Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-Restriction

Round 1

Reviewer 1 Report

Review of Satellite Fog Prediction at Dawn and Dusk… by Ran et al

 

This paper describes an implementation of a deep learning algorithm (DDF-Net) that uses a terrain restriction to predict the presence of fog within a region of China.  The results from the satellite images and algorithm are compared to ground-truth data and to the results of another modeling approach and the DDF-Net algorithm is described in some details.

I think the paper could possibly be publishable, but I have a number of concerns I would need to see addressed prior to recommending it for publication.

First, I have some concerns that the authors categorized visibilities of 1000 m-3000 m with RH>90 % as light fog.  Fog by definition is associated with visibilities of less than 1000 m.  Instead, I suggest they call it “haze” in their subsequent analyses.  Interpretation of this feature as “light fog”  will inappropriately bias results.

I also have a difficult time reading and interpreting some of the figures.  Figure 5, for instance, has very small text that is challenging to decipher at the scale of a printed page.  Even when magnified, however, and with the accompanying text, I could not follow many of the details that the authors present.  For instance, the authors did not explain the Relu function, contextual feature information connections, and semantic space context information, all of which are important to understand the algorithm.  So, I had trouble making sense and significance of Figure 5.  If greater description is not warranted for this paper, please provide more citations for this information.

Additionally, in the chart in Table 3, how did the authors determine the ground observations of fog versus no-fog, since there were 3 (and I suggest 4) possible categories of fog from dense to light?  Did they sum all of the “dense fog”, “medium fog,” and “light fog” categories to do so?  If so, they should state it, since I think that it is required for this analysis that they classify just whether fog exists or not.  I also suggest that they remove what I would call “haze” (as stated above) from the fog category.

Furthermore, I found Figures 7-10 cryptic.  I assume the observation points are the dots shown?  They are very difficult to see.   Also, I did not understand why both false color and fog detection results are shown.  What is the significance of the false color images?  Aren’t we mainly interested in the fog that is predicted from these images?  I had a difficult time seeing how well predictions matched (or did not match) the ground truth data from these images.   If the false color images really don’t add to the discussion, I suggest removing them and only using the fog prediction results coupled with the point observations and in making these figures larger so that the comparisons are clearer.  Additionally, a graph or table that shows the percentage of observations that match or do not match a given prediction of light fog, medium fog, heavy fog, and haze for both models would provide more substance on claims as to which algorithm is more effective in different regions, times, and conditions.  As it is, I have a difficult time assessing any comparisons between the models and the data as shown in Figures 7-10.

In addition to these overarching issues, there were numerous grammatical problems and inconsistencies in the writing that sometimes made it difficult to follow the authors’ points and sometimes resulted in ambiguity.  I list many of those here, though I am hopeful that an editor can catch others.

 

Line 29-30:  airborne pollutants -> anthropogenically-generated chemicals.

Sentence on Line 31 is awkward.  I suggest.  “Therefore, fog detection is crucial to effectively support traffic planning and to provide information and bulletins for reducing risks to human health.”

Line 44:  “which is” -> “for which it is”  Do the authors means that it is difficult to isolate TIR and MIR at night or during the day?  They need to clarify that because as it currently is written, it seems contradictory.

Line 60:  phrase beginning with “while” is a run-on sentence.  Please fix

Line 67:  “lacking most” -> “lacking in most”

Line 68:  cloud forms -> clouds form

Line 106:  “locates” -> “is located in”

Line 109:  “In the eastern…”  not sure what authors intended to say, but this is grammatically problematic.

Line 109:  “it locates” –> “it is located”

Line 110: “with the average altitude drops…” ->  “where the average altitude drops…”

Line 134:  “hazy” -> “haze”

Line 144:  “Mask result” ->  I am not clear on the meaning of this term.  Please eiher describe it further or cite a reference that does.

Line 152:  “more abstract information” -> this is vague.  What information do the authors refer to more specifically?

Line 168 ->  I see no reference to “Scale” in Figure 3.  I have no idea what it refers to.

Figure 3:  I am unclear what the dimensions shown, e.g. C/16x1x1 refer to and why they are shown in the figure.

Right after Line 187:  This should be Equation (3) not Equation (2).

Lines 188-189:  Line ending with ‘calculates’ is a sentence fragment.

Line 256:  Not sure what is meant by “divided by 7:3”  What is 7:3?

Line 262-263:  “after the number of rounds” -> “after the number of rounds designated as Epoch”  (at least I think that is the authors’ intent?)

Line 272: “Repeat the above 1-2 process to iterated” -> “Processes 1-2 are iterated until…”

Line 273:  “a certain value, the model converges...”  -> “a certain value.  This means that the model converges…”

Line 302-303:  This seems like circular logic.  It implies that the satellite data are the observations that are compared with the model.  Am I misunderstanding, or isn’t the satellite data what is driving the model?  If this is the case, then the fact that the satellite data shows increasing or decreasing fog should not suggest whether the algorithm is working or not – it should be driving the algorithm, right?  If I am correct, please change this.  If I am not correct, please adjust the text to explain this.

Line 304:  “The fog area was fragmented due to low cloud”  Not sure what the authors mean.  Do they mean that the fog was interspersed with low clouds?

Line 308: “eastern” ->  “eastern portion of the figure”?

Line 309:  “data shows” -> “data show” (because data are plural)

Line 312:  “refined well because the…” -> “defined well because of the …”?  (Is this the authors’ intent?)

Line 366:  “the previous studies” -> “those of the previous studies”

Line 371:  “furtherly” -> “further”

Lines 373 and 374:  “accurate” -> “accurately”

Line 382:  “is more clearer” -> “is clearer”

Line 401:  “it developed” -> “we developed”

Lines 406-407:  I am not sure what the authors mean by:  “ Both the DDF-Net and U-net have false discrimination under the fog dissipation, medium/high clouds and low cloud”  Do they mean that the algorithms have difficulty identifying fog when these phenomena are occurring?  If so, I would rewrite this sentence as “Both the DDF-Net and U-net are challenged in detecting fog during the periods of time when the fog is dissipating and when there are clouds.”

