the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Oct 2023
Cloud Detection from Multi-Angular Polarimetric Satellite Measurements using a Neural Network Ensemble Approach
Abstract. This paper describes a neural network cloud masking scheme from PARASOL (Polarisation and Anisotropy of Reflectances for Atmospheric Science coupled with Observations from a Lidar) Multi-Angle Polarimetric measurements. The algorithm has been trained on synthetic measurements and has been applied to the processing of one year of PARASOL data. Comparisons of the retrieved cloud fraction with MODIS (Moderate Resolution Imaging Spectroradiometer) products show overall agreement in spatial and temporal patterns but the PARASOL-NN retrieves lower cloud fractions. Comparisons with a goodness-of-fit mask from aerosol retrievals suggest that the NN cloud mask flags less clear pixels as cloudy than MODIS (∼3 % of the clear pixels, versus ∼15 % by MODIS). On the other hand the NN classifies more pixels incorrectly as clear than MODIS (∼19 % by NN, versus ∼15 % by MODIS). Additionally, the NN and MODIS cloud mask have been applied to the aerosol retrievals from PARASOL using the Remote Sensing of Trace Gas and Aerosol Products (RemoTAP) algorithm. Validation with AERONET shows that the NN cloud mask performs comparably with MODIS in screening residual cloud contamination in retrieved aerosol properties. Our study demonstrates that cloud masking from MAP aerosol retrievals can be performed based on the MAP measurements themselves, making the retrievals independent of the availability of a cloud imager.
- Preprint
(14012 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on amt-2023-145', Anonymous Referee #1, 20 Dec 2023
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-145/amt-2023-145-RC1-supplement.pdf
- AC2: 'Reply on RC1', Zihao Yuan, 14 Feb 2024
-
RC2: 'Comment on amt-2023-145', Anonymous Referee #2, 21 Dec 2023
Yuan et al presented a neural network approach to conduct cloud detection using multi-angle polarimetric measurements, and applied the model over a year of PARASOL data. Performance was evaluated by comparing with MODIS cloud fraction and goodness of fitting using an aerosol retrieval algorithm. The results showed the NN cloud fraction is very effective at least comparable to the goodness of fitting results. With the high computational efficiency, the NN model is promising to be applied in future satellite missions with multi-angle polarimetric observations, and help both aerosol and cloud studies. Overall, the manuscript is well written with detailed discussions on the methodology and practical applications. I would recommend the publication of this work with suggestive comments below.
Major comments:
- I don’t see much discussion on the accuracy of the cloud mask related to the measurement uncertainty itself, which, in my option, should be very important. The cloud fraction threshold which impact aerosol retrievals may also depend on the accuracy of the measurements.
- Almost all the measurements are used to derive a single cloud fraction. Cloud fraction seems easier to determine than cloud and aerosol microphysical properties, I wonder whether less measurements can still achieve reasonable performance. If less measurement can be used, maybe NN with smaller size and faster speed can be developed, or applied to fewer angles to gain more flexibility?
Detailed comments:
P3, L62, “Given that the results for the cloudy pixels (∼80% of all pixels)”, 80% sounds too much, any reference?
P3, L70: “MODIS cloud mask, … is based on input signals from visible and infrared bands, which detect the high, spectrally flat reflectance and low brightness temperature feature of clouds.”, The color on figure 4 over water is clearly not spectrally flat, just wonder why it is picked up by the MODIS cloud mask?
P3, L80, it may sound trivial, how the cloud fraction is defined? Cloud can cover partially in space or transparently over the pixel.
P4, L106, why a minimum of 14 out of 16 angles are used? Can less angle be used?
P4, L 106, What is the accuracy of the PARASOL measurements?
P5, L132, “The cloud fraction (referred to as MODIS cloud fraction hereafter) is calculated as the fraction of confidently- and uncertain-cloudy-flagged 1-km-resolution MODIS pixels within a 6km°ø6km PARASOL grid.”
Should different weight be applied to the confidently cloudy and uncertain cloudy mask in calculating the cloud fraction? Since later the authors reported MODIS cloud fraction values are larger, would this be part of the reason?
