Optimal Inversion of Conversion Parameters from Satellite AOD to Ground Aerosol Extinction Coefficient Using Automatic Differentiation

Round 1

Reviewer 1 Report

The author addressed most of my comments well, but I believe some further improvements are necessary to make the paper ready for publication. Please find my specific comments below.

1.

Perhaps I just didn’t find the right part, but it appears to me that the revision didn’t quite address one of the main recommendations of my initial review:

“Most importantly, I am not quite sure about the importance of accuracy improvements the newly presented “optimal GAC” method offers over the previously published “empirical GAC” method. I wonder about this, because Table 3 shows that the “optimal GAC” methods increases the correlation with truth from 0.56 to 0.58, and reduces the root mean square error from 59 um/m3 to 58 um/m3. I don’t know if this improvement is significant in practical applications and if the benefits are worth the effort required for implementing the new technique. It would be helpful if the paper offered some guidance on this.”

Although the accuracy improvements are slightly larger in the revised version of Table 3 (indicating that R increases from 0.54 to 0.58 and RMSE drops from 0.59 to 0.56), I still feel that it is important to either

(a) mention in the abstract and in the discussion (or the conclusions) that the new method offers fairly modest accuracy improvements over existing methods by adding a word like “slightly” or “modestly” in Lines 25 and 430, for example to make the text “slightly improved” and “slightly better”, or

(b) explain (if this is correct) that, while the accuracy improvements offered by the new method are modest in Table 3, the main importance of accuracy improvements offered by the new method lies in the finding mentioned in Line 374 (that the new method improved the test R2 from 0.78 to 0.90 in the downstream validation of PM2.5 estimation), or

(c) discuss why the presented accuracy improvements are important in practice, even though the improvements in RMS errors and correlation coefficients in Table 3 appear fairly modest.

2.

Line 14: The text “the Terra and Aqua sensors aboard MODIS” got it backward: Actually, MODIS is the sensor and Terra and Aqua are two satellites that carry MODIS sensors. Accordingly, the text should be changed to “the MODIS sensors aboard the Terra and Aqua satellites”.

3.

Lines 15-16: The text “distributed along the vertical height of plane atmosphere” seems awkward and imprecise; I recommend a rewording to something like “integrated over all altitudes in the atmosphere”.

4.

Line 16: The words “related with” should be replaced by “related to”.

5.

Line 18: The word “an” should be deleted.

6.

Line 24: The acronym “MSE” should be defined.

7.

Line 29: I recommend adding “an” or “the” at the beginning of the line, and removing “the” from “the other”.

8.

Line 276: I recommend replacing “or/and” by the more usual “and/or”.

9.

Line 381: I suggest replacing “inverse to seasonal” by “opposite of the seasonal”.

10.

Line 383: I recommend replacing “for” by “within the”.

11.

Line 392: The word “averagely” should be replaced by “on average”.

12.

Line 426: I recommend adding “are” or “may be” right after “thus”.

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Review of manuscript “remote sensing-692653” titled:

“Satellite AOD to Ground Aerosol Coefficient Using 3 Automatic Differentiation”

This manuscript presents a new method to retrieved the PM_2.5 from the integrated AOD, generated operationally from MODIS 1 km product using MAIAC retrieval, which is the most recent and robust product. Not only PM_2.5 is an important parameter to retrieve from the widely used and widely covered MODIS data but the presented method of retrieval from the AOD is sound. The technique is based on using the automatic differentiation with its application in deep learning framework.

In this revised version, the authors crafted the presentation in coherent and convincing order and the writing is fairly well. The objective is clearly presented, the Introduction lays the ground and justification for the study and the analysis is well performed. One part I particularly like in this section is the explanation for why the ground level pollution is heavier in winter while the satellite AOD shows higher values in summer. I reviewed a few manuscripts about pollution in same region but did not find a reference to this observation.

The results were verified, and compared well, using measurements from 102 stations in the study region. Overall the method is suitable, and impressive, and the results are interesting.