 

 

 

Author Response

Response to Comments of Reviewer 1

 

Dear Reviewer

 

Thank you for your letter and comments concerning our manuscript entitled “Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-restriction” (No. remotesensing-1818786). We were encouraged by the positive comments regarding the usefulness of our research. The revised manuscript now includes appropriate modifications and/or improvements which have been made possible from your comments. Revised portions are marked in red in the revised manuscript. The main corrections and responds to comments are as follows:

 

Comment 1: This paper describes an implementation of a deep learning algorithm (DDF-Net) that uses a terrain restriction to predict the presence of fog within a region of China. The results from the satellite images and algorithm are compared to ground-truth data and to the results of another modeling approach and the DDF-Net algorithm is described in some details.

I think the paper could possibly be publishable, but I have a number of concerns I would need to see addressed prior to recommending it for publication.

Response: Thanks very much for your positive evaluation and valuable comments. We read them carefully and revised the manuscript profoundly. The revised portions are marked in red in the revised manuscript.

 

Comment 2: First, I have some concerns that the authors categorized visibilities of 1000 m-3000 m with RH>90 % as light fog. Fog by definition is associated with visibilities of less than 1000 m. Instead, I suggest they call it “haze” in their subsequent analyses. Interpretation of this feature as “light fog” will inappropriately bias results.

Response: Thanks for your comment. We absolutely agree you that the definition of fog is associated with visibilities of less than 1000 m. In this study, we followed the criteria and classified the fog as strong, dense and haze ones as visibility ranges between 0~200, 200~500 and 500~1000 meters. Meanwhile, the China Meteorological Administration (CMA) promulgated the national standard of Grade of fog forecast (GB/T 27964-2011GB/T), which also defined haze when the horizontal visibility ranges from 1.0 to 10.0 km. Furthermore, it is less harmful to traffic when the visibility is greater than 3 km. Therefore, we defined as the haze as visibility ranges from 500~3000 meters. We explained it in lines 143 to 150 of the revised manuscript.

 

Comment 3: I also have a difficult time reading and interpreting some of the figures. Figure 5, for instance, has very small text that is challenging to decipher at the scale of a printed page.  Even when magnified, however, and with the accompanying text, I could not follow many of the details that the authors present. For instance, the authors did not explain the Relu function, contextual feature information connections, and semantic space context information, all of which are important to understand the algorithm. So, I had trouble making sense and significance of Figure 5. If greater description is not warranted for this paper, please provide more citations for this information.

Response: Thanks for your comment. We improved the Figure 5 according to your suggestion in the revised version. We also explained the figure in detail. Specifically, Relu function is used to improve the model’s ability to fit nonlinear relationships. The contextual feature information connections refer to a certain relationship between pixels and its surrounding pixels. Finally, semantic space context information can be characterized as the fused information of the shallow features extracted in the encoder and the high-level semantics extracted in the decoder. Please see the red portions in lines 242 to 254. Furthermore, we added references [21], [24] and [37] to help the readers understand and capture more information of the model. Please see lines 227 to 232.

 

Comment 4: Additionally, in the chart in Table 3, how did the authors determine the ground observations of fog versus no-fog, since there were 3 (and I suggest 4) possible categories of fog from dense to light?  Did they sum all of the “dense fog”, “medium fog,” and “light fog” categories to do so?  If so, they should state it, since I think that it is required for this analysis that they classify just whether fog exists or not.  I also suggest that they remove what I would call “haze” (as stated above) from the fog category.

Response: Thanks very much for your comments. We sum all of the “dense fog”, “medium fog,” and “light fog” categories to validate the algorithm. It would be very useful to evaluate the accuracies under different fog types. However, the observation points are insufficient to be divided into so many sub-datasets. Therefore, a small disturbance of the ground samples would generate great uncertainty on the validation. We will continue the collection of ground observation to validation the performance of our algorithm in our future researches. We explained it in lines 402 to 407 of the revised manuscript.

 

Comment 5: Furthermore, I found Figures 7-10 cryptic. I assume the observation points are the dots shown?  They are very difficult to see. Also, I did not understand why both false color and fog detection results are shown. What is the significance of the false color images? Aren’t we mainly interested in the fog that is predicted from these images? I had a difficult time seeing how well predictions matched (or did not match) the ground truth data from these images. If the false color images really don’t add to the discussion, I suggest removing them and only using the fog prediction results coupled with the point observations and in making these figures larger so that the comparisons are clearer. Additionally, a graph or table that shows the percentage of observations that match or do not match a given prediction of light fog, medium fog, heavy fog, and haze for both models would provide more substance on claims as to which algorithm is more effective in different regions, times, and conditions. As it is, I have a difficult time assessing any comparisons between the models and the data as shown in Figures 7-10.

Response: Thanks very much for your valuable comments. We removed the false color and improved the Figures 7-10 according to your comments. It is very useful to evaluate the accuracies under different conditions fog types. However, the observation points are insufficient to be divided into so many sub-datasets. As shown in Tables 3~6, the numbers of ground observation labeled by fog are no more than 40. Therefore, a small disturbance of the ground samples would generate great uncertainty on the validation. We will continue the collection of ground observation to validation the performance of our algorithm in our future researches. We discussed it in lines 402 to 407 of the revised manuscript.

 

Comment 6: In addition to these overarching issues, there were numerous grammatical problems and inconsistencies in the writing that sometimes made it difficult to follow the authors’ points and sometimes resulted in ambiguity. I list many of those here, though I am hopeful that an editor can catch others.

Response: Thanks very much for your valuable comments. We checked the grammar throughout the manuscript and revised them according to your suggestions.

 

Comment 7: Line 29-30:  airborne pollutants -> anthropogenic ally-generated chemicals.

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 31 of the revised version.

 

Comment 8: Sentence on Line 31 is awkward.  I suggest.  “Therefore, fog detection is crucial to effectively support traffic planning and to provide information and bulletins for reducing risks to human health.”

Response: Thanks for your valuable suggestion. We changed the sentence according to your comment. Please see Line 32 to 34 of the revised version.