P5, L149: one shape of ice crystal is used, does the shape matters? Would it possible to distinguish ice cloud or water cloud fraction maybe in future work?
P6, L178, “We do not consider situations that are partly covered by both ice and liquid clouds.”. With the large pixel size of 6km x 6 km, there could be higher chances to observe partially covered cloud or cloud edge. I wonder whether the NN can be applied to every angle of the multi-angle observations, which may help detect partially covered cloud? There are some works which seems study such scenarios but with less efficiency (example: Gao et al, 2021: https://doi.org/10.3389/frsen.2021.757832), a more flexible NN cloud mask can be a remedy.
P7, Table 1, what does “RemoTAP” mean for distribution? You may already explain other-where, but it would be useful to include some information in the caption.
P8, L189, does smaller size of NN is needed when trained on parts of the total data?
P8, L205, “The neural network architecture consists of three hidden layers with 80 neurons each layer for over ocean scenes and 40 neurons for over land scenes.”
I would expect cloud detection is harder over land due to its complexity, but here smaller NN is used over land comparing with the ocean one. Can you explain more?
P9, L233, “Therefore,we suggest that a threshold of 0.05 for the retrieved cloud fraction should be good for aerosol retrieval.” I wonder whether the threshold should relate to the measurement uncertainty. Or how the measurement uncertainty relates to the accuracy of the cloud fraction accuracy?
P15, L309, “the AOD error should be smaller than 0.03 or 10%”, can you confirm the definition of the error? As retrieval-truth or 1-sigma uncertainty etc?
Page 17, Figure 8, the title of the figure seems indicate goodness-of-fit has been applied for all three columns of the plots, but it seems not the case as indicated in the caption. Can you verify?
Page 19, Figure 9, first row, the maximum percentage is around 65%, which is roughly the percentage within 1-sigma of a Gaussian distribution. Is this the correct explanation?
Citation: https://doi.org/10.5194/amt-2023-145-RC2 -
AC1: 'Reply on RC2', Zihao Yuan, 14 Feb 2024
Response to reviewer 2
We would like to thank the reviewer for his/her important comments and suggestions.
Major comments:
- I don’t see much discussion on the accuracy of the cloud mask related to the measurement uncertainty itself, which, in my option, should be very important. The cloud fraction threshold which impact aerosol retrievals may also depend on the accuracy of the measurements.
Response: It is hard to predict how the (required) accuracy of the cloud fraction is related to measurement accuracy. On the one hand, the measurement accuracy may affect the capability of the goodness-of-fit mask that is applied after the aerosol retrieval, i.e. the higher the measurement accuracy, the better the goodness-of-fit mask would work. On the other hand, for a high measurement accuracy, also the accuracy of retrieved aerosol properties is better and the effect of residual cloud may be more important relative to the aerosol retrieval accuracy. Another effect would be that the cloud fraction can be more accurately retrieved when the measurement accuracy is higher. Given the complexity of this discussion we believe it is a study on its own. We added in the conclusion the phrase:
“Application of the NN cloud screening approach to these new instruments will provide insight in the sensitivity of the approach to measurement uncertainty, number of viewing angles, and number of spectral bands.” Which provides an outlook to future studies.
- Almost all the measurements are used to derive a single cloud fraction. Cloud fraction seems easier to determine than cloud and aerosol microphysical properties, I wonder whether less measurements can still achieve reasonable performance. If less measurement can be used, maybe NN with smaller size and faster speed can be developed, or applied to fewer angles to gain more flexibility?
Response: The speed of the NN is not a concern at all for the current size of the NN (~12h to train and ~48h to retrieve the 2008 full year cloud fraction from PARASOL) so this would not be a reason to reduce the number of angles. We prefer to use all information available to perform the cloud screening. We do not see why there would be more flexibility when using less angles.
Detailed comments:
P3, L62, “Given that the results for the cloudy pixels (∼80% of all pixels)”, 80% sounds too much, any reference?
Response: In the context, cloudy pixels are defined in perspective of aerosol retrievals. Based on our previous global aerosol retrievals, around 80% of pixels are cloud-contaminated and thus not able to produce a reasonable aerosol retrieval result. Note that the percentage of cloudy pixels strongly depends on the sensor resolution. The 80% value is consistent with the analysis provided by Krijger et al. (2007; https://acp.copernicus.org/articles/7/2881/2007/acp-7-2881-2007.html) which is added to the paper as reference.
