the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 29 Aug 2023
Simultaneous retrieval of aerosol and ocean properties from PACE HARP2 with uncertainty assessment using cascading neural network radiative transfer models
Abstract. The UMBC Hyper-Angular Rainbow Polarimeter (HARP2) will be onboard NASA’s Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission, scheduled for launch in January 2024. In this study we systematically evaluate the retrievability and uncertainty of aerosol and ocean parameters from HARP2 multi-angle polarimeter (MAP) measurements. To reduce the computational demand of MAP-based retrievals and maximize data processing throughput, we developed improved neural network (NN) forward models for space-borne HARP2 measurements over a coupled atmosphere and ocean system within the FastMAPOL retrieval algorithm. A cascading retrieval scheme is further implemented in FastMAPOL, which leverages a series of NN models of varying size, speed, and accuracy to optimize performance. A full day of global synthetic HARP2 data was generated and used to test various retrieval parameters including aerosol microphysical and optical properties, aerosol layer height, ocean surface wind speed, and ocean chlorophyll-a concentration. To assess retrieval quality, pixel-wise retrieval uncertainties were derived from the Jacobians of the cost function and evaluated against the difference between the retrieval parameters and truth based on a Monte Carlo error propagation method. We found that the fine-mode aerosol properties can be retrieved well from the HARP2 data, though the coarse-mode aerosol properties are more uncertain. Larger uncertainties are also associated with a reduced number of available viewing angles, which typically occurs near the scan edge of the HARP2 instrument. Results of the performance assessment demonstrate that the algorithm is a viable approach for operational application to HARP2 data after PACE launch.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(5997 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(5997 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-1843', Anonymous Referee #1, 15 Sep 2023
For the upcoming launch of the HARP2 instrument on the PACE mission, Gao et al. evaluate the retrieval ability and uncertainty of aerosol and ocean parameters. To reduce the computational demand of the retrievals and maximize data processing throughput, they developed improved neural network forward models. To this end, a cascading retrieval scheme is implemented in the retrieval algorithm, which leverages a series of neural network models of varying size, speed and accuracy to optimize performance. Using the new retrieval scheme, one day of global synthetic data was retrieved and the quality assessed. The authors find that the fine-mode aerosol properties can be retrieved well, but the coarse-mode aerosol properties are more uncertain.
The manuscript is generally well written and due to the expertise of the authors I also have no doubts on the presented results. However, I think in order that the readers that have not read your previous publications can follow the study presented here, some more information throughout the paper is needed as provided by my comments in the following.
General comments:
- You use plural for the usage of the neural network model (e.g. title and abstract). However, I was wondering if you do not rather have one neural network model, but run it with different initial parameters and thus perform several neural network model “simulations”?
- You present here an improved version of your current retrieval algorithm for the MAP instrument. First, I thought it is the implementation of the neural network model itself, but then later you write that you improved the neural network model. So what neural network model e.g. has been used previously? What actually are the differences between the old and new retrieval algorithm? That does not become entirely clear and should be improved throughout the manuscript.
- Motivation for the improvement of the retrieval algorithm is the reduction of the computational demand and to maximize data processing throughput. However, nowhere in the manuscript I could find a statement how large the improvement actually was. How faster is the new retrieval algorithm?
- Similar to previous comment, but now on the retrieval results. What are the differences in e.g. accuracy of the old and new retrieval algorithm?
- The retrieval tests have been done here for a synthetic data set. What can you expect when HARP2 is launched? What are the uncertainties and difficulties/challenges for the upcoming real retrieval? How close is your set-up to the real atmosphere/atmospheric conditions?
- Before submission of the revised version the manuscript should be more carefully checked. There are a lot of technical errors that could have been avoided and removed by the authors before submission of the manuscript.
Specific comments:
P1, title: Is singular really correct? Are you using one neural network model and preform several simulations, or are you really use several neural network models?
P2-4: The introduction is too long and not really easy to follow. Text from P3, L61 to L79 should be significantly shortened. Here you actually describe the differences between the old and new retrieval scheme, but the details belong rather to the method section than to the introduction. Further, the description of your new retrieval code after shortening (2-3 sentences) should appear rather at the end of the introduction.
Another comment on the introduction or the manuscript in general: Do you really need that many references? Especially reading the introduction becomes quite tough. It is not necessary to reference every study that ever has been published. Having a reference list of 7 pages for a technical paper is quite a lot and in my opinion somewhat too much.
P5, L117: Where did you get these expected values from? Where these derived in the present study or documented somewhere else? Add reference?
P12, L247: Has the day of 21 March 2022 chosen by purpose (reason?) or arbitrarily? How would the results look like for another day, especially another day in another season?