 

Comment 9: Line 44: “which is” -> “for which it is” Do the authors means that it is difficult to isolate TIR and MIR at night or during the day? They need to clarify that because as it currently is written, it seems contradictory.

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 46 of the revised version. We are sorry for the ambiguous description. We changed the sentence as “Considering the great difference of the dual-channel brightness temperature difference (BTD) between fog, surface and clouds, BTD is widely used in nighttime fog detection [9-14]. “. Please see Line 43 to 45 of the revised version.

 

Comment 10: Line 60: phrase beginning with “while” is a run-on sentence.  Please fix

Response: Thanks for your comment. We changed it as “However, the method is difficult to monitor fog areas in real-time due to the low temporal resolution of the polar-orbiting satellites” (see lines 62 to 64). We hope the revised version could describe the method clearly.

 

Comment 11: Line 67: “lacking most” -> “lacking in most”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 70 of the revised version.

 

Comment 12: Line 68:  cloud forms -> clouds form

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 71 of the revised version.

 

Comment 13: Line 106: “locates” -> “is located in”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 117 of the revised version.

 

Comment 14: Line 109: “In the eastern…”  not sure what authors intended to say, but this is grammatically problematic.

Response: We are sorry for the ambiguous description. We changed it as “The eastern of the study area is located in…”. Please see Line 120 of the revised version.

 

Comment 15: Line 109: “it locates” –> “it is located”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 120 of the revised version.

 

Comment 16: Line 110: “with the average altitude drops…” -> “where the average altitude drops…”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 121 of the revised version.

 

Comment 17: Line 134: “hazy” -> “haze”

Response: Thanks for your comment. We deleted the word after comprehensive consideration in the revised version.

 

Comment 18: Line 144: “Mask result” -> I am not clear on the meaning of this term. Please either describe it further or cite a reference that does.

Response: We are sorry for the ambiguous description. We changed the word as “target result

”. We clarified it in line 158 of the revised manuscript.

 

Comment 19: Line 152: “more abstract information” -> this is vague.  What information do the authors refer to more specifically?

Response: We are sorry for the ambiguous description. The “more abstract information” means the higher-level semantics. We changed the word as “the higher-level semantics” (see line 166 of the revised manuscript). We clarified it in line 166 to 167 of the revised manuscript.

 

Comment 20: Line 168 -> I see no reference to “Scale” in Figure 3. I have no idea what it refers to.

Response: We are very sorry for the ambiguous description and thanks for your comment. The “Scale” means “weighted”, We changed the sentence as “and 1 channel weighted operation” (see lines 193 to 194 of the revised manuscript), and we redraw Figure 3. The weighted attention feature  can be acquired through the weight factor W_p and the pixel value x_p of X at channel p (Eq. (4)). we explained it in lines 222 to 223 of the revised manuscript.

 

Comment 21: Figure 3:  I am unclear what the dimensions shown, e.g. C/16x1x1 refer to and why they are shown in the figure.

Response: Thanks for your comment. We explained it in lines 193 to 199 of the revised manuscript. Specifically, the Global pooling (GP) refers to the average of all the pixels of the feature map on each channel. FC1 is a fully connected layers with C/16 filters, which is calculated by the weighted summation. FC2 is also a fully connected layers with C filters, which allows the number of outputs to be consistent with the number of channels. Finally, the sigmoid normalizes these learned weights to be between 0-1 for dimensionless processing of different features.

 

Comment 22: Right after Line 187: This should be Equation (3) not Equation (2).

Response: We are very sorry for the writing mistake. We corrected it in line 219 of the revised manuscript.

 

Comment 23: Lines 188-189:  Line ending with ‘calculates’ is a sentence fragment.

Response: We are very sorry for the writing mistake. We replace that sentence with “where W_p represents the weight factor of channel p after function calculation of FC₂ and Sigmoid “. Please see Line 221 of the revised version.

 

Comment 24: Line 256: Not sure what is meant by “divided by 7:3” What is 7:3?

Response: We are sorry for the ambiguous description. We changed it as “…divided into 70% for training and 30% for validating” (lines 135 to 136 of the revised manuscript).

 

Comment 25: Line 262-263: “after the number of rounds” -> “after the number of rounds designated as Epoch” (at least I think that is the authors’ intent?)

Response: We are sorry for the ambiguous description. We changed it as “after the studies number of Epoch minus Decaying Epoch”. We corrected it in lines 313 to 314 of the revised manuscript.

 

Comment 26: Line 272: “Repeat the above 1-2 process to iterated” -> “Processes 1-2 are iterated until…”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 321 of the revised version.

 

Comment 27: Line 273: “a certain value, the model converges...”  -> “a certain value.  This means that the model converges…”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 322 of the revised version.

 

Comment 28: Line 302-303: This seems like circular logic. It implies that the satellite data are the observations that are compared with the model. Am I misunderstanding, or isn’t the satellite data what is driving the model? If this is the case, then the fact that the satellite data shows increasing or decreasing fog should not suggest whether the algorithm is working or not – it should be driving the algorithm, right? If I am correct, please change this. If I am not correct, please adjust the text to explain this.

Response: We are very sorry for the ambiguous description and thanks for your comment. We change this statement as “the detection results of satellite data can capture the generation and disappearance of fog, which is highly consistent with the ground observations”. Please see Line 353 to 355 of the revised version.

 

Comment 29: Line 304: “The fog area was fragmented due to low cloud” Not sure what the authors mean. Do they mean that the fog was interspersed with low clouds?

Response: We are very sorry for the ambiguous description and thanks for your comment. We change this statement as “The fog area was fragmented because it was interspersed with low cloud”. Please see Line 355 to 357 of the revised version.

 

Comment 30: Line 308: “eastern” -> “eastern portion of the figure”?

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 359 of the revised version.

 

Comment 31: Line 309: “data shows” -> “data show” (because data are plural)

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 361 of the revised version.

 

Comment 32: Line 312: “refined well because the…” -> “defined well because of the …”?  (Is this the authors’ intent?)

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 364 of the revised version.