P3, L70: “MODIS cloud mask, … is based on input signals from visible and infrared bands, which detect the high, spectrally flat reflectance and low brightness temperature feature of clouds.”, The color on figure 4 over water is clearly not spectrally flat, just wonder why it is picked up by the MODIS cloud mask?
Response: The MODIS cloud mask makes use of a number of thresholds and as a proxy of the spectrally flatness of the observations, the mask algorithm uses a ratio of reflectance at 865 and 670 nm. We do not have information which threshold combination has caused the dust scene to be flagged as cloudy by MODIS. Probably it is related to the spectral flatness (see above), the high signal level and a low brightness temperature (because of elevated dust). As a discussion on this bias of the MODIS cloud mask is beyond the scope of this paper, no changes were made to the paper related to this question.
P3, L80, it may sound trivial, how the cloud fraction is defined? Cloud can cover partially in space or transparently over the pixel.
Response: The cloud fraction is defined according to the independent pixel approximation (see Eq 1).
P4, L106, why a minimum of 14 out of 16 angles are used? Can less angle be used?
Response: Most PARASOL measurements have 14 angles. We added this as clarification to the revised version. We did not investigate the performance as number of viewing angles but for SPEXone (5 angles) we obtain similar performance as for PARASOL on synthetic measurements, but here the smaller number of angles may be compensated by a larger number of wavelengths.
P4, L 106, What is the accuracy of the PARASOL measurements?
Response: For the bands used in the study, we assumed intensity has a 2% relative noise and DoLP 0.007 absolute noise in the original version of the manuscript. Based on the questions of the reviewer, we did a quick test on noise settings and found a variable relative noise between 1-3% for intensity and a 0.012 absolute noise for DoLP gave a better performance in term of effectiveness. Therefore, in the revised version, we use the latter noise settings.
P5, L132, “The cloud fraction (referred to as MODIS cloud fraction hereafter) is calculated as the fraction of confidently- and uncertain-cloudy-flagged 1-km-resolution MODIS pixels within a 6km°ø6km PARASOL grid.”
Should different weight be applied to the confidently cloudy and uncertain cloudy mask in calculating the cloud fraction? Since later the authors reported MODIS cloud fraction values are larger, would this be part of the reason?
Response: We tested using only confidently-cloudy flag as cloudy in collocating MODIS cloud fraction from MYD35 data, but no significant differences are observed. The reason is that there are usually not many MODIS (MYD35) uncertain pixels in a PARASOL grid, e.g. on 01 Jan 2008, 90% PARASOL grids have less than 5 MODIS uncertain flags, so we won't expected there are obvious differences between the two different definition of MODIS cloud fraction.
P5, L149: one shape of ice crystal is used, does the shape matters? Would it possible to distinguish ice cloud or water cloud fraction maybe in future work?
Response: The hexagonal ice crystals with varying aspect ratios and surface distortions represent the scattering properties of ice crystals with variable complex shapes. This is demonstrated in the reference given (van Diedenhoven et al. 2020) and references therein. We modified the sentence to mention that the hexagonal ice crystals are used as proxies for ice crystals with variable complex shapes. Furthermore, based on the results in the article, the current settings are sufficient for the main focus of cloud masking for aerosol retrievals. Distinguishing ice / liquid cloud fraction is one of our on-going work.
P6, L178, “We do not consider situations that are partly covered by both ice and liquid clouds.”. With the large pixel size of 6km x 6 km, there could be higher chances to observe partially covered cloud or cloud edge. I wonder whether the NN can be applied to every angle of the multi-angle observations, which may help detect partially covered cloud? There are some works which seems study such scenarios but with less efficiency (example: Gao et al, 2021: https://doi.org/10.3389/frsen.2021.757832), a more flexible NN cloud mask can be a remedy.
Response: There are indeed situations where different angles see different cloud fractions because of inhomogeneity. Therefore, we include in our training set also samples where different angles have different cloud fraction. This increases the capability of the cloud screening (see l 185 of the revised version). We also added a reference to Gao et al, 2021 in the revised version.
