Remote Sensing Mapping of Build-Up Land with Noisy Label via Fault-Tolerant Learning

Round 1

Reviewer 1 Report

This manuscript employs a procedure based on fault-tolerant learning to produce build-up land cover maps developed with training data from lower resolution cartography. Authors used recent land cover products with 30 m and 10 m of spatial resolution (obtained with different procedures and datasets). They carried out an experiment that used training labels from the 30 m land cover map that were employed to classify Sentinel 2 images (10 m of spatial resolution). These results were validated against the global 10 m land cover map.

I consider that the purpose of the methodology is interesting because it may help to obtain higher resolution mapping products respect to the resolution of the base datasets employed for the development of the training dataset. The presentation of the manuscript is good.

I consider several problematic issues of the manuscript.

1. The conceptualization of the experiment may be improved. You are employing a training dataset from a 30 m spatial resolution global cartography obtained with optical images, you applied them to map built-up areas from 10 m spatial resolution satellite optical images, and you compare your results against a 10 m spatial resolution global cartography obtained with optical images and RADAR images. What is the accuracy of the global products specifically for your study area? Did you consider the possibility of also using RADAR images for your mapping procedure? I consider that you have to develop and independent validation dataset, because experimental conditions are uncertain.
2. Generalization of the experiment has been conducted within areas that, as authors suggests, are similar to Taoyuan county. Did you test your methodology in non-adjacent study areas with very different urbanization processes?
3. No discussion of the results has been provided and this is mandatory in any scientific manuscript.

I include some additional comments and suggestions.

Lines 130 to 135. The employment of the term “historical land cover products” is unclear. You are employing satellite images from September 2020 (further details of the images are needed) for mapping build-up areas and using land cover products (with different resolutions and produced with different methodologies) also developed with 2020 satellite images.

Line 149. Provide further details of the images employed for your mapping procedure. Did you apply atmospheric or topographic corrections?

Lines 153 to 160. You may also provide bibliographic references of both land cover products. The may also employed for the discussion of your results.

Figure 2 label. You may include in the label that the left map is from GLOBELAND30 and the right map is from WorldCover.

Line 167. You wrote: “Subsequently, 64*64 size patches were added to the training set with the selected pixels as the center”. Provide further details of this procedure (purpose, possible overlaying among patches…).

I missed the development of an additional validation dataset based on field knowledge and higher resolution images (e.g. 1m) in order to avoid potential inaccuracies of the data sources employed in this study. If you compare your results respect to a product with unknown specific accuracy for your study area (these global land cover maps are outstanding but globally validated), you are assuming some kind of uncertainty in your experiment.

Line 207. Remove this line.

Table 2. Include additional metrics (e.g. confusion matrix, user’s and producer’s accuracy, Cohen’s kappa coefficient) for a more comprehensive analysis of the results.

Table 4. Include additional metrics (e.g. confusion matrix, user’s and producer’s accuracy, Cohen’s kappa coefficient) for a more comprehensive analysis of the results. You may also include individual results for each county.

Line 417. No discussion section has been provided. This is mandatory. Analyze your methodology and results by comparing them with previous research, including proper references.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Reviewer’s Report on the manuscript entitled:

Remote Sensing Mapping of Build-up Land with Noisy Label via Fault-tolerant Learning

The authors proposed a noise-label learning method for fine-grained mapping of urban build-up land in a county in central China. The topic and results are generally interesting, but the length of the manuscript is short for an article type. The authors need to add at least 2 more pages. Please see below my comments.

General Comments:

Please expand the Introduction a bit further:

The Introduction part can be expanded by adding more recent articles:

Please add another paragraph and add the following references for land use and land cover (LULC) classifications:

Deep transfer learning applied to EuroSAT data for LULC classification:

https://doi.org/10.3390/s21238083

The reference above can also be added after EuroSAT in line 58.

Use of Landsat and MODIS for LULC classification:

https://doi.org/10.3390/s19224891

Deformable Convolutional Neural Network for Remote Sensing Scene Classification:

https://doi.org/10.3390/rs13245076

Please mention how the paper is organized:

At the end of the Introduction, please write how the rest of the manuscript is organized. For example, say “The rest of the paper is organized as follows. Section 2 describes the study region, datasets, and methodology. Section 3 demonstrates the results, etc.”

Please add a Discussion section:

Please add a Discussion section before the Conclusion section. In the discussion section, please briefly mention the objective of the paper and then discuss the results in the light of other similar studies by other researchers. This section can be at least one full page long. Please also discuss the use of different spectral bands, such as RGB, NIR-Infrared, Red-edge, etc. or a combination of them for improving the classification accuracy (see the first article that I mentioned above for more details).

Please improve the quality of the figures:

The font size of the numbers and texts in all the figures should be enlarged. Please enlarge the dimension of the figures so that their width covers the width of the page but with about a one-inch margin from both sides. The font size should be enlarged so that the size of the text and numbers written inside the figures (including the legends, coordinates, axis labels, etc.) all be the same size and almost the same size as the font size of the captions of the figures. Furthermore, please ensure the quality/resolution of the figures is at least 300dpi.  

Specific Comments.

Section 2.2. Lines 154 and 159. Please move all the links to the References and cite them there instead. In the references please also add the access date for the links. Please see the MDPI guideline for formatting and style of references.

Please remove line 207, where it says 3.1 Subsection

The caption of Figure 3. “flowchart” is one word.

Line 324. Please define the acronym CNN. Please check to ensure all the acronyms are defined the first time they are mentioned in the manuscript. Also, in the caption of Table 1 please write what FC is.

Figures 4, 6, etc. Please add a legend to illustrate what those colours, yellow, green, blue, red, etc. are.

Thank you for your contribution

Regards,

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

The text is interesting and has great application significance. However, I have minor comments:

The study area should be expanded for example more information in the text for land use within this area.The authors should add the land use map for this area

I consider that the part related to the sources should be combined with the methods. The research area should be separate and should be expanded

Authors should also extend their research to the future.

In conclusion, it is worth writing more about the applicability of these studies.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

I would like to thank the authors for addressing my comments. The manuscript is significantly improved, and I recommend acceptance.

Please carefully proofread the article. For example, I notice that some of the references do not have an appropriate format.

References [30],[34],[37],[43],[57],[72]: It should be "IEEE Trans Geosci Remote Sens" instead of "Ieee T Geosci Remote"

Reference [44] and [47]: It should be "Sensors" instead of "Sensors-Basel"

Please follow the MDPI guideline for formatting the references.

Thank you for your contribution