Methane retrieval from airborne HySpex observations in the short-wave infrared
Abstract. A reduction of methane emissions could help to mitigate global warming on a relatively short time scale. Monitoring of local and regional anthropogenic CH4 emissions is crucial in order to increase our understanding of the methane budget which is still subject to scientific debate.
The study compares various retrieval schemes that estimate localized CH4 emissions from ventilation shafts in the Upper Silesian Coal Basin (USCB) in Poland using short-wave infrared nadir observations of the airborne imaging spectrometer HySpex. The examined methods are divided into nonlinear and linear schemes. The former class are of iterative nature and encompass various nonlinear least squares setups while the latter are represented by the Matched Filter (MF), Singular Value Decomposition (SVD) and Spectral Signature Detection (SSD) algorithms. Particular emphasis is put on strategies to rem- edy the problem of albedo related biases due to correlation with broad band absorption features caused by the hyperspectral instrument's low spectral resolution.
It was found that classical nonlinear least squares fits based on the Beer InfraRed Retrieval Algorithm (BIRRA) suffers from surface-type dependent biases. The effect is more pronounced for retrievals from single spectral intervals but can be mitigated when multiple intervals are combined. The albedo related correlation is also found in the BIRRA solutions for the separable least squares. A new BIRRA setup that exploits the inverse of a scene's covariance structure to account for reflectivity statistics significantly reduces the albedo bias and enhances the CH4 signal so that the method infers two- to threefold higher methane concentrations.
The linear estimators turned out to be very fast and well suited to detect enhanced levels of methane. The linearized BIRRA forward model turned out to be sensitive to the selected retrieval interval and in the default setup only works for very narrow windows. Other well established linear methods such as the MF and SVD identified the methane pattern as well and largely agree with the BIRRA fitted enhancements hence the methods allow quantitative estimates of methane. The latter two methods yielded increased performance when the scene was further divided into clusters by applying k-means in a preprocessing step. Methane plumes detected with the simple SSD method were faint and found rather sensitive to the polynomial used to compute the method's residuum ratio.
Philipp Hochstaffl et al.
Status: final response (author comments only)
RC1: 'Comment on amt-2022-271', David R. Thompson, 12 Dec 2022
- AC1: 'Reply on RC1', Philipp Hochstaffl, 21 Apr 2023
RC2: 'Comment on amt-2022-271', Anonymous Referee #2, 15 Jan 2023
- AC2: 'Reply on RC2', Philipp Hochstaffl, 21 Apr 2023
RC3: 'Comment on amt-2022-271', Anonymous Referee #3, 01 Feb 2023
- AC3: 'Reply on RC3', Philipp Hochstaffl, 21 Apr 2023
Philipp Hochstaffl et al.
Philipp Hochstaffl et al.
Viewed (geographical distribution)
This article applies several different CH4 retrieval schemes to HySpex imaging spectrometer data from an anthropogenic point source. It compares a nonlinear algorithm, involving spectrum fitting using a nonlinear radiative transfer model, to various linear schemes. It also compares several algorithm variants, including pre-clustering data with k means and accounting for the covariance of surface reflectance in the nonlinear model.
There are some clear and obvious achievements. The algorithm descriptions are incredibly comprehensive, with enough detail to serve as a reference for investigators coding their own implementations. The manuscript deals with an important and timely topic, adding to the growing literature on point source GHG detection from coarse-spectral-resolution imaging spectrometers. It independently validates these approaches, and adds some new data sets to the mix.
This said, I have some recommendations for how the manuscript might be improved.
My first major recommendation is to clarify the thesis statement. After reading the article, I'm still a bit unclear on the fundamental contribution. The manuscript focuses mostly on retrofitting the BIERRA algorithm for CH4 point source detection at coarse spectral resolution. However, by the end of the manuscript it is unclear what advantages this offers beyond the state of practice Matched Filter methods or other very similar nonlinear model-fitting methods in common use (like the Thorpe et al. IMAP-DOAS approach). The affect of albedo on nonlinear CH4 retrievals is great, but it has been investigated even more thoroughly before - see for example Ayasse et al. 2019 (https://doi.org/10.1016/j.rse.2018.06.018). I think the authors could do a better job of calling out what is new and significant about the BIERRA approach.
My second major recommendation is to have a quantitative performance comparison. The current assessment is fairly subjective, related to the quality of the plume image and the visual appearance of background interference. Couldn't the background variability outside the plume be used to quantify the detection noise for each method? And couldn't the strength of the plume enhancements then be used to create an SNR score or statistical confidence? As a part of this effort, it would be great to translate all of the plume maps into similar units. Currently maps appear variously as ppbv, alpha-CH4, and "enhancement factors," which makes it difficult to inter-compare. It should be possible to translate any one of the linear detection algorithm results into an equivalent CH4 mass enhancement, and compare the effective plume-to-background detection SNR of each of the algorithms.
1. Almost all of the prior literature cited on CH4 point source detection, and the vast majority of the imaging spectroscopy community working at these spectral resolutions, plot spectra in wavelength rather than wavenumber. Setting aside the question of which convention is more convenient or appropriate from a technical perspective, it would certainly be easier for the majority of the readership to quickly understand the figures if wavelengths were used. This would make the instrument sampling evenly-spaced in the horizontal direction.
2. On line 61, are there citations for CarbonMapper or CO2Image missions? I think the claim that CarbonMapper operates at higher spectral resolution than average land surface imaging spectrometers (5-10nm sampling) could be incorrect.
3. On line 65, in the literature review of airborne CH4 point source campaigns, consider also the studies by Frankenberg et al. 2016 (https://doi.org/10.1073/pnas.1605617113) and Duren et al. 2019 (https://doi.org/10.1038/s41586-019-1720-3) which were earlier and larger.
4. Figure 2 (a) seems to be missing some lines. "Grass" is misspelled.
5. On line 183, can you provide any more specifics about the low order polynomial used? What was its degree and where was it centered? The details are significant because, as you note, the surface reflectance is often quite complex over these wide spectral ranges. Changes in the reflectance representation can have huge changes on albedo sensitivity.
6. On line 195, the section comparing different least squares solvers seems ancillary. Least squares solvers are a commodity
7. Figure 10. I'm not sure what this is supposed to show. Could it be removed?
8. On Figure 12, the enhancement within the plume appears completely saturated, which makes it difficult to assess. Can you rescale the colormap to make it more similar to the other plume images?
9. The discussion and conclusion is good, but would be further strengthened by quantitative claims about where and how the different algorithms outperform each other.