the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 03 Jul 2026
Operational XCH4 Retrievals from MethaneSAT: Demonstrating Sensor Performance for Constraining Regional Methane Emissions
Abstract. MethaneSAT, launched in March 2024, was designed to quantify regional methane (CH4) emissions, with a primary focus on the oil and gas (O&G) sector. The satellite mission bridges the gap between coarse-resolution global flux mappers and high‑resolution plume imagers by combining fine spatial resolution (~110 x 400 m2 pixels at nadir), high spectral resolution (~0.23 nm FWHM), and a ~220 km swath at nadir. In this study, we present the first operational retrievals of CH4 column-averaged dry-air mole fractions (XCH4) using the CO2‑proxy method. We assess the instrument and retrieval algorithm performance relative to the mission's precision and accuracy requirements needed to constrain CH4 emissions at the scale of individual O&G basins.
We focus on evaluating the retrieval against four major potential sources of systematic error: albedo dependent CH4/CO2 column retrieval sensitivity differences, cross-track stripe biases, aerosol-induced light path errors, and subscene CO2 variability. Tuning the a priori covariance matrix suppresses albedo-dependent errors caused by the influence of priors to sub ppb levels. Through a careful selection of homogeneous validation targets, we show that XCH4 cross-track biases strongly correlate with changes in the apparent instrument spectral response function (ISRF). We develop a stripe-correction algorithm using retrieved ISRF variations in a regression model combined with wavelet-Fourier filtering to reduce stripe noise from approximately 15 ppb standard deviation to near the random noise limit.
Analysis of the homogeneous validation scenes shows that single-pixel precision is approximately 30 ppb for conditions from a typical bare-ground O&G scene (0.4 albedo, 30° SZA), corresponding to ~3 ppb at 2 x 2 km2, and thus well within the mission requirement of 3 ppb at 5 x 5 km2. This requirement is also met for all targeted viewing geometry/albedo combinations. Comparisons with XCH4 from TROPOMI show excellent agreement, with a mean bias of 0.1 ppb and a regression slope of 0.99 when using XCO2 from the CAMS greenhouse gas forecast as the prior in place of the GINPUT prior used operationally. The spatial pattern of Permian basin XCH4 enhancements between the two instruments is highly consistent.
Observations of a Pseudo-Invariant Calibration site in Libya confirm that the CO2-proxy approach effectively mitigates biases induced by cloud and aerosol scattering, with subscene XCO2 variability emerging as the most challenging remaining error source; gradients of a few ppm can lead to XCH4 errors comparable to typical basin enhancements, though such conditions appear infrequent in both MethaneSAT observations and CAMS GHG forecast simulations. Overall, MethaneSAT retrieves XCH4 with the precision and accuracy required for basin-scale emissions inversion. These results suggest that the CO2-proxy approach remains the most viable retrieval approach for regional CH4 emission mapping, with improved treatment of subscene XCO2 variability representing the key priority for future work.
- Preprint
(52052 KB) - Metadata XML
-
Supplement
(4789 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-3365', Anonymous Referee #1, 20 Jul 2026
-
RC2: 'Comment on egusphere-2026-3365', Anonymous Referee #2, 04 Aug 2026
General Comments
This study presents the first operational CH4 retrievals from MethaneSAT using the CO2-proxy method. It assesses the performance of the retrieval algorithm against mission requirements and provides an explicit analysis of several potential sources of systematic error. The paper represents an important contribution to the field and is well suited to the scope of Atmospheric Measurement Techniques. It can be published once the following points have been adequately addressed and clarified.
Specific Comments
Section 3.1
Quality filtering is based on the O2 band at 1.2µm. Are dayglow emissions (adding radiance that is not reflected sunlight) somehow considered/modelled to achieve the required accuracy needed for quality filtering? Isn't dayglow a problem for dark surfaces and/or high solar zenith angles? Please elaborate a bit more on the topic.
Is ΔXCO2 really needed as an additional quality criterion in a proxy retrieval? The assumption is that CH4 and CO2 are spectrally close and therefore react in virtually the same way to changes in path length, so that the effect becomes negligible in the ratio.