P7, Table 1, what does “RemoTAP” mean for distribution? You may already explain other-where, but it would be useful to include some information in the caption.
Response: “RemoTAP” means the properties are randomly taken from RemoTAP global aerosol retrieval for the year 2008, as explained in line 159 of article. We added a short explanation in the caption of the table in the revised version.
P8, L189, does smaller size of NN is needed when trained on parts of the total data?
Response: We cannot give a universal answer to this question. Actually, the best size of NN depends more on the task itself (e.g. task complexity).
P8, L205, “The neural network architecture consists of three hidden layers with 80 neurons each layer for over ocean scenes and 40 neurons for over land scenes.”
I would expect cloud detection is harder over land due to its complexity, but here smaller NN is used over land comparing with the ocean one. Can you explain more?
Response: The choice of NN is based on the performance on real measurements (with comparison to goodness-of-fit mask, i.e. experiment in section 4.2), which can reflect the generalization ability of the NN. Based on this performance we found smaller NN is better for retrievals over land than over ocean. An explanation can be that a larger NN usually has larger risk in overfitting data and thus struggling to generalize to new data. Therefore, sometimes it can be possible that a smaller NN behaves better for the more complicated over land case.
P9, L233, “Therefore,we suggest that a threshold of 0.05 for the retrieved cloud fraction should be good for aerosol retrieval.” I wonder whether the threshold should relate to the measurement uncertainty. Or how the measurement uncertainty relates to the accuracy of the cloud fraction accuracy?
Response: See our response above to the general comment.
P15, L309, “the AOD error should be smaller than 0.03 or 10%”, can you confirm the definition of the error? As retrieval-truth or 1-sigma uncertainty etc?
Response: AOD error is here defined as the absolute value of ‘true-retrieved’ . We add the explanation at I 335 of the revised version.
Page 17, Figure 8, the title of the figure seems indicate goodness-of-fit has been applied for all three columns of the plots, but it seems not the case as indicated in the caption. Can you verify?
Response: The goodness-of-fit mask is always applied on top of the NN and MODIS in this section of assessing their effect on aerosol retrievals. We revised the titles in the section to avoid this confusion. Figure 8 is moved to supplement.
Page 19, Figure 9, first row, the maximum percentage is around 65%, which is roughly the percentage within 1-sigma of a Gaussian distribution. Is this the correct explanation?
Response: Indeed if the 1-sigma uncertainty would be the same as the GCOS requirement this would result in 67% of the pixels within the requirement. This only is the case for SSA in our retrievals (although please note the number of validation points for SSA is low).
Citation: https://doi.org/10.5194/amt-2023-145-AC1
Status: closed
-
RC1: 'Comment on amt-2023-145', Anonymous Referee #1, 20 Dec 2023
The comment was uploaded in the form of a supplement: https://amt.copernicus.org/preprints/amt-2023-145/amt-2023-145-RC1-supplement.pdf
- AC2: 'Reply on RC1', Zihao Yuan, 14 Feb 2024
-
RC2: 'Comment on amt-2023-145', Anonymous Referee #2, 21 Dec 2023
Yuan et al presented a neural network approach to conduct cloud detection using multi-angle polarimetric measurements, and applied the model over a year of PARASOL data. Performance was evaluated by comparing with MODIS cloud fraction and goodness of fitting using an aerosol retrieval algorithm. The results showed the NN cloud fraction is very effective at least comparable to the goodness of fitting results. With the high computational efficiency, the NN model is promising to be applied in future satellite missions with multi-angle polarimetric observations, and help both aerosol and cloud studies. Overall, the manuscript is well written with detailed discussions on the methodology and practical applications. I would recommend the publication of this work with suggestive comments below.
Major comments:
- I don’t see much discussion on the accuracy of the cloud mask related to the measurement uncertainty itself, which, in my option, should be very important. The cloud fraction threshold which impact aerosol retrievals may also depend on the accuracy of the measurements.
- Almost all the measurements are used to derive a single cloud fraction. Cloud fraction seems easier to determine than cloud and aerosol microphysical properties, I wonder whether less measurements can still achieve reasonable performance. If less measurement can be used, maybe NN with smaller size and faster speed can be developed, or applied to fewer angles to gain more flexibility?