 

Comment 33: Line 366: “the previous studies” -> “those of the previous studies”

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 424 to 425 of the revised version.

 

Comment 34: Line 371: “furtherly” -> “further”

Response: Thanks for your comment. We changed the word as your suggestion. Please see Line 429 of the revised version.

 

Comment 35: Lines 373 and 374: “accurate” -> “accurately”

Response: Thanks for your comment. We changed the words as your suggestion. Please see Line 431 of the revised version.

 

Comment 36: Line 382: “is clearer” -> “is clearer”

Response: We are sorry for the writing mistake. We changed the phrase as your suggestion. Please see Line 440 of the revised version.

 

Comment 37: Line 401: “it developed” -> “we developed”

Response: Thanks for your comment. We changed the words as your suggestion. Please see Line 461 of the revised version.

 

Comment 38: Lines 406-407:  I am not sure what the authors mean by: “Both the DDF-Net and U-net have false discrimination under the fog dissipation, medium/high clouds and low cloud” Do they mean that the algorithms have difficulty identifying fog when these phenomena are occurring? If so, I would rewrite this sentence as “Both the DDF-Net and U-net are challenged in detecting fog during the periods of time when the fog is dissipating and when there are clouds.”

Response: Thanks for your comment. We changed the sentence as your suggestion. Please see Line 466 to 467 of the revised version.

 

Author Response File: Author Response.docx

Reviewer 2 Report

 

Review of paper:

Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning
Algorithm under Terrain-restriction

By Ran et al.

Submitted to Remote Sensing

Manuscript remotesensing-1818786

 

Recommendation: Major revisions

General comments:

The manuscript presents results from use of machine learning algorithms aiming to improve the identification of fog at dawn and dusk from multi-channel satellite imagery. The study provides some indication of an improvement of results based on case studies.

The manuscript would greatly benefit from a number of significant improvements. I would invite the authors to consider the following:

1. Since this is not a journal which specializes in machine learning (ML), no assumption should be made about the level of knowledge of the reader on the subject. Therefore, the way ML is used on this particular study should be clearly described (in layman’s terms as much as possible) so the reader gains a broad understanding of what is done. I realize this is significant challenge, but an attempt should be made. In the current text, I did not understand how these methods are applied and why particular algorithms have been chosen over others. A number of algorithms are mentioned, but it remains unclear to me which ones are actually used and why.
2. The description of the methods developed in this paper should be more clearly described. For instance, what are the inputs (station visibility? Yes/no fog detection, etc), and the output? Please describe. Also, it remains unclear how information about the SZA is included, as well as information on terrain from the DEM, the two elements that I understood are supposed to be the focal points of the study. The mention of these items in the list of “future efforts” in Section 5 leave me confused. Clarification on all these items are necessary.  
3. The authors should discuss the relevance of the results toward real-world application, as results presented are based on a small sample of cases. This is particularly a sensitive issue with ML and related training of the algorithms for one thing, but also the evaluation of the results over a small number of independent cases.

 

 

Specific comments and recommended revisions:

1.    Line 28: Why would fog be more likely as dusk and dawn specifically? In particular, why at dusk compared to later in the middle of the night? Please clarify.

2.    Line 61: “… low temporal resolution of the satellites”. Perhaps you are specifically referring to polar-orbiting satellites? Please add the qualifier.

3.     Line 61: add “geostationary” before Fengyun-4A.

4.    Line 77: What are the “targets” here? Please describe.

5.    Lines 91-92: And how are these problems addressed in this particular application of ML? Please discuss briefly.

6.    Line 96: What does “fuse” mean here? Please explain and describe how this method is integrated.

7.    Line 106 and elsewhere: replace “locates” by “is located”.

8.    Line 112: source of water vapor, from what? Lakes? Industry? Please clarify.

9.    Line 122: What does “14 of 7” mean?

10. Lines 124-125: “training and validation” is redundant. Please re-phrase.

11. Lines 128-129: What is the temporal resolution of the ground truth data? Please clarify.

12. Line 152 “more abstract information”: Such as? Please describe in more detail.

13. Line 153: what does “downward” mean in the present context? Please explain.

14. Line 157: the “true target mask” remains undefine at this point. Please rectify.

15. Line 161: What does “Relu” mean? Please define in the caption of Fig. 2.

16. Line 254: Please provide the reader with reference to the “Adams descent” and “BN”, or describe. What does BN stand for?

17. Line 318, caption of Figure 7: please mention that symbols represent surface observations at stations. Not hourly? why available just at 8am.

18. Figures 9 and 10: “our method” should be labelled “DDF-Net”. IS that correct? Please clarify.

19. Line 423-424: Why better detection in mountainous areas and inclusion of SZA left for future efforts? Not the main points of this work (as suggested by the title of the paper?

 

 

 

Author Response

Response to Comments of Reviewer 2

 

Dear Reviewer

 

Thank you for your letter and comments concerning our manuscript entitled “Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-restriction” (No. remotesensing-1818786). We were encouraged by the positive comments regarding the usefulness of our research. The revised manuscript now includes appropriate modifications and/or improvements which have been made possible from your comments. Revised portions are marked in red in the revised manuscript. The main corrections and responds to comments are as follows:

 

1. General comment

Comment 1: The manuscript presents results from use of machine learning algorithms aiming to improve the identification of fog at dawn and dusk from multi-channel satellite imagery. The study provides some indication of an improvement of results based on case studies.

The manuscript would greatly benefit from a number of significant improvements. I would invite the authors to consider the following:

Response: Thanks very much for your positive evaluation and valuable comments. We read them carefully and revised the manuscript profoundly. The revised portions are marked in red in the revised manuscript.

 

Comment 2: Since this is not a journal which specializes in machine learning (ML), no assumption should be made about the level of knowledge of the reader on the subject. Therefore, the way ML is used on this particular study should be clearly described (in layman’s terms as much as possible) so the reader gains a broad understanding of what is done. I realize this is significant challenge, but an attempt should be made. In the current text, I did not understand how these methods are applied and why particular algorithms have been chosen over others. A number of algorithms are mentioned, but it remains unclear to me which ones are actually used and why.