What is the ΔXCO2 threshold used for quality filtering?Section 3.2
The argument not to use profile scaling is not fully convincing. I think a retrieval response exceeding unity is no problem, because such behaviour is fully characterised by the averaging kernels, which are specifically provided to describe the retrieval sensitivity and allow proper interpretation of the results. It is true that a scaling factor retrieval is structurally incapable of representing changes in vertical shape. But as written in the text (L157), the retrieved profiles are highly sensitive to the assumed a priori covariance. This suggests that much of the apparent flexibility is supplied by the assumed covariance model rather than by information contained in the measurements. As a result, it is not evident that profile retrievals are inherently superior to profile scaling; they simply replace one explicit assumption (fixed profile shape) with another (assumed covariance). Please elaborate a bit more on this trade-off (between minimising assumptions about the profile shape and minimising dependence on an assumed covariance model) in the manuscript. How much independent vertical information (DoFS) do the measurements actually contain on average (quality cut-off is at 0.6 according to L138)?
This concern seems to be particularly relevant when the prior covariance is parameterised by adjusting only scalar scaling factors in the tuning process, while the covariance structure is fixed (Equation 3). Doesn't changing the scaling factors simply alter the balance between the measurement and the prior? How sensitive are the retrieval results to the choice of γch4 and γco2?Section 3.4
The stripe error of the PLS+Wavelet combined destriping approach falls below the level expected from purely random, normally distributed noise. This is attributed to the suppression of Fourier modes due to the Wavelet destriping (L424). Although this may have a small effect, I do not think that significant real features are suppressed, provided that σf has been chosen appropriately. I would be more inclined to suspect that the 'over-correction' is more related to the previous PLS regression. Wouldn't a wavelet correction alone be enough? Have you tried to examine what the result would look like? A wavelet-only correction would also have the advantage that the PLS-calibration on homogeneous scenes would not be required, which is an additional source of uncertainty.Section 4.1
Why is the precision on L447 reported on a 2°×2° grid and not on 5°×5° as the other estimate and the mission requirement?Section 4.2
L452: As there are at least three different TROPOMI products, please state explicitly "... operational TROPOMI XCH_4 retrievals of version ...".
L455: Is there a reference that TROPOMI exhibits larger systematic biases than GOSAT? I think, concerning station-to-station bias (stddev of the individual biases at TCCON sites), both instruments are very similar and close to the TCCON accuracy.
L459: The station-to-station bias (5.1 ppb in Lorente et al., 2023) is the much more important measure than the mean bias (-5.3 ppb in Lorente et al., 2023). I do not find a -5.6 ppb in the cited paper.
Figure 11: The MethaneSAT retrievals differ quite noticeably in detail for the two priors used. Just how different are the two prior data sets from one another?
L491: You write that CAMS likely cannot properly capture the true subgrid XCO2 variations at the scale of a MethaneSAT target. But can GGG2020 actually resolve those variations? Or actually not (better not at all than wrong)? Isn't it a general problem if the prior sensitivity is so high that the result depends on this capability? Or are you not talking about the prior of the CO2 retrieval, but rather the modelled XCO2 scaling as used in Equation 1? Please clarify in the text what is meant by 'prior' in each instance, as modelled CO2 is incorporated into the XCH4 retrieval in two different ways.
Figure 12: The colour scales for the second and third column seem to be saturated; please increase the spanned value range appropriately. Why isn't the same colour scale used for all three columns?Section 4.3
Figure 13: The proxy approach seems to work well for low-level clouds. But which scenes would the quality filter described in Section 3.1 have filtered out in the map anyway? Would there be any problematic scenes left at all? What about increases in methane that may be partially masked by clouds? (XCH4 [CO2 proxy] cannot see below the clouds)
L536: Is it really dust, it looks more like small patchy clouds? A similar figure for a strong desert dust aerosol event would also be interesting. The scene in Figure 15 also looks quite clean regarding aerosols.Section 4.4
Also here, please make clearer what is conceptually meant by prior. This section is about the modelled CO2 scaling 1:1 the retrieved XCH4 according to Equation 1 (and not about the prior for the CO2 retrieval itself), right? It would be helpful to use two different terms throughout the entire manuscript to distinguish between the two applications currently referred to as prior, even if the same data set is used.