I would encourage the authors to pursue applications of this work to study the seasonal and interannual variability of PM_2.5 in the same and other regions in the world.

I would recommend publication. The following minor suggestions do not mean request for another revision.

Due to my limited familiarity with the deep learning approach I could not review or verify the correctness of equations 10 and 11 as well as Fig. 3 and Table 1. Table 3 and Fig. 6 are the highlight results that proves the utility of the optimal GAC approach.

Some minor suggestions:

I would prefer “simulated GAC instead”of “GAC simulated”

Line 313: the author can mention the source of the PM_2.5 concentration.

Line 362: Figure 7 does not really show seasonal pattern. It shows higher GAC and PM_2.5 in winter but no identifiable pattern during the rest of the year.

 

In the caption of Fig. 8, add (c and f) at the end of the line.

Line 408 make it “amount of monitoring data sources”

Line 410 make it “such as” instead of “like”

Line 414 make it “the potential difficulty is the significant inconsistency …”

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Please see attached review. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The manuscript lacks the validation of the results. It also lacks the method to consider the aerosols in multiple layers, and it should be rejected. 

Reviewer 2 Report

The article discusses the methodology for the conversion from satellite AOD to ground aerosol coefficient. Generally, the submitted paper is well-written and addresses a very interesting and intensively developing field of science, including the use of satellite techniques in the assessment of air pollution by particulate matter, which is recognized as one of the most important environmental problems not only in China but around the world. The latest publications within this field are cited adequately.

In my opinion, the thematic background, methodology, results and conclusions have been described extensively. The article is very orderly and in principle I have no objections to its layout and substantive issues. A request to the authors to supplement the subsection 2.1 (Study region) with short description of climatic conditions and dominant emission sources. Too long sentences are often used (e.g. section 4, rows 318-329), which can disturb the transparency of the article and to some extent impede its reading.

Reviewer 3 Report

Review of manuscript remotesensing-600288

“Optimal Inversion of Conversion Parameters from Satellite AOD to Ground Aerosol Coefficient using Automatic Differentiation”

 

The manuscript introduces a new method, composed form elements already established in previous studies, to retrieve the ground aerosol parameter PM_2.5 from the atmospheric column estimation of AOD from MODIS. Considerable effort has been put in to develop this method. The manuscript introduces the formulation, the method of the solution and some comparison data. The results shown in Fig. 8 is the highlight results. It places more focus on the method than the results. But that should be OK since it is a new method that has potential for broad applications.

 

The work is worth publication in Remote Sensing but after two tasks of major revision: The first is about the writing. It needs considerable revision. I would recommend the author to use the professional editing service provided by MDPI. Many sentences are too long and difficult to follow, with many grammatical and style mistakes. The second is the description in section 2.4. This a crucial step in the method that starts in the paragraph “With (8) defined as the estimated …” (Line 210). It could not follow this description. It does not mean it is wrong but it is far from being clear.

 

The method addresses an important parameter to retrieve (mainly because of its health impact) from the readily-available, fine spatial and temporal coverage of MODIS and similar sensors. It has potential for applications though it needs more validation, and preferably comparison against CALIPSO data in a future study.

 

Here are some minor comments:

Abstract

Line 24: specify what ground aerosol coefficient mean

 

Introduction

Line 48: include unity for PM2.5

Line 52: why “particularly before the launch of MODIS”?

Line 63-65: explain why AOD is high in summer and low in winter. This depends on the region.

Line 89: The sentence “since PBLH determines the altitude of the vertical profile, it can be denominator factor in this conversion”. What does that mean?

Line 86-91: this is an example of an awkward sentence. What does “to particularly capture the influence of other cofounders” mean?

Line 100: What does “the results evaluated in comparison with the empirical conversion method” mean?

Figure 1: the font is too small. I could not read any label.

 

2.2 Dataset:

Line 119: what does “the pre-projection of the 1954 Beijing coordinate system ..” mean? Why did you apply to the images?

