Integrated Framework for Unsupervised Building Segmentation with Segment Anything Model-Based Pseudo-Labeling and Weakly Supervised Learning

Round 1

Reviewer 1 Report

please see the attachment

Comments for author File: Comments.pdf

none

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

(1) The direct incorporation of figures from other researchers by the author in Figure 2 is deemed inappropriate. Kindly include proper citations and explanations to address this issue.

(2) From the visual results in Figure 8, it is observed that the smoothness around the edges of buildings is better when using pseudo-labels directly obtained from the SAM model, whereas the extraction results of our proposed method exhibit significant salt-and-pepper noise. Addressing this issue is crucial for practical applications. Additionally, the visualized results in the manuscript appear blurry, making it challenging to discern differences in recognition outcomes among various methods and to showcase the advantages of our approach. It is recommended to incorporate supplementary local detail images for a more effective comparative analysis.
(3) The manuscript employs the mIoU metric, where, for binary semantic segmentation tasks, mIoU represents the mean Intersection over Union (IoU) accuracy of the target class and the background class. However, in Table 4, the authors separately apply the mIoU metric to building, vegetation, and road. How is mIoU computed in these cases? Would it be more appropriate to directly use the IoU metric?

(4) The experiments in this paper are conducted solely on a single dataset, which may limit the ability to demonstrate the effectiveness of the proposed method. It is recommended to conduct experiments on publicly available building extraction datasets with varying resolutions, such as WHU and Massachusetts, to provide a more comprehensive validation of the proposed approach.

(5) The figure captions for both Figure 8 and Figure 9 contain inaccuracies and require correction. Please make the necessary revisions.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

1.     There is too much description of SAM in the abstract. Provide a more detailed description of your proposed edge-driven model.

2.     Don’t discuss SAM in the Materials and Methods section. It breaks the readability flow in the manuscript.

3.     It would be better to provide more details about choosing the NDVI index.

4.     A detailed description of the thresholding algorithm is stated in line no 136.

5.     How would you deal with cloud based NDVI index? Has the thresholding algorithm provided any impact cloud based data?

6.     Please improve the quality of Figures 5 and 6 for better understanding.

7.     A described figure of the overall framework of the system should be included before the “Materials and Methods” section.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The modified manuscript has responded to the reviewers' comments point by point and has significantly improved.