissions and boundary conditions, xA is the prior estimate, y is the vector of observations, K = ∂y/∂x is the Jacobian matrix relating emissions to concentrations in GEOS-Chem, SA is the prior error covariance matrix, SO is the observational error covariance matrix, and γ is a regularization parameter to prevent overfit to the observations. SA and SO are taken as diagonals for lack of better information, and γ < 1 is needed to correct for error correlation between super-observations that is not accounted for in SO.

The optimal estimate, x^, is derived analytically by solving dJ/dx = 0 (Rodgers, 2000; Brasseur and Jacob, 2017), yielding: 2x^=xA+γKTSO-1K+SA-1-1γKTSO-1y-KxA. The analytical solution provides explicit error characterization through S^, the posterior error covariance matrix: 3S^=γKTSO-1K+SA-1-1. The averaging kernel matrix A = ∂x^/∂x = I-S^SA-1 describes the sensitivity of the inversion to the true state. The diagonal elements of A, aii, are called averaging kernel sensitivities and measure the ability of the observations to infer emissions on the native state vector grid independently of the prior estimate, ranging from 0 (no ability) to 1 (complete ability). Emissions can still be quantified from the observations when averaging kernel sensitivities are low by spatial aggregation. The sum of averaging kernel sensitivities (trace of A) defines the degrees of freedom for signal (DOFS), estimating the total independent number of pieces of information from the observations.

The posterior error covariance matrix S^ does not account for uncertainties in inversion hyperparameters (prior error standard deviation, observational error standard deviation, regularization parameter, error distribution, etc.) which dominate the overall error (Chen et al., 2022b). As a more conservative estimate of uncertainties on our posterior emission estimates, we generate an ensemble of estimates with varied inversion hyperparameters, considering both normal and lognormal probability density functions (pdfs) for the prior emission error estimates (Table 1). Lognormal error pdfs may better characterize the heavy tail of emissions, particularly for the oil/gas sector (Yuan et al., 2015; Cui et al., 2019), and have the advantage of enforcing positivity in the solution. Lognormal error pdfs are accommodated in Eq. (1) by solving for ln⁡x instead of x, which makes the forward model non-linear and requires solving for ln⁡x^ iteratively. We solve this problem with the Levenberg–Marquardt algorithm (Chen et al., 2022b). The optimization is then for the median of x rather than the mean, which requires median-mean conversions as described by Hancock et al. (2025). To account for errors in the boundary conditions, we optimize each cardinal domain edge as part of the inversion following Nesser et al. (2024). In practice, we find edge corrections to range in magnitude from 0.5 to -21 ppb for 2024 with the highest corrections on the western edge. From the combinations of inversion hyperparameters in each column of Table 1, we generate an ensemble of 36 normal and 6 lognormal estimates. These combinations were chosen from an original pool of 72 inversion members that used a range of values from previous TROPOMI inversions (Nesser et al., 2024; Varon et al., 2022; Chen et al., 2022b), such that the prior terms of the posterior cost function match the expected normalized chi-square value of 1 to within the range 0.5 to 1.5, indicating a successful fit while avoiding overfit (Lu et al., 2022). We report the mean from the ensemble members (equal weighting for normal and lognormal pdfs) as our best estimate and the range across ensemble members as our uncertainty. We calculate our 2019–2024 emission trends using Ordinary Least Squares (OLS) regression on the mean estimates for each individual year and report the corresponding standard error from the regression. We also calculate trends for individual ensemble members and find that the regression standard error is typically a more conservative estimate of the trend uncertainty (Fig. S2).

We apply the GEOS-Chem chemical transport model (version 14.4.1) as the forward model to relate emissions to atmospheric concentrations (Maasakkers et al., 2019) as expressed by the Jacobian matrix K in Eq. (1). The model is driven by NASA GMAO GEOS-FP meteorological fields at 0.25° × 0.3125 resolution (Lucchesi, 2017). We use the nested version of the model (Kim et al., 2015; Zhang et al., 2015) with a simulation domain of 19.25 to 54.75° N and 61.5625 to 130° W (domain of Fig. 1). Simulations are run for the full calendar year for each inversion year 2019–2024. The model includes methane sinks from atmospheric oxidation and soil uptake (Maasakkers et al., 2019). These are not directly optimized in the inversion but indirectly through the optimization of boundary conditions.

