| 17 Jul 2026
Evaluating CH4 retrieval methods for hyperspectral images from EnMAP
Abstract. The Environmental Mapping and Analysis Program (EnMAP) satellite carries a hyperspectral instrument that is able to measure methane (CH4) plumes of localized hotspot sources from which facility-scale emission rates can be estimated. Here, we implement and evaluate three retrieval techniques for scenes with different complexity and source strengths and discuss the differences and implications for emission estimation. These techniques are (i) RemoTeC, a physics-based retrieval that relies on pixel-wise radiative transfer calculations; (ii) a matched-filter retrieval that exploits the spectral covariance of a scene to detect CH4 enhancements; and (iii) a hybrid retrieval that incorporates spectral covariance information into RemoTeC, thereby combining the strengths of approaches i) and ii).
RemoTeC and the hybrid method yield similar emission estimates, with the hybrid method exhibiting lower statistical noise. The enhancements retrieved by the matched filter have the lowest statistical noise. For sources with emission rates larger than ∼ 3 t h−1, the matched filter yields source strength estimates similar to the physics-based retrievals. For weaker sources, the matched filter estimates larger emission rates than the physics-based retrievals, with deviations being on the order of the emission rate itself. While the matched filter effectively suppresses regular and recurring spectral albedo structures, its performance degrades in the presence of statistically rare spectral features. In contrast, the hybrid retrieval more reliably accounts for such albedo-induced artifacts.
Our results demonstrate that retrieval methodology can significantly influence methane emission estimates, particularly in challenging scenes. The matched filter is well suited for rapid quantification of strong emission sources, whereas the physics-based approaches provide greater robustness under difficult observational conditions and for weak emitters. The hybrid retrieval offers the best overall performance by combining the mechanistic rigor of radiative transfer modeling with the noise-reduction benefits of covariance-based methods.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Atmospheric Measurement Techniques.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
The study by Scheidweiler et al. intercompares three methods for the retrieval of methane concentration enhancements from hyperspectral satellite data. The selected retrievals are the widely-used matched-filter (MF) data-driven method, the RemoTeC full-physics retrieval, which is well-known to the authors, and a variation of RemoTeC using spectral covariance information to constrain the retrieval. A number of datasets acquired by the hyperspectral mission EnMAP over methane hotspots around the world were processed with those three retrieval methods. The results were intercompared in terms source strength quantification, retrieval artefacts, noise performance, and some other observation-related challenges.
In my opinion, this is a well-thought study, with several important findings and conclusions despite a somewhat late timing for the rapidly evolving field of methane plume retrievals. The document is also well written and presented, and I have definetely enjoyed reading it.
The only major comment I have is related to the low/high offsets found for the MF with respect to the other two retrievals (e.g. L9 “For sources with emission rates larger than ∼ 3 t h−1 , the matched filter yields source strength estimates similar to the physics-based retrievals. For weaker sources, the matched filter estimates larger emission rates than the physics-based retrievals”). Those trends do not seem to agree with the expectations, namely that the MF based on Foote et al. underestimates for large XCH4 due to the linear assumption, while it works fine for smaller emission rates. Could it be that the RemoTeC and hybrid retrievals also present systematic errors that partly explain the trends obseved in the MF? To investigate this issue, one could use publicly-available controlled release data from the controlled release experiments run by Stanford U. in the last years (https://www.researchsquare.com/article/rs-9110475/v1, https://amt.copernicus.org/articles/17/765/2024/). Taking controlled release data as a reference, one could extract more information about the potential high/low biases of each methods, for Qs <1.5 t/h. I would definitely encourage the authors to add this exercise to their study.
Other minor comments:
L51: I don’t think soil moisture is a primary variable for “point source imagers” (or for any optical instrument). Perhaps “mineral composition, vegetation biophysical variables, and water components” would be a better list.
L60: “have a spectral resolution of 5-10 nm, that does not allow…”
L62: I don’t think that physics-based algorithms are “commonly used” with high resolution hyperspectral instruments
L72: Guanter et al. 2015 does not deal with trace gas retrievals
L85-86: I don’t understand “ the horizontal extent of the emission plume is resolved by the high spatial resolution of the image, requiring less accuracy than for area mappers of background CH4”. Scattering will still affect the retrieved XCH4 values within the plume, I think?
L115: the “albedo polynomial” has not been introduced yet
L127: I think “column” is more widely used than “line” in this context
L148: on optically-thick lines: the log-version of the MF (e.g. Pei et al. https://www.sciencedirect.com/science/article/abs/pii/S0034425723002031) has been shown to compensate for the linear assumption in the basic MF. Could you consider to include some tests with th log-MF version (see the related major comment above)
L157: I would write “column-wise” for operations per across-track sample.
L183: the log-version of the MF would compensate for this
L189: is the same 0.95 factor applied to the \Sigma_f of the MF retrieval?
L252: Valverde et al. write it as “Darvaza”
L305: given the central point of the covariance matrix in the performance of the MF and hybrid retrievals, I think it would be useful to see how they change for different surfaces with “good” and “bad” behavior
Fig. 5, caption: how is the uncertainty associated with the Q values calculated?
Fig. 6 and related text: I was confused by the references to an “albedo study” that hadn’t been described yet. It may be worth considering to add an extra figure later with only the results of the albedo study.
L334: “since it uses”
L363: perhaps Guanter et al., 2021, no 2015