Technical correctionsL90: the the
Citation: https://doi.org/10.5194/egusphere-2026-3365-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 103 | 98 | 16 | 217 | 34 | 16 | 17 |
- HTML: 103
- PDF: 98
- XML: 16
- Total: 217
- Supplement: 34
- BibTeX: 16
- EndNote: 17
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General comments
A strong and important paper documenting the first operational MethaneSAT XCH4 retrievals. The stripe-correction and validation work is careful overall. Comments below are mostly requests for clarification or additional detail; points 2 and 7 affect the headline accuracy claims and I'd like to see them addressed.
Specific comments
1. L4, spatial resolution. Native pixels are ~110×400 m² but precision is reported at 2×2/5×5 km at L14-16. A sentence on why this resolution was chosen, presumably to support cloud/aerosol screening and stripe/ISRF characterization rather than the reported precision itself, would help readers.
2. L19-20, Section 4.2, CO2 prior choice. The mean bias against TROPOMI is similar under either prior (~0.1 ppb), but the regression slope and hemispheric behavior differ substantially: under the operational GGG2020 prior the slope is 1.069, driven by an 8 ppb low bias in the Southern Hemisphere (vs. +2 ppb in the North); under CAMS the SH bias drops to 1.4 ppb and slope to 0.991. The abstract's slope figure of 0.99 is explicitly tied to the CAMS substitution. Please state more clearly in the abstract/conclusions which prior underlies which reported number.
3. L106-108, dataset selection. Please clarify how the 340 scenes were selected from the 1160 processed to L2, and ideally show how they compare to the full dataset in albedo/SZA/region, so readers can judge how representative the validation statistics are.
4. Section 3.2, prior covariance tuning. γ_CO2 = 10 is selected to give "near unity PBL averaging kernel responses" in Fig. 2, a reasonable but qualitative criterion. A brief comment on sensitivity of the reported precision/bias numbers to this choice would strengthen the section.
5. Section 3 ordering. The DoFS-based quality filter in 3.1 depends on the retrieval's averaging kernels, which in turn depend on the prior covariance tuning introduced afterward in 3.2. Consider reordering for clarity.
6. Eq. 13-14, Figure 4. The precision estimator assumes "natural XCH4 variability occurs over spatial scales larger than a few pixels" (L270), which may not hold near plume edges. Since this feeds into homogeneous-scene selection and the stripe correction training set, a brief comment on sensitivity here would be useful.
7. Section 3.4, destriping. PLS alone reduces σ_b to 6.1 ppb; adding the wavelet step pushes it below the 3.5 ppb theoretical noise floor, which the text attributes to the band-pass filter suppressing some real Fourier modes along with the stripe noise. A wavelet-only run (without the PLS step) would help clarify how much each stage of the correction actually contributes to the final result.
8. Section 4, independent validation.The TROPOMI and L1/L2 self-consistency checks are useful but not fully independent of the retrieval framework. Was a comparison against controlled methane release data considered?
9. TCCON comparison. Since GINPUT/GGG2020 (used as the CO2 prior) is itself developed for TCCON (L130-131), a direct MethaneSAT–TCCON comparison could help clarify the prior-sensitivity question in point 2, where coverage allows.
10. Section 4.3, aerosol dependence. The Libya PICS analysis covers one aerosol regime (bright surface, mineral dust). It would help to know whether targets with other aerosol types are available to test the robustness of the CO2-proxy approach more broadly.
11. Data availability. Given the paper's role as a mission-performance demonstration, consider releasing the 340-scene validation subset (with homogeneous-scene flags) as a standalone dataset alongside the gated access form.
Minor
• Eq. 6: briefly define h in words on first use.
• L195-199: clarify whether the averaging kernels in Fig. 2 include the full state vector or just CH4/CO2 layers.
• Fig. 12: correlation coefficients per scene range 0.46-0.86; worth stating whether cloud-contaminated scenes are excluded from any summary statistics or only flagged qualitatively.