Prior emission estimates for CONUS are summarized in Table 2. Anthropogenic emission estimates by sector are from the monthly GHGI at 0.1° × 0.1° grid resolution (Maasakkers et al., 2023) produced annually from 2012 to 2020. 2020 GHGI values are used as prior estimates for subsequent years in the inversion. GHGI does not include emissions from hydroelectric reservoirs, which we add from the Reservoir Methane Emissions inventory (ResME; Delwiche et al., 2022) and account for 7 % of anthropogenic emissions. Prior anthropogenic emission estimates for Mexico and Canada are from Scarpelli et al. (2020, 2022), who spatially allocate the UNFCCC reports on a 0.1° × 0.1° grid. Prior anthropogenic emission estimates for other countries in the model domain are IMI defaults (Estrada et al., 2025). For wetlands, we use the 2019 mean monthly estimates generated from the nine high-performance members of WetCHARTs v1.3.1 (Bloom et al., 2021) as prior estimate for all inversion years. Other minor natural sources include daily open fires from the Global Fire Emissions Database (GFED4) (Randerson et al., 2017), geological seeps (Hmiel et al., 2020), and termites (Fung et al., 1991).

Prior estimates for a given year do not include information from posterior results for the previous year, as would be done in a Kalman filter, in order to apply consistent corrections to the bottom-up estimates from year to year. The inversion optimizes annual posterior estimates only, assuming that the relative seasonality in the prior emissions is correct. That seasonality is mainly driven by wetlands. Seasonal variability in the GHGI is small and mainly driven by temperature-dependent emissions from livestock manure.

We apply the IMI's smart clustering algorithm (Estrada et al., 2025) to generate a multi-resolution state vector that aggregates native grid cells in areas with weak emissions and low observation density, while maintaining native 0.25° × 0.3125° grid resolution in areas with strong emissions and high observation density. The same state vector clustering is used for all annual inversions and is based on the average observation density for 2019–2024 when applying the algorithm. We force clusters to respect state boundaries to avoid aggregation error on the calculation of state total emissions.

The number of state vector elements defining the resolution of the inversion should maximize the DOFS while remaining computationally affordable. We determined the optimal number before running the inversion by estimating the DOFS for varying state vector sizes in the IMI preview as described by Estrada et al. (2025). We choose a state vector dimension of 2600 as this is where the gain in the DOFS plateaus. Additionally, we impose a maximum cluster size of 10 grid cells per element to maintain relatively high resolution across the domain.

The inversion returns posterior annual emissions on the 0.25° × 0.3125° grid. Results are then aggregated nationally, by state, or by sector. Sectoral attribution is done by applying the posterior/prior emission ratio in each grid cell to correct the prior sectoral emissions in that grid cell. Spatially aggregated posterior emissions by sector are obtained by applying the summation matrix W with rows containing the relative contributions from individual sectors to emissions from each state vector element (Nesser et al., 2024). This reliance on the distribution of the prior emissions can lead to errors in sectoral attribution of the posterior/prior correction in regions where there is spatial overlap between sectors not resolved by the inversion. We characterize this error following Hancock et al. (2026) with the reduced averaging kernel matrix Ared: 4Ared=WAW*, where W* is the Moore–Penrose pseudoinverse matrix.

We evaluate our inversion results by comparing the GEOS-Chem simulation with posterior emissions (Fig. 1) to the TROPOMI observations used in the inversion and to independent observations from surface, aircraft, and tower observations in the NOAA CH₄ GLOBALVIEWplus v7.0 data product (Schuldt et al., 2024). Following Lu et al. (2022), NOAA observations are sampled during daytime hours (10:00–16:00 local time) and only include observations within 3 standard deviations from the daily mean, or 2 standard deviations if the standard deviation is greater than 30 ppb. We compare the model to the annual ensemble mean observations for each site for 2019–2023. Figure 2 shows improvements in RMSE and mean bias in the TROPOMI residuals relative to the prior estimate (panels a and b), a shift of the residual distribution toward zero (panel c), and a general whitening of noise across the domain, indicating improved agreement with the observations. Independent observations (panel d) show a 5.6 ppb improvement in mean bias relative to the prior estimate and a 5.3 ppb improvement in RMSE. The coefficient of determination (R2) increases from 0.61 to 0.74. Independent observation sites with the highest remaining bias tend to be close to high emitting regions (e.g. Permian basin, Appalachian basin), where local influences in surface air may be difficult to reproduce at the 25 km model resolution.

