High-Resolution Urban Land Mapping in China from Sentinel 1A/2 Imagery Based on Google Earth Engine

Round 1

Reviewer 1 Report

This paper develops a new methodology that combines optical and SAR data in order to identify urban lands at high resolution. Authors use temporally aggregated Sentinel-1 data to identify potential urban land, and then Sentinel-2 data is used improve the identification by discarding highly confused land covers in respect to urban cover. The methodology is novel, the paper is well written, and I believe this could be considered for publication in Remote Sensing if certain issues can be addressed. However, I have some concerns with some of the methods, and some sections need to be clarified.

General comments:
Although the authors did a great job creating a whole new methodology for accurate identification of urban cover, it is difficult to infer whether the higher accuracies (than previous maps) are due to the novel methodology, or to the higher spatial resolution. Classic approaches use training data to train a classificatory algorithm. The authors collected “ground truth” data from Google Earth, which could have been used to train a classificatory algorithm using combined S1 and S2 data. Did the authors consider this approach? Those approaches don’t need to set thresholds, and can be really fast with Google Earth Engine. It is difficult to understand what is causing the improvement in accuracy from previous methods: combination of sensors? Methodology? Or just the higher resolution? If authors don’t consider to apply a simple classification, they should at least mention why their methodology might improve the accuracy of a classification with a combination of S1/S2 data.

A general restructuring is needed. Some discussion appears on results, some introduction appears in discussion, and so on. Authors might consider mixing results and discussion, section by section (if editor allows). See specific comments on this.

Specific comments:
L47-54: This should be discussed in the methods. This does not help the reader understand the aim of the study.
L55-71: I don’t understand the point of this paragraph. I think is very important to know why the authors didn’t use a classification approach. The first line is just listing classificatory algorithms, not methodologies to extract urban cover. There are too many topics in this paragraph. Algorithms, methodologies, extent of the study, representative vs operational. This paragraph is key for the reader to understand why you don’t use a classical supervised classification approach. Please clarify each topic addressed in this paragraph.
L163: Why did you use VV polarization only?
L179: Why theses thresholds? Show results or references to justify these values.
L184: Why NDVImean will not work but NDVI will? Please clarify this whole paragraph. Also explain that NDVI is extracted using S2, and how, etc.
L180-: Sentinel2 data for 2015 has co-registration problems (Yan et al, 2018, Rem Sens Environ 215:495-506). Please discuss whether this issue is affecting your results and why.
L199: Again, please justify the use of threshold values.
L207-2014. Determination of thresholds is not clear. Are the backscatter coefficient averaged across a whole ecosystem region? Only for sampled cities? Is the threshold selected based on averaged values across the whole region?
L244: Why is the filter applied after obtaining the TUL? Wouldn’t a filter improve the identification of the urban thresholds from the beginning?
L263: Points don’t have area, so I guess the 10m resolution refers to the pixels of your created map, when overlapped with the random points. Please correct.
L263-280: Any sort of validation that involves datasets of different accuracies will face the challenges that you encounter in here. Please discuss the resolution effects on validation and comparison between urban cover datasets.
 L281:292: I think this paragraph should be placed before the previous one.
L297; What are “statistical files”?
L298: These provinces/regions have not been mention before, so I don’t understand why this is explained here. Please provide background or delete.
L314: How are these areas generated from the ground truth?
L354: Discuss why this low accuracy on discussion maybe?
L370-408: You are discussing the results in here. Please move this discussion to discussion section, or create a combined Results/Discussion section (ask Editors beforehand if you opt to do that?).
L371: Why do you say that globeLand30 is overestimating? Do you have the results to show that? How can you be sure that it is not your dataset which is being underestimating in this case? Also, couldn’t it be a resolution issue? Maybe most part of the pixels in this dataset are urban area, so they have been classified as urban. But in other areas with lower density, these dataset would underestimate urban cover. Please don’t infer conclusions from data not shown in your paper.
The same applies to the rest of this itemized paragraph.
Figure 13: Please add row numbers in order to follow results.
L440: I agree that computation of the segmentation can be faster than classification algorithms. But obtaining the thresholds could be much more time consuming. Specially using GEE, classification can be very fast and done for huge areas. I think your method might be really accurate, but I am not sure that time efficiency is the main advantage. Please clarify this.
L444-445: Authors are drawing conclusions from non-shown results. Please show results supporting this claim, or reference past work.
L455-459: Again, authors draw conclusions from non-shown results. Please explain why combining optical and SAR data is achieving this, and how your results support this.
L464-469: The greenhouse problem has not been introduced, and not shown in results. It is difficult to understand why this is discussed here. No references either.
L494: If the methodology is not “rigorous”, perhaps should not have been used? I would rephrase this sentence, discussing about “the limitations” of the method (e.g., subjectivity).
L503: How accurate is “acceptable”? Please provide exact number.