Detailed comments:
P3, L62, “Given that the results for the cloudy pixels (∼80% of all pixels)”, 80% sounds too much, any reference?
P3, L70: “MODIS cloud mask, … is based on input signals from visible and infrared bands, which detect the high, spectrally flat reflectance and low brightness temperature feature of clouds.”, The color on figure 4 over water is clearly not spectrally flat, just wonder why it is picked up by the MODIS cloud mask?
P3, L80, it may sound trivial, how the cloud fraction is defined? Cloud can cover partially in space or transparently over the pixel.
P4, L106, why a minimum of 14 out of 16 angles are used? Can less angle be used?
P4, L 106, What is the accuracy of the PARASOL measurements?
P5, L132, “The cloud fraction (referred to as MODIS cloud fraction hereafter) is calculated as the fraction of confidently- and uncertain-cloudy-flagged 1-km-resolution MODIS pixels within a 6km°ø6km PARASOL grid.”
Should different weight be applied to the confidently cloudy and uncertain cloudy mask in calculating the cloud fraction? Since later the authors reported MODIS cloud fraction values are larger, would this be part of the reason?
P5, L149: one shape of ice crystal is used, does the shape matters? Would it possible to distinguish ice cloud or water cloud fraction maybe in future work?
P6, L178, “We do not consider situations that are partly covered by both ice and liquid clouds.”. With the large pixel size of 6km x 6 km, there could be higher chances to observe partially covered cloud or cloud edge. I wonder whether the NN can be applied to every angle of the multi-angle observations, which may help detect partially covered cloud? There are some works which seems study such scenarios but with less efficiency (example: Gao et al, 2021: https://doi.org/10.3389/frsen.2021.757832), a more flexible NN cloud mask can be a remedy.
P7, Table 1, what does “RemoTAP” mean for distribution? You may already explain other-where, but it would be useful to include some information in the caption.
P8, L189, does smaller size of NN is needed when trained on parts of the total data?
P8, L205, “The neural network architecture consists of three hidden layers with 80 neurons each layer for over ocean scenes and 40 neurons for over land scenes.”
I would expect cloud detection is harder over land due to its complexity, but here smaller NN is used over land comparing with the ocean one. Can you explain more?
P9, L233, “Therefore,we suggest that a threshold of 0.05 for the retrieved cloud fraction should be good for aerosol retrieval.” I wonder whether the threshold should relate to the measurement uncertainty. Or how the measurement uncertainty relates to the accuracy of the cloud fraction accuracy?
P15, L309, “the AOD error should be smaller than 0.03 or 10%”, can you confirm the definition of the error? As retrieval-truth or 1-sigma uncertainty etc?
Page 17, Figure 8, the title of the figure seems indicate goodness-of-fit has been applied for all three columns of the plots, but it seems not the case as indicated in the caption. Can you verify?
Page 19, Figure 9, first row, the maximum percentage is around 65%, which is roughly the percentage within 1-sigma of a Gaussian distribution. Is this the correct explanation?
Citation: https://doi.org/10.5194/amt-2023-145-RC2 -
AC1: 'Reply on RC2', Zihao Yuan, 14 Feb 2024
Response to reviewer 2
We would like to thank the reviewer for his/her important comments and suggestions.
Major comments:
- I don’t see much discussion on the accuracy of the cloud mask related to the measurement uncertainty itself, which, in my option, should be very important. The cloud fraction threshold which impact aerosol retrievals may also depend on the accuracy of the measurements.
Response: It is hard to predict how the (required) accuracy of the cloud fraction is related to measurement accuracy. On the one hand, the measurement accuracy may affect the capability of the goodness-of-fit mask that is applied after the aerosol retrieval, i.e. the higher the measurement accuracy, the better the goodness-of-fit mask would work. On the other hand, for a high measurement accuracy, also the accuracy of retrieved aerosol properties is better and the effect of residual cloud may be more important relative to the aerosol retrieval accuracy. Another effect would be that the cloud fraction can be more accurately retrieved when the measurement accuracy is higher. Given the complexity of this discussion we believe it is a study on its own. We added in the conclusion the phrase:
“Application of the NN cloud screening approach to these new instruments will provide insight in the sensitivity of the approach to measurement uncertainty, number of viewing angles, and number of spectral bands.” Which provides an outlook to future studies.