Line 122: not clear what you mean by fusing Aqua and Terra MODIS. Their overpasses occur at different times.

 

2.3 Simulation of ground aerosol extension coefficient

The formulation is presented well

Line 150: use comma instead of “and”

 

2.4 Solution by Automatic Differentiation

The explanation of the method is difficult to follow. I could not understand it. While the reader can check Reference 31, it is better to include a brief and clear description of the method. If necessary, include a figure to support the explanation.

Figure 3 has poor fonts. Needs improvement

I did not check the correctness of the information in Table 1.

 

2.5 Validation

Please clarify again the source of the data samples. This makes it easier for the reader to grab the information without going back in the text.

Line 234: I do not see 4 methods. I see only 3.

 

3.1 Description statistics

In table 2 you should provide and use statistics for winter and summer since it is mentioned that AOD and PM2.5 are different between these 2 seasons.

Line 258: “20,000 epochs of training”? what do you mean?

 

3.3 Comparison and validation

What is the difference between this section and the “2.5 Validation” section?

I do not see validation in this section. For validation you need ground measurement of PM2.5. I would suggest eliminate “validation” from the title

Figure 7 is not clear. There are Chinese characters on the x-axis and the font is extremely small.

 

3.4 Surface of simulated ground aerosol coefficient and PM2.5

It is not “Surface of …”. You can use instead “Spatial distribution of simulated …”

This section has only 7 lines. The author may keep it but should add more comments on the results in Fig. 8. As I mentioned, this is most interesting figure and it deserves comments on the spatial distribution of PM2.5 in relation to landscape, human activities and perhaps other factors.

 

Discussions:

This section needs considerable work to reorganize the information. It may be cancelled and the information moved to the Results section and the Conclusion section.

 

Conclusion

The only statement that summarizes the results is in lines 375-378. It is not clear. The manuscript has more information about the results that is worth inclusion here.

 

Reviewer 4 Report

This study is about conversion parameters from satellite AOD to ground aerosol coefficient. Before considering for publication, I have some detailed comments listed below to help improving this article.

1, Abstract, the importance of AOD and the complicated relationship between column AOD and ground aerosol properties are well known and should not be the focus of the abstract. Instead, the abstract should highlight the method, the unique feature of the method. The evaluations of the results for the case study. Quantification of the results is very important for scientific article, or else, how good/reasonable of this method is unknown. Need to address some technique detail, such as, what time period is studied? What kind of ground observations are used to evaluations? And what are the statistics results comparing against some ground measurement?

2, Figure 7, it would be good to overlay 3 plots on one plot to help to identify the performance of the method. 3 different colors on one plot would give a clearer idea of the comparisons and further analysis.

3, Need to elaborate more of the results with comparing with previous literature. A lot of studies did similar practice using different method. What are their results/limitations? What are the current studies’ limitations?

4, The conclusion section should be a more detailed summary comparing with the abstract section. Please consider to combine the current conclusion section with the discussion section and highlight the next step to improve the current method.

Reviewer 5 Report

In my view, this paper presents work that is worthy of publication. Its contribution lies in describing and testing a new technique for using satellite aerosol data in estimating surface air quality (specifically, the concentration of small particles, PM2.5). The methodology is sound and the presentation is clear and concise. Apart from some minor wording issues, I only have a few significant concerns. Most importantly, I wonder about the interpretation of findings and the overall evaluation of the proposed new technique. Please find my detailed comments below.

Major issues:

1.

