the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Sep 2024
Robustness of atmospheric trace gas retrievals obtained from low spectral resolution Fourier-transform infrared absorption spectra
Abstract. For atmospheric trace gas columns retrievals obtained from ground based Fourier-transform interferometer spectra we study the sensitivity of the retrieval processing chain to changes in the number of points in the recorded interferograms. Shortening an interferogram will alter the leakage pattern in the associated spectrum and we demonstrate that the removal of a relatively small number of points from the interferogram edges creates a beat pattern in the difference of the associated spectra obtained from the original and shortened interferogram. For low-resolution interferometers the beat pattern in the spectra may exceed the noise level and the effect on atmospheric gas column retrievals may be large. Sensitivity of the retrieval algorithm to the length of the underlying interferogram can be reduced by applying a non-trivial apodization such as Norton-Beer. A case study shows the effect on formaldehyde retrievals obtained from low-resolution spectra in Sodankylä and Kolkata.
- Preprint
(856 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-2764', Anonymous Referee #1, 17 Oct 2024
In their paper "Robustness of atmospheric trace gas retrievals obtained from low spectral resolution Fourier-transform infrared absorption spectra" Langerock et al. investigate the effects of removing sample points in recorded interferograms from Fourier-Transform infrared spectrometers, which are widely used in ground-based remote sensing. Thus the topic fits the scope of AMT and the results of the paper are of importance to the processing chains of ground-based remote sensing networks, such as NDACC, COCCON and TCCON.
However, I feel several important details need to be addressed:
- It is not clear why the interferogram should be truncated at all and if for some reason it is, why can't the correct number of points be used to determine the actual length of the interferogram. Clearly it is possible to use interferograms recorded at different maxOPD. How is this different?
- A discussion on the impact of the proposed apodization on the information loss in the retrieval is missing. Amato et al. 1998 only show that the retrieval is unaffected by the choice of the apodization function if a weak function is used. "In practice, heavy apodization moves the truncation point in the interferogram closer to the origin, and therefore it decreases the maximum path delay L, which inevitably causes a loss of information in the spectral range." (Amato98)
What happens if the apodization is strong enough to suppress the information under the noise limit at the end of the interferogram? Doesn't the difference in Fig 6a between (BX, d=0) and (NBS, d=0) show that there is an impact of the apodization without the issue of missing points?- Section 2 line 38 - I assume instead of time domain the spatial domain is meant? As in spectrum is spatial frequency and interferogram is derived from movement in space? So you are talking about DFT( DFT(I)-DFT(I_bandpass)) ? This part needs clarification/rephrasing. The time domain analogy is used several times and should be changed to avoid confusion.
- line 50 - and Fig 3 - I don't understand the ramp function. Is the goal to reduce the effect of the missing points by setting the long side of the interferogram gradually to zero? If so, shouldn't it be then 1 at the start and going down to zero. And you claim in line 50, that applying the ramp removes the impact of the point removal. Fig 3b shows otherwise, the beat is still at 2 %. In general the description of the actual steps done could be beneficial for understanding.
- Fig 6 and paragraph line 108 - It is reasonable to assume, that a stronger Norton-Beer apodization mitigates the effects, but at the same time you would loose information. Fig. 6a in particular shows that the retrieved H2CO columns are very different, but a discussion on the causes is missing. Is it due to the described beat pattern on the spectrum or due to loss of information due to a strongly apodized interferogram? I would argue, that especially in a low resolution scenario, the information contained in the tail end of the interferogram contains much of the information used in the retrieval, esp. information on the spectral lineshape. This effect would be smaller with high res interferograms/spectra. This is connected to the second comment above.
A minor detail:
line 14: TCCON uses the term dry-air mole fraction (DMF) instead of vertically integrated concentrations, as it better represents the approach of dividing by the O2 column.