P18, L319: “Based in the retrieval results shown in Fig 6-8.” This sentence is not complete. Please correct.
P18, L323: Something is missing in this sentence. Maybe “that”? Should it read “To verify that the theoretical retrieval………”?
P18, L332 approx.: Here, the same text is repeated a second time with slight differences. Please omit one version.
P19, L340: The abbreviation ”RSP” has not been introduced.
P19, L345: m/s should be written as m s-1 (according to Copernicus guidelines).
P21, L374: ….radiative transfer behind and……. The sentence makes no sense. Please correct/rephrase.
P21, L377: You use computer performance as a motivation for your study and also mention this in the conclusion, but nowhere throughout the text it is discussed if you actually derive an improvement and how large this improvement is.
P21, L380: Here you provide the actual time needed to process one 5 min granule. However, more interesting it would be how much time is needed to process one day. On page P12, L251 it is stated that 150 granules in 15 orbits are yielded. If the processing of one granule takes 5 h, then processing of one day would take 825 h!? If yes, that would be still incredibly long and maybe much too long for retrieving global data from MAP. So in that case I would not call it feasible at all.
P21, L390: Abbreviation BRDF has not been introduced.
P21, L395: The uncertainties (e.g.numbers, magnitudes) should be given in the conclusions.
General comment on the text: Too many self-citations. You do not need to cite one of your publications in every second sentence. From introduction it became clear that you have done a lot of work already before writing up this study. So reduce the number of occasions and use references to your previous studies only where really necessary.
Technical corrections:
P1, L6: Start sentence with “To this end,” and delete further.
P1, L13: delete “also”
P2, L45: Closing parenthesis after the reference is missing (you need a second one since this text part is in parentheses)
P3, L52: “model” here obsolete -> delete
P3, L57: space between “retrievals” and full stop obsolete.
P3, L76: space between “networks” and the reference “Gao et al.” missing.
P4, L91: chlorophyll a -> chlorophyll-a
P4, L104: The abbreviation DoLP has not been introduced.
P5, L113: “with” should be rather read “whereby”. Maybe it would be better to split this sentence into two sentences.
P5, L116: the “m” in the sigma should be in subscript.
P6, Figure 1 caption: but not -> but are not
P7, L156: chlorophyll a -> chlorophyll-a and chla -> chl-a. Further, units should be written with upright font.
P7, L163: I would suggest to write “Chl-a” instead “Chla” throughout the manuscript. Check all occasions and correct these.
P8, L172: Sentence incomplete. Something is missing here; maybe “is used”?
P8, L172: Add comma after “Note”.
P9, L177: What do you mean with Sunglint? Do you mean “sunlit”? This should be corrected throughput the manuscript.
P10, L230: DOLP -> DoLP
P11, Figure 2 caption: DOLP -> DoLP
P11, L244: Correct reference “J. P., 1987”.
P12, L253: 40 o -> 40°
P12, L254: space before the comma obsolete.
P13, L260 and 263: remove obsolete space before the respective full stop of the sentence.
P14, L269: Add comma after “Note”.
P14, L275: sunlingt -> sunlit
P14, Figure 2 caption: Sec 2 -> Sect. 2
P15, L292: Add comma before “respectively”.
P15, L301: Fig 5 -> Fig. 5
P16, L310ff: Units should be in upright font (according to my knowledge of the Copernicus guidelines) and add a full stop between “Fig” and the respective figure number.
Citation: https://doi.org/10.5194/egusphere-2023-1843-RC1 - AC1: 'Reply on RC1', Meng Gao, 07 Oct 2023
-
RC2: 'Comment on egusphere-2023-1843', Anonymous Referee #2, 18 Sep 2023
General comments:
This study trained a new NN model through measurement uncertainty-aware training and Training data augmentation. The new NN model was used to generate pseudo HARP2 observations and retrieve both aerosol and ocean properties. The methods and results are reasonable, and the manuscript is well written. I have only a few confusions that needs to be clarified.
Specific comments:
- For Equation (3), some terms are not explained. I think the terms with f superscript is NN simulation and the terms without f superscript are pseudo-observations. Please confirm it or correct me.
- For Equations (5) and (6), every term should be explained. Is the uncertainty of DoLP a constant (0.005)? If so, what is difference between Equation (6) and conventional MSE cost function. It seems Equation (6) is just a conventional MSE cost function multiplied by a constant. If uncertainties of DoLP are not a constant in Equation (6), how they are quantified?