The averaging kernel sensitivities in Fig. 1 indicate regions where emissions are most informed by the observations versus the prior estimate (1 = fully, 0 = not at all). They are highest where prior emissions are high (because the prior error standard deviation is then high) and where the observation density is high. The values in Fig. 1 are for 2024 but observation density varies little between years and so do the averaging kernel sensitivities. Emissions for grid cells with low averaging kernel sensitivities cannot be quantified at 25 km resolution but can still be quantified as national, state, and sectoral levels as measured by the reduced averaging kernel matrix (Eq. 4). Statewide reduced averaging kernel sensitivities are listed in Table S1 in the Supplement. The national sectoral averaging kernel matrix is shown in Fig. S3 row by row to evaluate the ability of the inversion quantify emissions from individual sectors independently from the others on the national scale. Observations contribute substantial information for all sectors with averaging kernel sensitivities ranging from 0.3–0.7 (highest for oil/gas and coal) but still allowing for significant contribution from the prior estimate. Off-diagonal terms of the averaging kernel matrix indicate that the posterior for a given sector is influenced by the corrections to other sectors as would be the case for overlap. The mainly urban “other” emissions thus influence the posterior estimates for oil/gas (mainly from downstream distribution sources), livestock, and landfills. The remaining sectors are cleanly separated. Corrections to boundary conditions have averaging kernel sensitivities near unity and do not affect significantly the corrections to emissions.

Table 2 summarizes sector-resolved methane emissions for CONUS. Our mean posterior estimate of total methane emissions is 47 Tg a⁻¹ with 10 Tg a⁻¹ from natural sources and 37 Tg a⁻¹ from anthropogenic sources for the 2019–2024 period. This estimate is 34 % higher than reported in the GHGI and 64 % higher for oil/gas emissions. Livestock and landfills are adjusted upward by 29 % and 30 %, respectively. Our results are consistent with other GOSAT and TROPOMI inverse studies, which find anthropogenic emissions in the range 30.0–42.7 Tg a⁻¹ (Turner et al., 2015; Maasakkers et al., 2021; Lu et al., 2022; Worden et al., 2022; Lu et al., 2023; Nesser et al., 2024). Results for individual sectors broadly agree with previous studies (Fig. S4) regardless of differences in time period, satellite product, prior inventory, and inversion methodology. There are discrepancies for coal and livestock with Turner et al. (2015), who used much sparser GOSAT observations, and for waste with Worden et al. (2022), who used much coarser resolution. Coal emissions have also decreased considerably since the 2009–2011 inversion years of Turner et al. (2015). Results for hydroelectric reservoirs show little departure from the ResME inventory used as prior estimate. Posterior emission estimates for wetlands and other natural sources also show little departure from the prior estimates.

Additionally, we compare our results in Table 2 to the Greenhouse Gas Inventory and Analysis (GHGIA; Desai et al., 2026), a recently released annual US inventory designed to be comparable with previous EPA GHGI inventories. This dataset does not include gridded information, but the national estimates and trends are broadly consistent with the GHGI. The underestimate identified in the GHGI still exists in the GHGIA. However, the GHGIA covers the period 1990–2024, allowing direct comparison of national trends with our posterior emission estimates.