Citation: https://doi.org/10.5194/egusphere-2024-2764-RC1 -
AC1: 'Reply on RC1', Bavo Langerock, 03 Dec 2024
- It is not clear why the interferogram should be truncated at all …Compared to the underlying theoretical infinite interferogram, the measured interferogram is always truncated due to the finite length of the interferometer and this truncation causes the leakage in the spectra [Herres 1984]. In the paper we study the effect of shortening the interferogram with a relatively low number of points and the associated change in the leakage pattern.The examples given in line 19 on page 1 should give a motivation why we considered the effect of shortening: eg a convolution of the interferogram may introduce a shortening with half the size of the convolution window. The Forman phase correction uses a convolution in the time(=spatial) domain to reduce asymmetry in the interferogram. The Forman phase correction is eg used in TCCON and the typical size of the Forman correction window is 2**10 points (see https://github.com/TCCON/GGG/blob/main/src/i2s/opus-i2s.example.in, line 194).We would like to emphasize that the question “why do shortening” is not really relevant. It is answered implicitly because we show that the retrieved targets are sensitive to shortening of the underlying interferogram. Using the actual number of points or slightly less may result in different retrieval results and to our knowledge there is no way to tell which retrieval result is closer to the truth.… and if for some reason it is, why can’t the correct number of points be used to determine the actual length of the interferogram. Clearly it is possible to use interferograms recorded at different maxOPD. How is this different?The max OPD setting for a measurement is related to the number of points in the interferogram because the distance in cm between two registered points is related to the laser wavenumber [Herres 1984]. The maxOPD (a rounded number in cm) is only a rough estimate for the actual number of points (laser wavenumber scale). The instruments’ software will determine how many points should be registered given a user defined max OPD.The zero filling of an interferogram [Herres 1984] will pad the interferogram with zeros to bring the number of points to a power of 2 (for the FFT algorithm). This means that a shortening of the interferogram will not have an effect on the number of points in the spectral domain. This comment was added to the introduction in the revised version.The standard retrieval proces requires the calculation of a modelled spectrum. To mimic the leakage pattern, the max OPD is used on the simulated spectrum (see the discussion on p6, from line 77, which is extended in the revised version). Here it is important to note that the retrieval software uses a smaller spectral window inside the full measured spectral range. The exact total number of points in the measured spectrum is not used in the retrieval iteration in order to keep oa the required computational resources manageable. The leakage pattern is introduced in the simulated spectrum by means of the max OPD which is only an approximation of the actual number of measured points in the interferogram.- A discussion on the impact of the proposed apodization on the information loss in the retrieval is missing.We agree that this is an important aspect to consider in the case of profile retrievals. However, the paper only discusses column retrievals where this is less crucial. Fig 7 should convince the readers that in the NBS apodized time series, the H2CO signal is well captured when compared to the 125HR NDACC columns. As the prior in the retrieval is constant, Fig 7 shows that the information loss due to apodization is not important for the column retrievals shown in the paper. We have added a comment on this in §4 in the revised version.- Section 2 line 38 - I assume instead of time domain the spatial domain is meant?In literature both “time” and “spatial” domain are used to indicate the other domain opposed to “spectral". See for example Forman 1966 where time domain is used. We have added a footnote to remind the reader about this.- line 50 - and Fig 3 - I don't understand the ramp functionThe multiplication of the interferogram with the ramp function is a standard technique in FTIR processing applied to single sided interferograms, see [Herres 1984]. Single sided interferograms are not symmetric (short vs long arm) and hence will always return complex spectra under the DFT. The technique ensures that the real part of the DFT of the ramp apodized interferogram coincides with the (unknown symmetric) double sided interferogram.- And you claim in line 50, that applying the ramp removes the impact of the point removal. Fig 3b shows otherwise, the beat is still at 2 %. In general the description of the actual steps done could be beneficial for understanding.in line 50 we claim that the ramp correction annihilates the effect of the removal of points on the left shorter arm only (where the ramp correction is close to zero). Removal of points from the right longer arm does have an effect as is shown in Fig 3b.- Fig 6 and paragraph line 108 … but a discussion on the causes is missingWe have formulated the cause in §4: it is related to how the modelled interferogram is truncated to the measured opd with an integer division. In the updated version we have extended this and emphasized in §6 that the modeling of the leaking pattern is not accurate enough. The abstract is re-phrased.- TCCON DMF versus column: In the introduction we wanted to emphasize that NDACC provides profile concentration while TCCON/COCCON provide vertically integrated measurements. In the revised version we now use DMF.Citation: https://doi.org/
