High-Resolution Mapping of Aerosol Optical Depth and Ground Aerosol Coefficients for Mainland China

Round 1

Reviewer 1 Report

For mainland China, the percentage of MODIS AOD missing is very high, which severely limits the application of AOD data. Although many methods have been used to estimate missingness of AOD, these methods are limited to local areas with limited test performance. This paper made great imputation of missing AOD, which is reliable as indicated by their high correlation with AERONET data. In addition, the paper made the conversion of AOD to ground aerosol coefficient, which in turn improved their power to the inversion of ground aerosol variables such as PM2.5 and PM10. Through the data sharing of estimated AOD and ground aerosol coefficient, this article undoubtedly makes an important contribution to the wide application of MODIS AOD products. This paper is well written and I recommend its publication in Remote Sensing. 
Minor comments: 
Line 23: no need for the abbreviation, AERONET since it is used once in Abstract; 
Line 526-527: “to improve the efficiency in the optimization of hyperparameters,”→“to improve the optimization of hyperparameters,” (duplicated efficiency)?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Satellite-derived aerosol optical depth (AOD) products tend to have large proportions of missing values due to clouds and high surface reflectance. In this study, the authors have generated the grid maps of high-resolution (1 km mesh), daily complete AOD, and surface aerosol extinction coefficient for mainland China from 2015 to 2018. Based on the AOD retrieved using the MAIAC (Multi-Angle Implementation of Atmospheric Correction) algorithm, a climate zoning factor is added to account for the variability in meteorology. It is shown that the use of an improved full residual deep network with attention layers is effective for imputing missing AOD in both temporal and spatial aspects. The resulting high-resolution grid maps of complete AOD and ground aerosol coefficients show an R2 of 0.78 compared with the AOD data from AERONET. Although this work has introduced some novel ideas for data imputation, the current description is insufficient for potential readers from the remote sensing field who wish to apply the proposed methodology in their projects. Further considerations of the following points would lead to improvements in the presentation.

(major)

L16-17 (The line numbers are quoted from the abstract, but the improvement should be addressed for the entire manuscript, not limited to abstract. The same applies to the issues pointed out below.)

“a climate zoning factor” – Does this imply that different regions of the nation were treated more or less separately? What was the exact relationship between the “18 tiles of MAIAC AOD” (Sec. S2) and the climate zones? Also, please note that the zoning shown in Table 5 should also be illustrated in a figure. For this purpose, Fig. S1 should be revised, which is depicting only three zones (middle temperate and frigid, warm temperate and sub-temperate, and tropical and sub-tropical).

L21 “an average R^2 of 0.90” – From this part, it is not clear exactly what parameter was validated with what validation data. Thus, the reader cannot understand why “the high test performance shows the reliability of AOD imputation” (L22). (The following issue at L23 seems to be related to this question, if this validation concerns the time series of AOD).

L23 “AERONET monitoring stations” – Please cite here how many stations were available for this study. A related question: It seems that the limited number of AERONET stations (just six sites from Fig. S1?) and their spatial sparseness may lead to the question of minimal reliability for discussing the result over the whole nation. (For example, only limited regions of climate zones were effectively validated, etc.) This issue is related also to the MAIAC validation itself.

L81-82 “Derived from AOD, the aerosol extinction coefficient on the ground ...” – There are a variety of instruments that can measure the aerosol extinction coefficient (or equivalent quantities) near the surface level (e.g., an integrating nephelometer, a visibility-meter, etc.). Therefore, the use of satellite AOD to derive the aerosol extinction coefficient near the surface level is an indirect method, with inevitably ill-posed features.

L115 Please explain more about the role of “attention layers”, counted here as one of the novel aspects of the present paper. The same applies to the “zoning indicator” (maybe in Sec.2.1, regarding Fig. S1).

L116-117 “important predictive covariates” – Please explain explicitly the parameters that are to be considered as “important predictive covariates” in the present paper. The same applies to “hyperparameters” (L122). The additional, clearer explanations related to L115-117 will be indispensable for the reader to understand the fundamental framework of the “adaptive method based on the improved full residual deep network” (L18 in the abstract).

L201-202 In relation to L116, please clarify if the parameters listed here contain all the “important covariates.” Besides, how are the MAIAC AOD data (non-missing part) treated in the adaptive imputation modeling? This essential aspect is not clear from the current description.