Response: Thanks for your valuable comment. We described and explained the methodology in detail according to your suggestion. Specifically, we improved the principles and processes of deep learning algorithm and the SE-Net Module, and also the development and application in the satellite fog detection. Please see the red portions of the sections 2.2~2.4 of the revised version. Furthermore, we improved the Figure 2~5 to help the readers under our model clearly and directly. Finally, we also added more references to help the readers gathering more information of the models. We hope the revised version could help the readers gain a broad understanding of what is done in our work.

 

Comment 3: The description of the methods developed in this paper should be more clearly described. For instance, what are the inputs (station visibility? Yes/no fog detection, etc), and the output? Please describe. Also, it remains unclear how information about the SZA is included, as well as information on terrain from the DEM, the two elements that I understood are supposed to be the focal points of the study. The mention of these items in the list of “future efforts” in Section 5 leave me confused. Clarification on all these items are necessary. 

Response: Thanks for your comment. We improved the method according to your suggestion. The inputs include the ground observation data, satellite data of H8/AHI (7 bands in Table 1) and the terrain data (DEM), while the outputs refer to spatial coverage of fog detection under different times. We explained it in lines 236 to 238 of the revised manuscript. The SZA and terrains act as the constraint conditions in our model. Specifically, SZA strongly affects the fog spectral characteristics, even with the same remote sensing band. For example, the VIS band has a strong effect on the low SZA regions, while it is ineffective in the nighttime regions. Meanwhile, the terrain strongly affects the fog formation and development, which would act as one of the key parameters of fog detection. We explained it in lines 287 to 293 of the revised manuscript. Two issues would be addressed in our future research. Firstly, other geographical parameters (i.e., latitude and surface temperature, etc.) would also affect the fog formation, which should be addressed in the algorithm to improve the accuracy of satellite fog detection. Secondly, a public fog dataset for each season would be developed and published for supporting the management of traffic and public health. We clarified it in lines 482 to 486 of the revised manuscript.

 

Comment 4: The authors should discuss the relevance of the results toward real-world application, as results presented are based on a small sample of cases. This is particularly a sensitive issue with ML and related training of the algorithms for one thing, but also the evaluation of the results over a small number of independent cases.

Response: Thanks for your comment. It is true that the performance of satellite detection algorithm strongly relies on the training and validation samples. In this study, we collected the ground observation from China Meteorological Administration to execute and validate our algorithm, which is one of the most accurate official data sets with most stations. We explained it lines 139 to 150 of the revised manuscript. Even this, however, there would be unavoidable uncertainty due to the numbers and locates of the ground observation. We will continue to collect the ground observation for improving the execution and validation in our future searches. We discussed it in lines 402 to 407 of the revised version.

 

2. Specific comments and recommended revisions

Comment 1: Line 28: Why would fog be more likely as dusk and dawn specifically? In particular, why at dusk compared to later in the middle of the night? Please clarify.

Response: Thanks for your comment. Solar radiation is weak at dawn and dusk, resulting in a high probability of fog formation because of the low surface air temperature and high vapor saturation. We clarified it in lines 28 to 29 of the revised manuscript.

 

Comment 2: Line 61: “… low temporal resolution of the satellites”. Perhaps you are specifically referring to polar-orbiting satellites? Please add the qualifier.

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see lines 62 to 64 of the revised version.

 

Comment 3: Line 61: add “geostationary” before Fengyun-4A.

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see line 64 of the revised version.

 

Comment 4: Line 77: What are the “targets” here? Please describe.

Response: We are sorry for the ambiguous description. The “targets” means the detection output of the deep learning (i.e., the fog detection of this study). We changed the word as “detection target” (Please see line 80 of the revised manuscript).

 

Comment 5: Lines 91-92: And how are these problems addressed in this particular application of ML? Please discuss briefly.

Response: Thanks for your valuable comment. We clarified it in lines 95 to 103 of the revised manuscript. Specifically, these problems can be optimized by improving the application of deep learning model. For example, the sensitivity of different SZA and terrain conditions to fog detection can be effectively alleviated by introducing the Squeeze-and-Excitation Networks (SE-Net), which could automatically obtain the weights of the contributions of the parameters. Uneven distribution of training data samples can be solved through the batch normalization of input data. Additionally, the training process could be optimized by designing the independent adaptive learning rates for different parameters and by using efficient optimization algorithms. However, the relevant researches are still rare, which need a further investigation.

 

Comment 6: Line 96: What does “fuse” mean here? Please explain and describe how this method is integrated.

Response: We are sorry for the ambiguous description. The “fuse” means integration. We changed the word as “integrated” (Please see Line 107 of the revised manuscript). As shown in Figure 5, the "CL-Relu-PL" structure is further integrated as "CL-BN-Relu-SE-Net-PL". We also explained how the method is integrated in lines 242 to 254 in detail.

 

Comment 7: Line 106 and elsewhere: replace “locates” by “is located”.

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see line 117 of the revised version.

 

Comment 8: Line 112: source of water vapor, from what? Lakes? Industry? Please clarify.

Response: Thanks for your comment. The climate of the study area is characterized by a rich source of water vapor from the Western Pacific. We clarified it in line 124 of the revised manuscript.

 

Comment 9: Line 122: What does “14 of 7” mean?

Response: We are sorry for the ambiguous description. The number 7 means the total bands of the data sets selected in this study. We corrected it as “…we select the data sets of bands 3, 5, 6, 7, 11, 13, and 14 for model training and validation” (Please see lines 133 to 134 of the revised manuscript). We hope the revised version could describe the data set clearly.

 

Comment 10: Lines 124-125: “training and validation” is redundant. Please re-phrase.

Response: Thanks very much for your valuable comment. We deleted the phrase as your suggestion in the revised manuscript. Please see lines 135 to 136 of the revised version.

 

Comment 11: Lines 128-129: What is the temporal resolution of the ground truth data? Please clarify.

Response: Thanks for your comment. The temporal resolution of the ground observation data is 3 hours. We clarified it in lines 141 to 142 of the revised manuscript.

