Exploring VIIRS Night Light Long-Term Time Series with CNN/SI for Urban Change Detection and Aerosol Monitoring

Round 1

Reviewer 1 Report

This manuscript demonstrated a new methodology of analyzing the time-series imagery of night-light radiance to assess aerosol in urban areas. 

The paper is well-organized and the well-executed. There are however some issues in presentation to improve and one aspect of the conclusion to clarify. 

The concerns to be attended to are listed in the following according to roughly the order of appearing in the text:

1. abstract.  line 12, "dataset overcome(s)" ------please check the grammar.

2. Figure 1 doesn't do a great job of explaining the process. I would suggest that you place a subject atop each block, like "data selection" for step 1, "feature factor" for step 2, similarity index" for step 3, "seed-matching" for step 4, "radiance of ROI". And it may be a good idea to use different shapes for the many blocks.

3. Figure 5, the caption shall list "9 years" as the duration of the time-series. Ans a simple harmonic fit to the oscillation would be imperative to show the level of variation and the exact period of the oscillation to suggest the pattern of yearly change.

4. The night-light images of LA, Beijing and Dehli shall be accompanied with the profiles of the respective city to match the contour of urban activities. 

5. Fig. 6 can use much bigger fonts for the x and y labels as well as the legends. Actually, the font sizes used for figures have been generally small and are better to be enlarged.

6. Figure 6 suggests lower aerosol optical density in winter months than in summer months for Beijing. I believe that  Beijing's winter months generally have lower air quality than the summer months, due primarily to the heating need that may rely more on coal products. A lower air quality would mean a higher aerosol optical density. Can you rectify your conclusion of lower AOD in the winter months of Beijing?

7. Fig7. please mark LA, Delhi,. and Beijing to the right of the respective traces.

With the time-series data of the three cities compared, what can you suggest about the AODs of the three cities?

8. The English can be improved.    

Author Response

Dear reviewer#1, please see our responses to your comments and suggestions in the attached file.  Thank you!

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript by Cao et al. (MS#: remotesensing-1761889) developed an approach of using VIIRS night light data for urban change detection and aerosol monitoring. The method proposed here is mainly based on a convolutional neural network similarity index (CNN/SI) between different images. Overall, the work is good and the manuscript was prepared well. However, I think the authors need to further clarify some important issues about the methodology, which may affect the reliability of the approach and induce some uncertainties in the time series generated here. I provide some questions and suggestions below, which may help the authors to improve the paper. I would suggest a major revision of the paper at this point.

Main points:

1. Methodology. As the authors indicated, the proposed approach needs to train the model for each city when it is applied. In this case, it is likely the cloud-mask approach is more effective than the CNN/SI approach especially when it is used for large-scale studies. Please add more discussions about this question.

2. L120-130. Did the authors removed all the cloud-contaminated images? Will this largely reduce the data availability, especially for some cities which tend to have many clouds in some seasons? Compared the proposed approach, the cloud-mask approach only removes the cloud-contaminated pixels, which means the rest pixels can still be used for analyses.

3. L121-122. If a city is expanding rapidly during a period, the SI of the city night light images between two-time intervals might be low even the two images were retrieved under clear-sky conditions. How did the method consider this kind of situation?

4. L136-139. Did you removed all the images that having a minimum radiance >0.6nW/(cm2-sr)? Does this again largely reduce the data availability?

5.  L221-227. How did you use the SI values to filter out the cloud and lunar illumination-contaminated images? Did you select a threshold of SI? If yes, how did you determine the threshold and why is it reasonable?

Minor points:

1. The legends and text of some figures are two small and not clear. Please enlarge the font sizes. Please also improve the image quality of the figures.

2. L401. Did you meant “increase in L”? Please check.

Author Response

Dear reviewer#2, please see our responses to your comments and suggestions in the attached file.  Thank you!

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

congratulations on the good work. The paper in general is well organized and reasoned. It is well written in the introductory phase, but then the analysis phase must be implemented. I would implement only some parts and figures, and I would present a summary table of the results.

In particular:

Figure 3: a scale must be added to evaluate the spatial resolution 750x750m with the coordinates, in addition a legend must be added to evaluate the radiance.

Figure 4: (see Figure 3) also on 05/07/2017 the moon was at 86% (the effect of the moon is evident in the image) while the authors speak only of images in the new moon.

Check the algorithm well in order not to invalidate the results and replace the image of Delhi.

Section 3.1

It would be useful to add an image classified as clear sky and then discarded by the algorithm in order to understand when the authors are talking about images 7 and 375 (line 307).

Section 3.2

This is the section that in my opinion needs major revisions

Something is wrong with the results in Figure 5 and 6. Los Angeles in Figure 5 has values of about 20 nW / cm2sr and in Figure 6 about 12 nW / cm2sr

The spatial resolution used and what kind of images are used must be added in the caption: daily, monthly ...

I recommend checking and explaining

I would add the linear regression before and after the algorithm (Figure 5).

There are several typos of this type (Line 349-350) e.g. inves-tigated

Figure 6 It is taken directly from the AERONET site, but is not related to the text. First, Level 1 does not use thresholds for cloud cover: as you can see, in fact, we have AOD peaks greater than 2. I recommend using Level 1.5: they are fewer images, but cleaner.

It must be specified that they are daily data (plus data for each day), while the VIIRS analysis is nocturnal (1 image per night). The data for the three sites must be downloaded and a correlation sought, perhaps graphing together the two time series (VIIRS-AOD).

Why do the authors show AOD at multiple wavelengths? Furthermore, wavelengths outside the range of VIIRS (Figure 8).

The diffusion of light by aerosols is a very complex phenomenon and should be better argued. I recommend reading and mentioning some publications of:

Kocifaj, M., 2007, Applied Optics, Vol. 46, 3013-3022

Kocifaj, M., 2008, Applied Optics, Vol. 47, 792-798

Kocifaj, M., 2009, Applied Optics, Vol. 48, 4650-4650

Kocifaj, M., Ladislav Komar, L., 2016, MNRAS, 458, 438

Kocifaj, M., 2018, Journal of Quantitative Spectroscopy and Ra-

diative Transfer, Vol. 206, 260-272

This is to make the function more credible and better argued ( Line 394 ).

So written it doesn't have much meaning: it has to be reworked or eliminated.

The radiance is called L then R (Line 401).

The correlation between VIIRS and AOD has already been analyzed in publications and should be mentioned:

Satellite measurements of artificial light at night: aerosol effects Cavazzani et al.

Relationship between VIIRS, AOD and lunar cycles.

Figure 8 Atmospheric Spectral Transmittance it must be related to aerosols, all absorption lines must be indicated: not only O2 and H2O.

Summary

In this section or in the previous one, a table should be inserted which indicates the precise number of images analyzed for each site, what kind of averages were carried out for figures 5 and 7.

The images must be discarded by the algorithm and if possible, the time of the year (e.g. in which season the algorithm discards more images). This would make all the work better.

I still congratulate the authors for the nice idea, but I recommend deepening the data analysis part (the errors highlighted could indicate excessive haste in submitting), or show it better. Especially in relation to aerosols.

 I also recommend visiting the site:

https://lighttrends.lightpollutionmap.info/

I enclose the analysis done for the Los Angeles area (80x80km), the results are similar to those of the authors.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer #3,  please find our responses to your comments and suggestions in the attached file.  Thank you!

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors have revised the paper according to the reviewers' comments. Most of my previous concerns have been addressed. I think the paper has been largely improved and is publishable at this point. 

Reviewer 3 Report

Dear Authors,

the paper seems to me to have improved sufficiently. Make a careful final revision of English.

Congratulations