L361 “Figure 4 shows the time series of test R2 and RMSE for the 1461 models” – Please explain exactly with what data set the MAIAC AOD after the imputation is compared in Fig. 4. This question is related to the question stated just above (L201-202).

L367, 386: In Fig. 4, there are short-term occasions in which R^2 deteriorates considerably (shown as sharp dips). What are thought to be the major reasons for these degradations? Are they associated with some noticeable aerosol events in a particular climate zone? Please add some brief explanations.

L406 “The optimal g in Eq. (3) is 0.21” – the characteristics of aerosol growth critically depend on the aerosol components. Therefore, it is not adequate to assume a single parameter for such a vast region with different climatology.

 

(minor)

L28 “such as PM_2.5 and PM_10” – this description gives an impression that both of these parameters are considered, but obviously, it is not the case.

L32-33 “wood and biofuels” – What are aerosols from “wood”? (pollens? organic particles?)

L36 “a diameter of ≤10 \mu_mm (PM10) or with a diameter of ≤2.5 \mu_m (PM2.5)” – somewhat misleading since the detection-based definition is 50% capturing probability at the claimed diameters due to the aerodynamic effects.

L43 “such as PM10, PM2.5” – “such as sampling stations of PM10 and PM2.5”

L66 “Fan, Xia and Chen [12]” – “Fan et al. [12]” (family names are cited only up to two authors)

L103 It would be better to rephrase the sentence so that “high variability” appears only once.

L106-107 “from 2014 to 2018” – in the abstract, “from 2015 to 2018”. Please be consistent throughout the paper.

L110-111 “the unavailability of publicly available high-resolution parameters” – “the lack of publicly available high-resolution parameters”.

L129 “PBLH” should be spelled out here (not L241).

L144 “and industrial sources etc.” – “and industrial sources, etc.”

L158 “surface bidirectional reflectance so it works on both ...” – “surface bidirectional reflectance. Thus, it works over both ...”

L178 “is global reanalysis data product” – “is a global reanalysis data product”.

L265, 275, 304: Please add a comma (or a period) after each equation since equations should be part of a sentence.

L305, L479, etc. Probably not “GAE” but “GAC”.

L388, etc. Please indicate the panel indexes such as (a) and (b) to all the panels in Figs. 5 -8.

L396 “cloud sky”– “cloudy sky” (twice)

L404 Fig. 6(a) is a scatter plot, not a boxplot.

L414 “a better predictive covariate for estimating PM10 or PM2.5” – It is not clear how PM10 and PM2.5 are discriminated.

L422 Figs 8-12 and Table 5: The unit of GAC (probably m^-1) should be shown explicitly in the graph axis, panel legend, or Table caption.

L425 “Figures 11–14” – “Figures 9-12”.

L430-431 Why “the land surface reflectance of true color shows that cloud cover was a dominant factor for the extensive missingness of MAIAC AOD”? Probably it would be more direct to mention that highly reflective surface coverage is rather limited.

L459 In relation to Fig. 7 (and also Fig. 13), it would be meaningful to refer to the influence of both desert dust and anthropogenic aerosols.

L468 In relation to Fig. 13, the explanation of the four cities should appear here, not later (L498).

L531 What is “a natural and smoothing transmission”? Please paraphrase this part.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Review Remote Sensing

Mapping of the High-Resolution Aerosol Optical Depth and the Ground Aerosol Coefficients for Mainland China

Summary: This is a good study with a fairly large scope, where they estimated mapped datasets of high-resolution (1x1 km2) daily Aerosol Optical Depth AOD and ground aerosol extinction coefficients of mainland China from 2014 to 2018. They also “combined ground monitoring data with meteorological reanalysis data, topography, coordinates and time index to generate reliable high spatiotemporal resolution surfaces of meteorological factors. Overall, the study seems to be a good contribution to the literature in this area.

Overall, I think this study is solid scientifically and a good contribution to improving satellite estimates of aerosols concentrations.