Figure 3 shows the change in posterior emission estimates for each year relative to 2019 and the sector-specific trends for 2019–2024. Trends are derived through Ordinary Least Squares (OLS) regression of the mean sectoral emissions for each year. Uncertainties are the corresponding standard error of estimates from the regression line. Total anthropogenic emissions are flat in our posterior estimate (0.0 ± 1.0 % a⁻¹) in conflict with the GHGIA trend (-1.8 ± 0.3 % a⁻¹), but consistent with global inversions of TROPOMI data (He et al., 2026; Pendergrass et al., 2025). However, this overall flat trend reflects offsetting trends from different sectors. Emissions from fuel exploitation declined over 2019–2024 by -1.1 ± 0.9 % a⁻¹ for oil/gas and -2.3 ± 1.3 % a⁻¹ for coal. The decrease of fossil fuel emissions supports the declining trends found in the GHGIA for oil/gas (-3.5 ± 0.6 % a⁻¹) and coal (-2.7 ± 1.2 % a⁻¹), which reflect lower emission intensities (Desai et al., 2026) and reduced production from underground mines (Penn et al., 2026). Livestock emissions including enteric fermentation and manure increase in our posterior estimate (1.8 ± 1.3 % a⁻¹), despite falling cattle populations (8 % since 2019; USDA NASS, 2025) that drive the GHGIA trend (-1.2 ± 0.2 % a⁻¹). However, the trend in our estimate is driven by a single year (2024). We do not find a significant trend in landfill emissions (0.5 ± 1.4 % a⁻¹), in contrast to a declining trend reported by the GHGI (and GHGIA) that Balasus et al. (2025) show to be an artifact from a switch in landfill emission models. For rice agriculture, we see a large decreasing trend (-9.1 ± 2.0 % a⁻¹), which may be due to increasing adoption of water-saving practices (Hardke et al., 2024) not accounted for in the GHGIA trend (1.6 ± 1.6 % a⁻¹).

Methane intensity from the oil/gas sector is commonly defined as the total methane emissions along the oil/gas supply chain per unit of dry production of methane (Alvarez et al., 2018). We compute it using dry gas production data from the U.S. Energy Information Administration (EIA) assuming 85 % methane content (EPA, 1998). We find that the methane intensity decreased from 2.3 % to 1.9 % over the 2019–2024 period (Fig. 4). This continues the decreasing trend identified by Lu et al. (2023) for 2010–2019 and the consistent 2012–2020 trend in the GHGI. Our computed intensities are further consistent with the 2019 values computed using GOSAT and NOAA observations for 2019 (Lu et al., 2023) and for 2019 and 2023 values computed using TROPOMI (Shen et al., 2023; East et al., 2025). Lu et al. (2023) and Varon et al. (2023) found that declining intensities were due to increasing production with stable emissions, implying that emissions were decoupled from production. Here, we further find that actual emissions decrease while production increases, demonstrating recent improvements in emission management. Trends in emissions and intensities for individual oil/gas basins are discussed in Sect. 3.4.

Figure 5 shows anthropogenic emission totals and trends for individual US states and Table 3 summarizes data for the top 10 emitting states, which together account for 55 % of total anthropogenic US emissions. Data for all states are in Table S1. Averaging kernel sensitivities for total emissions from individual states in Table S1 range from 0.07 (Rhode Island) to 0.72 (California), reflecting differences in both state sizes and magnitudes of emissions. Emissions from Rhode Island (the smallest US state) are weak and cannot be effectively separated from those in neighboring states. Averaging kernel sensitivities for the top 10 states in Table 3 are 0.46 for Illinois and 0.58–0.72 for other states, indicating a high level of state-specific observational information from the inversion.

Texas is the largest emitting state, responsible for 19 % of national anthropogenic emissions, primarily from oil/gas activity. Six of the top 10 states are dominated by oil/gas emissions, and all six show upward oil/gas adjustments relative to the GHGI: Texas (+121 %), Oklahoma (+100 %), Pennsylvania (+47 %), Ohio (+41 %), New Mexico (+98 %), and Louisiana (+69 %). Texas, Oklahoma, and New Mexico exhibit particularly high oil/gas sector methane intensities (Fig. 5). In general, we find that states with dominant fossil fuel emissions show declining trends and states with dominant livestock emissions show increasing trends. Texas, California, and Nebraska are the largest livestock emitters, reflecting their large animal populations, but national trends are disproportionately influenced by livestock emissions increases in Iowa (8.3 ± 4.8 % a⁻¹), Kansas (7.0 ± 2.1 % a⁻¹), and Missouri (5.4 ± 4.3 % a⁻¹). These increases cannot be explained by herd size, as livestock populations remained stable in Iowa, although with higher hog populations (USDA NASS, 2025), and declined in Kansas and Missouri. Dairy cattle numbers rose in Iowa and Kansas but insufficiently to account for the observed emission trend. Rice emissions are concentrated in Arkansas and Louisiana and drive the decline in Arkansas' methane emissions, possibly driven by changing agricultural practices (Hardke et al., 2024) with methane reduction benefits (Runkle et al., 2019; Karki et al., 2021), though attribution is complicated by declining summer precipitation during 2019–2024 (National Centers for Environmental Information (NCEI), 2025).