- Almost all the measurements are used to derive a single cloud fraction. Cloud fraction seems easier to determine than cloud and aerosol microphysical properties, I wonder whether less measurements can still achieve reasonable performance. If less measurement can be used, maybe NN with smaller size and faster speed can be developed, or applied to fewer angles to gain more flexibility?
Response: The speed of the NN is not a concern at all for the current size of the NN (~12h to train and ~48h to retrieve the 2008 full year cloud fraction from PARASOL) so this would not be a reason to reduce the number of angles. We prefer to use all information available to perform the cloud screening. We do not see why there would be more flexibility when using less angles.
Detailed comments:
P3, L62, “Given that the results for the cloudy pixels (∼80% of all pixels)”, 80% sounds too much, any reference?
Response: In the context, cloudy pixels are defined in perspective of aerosol retrievals. Based on our previous global aerosol retrievals, around 80% of pixels are cloud-contaminated and thus not able to produce a reasonable aerosol retrieval result. Note that the percentage of cloudy pixels strongly depends on the sensor resolution. The 80% value is consistent with the analysis provided by Krijger et al. (2007; https://acp.copernicus.org/articles/7/2881/2007/acp-7-2881-2007.html) which is added to the paper as reference.
P3, L70: “MODIS cloud mask, … is based on input signals from visible and infrared bands, which detect the high, spectrally flat reflectance and low brightness temperature feature of clouds.”, The color on figure 4 over water is clearly not spectrally flat, just wonder why it is picked up by the MODIS cloud mask?
Response: The MODIS cloud mask makes use of a number of thresholds and as a proxy of the spectrally flatness of the observations, the mask algorithm uses a ratio of reflectance at 865 and 670 nm. We do not have information which threshold combination has caused the dust scene to be flagged as cloudy by MODIS. Probably it is related to the spectral flatness (see above), the high signal level and a low brightness temperature (because of elevated dust). As a discussion on this bias of the MODIS cloud mask is beyond the scope of this paper, no changes were made to the paper related to this question.
P3, L80, it may sound trivial, how the cloud fraction is defined? Cloud can cover partially in space or transparently over the pixel.
Response: The cloud fraction is defined according to the independent pixel approximation (see Eq 1).
P4, L106, why a minimum of 14 out of 16 angles are used? Can less angle be used?
Response: Most PARASOL measurements have 14 angles. We added this as clarification to the revised version. We did not investigate the performance as number of viewing angles but for SPEXone (5 angles) we obtain similar performance as for PARASOL on synthetic measurements, but here the smaller number of angles may be compensated by a larger number of wavelengths.
P4, L 106, What is the accuracy of the PARASOL measurements?
Response: For the bands used in the study, we assumed intensity has a 2% relative noise and DoLP 0.007 absolute noise in the original version of the manuscript. Based on the questions of the reviewer, we did a quick test on noise settings and found a variable relative noise between 1-3% for intensity and a 0.012 absolute noise for DoLP gave a better performance in term of effectiveness. Therefore, in the revised version, we use the latter noise settings.
P5, L132, “The cloud fraction (referred to as MODIS cloud fraction hereafter) is calculated as the fraction of confidently- and uncertain-cloudy-flagged 1-km-resolution MODIS pixels within a 6km°ø6km PARASOL grid.”
Should different weight be applied to the confidently cloudy and uncertain cloudy mask in calculating the cloud fraction? Since later the authors reported MODIS cloud fraction values are larger, would this be part of the reason?
Response: We tested using only confidently-cloudy flag as cloudy in collocating MODIS cloud fraction from MYD35 data, but no significant differences are observed. The reason is that there are usually not many MODIS (MYD35) uncertain pixels in a PARASOL grid, e.g. on 01 Jan 2008, 90% PARASOL grids have less than 5 MODIS uncertain flags, so we won't expected there are obvious differences between the two different definition of MODIS cloud fraction.