Author Response

Responses to reviewers for “High-resolution urban land mapping in China from Sentinel 1A/2 imagery based on Google Earth Engine”, Remote Sensing 458159

 

Dear editor and reviewers,

 

We are grateful for your constructive comments and suggestions. Following these comments, the original manuscript has been carefully revised. In this letter we responded each comment by describing the major changes in the revised manuscript. Changes are highlighted in yellow. Grammatical and spelling corrections are not highlighted to ensure readability.

Responses to Reviewer #1:

Comments and Suggestions for Authors

This paper develops a new methodology that combines optical and SAR data in order to identify urban lands at high resolution. Authors use temporally aggregated Sentinel-1 data to identify potential urban land, and then Sentinel-2 data is used improve the identification by discarding highly confused land covers in respect to urban cover. The methodology is novel, the paper is well written, and I believe this could be considered for publication in Remote Sensing if certain issues can be addressed. However, I have some concerns with some of the methods, and some sections need to be clarified.

Response: Thank you very much for your comments and encouragements.

 

General comments:

(1)Although the authors did a great job creating a whole new methodology for accurate identification of urban cover, it is difficult to infer whether the higher accuracies (than previous maps) are due to the novel methodology, or to the higher spatial resolution. Classic approaches use training data to train a classificatory algorithm. The authors collected “ground truth” data from Google Earth, which could have been used to train a classificatory algorithm using combined S1 and S2 data. Did the authors consider this approach? Those approaches don’t need to set thresholds, and can be really fast with Google Earth Engine. It is difficult to understand what is causing the improvement in accuracy from previous methods: combination of sensors? Methodology? Or just the higher resolution? If authors don’t consider to apply a simple classification, they should at least mention why their methodology might improve the accuracy of a classification with a combination of S1/S2 data.

Response: Thank you for your comments and questions. In our study, we also considered the classic pixel-based classification method (CART, SVM or Random Forest). Although traditional classification methods have been widely applied, there are challenges and limitations for applying at regional and global scales, e.g., subjective scene-to-scene data analysis, time consuming, and complicated computing. Especially for collecting the training and testing samples, based on VIS (Vegetation, impervious surface, and soil), the training and testing sample corresponding to the four land cover types (high albedo, low albedo, vegetation and bare soil) must be manually generated in order to map the urban land product in China. Therefore, this is a huge work for collecting training and testing samples. In addition, in order to obtain product that meet the accuracy requirement, it is also necessary to perform classification according to different regions in China. Furthermore, training samples also need constant iteration and modification. Therefore, comparing with classic classification method, index-based threshold segmentation approach combined with different data sources is finally selected and used to conduct our urban land mapping in China.

At present, most of urban land products at reginal or global scale (e.g., GlobeLand30, GHSL) are derived from medium-/low-resolution optical imagery, resulting in low accuracy of urban extraction, especially at low latitudes because of imagery quality. In our study, a fast and effective urban land extraction method using multi-temporal and ascend/descend orbit SAR data and optical imagery with 10-m resolution is used to delineate urban land distribution at national scale. Comparing with shared urban products, the accuracy of our product is higher. In revised manuscript, we introduce the challenges and limitations of urban land mapping at regional or global scale using single data source in Introduction section; and we also introduce the effectiveness of our method by combining use of optical and SAR data in Discussion section.

 

 (2) A general restructuring is needed. Some discussion appears on results, some introduction appears in discussion, and so on. Authors might consider mixing results and discussion, section by section (if editor allows).

Response: Thank you very much for your comments. According to your suggestion, we rewrite our Introduction, Results and Discussion sections.