Twenty-one states within CONUS produce their own greenhouse gas inventories, independent of the GHGI, using either custom-built frameworks or the EPA State Inventory Tool (SIT) with state-specific inputs (EPA, 2025). Here, we examine the available state inventories for the top ten emitting states, plus Colorado in support of their strong regulatory efforts (Colorado General Assembly, 2019). We find that state-specific data do not always lead to improvements over the GHGI. Texas released its first state inventory in 2025 using a combination of national and state information (Allen et al., 2025). Its estimate is lower than both the GHGI (-32 %) and our estimate (-64 %), driven by an underestimate of oil/gas emissions. Louisiana's inventory (Dismukes, 2021), based on the EPA SIT tool, is considerably lower than both the GHGI and our estimate, largely due to differences in reported oil/gas and landfill emissions. Pennsylvania is one of the highest producers of natural gas and has long shown very low methane intensity in the oil/gas sector, as seen in Fig. 5 and in other studies (Cardoso-Saldaña and Allen, 2021; Lu et al., 2023). The state inventory for Pennsylvania (Pennsylvania DEP, 2025) uses the EPA SIT but applies Appalachian-specific emission factors and incorporates more information on well type and counts from private data. It reports oil/gas emissions 30 % higher than the GHGI, aligning with the range of uncertainty in our estimate. However, despite close agreement with our work for oil/gas, Pennsylvania inventory totals are slightly lower than the GHGI estimate due to very low reported landfill emissions, which are substantially lower than both the GHGI and our estimate. This discrepancy may arise from overestimated landfill gas recovery efficiencies used to calculate total emissions (Balasus et al., 2025; Wang et al., 2025).

California has adopted targets to reduce 75 % of landfill emissions by 2025 relative to 2014 and 40 % of manure management emissions by 2030 relative to 2013 (California Legislature, 2016), alongside comprehensive oil/gas regulations (California Air Resources Board (CARB), 2017). The California Air Resources Board (CARB) generates annual emission estimates using custom methodology and state specific activity data to track state emission goals. California is the only state among the top 10 where our inversion adjusted emissions downward from the GHGI (-7 %) with particularly large decreases for oil/gas (-20 %) and landfills (-30 %). Comparison with the CARB inventory (California Air Resources Board (CARB), 2025) shows closer agreement across all sectors except oil/gas, reflecting the accuracy of non-fossil California data sources and reporting programs (Appuhamy and Kebreab, 2018). Comparisons of our results for individual sectors with the California inventory are shown in Fig. S5.

New Mexico has mandated a < 2 % methane intensity for oil/gas operations by 2026, frequent leak detection and repair (LDAR), and restrictions on flaring and venting beginning in 2021 (New Mexico Oil Conservation Commission, 2021). The state is home to a subset of the Permian basin, where high methane intensities have previously been documented (Zhang et al., 2020; Schneising et al., 2020; Liu et al., 2021; Chen et al., 2022a; Shen et al., 2022; Varon et al., 2023). The state's inventory (Bharadwaj et al., 2024), based on detailed equipment-level data, yields higher emissions than the GHGI (+10 %) but remains below our estimate (26 % higher than GHGI). We find a substantial decline in methane intensity from 4.3 % in 2019 to 1.7 % in 2024, indicating that New Mexico has already achieved its target of < 2 %, though most reductions occurred before the 2021 regulations. This decline in intensity is consistent with studies over the New Mexico portion of the Permian basin. Varon et al. (2025) found a decrease from 4.5 % to 2.1 % during 2019–2023. Aerial surveys over the area show a similar downward trend, with Chen et al. (2022a) reporting intensities of 9.4 % in 2018–2020 and Donahue et al. (2026) reporting 2.1 % in 2024.

Colorado has targeted a 50 % reduction in greenhouse gas emissions by 2030 relative to 2005 (Colorado General Assembly, 2019), supported by oil/gas regulations including quarterly and semiannual LDAR, a ban on routine flaring, and a 2029 phaseout of pneumatic devices (Colorado Department of Public Health and Environment (CDPHE), 2025). Further emission controls on landfills were enacted in late 2024. We find downward adjustments for Colorado relative to the GHGI across all sectors except oil/gas, including livestock (-17 %), landfills (-23 %), and coal (-52 %). The oil/gas sector is 12 % higher than the GHGI estimate and corresponds to a mean oil/gas methane intensity of 1.1 %. We find an 18 % downward adjustment for anthropogenic emissions relative to the most recent Colorado Department of Public Health inventory (Colorado Department of Public Health and Environment (CDPHE), 2026), but both inventories are within the bounds of the inversion ensemble. The updated state inventory substantially improves agreement for oil/gas emissions, which were previously overestimated in earlier CDPHE inventories due to outdated leakage rates (Twyman et al., 2024; Nesser et al., 2024). With the new inventory, our estimate is within 10 % of the CDPHE oil/gas estimate.

Figure 6 summarizes the oil/gas emissions and trends for the top 8 emitting oil/gas fields in CONUS. The annual mean reduced averaging kernel sensitivities for these fields are all 0.60 or above, except for Eagle Ford (0.44). Total methane emissions are dominated by the oil/gas sector, so that there is little error in attribution posterior/prior corrections to that sector. The Permian and Appalachian fields represent, respectively, 21 % and 14 % of CONUS oil/gas sector emissions with mean emissions of 2700 (2100–3400) Gg a⁻¹ and 1700 (1500–2300) Gg a⁻¹. Although the Appalachian's methane intensity is relatively low, its high production volume makes it one of the highest total emitters among US oil/gas fields. The Permian estimate is in the same range as other top-down estimates from Shen et al. (2022), Varon et al. (2023), Lu et al. (2023), and Omara et al. (2024). We find an average underestimate in the GHGI of 83 % across the top 8 oil/gas production fields with the largest adjustment in the Permian (+202 %) and the lowest adjustment to the Uinta Piceance (+28 %). We calculate methane intensities with the integrated emissions from the oil/gas supply chain, but upstream emissions from production are the dominant source in the top 8 oil/gas basins, so that methane intensity changes little if midstream or downstream emissions are excluded (Fig. S6). We find particularly high methane intensities in the Barnett (3.1 %), Permian (2.3 %), and San Juan (2.2 %). Oil-dominant fields, like the Permian, may exhibit high methane intensities due to insufficient gathering, processing, and midstream pipeline equipment for capturing associated gas from oil production (Lu et al., 2023; Omara et al., 2024; Varon et al., 2025). The Barnett and San Juan are gas-dominant fields that share common features: development peaked over a decade ago, production has since declined, and new investment has been limited. In contrast, gas-dominant fields with active development have some of the lowest intensities, including Haynesville (0.7 %), Appalachian (0.8 %), Anadarko (1.6 %), and Uinta-Piceance (1.7 %). The oil-dominant Eagle Ford, which also has low methane intensity (0.9 %), has rising gas production (and declining oil production) during the study period, suggesting increased investment in gas infrastructure.

All high-emitting production fields experience declining (Anadarko, Haynesville, and Eagle Ford) or flat emissions trends, consistent with findings for Texas oil/gas fields reported by Varon et al. (2025). The Haynesville and Eagle Ford fields increased production while reducing emissions, demonstrating improvements in emissions management. The San Juan and Barnett are the only fields in the top 8 with rising methane intensity. Their trends may reflect fixed emissions from aging infrastructure comprising a growing share of total emissions as gas production decreases.