 

Comment 12: Line 152 “more abstract information”: Such as? Please describe in more detail.

Response: We are sorry for the ambiguous description. The “more abstract information” means the higher-level semantics. We changed the word as “the higher-level semantics”. Please see lines 166 to 167 of the revised manuscript.

 

Comment 13: Line 153: what does “downward” mean in the present context? Please explain.

Response: Thanks for your comment. The PL is implemented to sample the channel feature information downward, which divides the input image into several rectangular areas to get the maximum value of each sub-area. Through this way, it will continuously reduce the spatial size of the data, which could reduce the parameters and improve the computation. We clarified it in lines 168 to 171 of the revised manuscript.

 

Comment 14: Line 157: the “true target mask” remains undefined at this point. Please rectify.

Response: We are sorry for the ambiguous description. The “true target mask” means the actual fog coverage (Please see Line 179 to 180 of the revised manuscript), which is usually generated by the combination of visual interpretation and ground observation that is used as the label data for training or validating the algorithm. We clarified it in lines 183 to 185 of the revised manuscript.

 

Comment 15: Line 161: What does “Relu” mean? Please define in the caption of Fig. 2.

Response: Thanks for your comment. Relu is used to achieve non-linear transformation between the CL and PL layers. We clarified it in lines 173 to 176 and in the caption of Figure. 2 of the revised manuscript.

 

Comment 16: Line 254: Please provide the reader with reference to the “Adams descent” and “BN”, or describe. What does BN stand for?

Response: Thanks for your comment. We provide references for Adams and Batch Normalization (BN) respectively. BN is a normalization of the input data to account for the sudden change in the distribution of the data. We clarified it in lines 304 to 306 of the revised manuscript.

 

Comment 17: Line 318, caption of Figure 7: please mention that symbols represent surface observations at stations. Not hourly? why available just at 8am.

Response: Thanks very much for your valuable comment. We added the symbols represent surface observations at stations as your suggestion in the revised manuscript in Figures 7~10. Please see lines 373 to 375 of the revised version. The temporal resolution of the ground observation data is 3 hours. The data include observations 8 times per day: 02:00 local time (LT), 05:00 LT, 08:00 LT, 11:00 LT, 14:00 LT, 17:00 LT, 20:00 LT, and 23:00 LT. Among them, the data at 8:00 (summer: 5:00) and 17:00 (summer: 20:00) are used to evaluate the accuracy of the DDF-Net. We clarified it in lines 141 to 143 of the revised manuscript.

 

Comment 18: Figures 9 and 10: “our method” should be labelled “DDF-Net”. IS that correct? Please clarify.

Response: Thanks for your comment. We correct the phrase and repaint Figures 9 and 10 as your suggestion. Please see lines 448 to 449, and lines 456 to 457 of the revised version.

 

Comment 19: Line 423-424: Why better detection in mountainous areas and inclusion of SZA left for future efforts? Not the main points of this work (as suggested by the title of the paper?

Response: We are sorry for the ambiguous description. We removed the statement "better detection in mountains". Two issues would be addressed after the careful consideration and summarization of our work. Firstly, other geographical parameters (i.e., latitude and surface temperature, etc.) would also affect the fog formation, which should be addressed in the algorithm to improve the accuracy of satellite fog detection. Secondly, a public fog dataset for each season would be developed and published for supporting the management of traffic and public health. Please see lines 482 to 486.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Review of Satellite Fog Prediction at Dawn and Dusk… by Ran et al

 

While the authors did make substantial improvements to the paper based on both reviewers’ comments, I still have some significant concerns.  The primary one I have is that despite one of the reviewers’ questions about Figure 7-10, and the use of false color images, which the authors said they would remove, since they do not add to the discussion, the false color images are still there.   Why is this the case?  I do not understand why they are still there and what they add to the discussion.  I cannot tell from them where the fog is or is not.  Additionally, in Figures 7 and 8, only two of the 8 images indicate ground observations.    I ask, then, what do the other 6 images in each figure (12 images in total) add to the discussion?  These figures are core to the paper and represent the results of pages of describing the models used.  Unless they are clarified and extraneous information is either removed or justified, as per the previous review, I cannot recommend publication of this paper. 

Other comments follow:

Lines 27-28:  I do not think the point the authors make about solar radiation being weak at dawn and dusk sufficiently addresses one of the reviewers’ comments about why dawn and dusk versus the middle of the night are important for fog formation.   Of course, solar radiation is weak at dawn and dusk, but it is even weaker at night (namely, it is 0 at night).

Lines 80-81: results through”-> results (that is, the detection of fog) through

Line 96:  of deep learning model -> of the deep learning model

Line 103:  relevant researches are still rare, which needs a further investigation -> relevant research is still rare, thus further investigation is warranted.

Line 139-140:  which was -> these data were

Line 145:  and haze ones -> and hazy

Line 146:  meters. -> meters, respectively. 

Line 161:  include mapping -> include the mapping

Line 165 – I think the authors should clarify what is meant by “textural and spectral” information and why it is important.

Line 171:  By “improve the computation” do they mean “improve the computational speed” or is it something else?

Line 175:  to the value -> to values

Line 183:  the word “usually” implies that sometimes other methods are used.  Either remove that word or specify what those other methods are in the less-common cases.

Lines 190-191:  , which is adopted in our study -> .  We choose to adopt this approach in our study.

Line 195:  FC is a fully connected layers -> FC consists of fully-connected layers

196:  FC is also fully connected layers ->  FC also consists of fully-connected layers

Lines 195-196 :  I am confused how FC consists of both C and C/16 filters.  Please clarify. 