10.5194/egusphere-2024-2764-AC1
-
AC1: 'Reply on RC1', Bavo Langerock, 03 Dec 2024
-
RC2: 'Comment on egusphere-2024-2764', Anonymous Referee #2, 30 Oct 2024
In this study, Langerock et al. examine the impact of varying the number of data points in interferograms recorded by FTIR instruments, which are commonly used for atmospheric composition measurements in networks like NDACC, TCCON, and low-resolution instruments. They present a case study analyzing the effects of low-resolution spectra in Sodankylä and Kolkata. The findings are pertinent to ground-based remote sensing networks and also to studies using low-resolution data from aircraft, making the paper relevant to AMT’s scope. Below are some comments that I suggest need to be addressed.
The manuscript lacks a clear overall motivation for conducting this work. What prompted the authors to undertake this analysis? I recommend providing a more detailed explanation of the study's significance in the introduction, along with a clear and in depth description of current practices in FTIR observations (both high and low resolution). Specifically, do all instruments perform the DFT individually, or do they rely on standard routines or software? What are these softwares actually doing? what are these routines typically designed to accomplish?
I suggest including more details on the standard procedures typically used by low-resolution instruments; for instance, how the network COCCON and others handle the DFT. If feasible, I recommend presenting results based on these common practices and providing an overarching set of recommendations.
The examples and applications presented here rely on solar absorption, which is valuable; however, I also recommend assessing the results using control-based cell spectra. Do you observe consistent findings, and what impact might the instrument line shape have?
Abstract: Opening Sentence (the first sentence is quite broad) and not clear. Do you mean something like this: “This study examines the sensitivity of atmospheric trace gas column retrievals from ground-based Fourier-transform interferometer spectra to variations in the number of points in recorded interferograms”.
The abstract is currently brief, leaving room for further improvement. I strongly recommend adding a motivating statement to highlight the importance of this work, along with quantitative findings on the formaldehyde results for a more comprehensive summary.
The introduction mentions the FTIR resolutions used by NDACC, TCCON, and low-resolution instruments, but does not address the significance of these differences. Is there a notable impact of using different resolutions? Can low-resolution instruments achieve the same gas retrievals as high-resolution ones?
In the second paragraph of the introduction, instead of starting directly with the removal of points, I suggest first explaining what determines the number of points in an interferogram. This can be followed by a description of both the removal of points and the process of zero-filling, along with guidance on when each is recommended.
In Figure 1 the noise level in the HR-FTIR appears higher than in the low-resolution instrument in terms of intensity [a.u.]. Could you consider including the signal-to-noise ratio for each spectrum?
In the introduction, you present the difference between spectra measured with and without the removal of points from the interferogram tails, yet the motivation behind this step is not explained. Why is it necessary to remove points from the interferogram? How are the original spectra generated, and is there a standard procedure for this? Additionally, is specific software used for these calculations, or is it a common practice to generate spectra directly from interferograms? Clarifying these points would offer valuable context for the reader.
Figure 6. The magnitude and shape of the diurnal variation using BX differ noticeably; is there an explanation for this? I also suggest adding the diurnal variation of HCHO from the HR-FTIR for comparison and the quantitative comparison/results.
Figure 6. I do not see green points on either plots a or b.
P4, L66, suggest to change “clean site” with remote
P6, L99. Is Sodankyla an official NDACC station? I could not find the data online.
