Neural Network-Based Urban Change Monitoring with Deep-Temporal Multispectral and SAR Remote Sensing Data

Round 1

Reviewer 1 Report

This paper is worth for acceptance, novelty of the idea seems interesting and small changes need to be incorporated in order to enhance.
The article has been read carefully, and some crucial issues have been highlighted in order to be considered by the author(s).
All the acronyms should be defined and explained first before using them such that they become evident for the readers.
Most of the typos and incorrect grammars have been corrected, but it is still necessary to subject the paper to proofreading.
The paper needs to be restructured in order to be precise.The Introduction and related work parts give valuable information for the readers as well as researchers. In addition recent papers should be added in the part of related work.
As it is real time application oriented, authors should care over the outcome of the proposed framework by meeting the future requirements too.
Representation of figures needs to be improved.
Grammatical errors should be validated.
It would be good if security domains [1], such as backdoor sample, would be reflected in future research or related work.
[1] Kwon, Hyun, Hyunsoo Yoon, and Ki-Woong Park. "Multi-targeted backdoor: Indentifying backdoor attack for multiple deep neural networks." IEICE Transactions on Information and Systems 103.4 (2020): 883-887.

Author Response

Thank you for your review!

With regards to your feedback:

• We will double check the used acronyms and ensure introduction in order.
• We will send the article for another round of native language speaker review (please note that we use US English)
• "The paper needs to be restructured in order to be precise." We'd be happy to hear what you'd like to see changed.
• Table 1 contains an overview of (not only) recent other work. E.g. Manzoni et al. [34] is from 2021 or Ansari et al. [33] is from 2020. We have only given a non-exhaustive (representative) overview as we're constraint in space (the paper already is 30 pages) and this is not a survey.
• "Representation of figures needs to be improved." We'd be happy to hear what you'd like to see changed.
• This case study was part of the BLENDED project which runs the training/inference in a blockchain monitored environment to add security. We will consider your provided reference for the BLENDED project.
• "As it is real time application oriented, authors should care over the outcome of the proposed framework by meeting the future requirements too." We have already applied the trained network on other sites. We plan for another publication to focus more on the "real-time" analysis that our method supports.

Reviewer 2 Report

This paper is very interesting and highly applicative, dealing with the crucial issue of monitoring the dynamics of urban areas and urban landscapes making use of large data sets while training neural networks for automated image-interpretation. The paper is clearly written and relatively easy to read. The use of two different types of data sets like time series of multispectral optical and SAR images for remote sensing automated classification is very interesting and may prove very useful in the near future. The work is well organized and has a good presentation of the collected scientific material. In order to improve the effectiveness of the paper, the Authors may consider profitable to include a map for allocating the 3 study sites (Areas of Interest). However, I repute that the article can be published in present form.

Author Response

Thank you for your review!

We will add a map of AoIs as appendix.

Reviewer 3 Report

An interesting paper that provides great detail about the models used and the examples of the results.

Author Response

Thank you for your review!

Reviewer 4 Report

 The paper presents an interesting study on remote sensing based urban change detection. Authors described methodology and developed a complete automatic system for utilization of satellite images  for detection of urban changes. 

Urban changes are detected based on a 1 year or 6 months time window, stacking images from both optical and SAR sensors and proposing a deep learning architecture for detection of changes. The method is implemented on three study areas and for 2 time eras using different constellations of satellites based on their availability for the era period.

The paper employs several novel aspects : 

• usage of complementary satellites for different eras
• fusion of SAR and optical images
• stacking satellite images that can be co-registered 
• proposition of deep learning architecture for detection of urban changes
• dealing with big data chalenges

The same proposed deep learning architecture is trained for each of the two eras and tested on  selected study areas. 

My only remark is that it is a pity that no quantitative evaluation is performed. While qualitative analysis holds interesting discussion it would be useful to provide a quantitative description of the performance of the proposed method. I understand that since there are no ground truth labels, but only uncertain synthetic labels are used for training,  it is not possible to calculate precise numbers but the authors should consider using data from other sources (such as cadastral data)  for ground truth labels of urban change.

Minor remarks:

-it would be useful to add a map od study areas

-why do you use different reference systems for two eras (table 2)

-distinguish between terms prediction and detection of urban change. 

Author Response

Thank you for your review!

We will add a map of AoIs as appendix.

The reason we use two different reference systems was originally due to two different processing pipelines. We left two different reference systems because of the following benefits:

• It demonstrates that the choice of reference systems is not crucial to our method as long as the mission pairs of the same era uses the same reference system.
• In our opinion, it aids in reading as coordinate formats can be directly assigned to one era (every era has a different format and hence is distinguishable).

We will consider the distinction of terms "prediction" and "detection". However, this might not be always possible as for the Remote Sensing domain the term is "detection", while in machine learning/deep learning, we apply a "prediction" task. We try to clarify this.

With regards, to quantitative evaluation:
We have access to cadastre data from Liege [1] and attempted a quantitative evaluation. This was not possible due:

• Cadastre data has a resolution of one year
• Not all changes that happened on the ground are reflected in time in the cadastre data
• The cadastre data from Liege is incomplete: not all buildings are in the system and contains no roads or other impervious objects
• There was also a transformation issue in the cadastre data itself which leads to some buildings shifting from one year to another.

We have started to look into quantitative analysis but that requires a significant amount of additional time and a a significant amount other sources to build a ground truth with a high confidence.

For the underlying work, we primarily describe our method. There should be further optimizations and fine tuning in future papers which should then also have a quantitative analysis to compare.

[1]: https://finances.belgium.be/fr/particuliers/habitation/cadastre