- The performance of the NN model is not well validated. Figure 2 has shown the cost function of training and validation, but readers cannot tell if the accuracy of the NN model is sufficient for simulation and retrieval. In this study, observations are generated by the NN model and the NN model is used for retrieval. Thus, it is important to compare the performance of the NN model with that of the radiative transfer model.
Technical corrections:
Line 116: σm. m should be subscript.
Line 189: (Gao et al., 2021a) -> Gao et al. (2021a)
Citation: https://doi.org/10.5194/egusphere-2023-1843-RC2 - AC2: 'Reply on RC2', Meng Gao, 07 Oct 2023
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2023-1843', Anonymous Referee #1, 15 Sep 2023
For the upcoming launch of the HARP2 instrument on the PACE mission, Gao et al. evaluate the retrieval ability and uncertainty of aerosol and ocean parameters. To reduce the computational demand of the retrievals and maximize data processing throughput, they developed improved neural network forward models. To this end, a cascading retrieval scheme is implemented in the retrieval algorithm, which leverages a series of neural network models of varying size, speed and accuracy to optimize performance. Using the new retrieval scheme, one day of global synthetic data was retrieved and the quality assessed. The authors find that the fine-mode aerosol properties can be retrieved well, but the coarse-mode aerosol properties are more uncertain.
The manuscript is generally well written and due to the expertise of the authors I also have no doubts on the presented results. However, I think in order that the readers that have not read your previous publications can follow the study presented here, some more information throughout the paper is needed as provided by my comments in the following.
General comments:
- You use plural for the usage of the neural network model (e.g. title and abstract). However, I was wondering if you do not rather have one neural network model, but run it with different initial parameters and thus perform several neural network model “simulations”?
- You present here an improved version of your current retrieval algorithm for the MAP instrument. First, I thought it is the implementation of the neural network model itself, but then later you write that you improved the neural network model. So what neural network model e.g. has been used previously? What actually are the differences between the old and new retrieval algorithm? That does not become entirely clear and should be improved throughout the manuscript.
- Motivation for the improvement of the retrieval algorithm is the reduction of the computational demand and to maximize data processing throughput. However, nowhere in the manuscript I could find a statement how large the improvement actually was. How faster is the new retrieval algorithm?
- Similar to previous comment, but now on the retrieval results. What are the differences in e.g. accuracy of the old and new retrieval algorithm?
- The retrieval tests have been done here for a synthetic data set. What can you expect when HARP2 is launched? What are the uncertainties and difficulties/challenges for the upcoming real retrieval? How close is your set-up to the real atmosphere/atmospheric conditions?
- Before submission of the revised version the manuscript should be more carefully checked. There are a lot of technical errors that could have been avoided and removed by the authors before submission of the manuscript.
Specific comments:
P1, title: Is singular really correct? Are you using one neural network model and preform several simulations, or are you really use several neural network models?
P2-4: The introduction is too long and not really easy to follow. Text from P3, L61 to L79 should be significantly shortened. Here you actually describe the differences between the old and new retrieval scheme, but the details belong rather to the method section than to the introduction. Further, the description of your new retrieval code after shortening (2-3 sentences) should appear rather at the end of the introduction.
Another comment on the introduction or the manuscript in general: Do you really need that many references? Especially reading the introduction becomes quite tough. It is not necessary to reference every study that ever has been published. Having a reference list of 7 pages for a technical paper is quite a lot and in my opinion somewhat too much.
P5, L117: Where did you get these expected values from? Where these derived in the present study or documented somewhere else? Add reference?
P12, L247: Has the day of 21 March 2022 chosen by purpose (reason?) or arbitrarily? How would the results look like for another day, especially another day in another season?
P18, L319: “Based in the retrieval results shown in Fig 6-8.” This sentence is not complete. Please correct.
P18, L323: Something is missing in this sentence. Maybe “that”? Should it read “To verify that the theoretical retrieval………”?
P18, L332 approx.: Here, the same text is repeated a second time with slight differences. Please omit one version.
P19, L340: The abbreviation ”RSP” has not been introduced.
P19, L345: m/s should be written as m s-1 (according to Copernicus guidelines).
P21, L374: ….radiative transfer behind and……. The sentence makes no sense. Please correct/rephrase.
P21, L377: You use computer performance as a motivation for your study and also mention this in the conclusion, but nowhere throughout the text it is discussed if you actually derive an improvement and how large this improvement is.
P21, L380: Here you provide the actual time needed to process one 5 min granule. However, more interesting it would be how much time is needed to process one day. On page P12, L251 it is stated that 150 granules in 15 orbits are yielded. If the processing of one granule takes 5 h, then processing of one day would take 825 h!? If yes, that would be still incredibly long and maybe much too long for retrieving global data from MAP. So in that case I would not call it feasible at all.