Section 4 lacks details about the standard procedures typically employed for these measurements. For instance, what method does COCCON use for the DFT? I recommend including this information and, if possible, presenting results using the same retrievals as COCCON to provide a comprehensive and meaningful contribution to the community.
Is Figure 7 based on observations conducted in Sodankylä? It's intriguing that the magnitude depicted in this figure does not align with the values from either Sodankylä or Kolkata shown in Figure 6. For instance, the maximum values for Vertex 70 in Figure 7 are approximately 10, whereas the maximum values for Sodankylä and Kolkata in Figure 6 are around 3 and 40, respectively.
In my opinion, conclusion can be expanded with quantitative findings and overall effect in other gases.
Citation: https://doi.org/10.5194/egusphere-2024-2764-RC2 -
AC2: 'Reply on RC2', Bavo Langerock, 03 Dec 2024
- The manuscript lacks a clear overall motivation for conducting this work. What prompted the authors to undertake this analysis?The motivation is written in the introduction on page 1 line 19. Examples of interferogram shortening in the standard FTIR data processing are listed: eg convolution shortens the interferogram.
- I recommend providing a more detailed explanation of the study's significance in the introduction, along with a clear and in depth description of current practices in FTIR observations (both high and low resolution). Specifically, do all instruments perform the DFT individually, or do they rely on standard routines or software? What are these softwares actually doing? what are these routines typically designed to accomplish?
- I suggest including more details on the standard procedures typically used by low-resolution instruments; for instance, how the network COCCON and others handle the DFT.
- The examples and applications presented here rely on solar absorption, which is valuable; however, I also recommend assessing the results using control-based cell spectra. Do you observe consistent findings, and what impact might the instrument line shape have?
We understand from several comments that the referee would like to have more details in the paper about COCCON/TCCON processing, details of the FTIR data processing, expanding the results to other light sources (cell measurements), a discussion on the influence of increased resolution in the retrieval, etc.
We believe that these suggestions would be more appropriate for a review paper. This was however never the intention of the paper and we consider these suggestions therefore out of scope . Our purpose is to show by means of a case study that the retrieval processing chain may be sensitive to shortening of the interferogram (or equivalently, sensitive to changes in the leakage pattern) and to demonstrate how the retrieval processing can be made more robust by enabling apodization.- If feasible, I recommend presenting results based on these common practices and providing an overarching set of recommendations.We have added in the conclusions a more general recommendation, but due to the complexity with many different influence parameters we believe that an out of the box list of recommendations is not feasible.- Abstract: Opening Sentence (the first sentence is quite broad) and not clear. Do you mean something like this: “This study examines the sensitivity of atmospheric trace gas column retrievals from ground-based Fourier-transform interferometer spectra to variations in the number of points in recorded interferograms”.We have made an adjustment to this phrase.- The abstract is currently brief, leaving room for further improvement. I strongly recommend adding a motivating statement to highlight the importance of this work, along with quantitative findings on the formaldehyde results for a more comprehensive summary.We added a sentence in the abstract mentioning the increase in correlation for the Sodankyla case study.- The introduction mentions the FTIR resolutions used by NDACC, TCCON, and low-resolution instruments, but does not address the significance of these differences. Is there a notable impact of using different resolutions? Can low-resolution instruments achieve the same gas retrievals as high-resolution ones?It is mentioned that NDACC higher resolution measurements allow the retrieval of vertical profiles, while this is not achieved for TCCON/COCCON type of measurements. A detailed discussion on the different networks can be found in eg https://doi.org/10.5194/amt-12-5979-2019 or https://doi.org/10.5194/amt-9-2223-2016.- In the second paragraph of the introduction, instead of starting directly with the removal of points, I suggest first explaining what determines the number of points in an interferogram. This can be followed by a description of both the removal of points and the process of zero-filling, along with guidance on when each is recommended.
This remark is related to what we assume a reader is familiar with. This is made more clear in the introduction.- In Figure 1 the noise level in the HR-FTIR appears higher than in the low-resolution instrument in terms of intensity [a.u.]. Could you consider including the signal-to-noise ratio for each spectrum?