 

See specific comments on this.

(1) L47-54: This should be discussed in the methods. This does not help the reader understand the aim of the study.

Response: Here, we just discuss our classification standard and the definition of ‘urban land’. In the second paragraph in Introduction section, we summary the various methods used to identify the urban land.

(2) L55-71: I don’t understand the point of this paragraph. I think is very important to know why the authors didn’t use a classification approach. The first line is just listing classificatory algorithms, not methodologies to extract urban cover. There are too many topics in this paragraph. Algorithms, methodologies, extent of the study, representative vs operational. This paragraph is key for the reader to understand why you don’t use a classical supervised classification approach. Please clarify each topic addressed in this paragraph.

Response: Thank you very much for your suggestion. I also agree with your opinion that there are too many topics in this paragraph. According to your advice, we rewrite this paragraph and also introduce our reasons why we don’t use a classical supervised classification approach. Thanks again for your comments.

 

(3) L163: Why did you use VV polarization only?

Response: Thank you for your question. For Sentinel-1A, the SAR data with VV polarization can cover the whole China, while SAR data with HH polarization can only cover Antarctica and Greenland. Therefore, we used data with VV polarization in IW mode operated in 'ascending’ or 'descending' orbit in our study.

 

(4) L179: Why theses thresholds? Show results or references to justify these values.

Response: Thank you for your comments. These thresholds were empirical thresholds which have obtained by repeated experiment.

 

(5) L184: Why NDVImean will not work but NDVI will? Please clarify this whole paragraph. Also explain that NDVI is extracted using S2, and how, etc.

Response: Thank you for your comments. NDVI_max has demonstrated its effectiveness in distinguish impervious surfaces from other land covers in low-density IS region and we have provided the reference here. All the current mainstream urban land products have a similar low-accuracy problem in low-density urban regions, such as in peri-urban areas. NDVI aims at quantifying vegetation intensity. NDVI_max, representing the maximum composite of NDVI over time, has been used to improve the accuracy of urban extraction in low-density impervious surface regions. Therefore, we just consider use the NDVI_max here, not NDVI_mean and NDVI_min. To clarify thus paragraph, we removed the last sentence. In addition, according to your suggestion, we give the formula that how to calculate NDVI and NDVI_max.

 

(6) L180-: Sentinel2 data for 2015 has co-registration problems (Yan et al, 2018, Rem Sens Environ 215:495-506). Please discuss whether this issue is affecting your results and why.

Response: Thank you for your comments. In the GEE cloud platform, we have visually checked the original data of S2 and its producing results, there is no visually obvious misregistration.

(7) L199: Again, please justify the use of threshold values.

Response: Thank you for your comments. Ban et al. [25] suggested that the threshold value of slope (Tslope) was set to 15.

 

Ban, Y.; Jacob, A.; Gamba, P. Spaceborne SAR data for global urban mapping at 30 m resolution using a robust urban extractor. ISPRS Journal of Photogrammetry and Remote Sensing 2015, 103, 28-37. [Doi: 10.1016/j.isprsjprs.2014.08.004]

 

(8) L207-2014. Determination of thresholds is not clear. Are the backscatter coefficient averaged across a whole ecosystem region? Only for sampled cities? Is the threshold selected based on averaged values across the whole region?

Response: Thank you for your comments. We selected many cities for each ecological zone and then obtain different thresholds for these cities. We applied the median value of these thresholds in its master ecological zone.

 

(9) L244: Why is the filter applied after obtaining the TUL? Wouldn’t a filter improve the identification of the urban thresholds from the beginning?

Response: Thank you for your comments. Here we applied filter on extracted results not on original Sentinel data. This processing was used to remove ‘salt-and-pepper noise’ which is a common issue in pixel-based classification. And this processing was also a standard processing flow in previous paper.

 

(10) L263: Points don’t have area, so I guess the 10m resolution refers to the pixels of your created map, when overlapped with the random points. Please correct.

Response: Thank you for your comments, we have revised it. Each random validation point with the size of 10m×10m is coarser than that of the high-resolution Google Earth images.

 

(11) L263-280: Any sort of validation that involves datasets of different accuracies will face the challenges that you encounter in here. Please discuss the resolution effects on validation and comparison between urban cover datasets.

Response: Thank you for your comments. Our extracted is 10-m and others’ are 30 m. During the comparison, we keep the Google Earth image as our referenced map which is higher than 10 m. Higher-resolution image could provide much more spatial details than coarse imagery and this is the reason why our extracted products give much semantic details than others. However, higher resolution does not mean high extraction accuracy. Our higher accuracy is from our designed method not the datasets we use. I think it is fair to validate our product with 10m resolution and other products with 30m spatial size using higher-resolution Google Earth image.

 

(12) L281:292: I think this paragraph should be placed before the previous one.

Response: Thank you for your suggestion. We have revised it. This paragraph was moved to the previous one.

 

(13) L297; What are “statistical files”?

Response: Thank you for your comments. Here statistical files should be ‘statistical results’.

 

(14) L298: These provinces/regions have not been mention before, so I don’t understand why this is explained here. Please provide background or delete.

Response: According to your advice, this sentence is deleted and moved to the title of figure 12 in the revised manuscript.

 

(15) L314: How are these areas generated from the ground truth?

Response: Thank you for your comments. These blocks were randomly created from our extracted results and ground truth was from visual interpretation of Google Earth. Please refer to the following sentences.

 

Besides, for the rigorous comparison between the extracted products and the referenced ground truth data obtained from the visual interpretation of Google Earth images, we calculated the urban land area based on each randomly generated block of 900 by 900 meters (Figure 8). Nearly 20 blocks were randomly generated for each province and a total of 735 for all of China.

 

(16) L354: Discuss why this low accuracy on discussion maybe?

Response: Thank you for your comments. The lowest PA was found in Heilongjiang province, here the lowest accuracy is just relative, the accuracy still achieved around 90%. But for the lowest UA in Xinjiang, Xizang and Gansu provinces (just 60 -70%), we consider that many of the urban settlements in these areas are located in areas with very complex topography, such as loess high slopes, valleys covered with glacier, and the bare walls covered with bare rock, and the resolution of 10-m is not enough to identify them well, we will provide discussion in 4.2.

 

(17) L370-408: You are discussing the results in here. Please move this discussion to discussion section, or create a combined Results/Discussion section (ask Editors beforehand if you opt to do that?).

Response: Thank you for your suggestion. Some contents about discussion in this section are deleted and moved to Discussion section.

 

(18) L371: Why do you say that globeLand30 is overestimating? Do you have the results to show that? How can you be sure that it is not your dataset which is being underestimating in this case? Also, couldn’t it be a resolution issue? Maybe most part of the pixels in this dataset are urban area, so they have been classified as urban. But in other areas with lower density, these dataset would underestimate urban cover. Please don’t infer conclusions from data not shown in your paper. The same applies to the rest of this itemized paragraph.

Response: Thank you for your comments. For urban land extraction from medium-/low-resolution imagery, urban area commonly overestimated in areas with low-density urban features and underestimated in high-density urban areas. Here the overestimation or underestimation is just from visual checking which is a common comparison method for different products. And we also provide quantitative analysis in Figure 14 which support our conclusions. I think globeLand30 product is overestimated or underestimated because of the low resolution of Landsat image.

 

(19) Figure 13: Please add row numbers in order to follow results.

Response: Thank you for your comments. We have added the row numbers in figure 15.

 

(20) L440: I agree that computation of the segmentation can be faster than classification algorithms. But obtaining the thresholds could be much more time consuming. Specially using GEE, classification can be very fast and done for huge areas. I think your method might be really accurate, but I am not sure that time efficiency is the main advantage. Please clarify this.

Response: Thank you for your comments. The key point of our method is the determination of thresholds. We obtained these thresholds separately for each ecological zone. This process was not time consuming and was also easily performed than collecting training samples in classification. In addition, we also discuss that how to determine the thresholds is a key problem for our method.

 

(21) L444-445: Authors are drawing conclusions from non-shown results. Please show results supporting this claim, or reference past work.

Response: Thank you for your comments. We believe that shadows are difficult to completely eliminated, but problems such as foreshortening and overlap are well resolved. We will provide references and related example in revised manuscript. Please refer to the figure 12 in the revised manuscript.

