FedUKD: Federated UNet Model with Knowledge Distillation for Land Use Classification from Satellite and Street Views

Round 1

Reviewer 1 Report

In the study, a new approach based on U-Net on Land Use is proposed. Although the work is in a popular field, major revision is required. Here are my suggestions:

1) There are not enough references in the Introduction. You must provide references for information that you did not produce yourself. For example, what is the source of the following information? Please re-evaluate the entire manuscript in this way.

‘’Around 196 countries had entered the Paris climate accord in 2015, an international agreement that pledges to keep the global average temperature rise to below 1.5°C and to reduce greenhouse gas emissions.’’

2) The introduction part should be rewritten. The problem should be defined from broad to narrow scope.

3) Appropriate references to these studies should be presented at the end of the sentence beginning with In line 51, Several research work….

4) Line 71, reference required.

5) Literature research is very limited. There are many studies using deep learning for LULC. Literature research should be developed. Below, I recommend a few studies that have been published recently.

- Liu, T., Yang, L., & Lunga, D. (2021). Change detection using deep learning approach with object-based image analysis. Remote Sensing of Environment, 256, 112308.

- Atik, S. O., & Ipbuker, C. (2021). Integrating convolutional neural network and multiresolution segmentation for land cover and land use mapping using satellite imagery. Applied Sciences, 11(12),

-Wang, S., Chen, W., Xie, S. M., Azzari, G., & Lobell, D. B. (2020). Weakly supervised deep learning for segmentation of remote sensing imagery. Remote Sensing, 12(2), 207.

6) Line 237 and Line 238 have typos and formatting errors.

7) Section 4.1.1 is unnecessarily long. It can start directly from the second paragraph.

8) Experiments should be performed on at least 1 more dataset. It should be understood whether your results depend on the dataset.

9) Compare your results with the existing methods in the literature. Thus, the place of your results in the literature can be better understood.

10) Please make suggestions for future studies in Conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Renuga Kanagavelu and co-authors proposed a federated UNet Model with knowledge distillation (FedUKD) for land use classification in a distributed fashion. 

 

Although the FedUKD's accuracy is on par with centralized models, the authors' main contribution/novelty is to reduce communication cost by ~60 times through local knowledge distillation. 

 

Overall, the authors presented enough information of their motivation and experiment details. I would recommend its publication if the authors could response to the following questions:

 

1. The dataset in this research is of relatively small scale. There are only 3 federated clients with less than 100 images for training. The fact that the local knowledge distillation is effective to compress model by ~60 times in this small dataset does not necessarily means one can expect similar compression rate in a more complicated large-scale real-world dataset. Are the authors confident that their impressive model compression rate can be achieve in a much larger dataset? Have the authors already conducted compression rate vs dataset size analysis? 

 

2. The benefit of model compression by knowledge distillation is at the cost of additional computation cost locally at each clients. In this paper, the authors concluded that they are able to save ~100 MB in communication cost. However, it is at the cost of building knowledge distillation model training at all local clients. Considering the low communication cost in modern internet infrastructures, the computation costs at each clients will likely out-weight the communication over-head. Could the authors explain if their FedUKD could achieve an overall cost saving from both communication and computation perspectives? 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper is quite well written. I only have some minor comments for authors to address, as explained below:

1. Overview of proposed work can be further elaborated in Section 1.

2. Please highlight the technical contributions of current work in Section 1. 

3. Please explain the main differences between the proposed work and those published works used for Semantic Segmentation of Satellite Images.

4. For the sake of fairness, the limitation of proposed work should be discussed as well. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I think the necessary revisions have been made. The manuscript is acceptable as it is.

Author Response

We would like to express our gratitude for your valuable feedback on our manuscript. We have taken your comments and suggestions into consideration and have made the necessary revisions. We are pleased to hear that you find the manuscript to be acceptable in its current form and look forward to the opportunity to share our research with the wider community.