The noise is indeed much higher in the HR spectrum (almost a factor of 10). When measuring low-resolution IR spectra, a larger entrance aperture can be used in the FTIR spectrometer allowing more light to enter the interferometer and resulting in a higher signal-to-noise ratio, additionally, the HR instrument was set up to average 6 scans while the LR instrument averages 12 scans. Besides this there is also the fact that the noise in the LR spectra is lower because the interferogram signal is less noisy near the centre burst. This observation can also be related to the fact that the shortening of the LR interferogram has higher amplitude in the beat pattern (more energy closer to the centre burst) as mentioned in the conclusions. We believe the key message from Fig 1 is that shortening introduces a spectral difference that exceeds the noise level for the LR spectra. We prefer not to quantify the noise level, nor the amplitude of the beat pattern (there are too many influence parameters for the latter).- In the introduction, you present the difference between spectra measured with and without the removal of points from the interferogram tails, yet the motivation behind this step is not explained. Why is it necessary to remove points from the interferogram? How are the original spectra generated, and is there a standard procedure for this? Additionally, is specific software used for these calculations, or is it a common practice to generate spectra directly from interferograms? Clarifying these points would offer valuable context for the reader.This is related to a previous comment: the motivation is in the introduction.- Figure 6. The magnitude and shape of the diurnal variation using BX differ noticeably; is there an explanation for this? I also suggest adding the diurnal variation of HCHO from the HR-FTIR for comparison and the quantitative comparison/results.
We think the most plausible cause is the diurnal variation in the SNR. As we find it hard to quantify the influence of a shortening on the diurnal variation, we prefer to not to include it in the paper.- Figure 6. I do not see green points on either plots a or b.
It was intended to not show NBM and NBW in Fig 6(a) and (c): they make the plot too crowded and do not bring additional information.- P4, L66, suggest to change “clean site” with remotedone
- P6, L99. Is Sodankyla an official NDACC station? I could not find the data online.Indeed, Sodankyla is not an NDACC affiliated station. However, the instrument operates partly in NDACC mode and NDACC allows the publication of this non-consolidated data in the “rapid delivery” folder on NDACC (see https://ndacc.larc.nasa.gov/data under "Rapid Delivery data”). This is now explicitly mentioned in the revised version- Section 4 lacks details about the standard procedures typically employed for these measurements. For instance, what method does COCCON use for the DFT? I recommend including this information and, if possible, presenting results using the same retrievals as COCCON to provide a comprehensive and meaningful contribution to the community.It is not entirely clear what is meant. The DFT is a fixed algorithm, in which choices can be made (zero filling factor, phase correction, apodisation). In (Sha 2020 and 2024) the Sodankyla campaign is discussed and includes a comparison between the TCCON/COCCON retrieval software for this type of LR spectra.
- Is Figure 7 based on observations conducted in Sodankylä? It's intriguing that the magnitude depicted in this figure does not align with the values from either Sodankylä or Kolkata shown in Figure 6. For instance, the maximum values for Vertex 70 in Figure 7 are approximately 10, whereas the maximum values for Sodankylä and Kolkata in Figure 6 are around 3 and 40, respectively.Figure 7 shows the columns for all spectra in 2019 campaign at Sodankyla while Figure 6 only shows one day (August 24 2019), as indicated in the figure captions. The Vertex columns from Figure 6 are shown on the y-axis in Figure 7. We adapted the Fig.7 caption to explicitly mention it shows the Sodankyla campaign data.Citation: https://doi.org/10.5194/egusphere-2024-2764-AC2
-
AC2: 'Reply on RC2', Bavo Langerock, 03 Dec 2024
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 172 | 48 | 12 | 232 | 1 | 3 |
- HTML: 172
- PDF: 48
- XML: 12
- Total: 232
- BibTeX: 1
- EndNote: 3
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1