 

(22) L455-459: Again, authors draw conclusions from non-shown results. Please explain why combining optical and SAR data is achieving this, and how your results support this.

Response: Thank you for your comments. Spectral confusion between urban land and other land-covers, especially bare soils, remains an unsolved problem in optical images. In SAR data, most of bare soil with flat surface exhibit low backscatter very dissimilar to urban region facilitate the discrimination between them. Hence, the combined use of optical (separate urban and bare soil from other land cover) and SAR data (separate urban land from bare soil with flat surface) partially solved this problem. We have explained this issue and also give an example about this problem in revised manuscript. Please refer to the figure 13 in the revised manuscript.

 

(23) L464-469: The greenhouse problem has not been introduced, and not shown in results. It is difficult to understand why this is discussed here. No references either.

Response: Thank you for your comments. We have introduced the greenhouse issue in Figure 7 (h) and then we have specially discussed it in discussion section.

 

(24) L494: If the methodology is not “rigorous”, perhaps should not have been used? I would rephrase this sentence, discussing about “the limitations” of the method (e.g., subjectivity).

Response: Thank for your comments, we have revised this sentence as:

We only chose a global empirical threshold for masking water and mountain regions, which is a limitation in this study.

 

(25) L503: How accurate is “acceptable”? Please provide exact number.

Response: Thank for your comments. We have revised it as:

Our approach produced a large-area urban land map for China with great than 88.00% overall accuracy.

 

End of the response letter.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Congratulations on your clearly presented, methodologically sound, thematically important, and engaging paper. I include all my suggestions, grammar and syntax corrections, and concerns in the attached .pdf file.

Comments for author File: Comments.pdf

Author Response

Responses to reviewers for “High-resolution urban land mapping in China from Sentinel 1A/2 imagery based on Google Earth Engine”, Remote Sensing 458159

 

Dear editor and reviewers,

 

We are grateful for your constructive comments and suggestions. Following these comments, the original manuscript has been carefully revised. In this letter we responded each comment by describing the major changes in the revised manuscript. Changes are highlighted in yellow. Grammatical and spelling corrections are not highlighted to ensure readability.

 

Responses to Reviewer #2:

(1) Title: I would propose to change your title into: High-resolution urban land mapping in China from Sentinel-1A/2 imagery on Google Earth Engine. It reads better and more logical in my opinion.

Response: Yes. We fully agree with your suggestion that the title of our manuscript is changed to “High-resolution urban land mapping in China from Sentinel-1A/2 imagery on Google Earth Engine”.

 

(2) I would paraphrase into: average overall, producer's and user's accuracies are 88.03%, 94.50% and 82.22%, respectively.

Response: Thanks for remarking this, we have changed it into ‘…average overall, producer's and user's accuracies are 88.03%, 94.50% and 82.22%, respectively.’

 

(3) perhaps is not really scientific (here); I would omit it or replace it with possibly

Response: Yes. The word ‘perhaps’ is deleted in the revised manuscript.

 

(4) did not instead of didn't; didn't is more informal, not really suitable in a publication context.

Response: Thank you very much for your comment, we have revised it.

 

(5) pixel and object-based; reads better in my opinion.

Response: Thanks for your suggestion. ‘pixel-based and objected-based’ has changed into ‘‘pixel and objected-based’.

 

(6) permits; something and something else requires still singular.

Response: Thank you very much for your correction. We have modified this gramma mistake.

 

(7) popularization (nice word) and popular in the same sentence -> repeatibility; I would propose fashionable, favored, attractive and the likes.

Response: Thank you very much for your suggestion. In this sentence, the word ‘popular’ is changed into ‘attractive’.

 

(8) are implemented or have been implemented; are been implemented is erroneous.

Response: Thank you very much for your correction. The sentence ‘Many corresponding global urban land products and their analysis are also gradually been implemented’ is changed into ‘… are also gradually implemented’.

 

(9) I would propose "fusing multi-temporal…" (without the fusion at the end).

Response: Thank you very much for your advices, we have changed it.

 

(10) Regarding Figure 2,

a. What does your inset map show? Other territorial Chinese grounds?

b. You should add a north arrow and a scale in your large map.

c. Small fluctuating mountain, mountainous area, and mountainous areas with great ups and downs are confusing in terms of their differentiation because of the color choice. I would propose to choose a blue or black color for the mountainous area or the ma with great ups and downs class.