After reading the paper, I was left with the overall sense that the proposed new method offers only modest improvements over a previously published approach described in reference [15]. Most importantly, I am not quite sure about the importance of accuracy improvements the newly presented “optimal GAC” method offers over the previously published “empirical GAC” method. I wonder about this, because Table 3 shows that the “optimal GAC” methods increases the correlation with truth from 0.56 to 0.58, and reduces the root mean square error from 59 um/m3 to 58 um/m3. I don’t know if this improvement is significant in practical applications and if the benefits are worth the effort required for implementing the new technique. It would be helpful if the paper offered some guidance on this. To be clear, I think it is worth publishing the new technique even if the results reveal that it does not offer huge improvements over previous techniques; at least readers will know. Moreover, it may be worth mentioning if the new method offered any advantages beyond accuracy (for example in computational needs or flexibility), which may make it a good starting point for incorporating into PM2.5 estimations the impacts of meteorology, elevation, and traffic density (as mentioned in Lines 353-354). In any case, it seems important that the discussion and conclusions give readers a bit more guidance on the benefits, drawbacks, and inherent limitations of the new, “optimal GAC” method. (For example, as mentioned in the introduction, all discussed methods work best in areas dominated by regional aerosols residing in the boundary layer, and work less well in areas dominated by long-range aerosol transport at higher altitudes.)

2.

Table 3: It does not seem correct that for the “GAC” and “empirical GAC” methods, the table shows the same values in the “Mean” and “Range” columns. Also, it is suspicious that the values in the “Standard deviation” column are two orders of magnitude smaller for the “GAC” and “empirical GAC” methods than for the other two methods.

 

Minor issues:

Lines 53-54: I suggest replacing “Since the sensors of the MODIS Terra and Aqua

satellites were onboard in 1999 and 2002, respectively” by something along the lines of “Since the sensors of the MODIS Terra and Aqua satellites started collecting data in 1999 and 2002, respectively”.

Lines 59-60: I suggest replacing “at all the altitudes along the orientation” by something like “at all altitudes along the line of sight”.

Line 65: I suggest removing the hyphen between “high” and “ground”.

Line 81: I suggest deleting “the altitude of”, or inserting a word like “mean” in front of “altitude”.

Line 92: I suggest replacing the word ”make” by something else (the meaning of the current wording is not quite clear to me and does not sound right).

Line 100: The word “were” seems to be missing between “results” and “evaluated”.

Line 143: The wording “The study region’s surfaces of RH” is not quite clear; should this be “The study region’s surface RH”?

Line 179: I suggest replacing “deriving” by “coming”,“originating”, or “arising”.

Lines 200-201: I suggest replacing “qualified” by “suitable”. Also, a different wording is needed, as it is not clear what is meant by “a stable property of numerical computation”.

Line 211: It would help to clarify what exactly is meant by “good property”.

Line 223: It should be clarified what the “Ada optimizer” is; is it a component/option of Tensorflow? If so, is there a generic method name for people who don’t use Tensorflow?

Line 223: The text should be refined to indicate that Table 1 contains not only the nodes and the derivatives for backward propagation, but also the derivatives for forward propagation. Also, it could be helpful to point out that each U value represents an element in the equations presented earlier; it may even help to mention which equation we use in defining each U value and/or its derivatives.

Lines 234-235: The first sentence of this paragraph mentions “four methods”, whereas the second sentence talks about “these variables”. A consistency between the two sentences (i.e., between “methods” and “variables”) should be achieved somehow. Also, the second sentence seems to mention only three methods/variables. It would help to clearly identify and number method/variable #1, method/variable #2, etc. (Is the fourth one based on Pearson’s correlation?)

Line 242: For consistency with the subsequent text that specifies similar differences, it should be specified how much higher the mean AOD was in the north than in the south.

Line 249: What exactly is meant by inter-quantile? For example, is it the range between the 20^th and 80^th percentile of the data?

Line 258: I suggest replacing “solve” by “obtain” or something similar. More importantly, it should be explained what is meant by training “epoch”.

Line 276: A zero seems to be missing: if R increases from 0.53 to 0.54, R**2 increases by 0.01, and not 0.1.

Figure 7: The horizontal axis labels should be replaced as they use Chinese characters (which are not helpful to readers who don’t speak Chinese).

Lines 318-324: This six-line sentence is too long; it should be split into two or more shorter sentences.