P21, L390: Abbreviation BRDF has not been introduced.
P21, L395: The uncertainties (e.g.numbers, magnitudes) should be given in the conclusions.
General comment on the text: Too many self-citations. You do not need to cite one of your publications in every second sentence. From introduction it became clear that you have done a lot of work already before writing up this study. So reduce the number of occasions and use references to your previous studies only where really necessary.
Technical corrections:
P1, L6: Start sentence with “To this end,” and delete further.
P1, L13: delete “also”
P2, L45: Closing parenthesis after the reference is missing (you need a second one since this text part is in parentheses)
P3, L52: “model” here obsolete -> delete
P3, L57: space between “retrievals” and full stop obsolete.
P3, L76: space between “networks” and the reference “Gao et al.” missing.
P4, L91: chlorophyll a -> chlorophyll-a
P4, L104: The abbreviation DoLP has not been introduced.
P5, L113: “with” should be rather read “whereby”. Maybe it would be better to split this sentence into two sentences.
P5, L116: the “m” in the sigma should be in subscript.
P6, Figure 1 caption: but not -> but are not
P7, L156: chlorophyll a -> chlorophyll-a and chla -> chl-a. Further, units should be written with upright font.
P7, L163: I would suggest to write “Chl-a” instead “Chla” throughout the manuscript. Check all occasions and correct these.
P8, L172: Sentence incomplete. Something is missing here; maybe “is used”?
P8, L172: Add comma after “Note”.
P9, L177: What do you mean with Sunglint? Do you mean “sunlit”? This should be corrected throughput the manuscript.
P10, L230: DOLP -> DoLP
P11, Figure 2 caption: DOLP -> DoLP
P11, L244: Correct reference “J. P., 1987”.
P12, L253: 40 o -> 40°
P12, L254: space before the comma obsolete.
P13, L260 and 263: remove obsolete space before the respective full stop of the sentence.
P14, L269: Add comma after “Note”.
P14, L275: sunlingt -> sunlit
P14, Figure 2 caption: Sec 2 -> Sect. 2
P15, L292: Add comma before “respectively”.
P15, L301: Fig 5 -> Fig. 5
P16, L310ff: Units should be in upright font (according to my knowledge of the Copernicus guidelines) and add a full stop between “Fig” and the respective figure number.
Citation: https://doi.org/10.5194/egusphere-2023-1843-RC1 - AC1: 'Reply on RC1', Meng Gao, 07 Oct 2023
-
RC2: 'Comment on egusphere-2023-1843', Anonymous Referee #2, 18 Sep 2023
General comments:
This study trained a new NN model through measurement uncertainty-aware training and Training data augmentation. The new NN model was used to generate pseudo HARP2 observations and retrieve both aerosol and ocean properties. The methods and results are reasonable, and the manuscript is well written. I have only a few confusions that needs to be clarified.
Specific comments:
- For Equation (3), some terms are not explained. I think the terms with f superscript is NN simulation and the terms without f superscript are pseudo-observations. Please confirm it or correct me.
- For Equations (5) and (6), every term should be explained. Is the uncertainty of DoLP a constant (0.005)? If so, what is difference between Equation (6) and conventional MSE cost function. It seems Equation (6) is just a conventional MSE cost function multiplied by a constant. If uncertainties of DoLP are not a constant in Equation (6), how they are quantified?
- The performance of the NN model is not well validated. Figure 2 has shown the cost function of training and validation, but readers cannot tell if the accuracy of the NN model is sufficient for simulation and retrieval. In this study, observations are generated by the NN model and the NN model is used for retrieval. Thus, it is important to compare the performance of the NN model with that of the radiative transfer model.
Technical corrections:
Line 116: σm. m should be subscript.
Line 189: (Gao et al., 2021a) -> Gao et al. (2021a)
Citation: https://doi.org/10.5194/egusphere-2023-1843-RC2 - AC2: 'Reply on RC2', Meng Gao, 07 Oct 2023
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 359 | 140 | 22 | 521 | 14 | 12 |
- HTML: 359
- PDF: 140
- XML: 22
- Total: 521
- BibTeX: 14
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Cited
2 citations as recorded by crossref.
Bryan A. Franz
Peng-Wang Zhai
Kirk Knobelspiesse
Andrew Sayer
Xiaoguang Xu
Vanderlei Martins
Brian Cairns
Patricia Castellanos
Guangliang Fu
Neranga Hannadige
Otto Hasekamp
Yongxiang Hu
Amir Ibrahim
Frederick Patt
Anin Puthukkudy
P. Jeremy Werdell
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(5997 KB) - Metadata XML