Response:

a: We divided the Chinese map into different regions and calibrated the centroids and thresholds separately for each region. We applied the Chinese stratification scheme of urban ecoregions. This stratification scheme takes three elements into account including a biome designation characterizing general climate and vegetation, urban topology differences and the economic level defined by per capita gross domestic product (GDP).

b: Thank you very much for your advices, we have added a north arrow and a scale in the revised figure 2.

c: According to your suggestion, we have changed them to different colors.

 

(12) As long you are introducing abbreviations for both satellites - S1Q and S2Q - you could use them in the rest of the text to save space and ease reading :)

Response: Thank you very much for your advices, we have accepted your good idea. S1Q and S2Q are changed into S1 and S2.

 

(13) line 169: Won't the use of TOA reflectance values affect the performance of the employed indices and related thresholds due to the atmospheric effects/interferences in your composites? You should at least discuss this into your Discussion section.

Response: Thank you very much for your question. In our study, some indices, including MNDWI and NDVI, are used to extract urban land. During the calculating these indices, we do not conduct the atmospheric correction and terrain correction, which has a negative effect on computing these indices. However, it is difficult to conduct the atmospheric correction in the GEE cloud platform due to the difficulty of obtaining parameters. Most of researches directly used the TOA reflectance to extract urban information, please refer to the following reference. In addition, some contents about TOA reflectance without conducting atmospheric correction are discussed in the Discussion section.

Goldblatt, R.; You, W.; Hanson, G.; Khandelwal, A. Detecting the boundaries of urban areas in India: a dataset for pixel-based image classification in google earth engine. Remote Sensing 2016, 8, 8, 634. [Doi: 10.3390/rs8080634]

 

(14) Line 258: How have you assessed the density and evenness of the distribution of your validation points? Was it based on your area extents and/or expected user accuracies? Also, where did you generate them? Based on Figure 2 map?

Response: Thank you for your comments. Validation points are randomly generated with nearly 5000 for urban and 5000 for non-urban in each province in China using our own software developed by C++ language. These validation points are randomly selected based on our classification results instead of the Figure 2 map. The quantity and randomness basically satisfy the required density and evenness.

In revised manuscript, the sentence ‘Dense and evenly distributed validation points covering various landscapes were needed to assure the fairness and rationality of the validation results.’ is changed into ‘Dense and evenly distributed validation points covering urban areas and non-urban areas were needed to assure the fairness and rationality of the validation results.’

 

(15) It would be nice if you could show in one inset map the distribution of both validation points in a small region within your selected larger province.

Response: According to your suggestion, we have added a map of small region which shows the distribution of both validation points within our selected province (Yunnan Province). Please refer to the figure 6 in the revised manuscript.

 

(16) use semicolon after the first area and delete the parenthesis at the end of the sentence.

Response: Thank you very much for your correction. We have modified this mistake. According to the first reviewer’s suggestion, this sentence is deleted and moved to the title of figure 14 in the revised manuscript.

 

 (17) A figure should generally stand on its own; hence you should add an elaboration on what is PSA and ISA in this legend.

Response: Thank you very much for your comment. In the title of figure 11, we added an explanation concerning PSA and ISA.

 

(18) As regards Figure 13, as a general rule of thumb, it is optimal to display figures in the next page, maximum, following their mention in the text. However, I would understand if you keep as it is, as you are using the space between to introduce the reader to this nice figure.

Response: Thank you very much for your advice. we added a list of numbers in the left column which could help readers to understand.

 

(19) It would be more consistent with the other results to include the year (2015) of Liu et al. 2018 paper instead of the year of the paper :)

Response: Thank you very much for your correction. We have changed it into: Urban2015 (Liu et al., (2018)).

 

(20) still exists; your results also exist in the present tense.

Response: Thank you very much for your correction. We have modified this mistake.

 

 

 

 

 

 

 

End of the response letter.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors did a very good job improving the manuscript and I mostly agree with their replies to my comments and suggestions.

I believe the manuscript flow has greatly improved and that it reads better. Considering this, and that their novel methodology could be of great help for urban mapping from remote sensing, I think this paper should be accepted in its present form.