I have a few issues that I would like to see addressed before publication:

• There are large variations in meteorological variables in both space and time, and estimates for PBLH are notoriously bad during inversion periods when aerosol are high in the PBLH. Some discussion of the potential sources of error contributed by uncertainties in the meteorological analysis needs to be discussed, more that what is already brought up in Section 3.2.
• The gaps in the AOD such as shown in Figure 9 can be so large spatially. More discussion in the least of the difficulty in imputation when you have a large portion of china cloudy versus smaller gaps would be good.
• The satellite AOD peaks in the Spring, but the observed PM2.5 peaks in the winter. Some discussion of this is needed in addition to the discussion around line 409. I am concerned that the shallow surface polluted layers are not being captured by the satellite, but as the PBLH increases in depth in the early spring, while pollution remains high, is when the AOD becomes the highest. Please address further the limitations of the satellite approaches with respect to seasonal variations in the PBLH the paper. The linkage between the ground aerosol and the satellite aerosol and the PBLH is an important and difficult topic, and as much clarity and discussion of the potential issues would be good.

Minor Comments:

In the title, the “the” is not needed, and can be written more succinctly. I would change title to: “High-Resolution Mapping of Aerosol Optical Depth and Ground Aerosol Coefficients for Mainland China”

Line 13: Replace “beyond” with “greater than”

Line 14: Another “the” that is not needed, should be :”We generated grid maps”

Line 45: Please rewrite/expand this sentence into a few sentences. I know the author is trying to say that 1500 sensors is not enough to capture the spatial variability of sensors, but more info is needed. I assume most sensors are in cities and there are large gap areas, etc.

Line 75: I think it should be “ARE actually non-random”

Line 90: I would make it a bit clearer that imputation is “assignment by inference” for the spatio-temporal gaps in the satellite data, If I understand correctly, as some readers may be confused by the imputation term.

Line 106: Change “provided the map” to “ We generated spatial maps of high-resolution”

Line 129: I don’t think PBLH is defined anywhere and needs to be

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The author has made a good job to improve the manuscript substantially, the current form of which seems mostly worth publishing. Please consider the following additional points for preparing the final version. No further review is mandatory.

L14 “high-resolution daily complete AOD and ground aerosol coefficients” – “high-resolution, daily complete AOD and ground aerosol coefficients”

L19 “to reliably impute extensive missing AOD.” –“to impute extensively missing AOD in the whole study area consistently and reliably.”

L22 “The high test performance” – “This high test performance”

L36 “from the sun to the Earth” – “from the Sun to the Earth”

L86 “PM2.5 is particulate matter with a diameter of less than 2.5 \mu m on the ground.” – “PM2.5 is particulate matter near the ground, with diameters not exceeding more or less 2.5 \mu m.”

L88-90 “the aerosol extinction coefficient on the ground can be used as an indirect proxy variable to represent the distribution of ground aerosols such as PM2.5 compared to AOD [28,29].” – “the aerosol extinction coefficient on the ground can be used as an indirect yet better proxy variable to represent the distribution of ground aerosols such as PM2.5 compared to AOD [28,29].”

L94-95 “(simplified to ground aerosol coefficient, abbreviated as GAC), which is a better proxy variable of ground aerosol compared with AOD [29].” – “(simplified to ground aerosol coefficient, abbreviated as GAC).” (the latter part is just the repetition of the statement at L88-90.)

L112-113 “high-resolution(1x1 km2) daily complete AOD” – “high-resolution(1x1 km2), daily complete AOD”

L165 Please insert the following sentence around L165 or at an appropriate part nearby: “For imputation modeling, however, we trained a national model for each day, not regional models separately. “ It is better to describe this information not later (L235) but here.

L235 “For imputation modeling, we trained a daily-level national model rather than regional models for each day. The one-hot coding method [53] was ... “ – “For training a daily-level national model, the one-hot coding method [53] was ... “

L317 “Among the daily samples with AOD available, ... “ – “Among the daily samples with AOD available from the MAIAC data without cloud coverage, ...”

L322 “The spectrally interpolated AOD from AERONET was ... “ – “Furthermore, the spectrally interpolated AOD from AERONET was ...”

L336 The comma should appear just after the equation, not the equation number, as

“             [equation],                                                                        (3)

where AOD is ...”. The same applies to all the equations.

L473 “and aerosols, may lead to ... “ – “and aerosols may lead to ...”

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The author has carefully and thoughtfully addressed all my concerns and the paper is now ready for publication in my opinion. Please read through one last time for any awkward sentences. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

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

See attached. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

 

This study proposes to map high resolution (1 km) aerosol optical depth over China using deep residual network by imputing (filling missed values) of the MODIS MAIAC AOD.