Line 246:  as increasing -> by increasing

Line 313:  ac-cording -> according

Line 313:  formula (4) -> formula (9)

Line 323:  Input the new remote sensing data -> The new remote sensing data are input

Line 365:  mainly locates in the eastern -> is primarily located in the eastern portion of the figure

Lines 365-366:  reason might originate -> reason for this missing detection might originate

Line 366:  of the low cloud -> of low clouds

 

 

 

 

 

 

 

 

Author Response

Response to Comments of Reviewer 1

 

Dear Reviewer

 

Thank you for your letter and comments concerning our manuscript entitled “Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-restriction” (No. remotesensing-1818786). We were encouraged by the positive comments regarding the usefulness of our research. The revised manuscript now includes appropriate modifications and/or improvements which have been made possible from your comments. Revised portions are marked in red in the revised manuscript. The main corrections and responds to comments are as follows:

 

General Comments: While the authors did make substantial improvements to the paper based on both reviewers’ comments, I still have some significant concerns.  The primary one I have is that despite one of the reviewers’ questions about Figure 7-10, and the use of false color images, which the authors said they would remove, since they do not add to the discussion, the false color images are still there.   Why is this the case?  I do not understand why they are still there and what they add to the discussion.  I cannot tell from them where the fog is or is not.  Additionally, in Figures 7 and 8, only two of the 8 images indicate ground observations. I ask, then, what do the other 6 images in each figure (12 images in total) add to the discussion?  These figures are core to the paper and represent the results of pages of describing the models used.  Unless they are clarified and extraneous information is either removed or justified, as per the previous review, I cannot recommend publication of this paper.

Responses: Thanks very much for your valuable comments. We removed the false color to improve the Figures 7-10 according to your comments. Specifically, we removed the false color of the Figures 7 and 8, and added the spatial patterns of ground sites in Figure 7 (a), (b) and (d), and Figure 8 (a), (b) and (c). We also removed the false color of the Figures 9 and 10 in the revised version. We hope the revised version could address the patterns of fog clearly. 

 

Comment 1#: Lines 27-28: I do not think the point the authors make about solar radiation being weak at dawn and dusk sufficiently addresses one of the reviewers’ comments about why dawn and dusk versus the middle of the night are important for fog formation. Of course, solar radiation is weak at dawn and dusk, but it is even weaker at night (namely, it is 0 at night). 

Responses: We are sorry for the ambiguous description. We agree you that the solar radiation is weakest, which is helpful for fog formation. Beside in night, solar radiation is also weak in dawn and dusk, resulting in a relatively high probability of fog formation because of the low surface air temperature and high vapor saturation. We explained it in lines 28 to 30 of the revised manuscript.

 

Comment 2#: Lines 80-81: results through”-> results (that is, the detection of fog) through

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 81 of the revised version.

 

Comment 3#: Line 96:  of deep learning model -> of the deep learning model

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 96 of the revised version.

 

Comment 4#: Line 103:  relevant researches are still rare, which needs a further investigation -> relevant research is still rare, thus further investigation is warranted.

Responses: Thanks for your comment. We changed the description as your suggestion. Please see Line 103 to 104 of the revised version.

 

Comment 5#: Line 139-140:  which was -> these data were

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 140 to 141 of the revised version.

 

Comment 6#: Line 145:  and haze ones -> and hazy

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 146 of the revised version.

 

Comment 7#: Line 146:  meters. -> meters, respectively.

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 147 of the revised version.

 

Comment 8#: Line 161:  include mapping -> include the mapping

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 163 of the revised version.

 

Comment 9#: Line 165 – I think the authors should clarify what is meant by “textural and spectral” information and why it is important.

Responses: We are sorry for the ambiguous description. CL uses a specified number of convolution kernels in each layer to extract the feature information of each pixel of x, with the shallow layer includes the textural and spectral feature that are used as the basis classification. Please see Line 167 to 168 of the revised version.

 

Comment 10#: Line 171:  By “improve the computation” do they mean “improve the computational speed” or is it something else?

Responses: We are sorry for the ambiguous description. We changed it as “improve the computational speed”. Please see Line 173 of the revised version.

 

Comment 11#: Line 175:  to the value -> to values

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 177 of the revised version.

 

Comment 12#: Line 183:  the word “usually” implies that sometimes other methods are used.  Either remove that word or specify what those other methods are in the less-common cases.

Responses: Thanks for your comment. We removed the phrase "usually" as your suggestion. Please see Line 185 of the revised version.

 

Comment 13#: Lines 190-191:  which is adopted in our study -> We choose to adopt this approach in our study.

Responses: Thanks for your comment. We changed the description as your suggestion. Please see Line 193 to 194 of the revised version.

 

Comment 14#: Line 195:  FC is a fully connected layers -> FC consists of fully-connected layers

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 198 of the revised version.

 

Comment 15#: 196:  FC is also fully connected layers -> FC also consists of fully-connected layers

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 201 to 202 of the revised version.

 

Comment 16#: Lines 195-196: I am confused how FC consists of both C and C/16 filters.  Please clarify.

Responses: We are sorry for the ambiguous description. The FC includes two sub-modes of FC₁ and FC₂. FC₁ consists of fully-connected layers with C/16 filters, which is calculated by the weighted summation to reduce the number of parameters and improve the calculation efficiency. FC₂ consists of fully-connected layers with C filters, which assure the same numbers of outputs and channels. Please see Line 197 to 202 of the revised version.

 

Comment 17#: Line 246:  as increasing -> by increasing

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 254 of the revised version.

 

Comment 18#: Line 313:  ac-cording -> according

Responses: We are very sorry for the writing mistake. We corrected it in line 330 of the revised manuscript.

 

Comment 19#: Line 313:  formula (4) -> formula (9)

Responses: We are very sorry for the writing mistake. We corrected it in line 330 of the revised manuscript.

 

Comment 20#: Line 323:  Input the new remote sensing data -> The new remote sensing data are input

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 340 of the revised version.

 

Comment 21#: Line 365:  mainly locates in the eastern -> is primarily located in the eastern portion of the figure

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 383 of the revised version.

 

Comment 22#: Lines 365-366:  reason might originate -> reason for this missing detection might originate

Responses: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 384 of the revised version.

 

Comment 23#: Line 366:  of the low cloud -> of low clouds

Response: Thanks for your comment. We changed the phrase as your suggestion. Please see Line 384 to 385 of the revised version.

Author Response File: Author Response.docx

Reviewer 2 Report

I find the revised manuscript improved to some extent over the original submission. I appreciate the authors' effort in describing in more detail the methods used in the study. Although I still find a lot of it unclear, leaving the reader having a difficult time really understanding how the different pieces contribute in the application at hand. However I do realize that a clear and concise description of these deep learning methods poses a significant challenge. Therefore, even though I remain unclear about parts of this work, I will give benefit of the doubt. But I encourage authors to continue thinking about improved ways of describing the methods and providing the reader with more intuitive insights.

Author Response

Dear Reviewer

 

Thank you for your letter and comments concerning our manuscript entitled “Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-restriction” (No. remotesensing-1818786). We were encouraged by the positive comments regarding the usefulness of our research. The revised manuscript now includes appropriate modifications and/or improvements which have been made possible from your comments. Revised portions are marked in red in the revised manuscript. The main corrections and responds to comments are as follows:

 

Comment 1: I find the revised manuscript improved to some extent over the original submission. I appreciate the authors' effort in describing in more detail the methods used in the study. Although I still find a lot of it unclear, leaving the reader having a difficult time really understanding how the different pieces contribute in the application at hand. However I do realize that a clear and concise description of these deep learning methods poses a significant challenge. Therefore, even though I remain unclear about parts of this work, I will give benefit of the doubt. But I encourage authors to continue thinking about improved ways of describing the methods and providing the reader with more intuitive insights.

Response: Thanks very much for your positive evaluation and valuable comments. Thanks very much for your valuable comments. We checked and improved the methods throughout the section 2.3~2.4 in a concise way. Please see the red portions of the revised version. 

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

3^rd review of 1818786

 

 

 

Figures are now acceptable.  I recommend publication with the following topical adjustments:

 

Line 28:  “besides in night, solar radiation is also weak in dawn and dusk” ->  “besides during the night, solar radiation is also weak during the dawn and dusk” although, admittedly, this still does not state why the authors focus on dawn and dusk rather than the entire night.  I believe an additional sentence is warranted, even if it is because that is when visual data are available at dawn and dusk or whatever other reasons the authors have.  Please clarify that point.

The sentence beginning on Line 164 is far too long.  I suggest breaking it up as follows:

Place a period after Line 166.  That is, make “of each pixel x,” -> “of each pixel x”

Then:  “with the shallow layer includes the textural and spectral feature that are used as the basis classification, and the deeper layer represents the higher-level semantics, which is the feature information obtained after several convolutions (feature extraction) [35].” -> “The shallow layer includes the textural and spectral features that are used as the basis classification, and the deeper layer represents the higher-level semantics, which is the feature information obtained after several convolutions (feature extraction) [35].”  Note the spectral feature(s) correction too.

Line 321:  “parameters are optimized” -> “optimized parameters”?  Not sure if this is the authors’ intent, but the grammar does need to be fixed.

Multiple locations where it is stated “symbols of square, triangle, circle, cross  represent” ->  “the square, triangular, circular and cross symbols represent…”  this happens on several lines, eg lines 389 and 390, but others as well.  Also on Lines 404-406, 459, 465 and maybe other places too.  Please fix.

 

Author Response

Dear Reviewer

 

Thank you for your letter and comments concerning our manuscript entitled “Satellite Fog Detection at Dawn and Dusk Based on the Deep Learning Algorithm under Terrain-restriction” (No. remotesensing-1818786). We were encouraged by the positive comments regarding the usefulness of our research. The revised manuscript now includes appropriate modifications and/or improvements which have been made possible from your comments. Revised portions are marked in red in the revised manuscript. The main corrections and responds to comments are as follows:

 

General Comments: Figures are now acceptable. I recommend publication with the following topical adjustments

Responses: Thanks very much for your positive evaluation. We continued to revise and improve the manuscript according to your valuable comments. Please see the red portions of the revised manuscript.

 

Comment 1#: Line 28: “besides in night, solar radiation is also weak in dawn and dusk” ->  “besides during the night, solar radiation is also weak during the dawn and dusk” although, admittedly, this still does not state why the authors focus on dawn and dusk rather than the entire night.  I believe an additional sentence is warranted, even if it is because that is when visual data are available at dawn and dusk or whatever other reasons the authors have.  Please clarify that point.

Responses: Thanks very much for your comment. We corrected the sentence as your suggestion in line 28 of the revised manuscript. Furthermore, we explained the reason why to focus on dawn and dusk. The dense fog seriously reduces horizontal visibility, which adversely affects public traffic and human health (particularly in rush hours at dawn and dusk). We clarified it in lines 30 to 32 of the revised manuscript.

 

Comment 2#: The sentence beginning on Line 164 is far too long.  I suggest breaking it up as follows:

Place a period after Line 166.  That is, make “of each pixel x,” -> “of each pixel x”

Then:  “with the shallow layer includes the textural and spectral feature that are used as the basis classification, and the deeper layer represents the higher-level semantics, which is the feature information obtained after several convolutions (feature extraction) [35].” -> “The shallow layer includes the textural and spectral features that are used as the basis classification, and the deeper layer represents the higher-level semantics, which is the feature information obtained after several convolutions (feature extraction) [35].”  Note the spectral feature(s) correction too.

Responses: We highly appreciate your valuable comment, which help to improve the sentence in a clear and concise way. We corrected it as your suggestion. Please see lines 165 to 170 of the revised manuscript.

 

Comment 3#: Line 321: “parameters are optimized” -> “optimized parameters”?  Not sure if this is the authors’ intent, but the grammar does need to be fixed.

Responses: Thanks very much for your comment. We corrected it as your suggestion after a careful consideration. Please see line 322 of the revised manuscript.

 

Comment 4#: Multiple locations where it is stated “symbols of square, triangle, circle, cross represent” ->  “the square, triangular, circular and cross symbols represent…”  this happens on several lines, eg lines 389 and 390, but others as well.  Also on Lines 404-406, 459, 465 and maybe other places too. Please fix.

Responses: Thanks very much for your comment. We corrected the sentences according to your suggestions. Please see lines 390 to 391, lines 405 to 406, line 460, and line 466 of the revised manuscript.