P5, L149: one shape of ice crystal is used, does the shape matters? Would it possible to distinguish ice cloud or water cloud fraction maybe in future work?
Response: The hexagonal ice crystals with varying aspect ratios and surface distortions represent the scattering properties of ice crystals with variable complex shapes. This is demonstrated in the reference given (van Diedenhoven et al. 2020) and references therein. We modified the sentence to mention that the hexagonal ice crystals are used as proxies for ice crystals with variable complex shapes. Furthermore, based on the results in the article, the current settings are sufficient for the main focus of cloud masking for aerosol retrievals. Distinguishing ice / liquid cloud fraction is one of our on-going work.
P6, L178, “We do not consider situations that are partly covered by both ice and liquid clouds.”. With the large pixel size of 6km x 6 km, there could be higher chances to observe partially covered cloud or cloud edge. I wonder whether the NN can be applied to every angle of the multi-angle observations, which may help detect partially covered cloud? There are some works which seems study such scenarios but with less efficiency (example: Gao et al, 2021: https://doi.org/10.3389/frsen.2021.757832), a more flexible NN cloud mask can be a remedy.
Response: There are indeed situations where different angles see different cloud fractions because of inhomogeneity. Therefore, we include in our training set also samples where different angles have different cloud fraction. This increases the capability of the cloud screening (see l 185 of the revised version). We also added a reference to Gao et al, 2021 in the revised version.
P7, Table 1, what does “RemoTAP” mean for distribution? You may already explain other-where, but it would be useful to include some information in the caption.
Response: “RemoTAP” means the properties are randomly taken from RemoTAP global aerosol retrieval for the year 2008, as explained in line 159 of article. We added a short explanation in the caption of the table in the revised version.
P8, L189, does smaller size of NN is needed when trained on parts of the total data?
Response: We cannot give a universal answer to this question. Actually, the best size of NN depends more on the task itself (e.g. task complexity).
P8, L205, “The neural network architecture consists of three hidden layers with 80 neurons each layer for over ocean scenes and 40 neurons for over land scenes.”
I would expect cloud detection is harder over land due to its complexity, but here smaller NN is used over land comparing with the ocean one. Can you explain more?
Response: The choice of NN is based on the performance on real measurements (with comparison to goodness-of-fit mask, i.e. experiment in section 4.2), which can reflect the generalization ability of the NN. Based on this performance we found smaller NN is better for retrievals over land than over ocean. An explanation can be that a larger NN usually has larger risk in overfitting data and thus struggling to generalize to new data. Therefore, sometimes it can be possible that a smaller NN behaves better for the more complicated over land case.
P9, L233, “Therefore,we suggest that a threshold of 0.05 for the retrieved cloud fraction should be good for aerosol retrieval.” I wonder whether the threshold should relate to the measurement uncertainty. Or how the measurement uncertainty relates to the accuracy of the cloud fraction accuracy?
Response: See our response above to the general comment.
P15, L309, “the AOD error should be smaller than 0.03 or 10%”, can you confirm the definition of the error? As retrieval-truth or 1-sigma uncertainty etc?
Response: AOD error is here defined as the absolute value of ‘true-retrieved’ . We add the explanation at I 335 of the revised version.
Page 17, Figure 8, the title of the figure seems indicate goodness-of-fit has been applied for all three columns of the plots, but it seems not the case as indicated in the caption. Can you verify?
Response: The goodness-of-fit mask is always applied on top of the NN and MODIS in this section of assessing their effect on aerosol retrievals. We revised the titles in the section to avoid this confusion. Figure 8 is moved to supplement.
Page 19, Figure 9, first row, the maximum percentage is around 65%, which is roughly the percentage within 1-sigma of a Gaussian distribution. Is this the correct explanation?
Response: Indeed if the 1-sigma uncertainty would be the same as the GCOS requirement this would result in 67% of the pixels within the requirement. This only is the case for SSA in our retrievals (although please note the number of validation points for SSA is low).
Citation: https://doi.org/10.5194/amt-2023-145-AC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 336 | 111 | 17 | 464 | 15 | 16 |
- HTML: 336
- PDF: 111
- XML: 17
- Total: 464
- BibTeX: 15
- EndNote: 16
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1