 

The authors did a comprehensive analysis of the results, which is excellent. However, there are many unclear statements in the methodology blocking me from giving suggestions to editors. I will need revised version with a clearer method description to better tell the quality of the approach.

The author should write not only ‘what is done’ but more importantly ‘why it is done and why the method is used’.

Climate zoning is not clearly how it is used? Each climate zone training a unique deep learning model?

 

The most unclear part is section 3.3.2 which is confused. The author has derived a high resolution (1 km) and fully covered of AOD from section 3.3.2 using downscaling. Why the later step is even necessary?

 

Furthermore, there is more need to be clarified in Section 3.3.2, what downscaling method is used? How the Eq. (1) is combined with the downscaling method?

 

What is the difference between data fusion (3.3.3) and downscaling in (3.3.2) since they are both downscaling?

 

Other issues

Line 113 import -> important

The reanalysis data can NOT be called ‘satellite’ variables in the title of Section 3.2.1

Line 183 ADO

Figure 9 (a) is AOD or ground PM2.5, this is not consistent with the texts which said AOD

Line 597 has -> have 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The revised version of the paper is improved compared to the original version.

I still suggest to move some of the details of the methodology to the supplemental section. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

 

This method is still very confusing. I have to reject (and resubmission) this manuscript as the author has been given a chance to revise but failed. There are many data manipulations lack of justification and clarification. The most unclear part to me is the so called ‘downscaling’ and ‘fused’ input covariates. Data manipulation is not good to convince the readers and reviewers the method is correct, especially when they are not well described. I am confident that the author is technically very strong. However, this is a science journal paper rather than a technique report and needs more justification and clarity so that readers can follow and trust what the author did.

 

Please see detailed comments.

 

1. “The meteorology parameters such as air temperature, relative humidity, wind speed, 198 air pressure and PBLH play important roles in the forming, mixing, dispersion and trans-199 portation of aerosols [48] and their transformation to ground particulate matter [49].”

What are the original resolution of these variables?

 

2. “we used the residual deep network to generate the daily 1×1 km2 grids of me-202 teorological variables, which fused ground meteorological measurements, MERRA2 rea-203 nalysis data, coordinates and elevation [50,51]. We collected the daily meteorological 204 monitoring data from approximately 900 stations of the China Meteorological Data Ser-205 vice (http://data.cma.cn). We generated the high-resolution grids of the following meteor-206 ological variables, including daily air pressure (hPa), air temperature (◦C), relative humid-207 ity (%) and wind speed (m/s).”

 

Very unclear. Why not simply using cubic or bilinear interpolation to 1 km ?  residual deep network trained how? Using what as input (predictor)? And what as output 1 km variables for training? My understanding is that you have to collect 1 km data for training as Y. How to you collect these 1 km air temperature, relative humidity, wind speed, 198 air pressure and PBLH data ??

 

What is the advantage of this unclear and complicated residual deep network compared to simply cubic or bilinear interpolation? I bet if the authors did a simple bilinear interpolation, the results have little impact.

 

3. “By downscaling, we could avoid discontinuities or abrupt changes on 243 the fine spatial scale when fusing coarse-resolution reanalysis data with other variables”

You can also do this by simply bilinear or cubic interpolation.

 

 

4. “The iterative full residual deep algorithm was used to reduce the inconsistency and 247 induce natural spatial variation, thereby improving downscaling [51], so it was used to 248 downscale the coarse resolution AOD and PBLH data to the target resolution (1x1 km2). 249 The covariates of the coordinates and their derivatives (squares of x and y, and product of 250 x and y), and elevation were used to capture spatial variability of the variable to be 251 downscaled.”

What is exactly the input for training? Again what is Y for training? I assume you used the 1 km MAIAC for training? If so what is the difference with the following imputation?

For PBLH, how to collection 1km PBLH data for training?

 

5. OK, in supplementary #2, I saw that 50 km PBLH is used as training for downscaling. This confused me more. How can 50 km at training can achieve 1 k downscaling goal? This makes me even further curious why not simply using bilinear or cubic interpolation. To write a scientific paper, it is not like the more complicated, the easies it can be accepted. It is usually the other way around.

 

 

I would like to see the image visual difference between downscaling and simple bilinear interpolation and if possible use MAIAC AOD at 1km to evaluate the downscaled results quantitatively.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx