Segmentation and Connectivity Reconstruction of Urban Rivers from Sentinel-2 Multi-Spectral Imagery by the WaterSCNet Deep Learning Model

Round 1

Reviewer 1 Report

The paper implements a revised deep learning network model by introducing MSD path module, MSP block and the attention gate mechanism, which is used for urban river segmentation from multi-spectral images and river connection. Comparisons with available network models show that the new structure outperforms in accuracy and efficiency, particularly in handling small rivers in urban regions.

 

The author may take the following suggestions for further revisions

1. Manilla is in South East Asia.

2. Performance validation of segmentation and connection can better be conducted with real world data, other than only results from other non-optimized network models.

3. Influences of overlays on water surface, for example by tree canopies, shall be discussed.

4. More explanations about how attention gates are designed and implemented are expected.

5. For verification of such design of two step network model, the implementation codes shall be provided.

Misspelling and sentences can be improved.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript introduces a cascaded deep-learning network model named WaterSCNet, designed for both urban river segmentation and the simultaneous reconstruction of river connectivity obscured by road and bridge crossings. The authors propose two subnetworks within WaterSCNet, namely WaterSCNet-c and WaterSCNet-s, the former focusing on river connectivity reconstruction while the latter deals with river segmentation. WaterSCNet-c differs from WaterSCNet-s by lacking the MSD path module, yet it incorporates the MSP block and attention gate mechanism, which enhance its ability to capture texture relationships, multi-scale information, and global positional relationships among rivers.

The authors claim that the proposed WaterSCNet model achieves superior performance in river segmentation and connectivity reconstruction compared to existing models such as E-UNet, U-Net, SegNet, and HRNet. The performance evaluation is conducted using metrics including MCC, F1, Kappa, and Recall, which are widely accepted in the field of image segmentation and pattern recognition. The results suggest that the WaterSCNet model outperforms the classic U-Net model by a range of percentages in these metrics. Moreover, it also surpasses the last-ranked HRNet model, demonstrating its competitive edge.

While the claims made in the paper are intriguing, some aspects require further clarification and validation. The architecture and design choices for the WaterSCNet model need to be elaborated upon in greater detail. The reasoning behind the decision to exclude the MSD path module from the WaterSCNet-c subnetwork should be explicitly discussed, along with its potential impact on the overall performance. Furthermore, the methodology for selecting the specific comparison models and datasets needs to be justified to ensure a fair evaluation.

In conclusion, the proposed WaterSCNet model holds promise in the domain of river segmentation and connectivity reconstruction from Sentinel-2 imagery. The incorporation of advanced techniques such as the MSP block and attention gate mechanism appears to provide a distinct advantage. However, further investigation is required to address the raised concerns and to provide a more comprehensive understanding of the model's performance and capabilities.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

This is a meaningful study. Masking and shadows have always been major limitations in the application of optical remote sensing imagery for water surface extraction. It is an innovative approach to skip the unrecognized water areas and infer the complete river segment using probability maps and spatial features. However, I have some unclear points regarding this study and I hope the authors can address and provide additional explanations in the main text.

1.   If the "gap" in the river is not caused by a bridge but by a dam (or dry riverbed), will the reconstruction process using WaterSCNet-c mistakenly identify originally discontinuous river segments as continuous ones?

2.   In the River Connectivity Reconstruction Subnetwork, what specific indicators are included in the input and output layers used for neural network training? How long are the samples and how were they obtained?

3.   It is suggested to provide clarification on the minimum river width that can be identified by this method and the maximum bridge width that can be restored (e.g., in Figure 7b, when the distance between two river segments exceeds what value, it is considered discontinuous?). This will help readers understand the applicability of the method clearly.

4.  How were the segmentation labels and connectivity labels obtained?

5. Shadows from tall buildings are also important limiting factors in water body identification. Can the authors provide information on the applicability of this method in mitigating the impact of shadows on river continuity?"

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

This work carried out research on river segmentation and connectivity reconstruction based on the improved UNet. Compared with existing deep learning networks such as UNet, SegNet, and HRNet, the proposed model (WaterSCNet) achieved better results in both river segmentation and connectivity reconstruction. The overall expression of the article is generally clear. I have the following suggestions for modification.

1.      It is recommended to introduce UNet and its improved models, such as UNet++, and related work in image segmentation in the introduction. These works are linked as much as possible to the intelligent recognition of remote sensing images.

2.      It is suggested to introduce E-Unet in WaterSCNet-s in the introduction section. For example, why this deep learning framework is better suited for the dataset in this study? At the same time, the relevant research progress of 3D convolution and attention gate should be introduced.

3.      It is recommended to briefly explain the necessity of developing the river connection reconstruction algorithm WaterSCNet-c in the introduction.

4.      It is recommended to present the scientific question to be addressed in the introduction.

5.      The innovation of this research should be more highlighted.

6.      Introduction should not contain results or conclusions, such as "Experimental results show that the WaterSCNet model could achieve better performance than the E-UNet [32], U-Net [33], SegNet [34], and HRNet [35] models in both river segmentation and connectivity reconstruction.”

7.      Please describe in detail the process of forming the experimental data set by slicing in Figure 6. Does the process apply data augmentation?

8.      Select a typical module in the method section to introduce the number of parameters and the number of pixels (or connections).

9.      The paper lacks discussion of key innovations. For example, the underlying reasons why 3d convolutional modules and attention gates improve the prediction performance of the proposed model should be discussed.

 

10.  It is recommended to introduce future work in the conclusion section, for example, possible improvements or applications.

 

The overall expression of the article is generally clear.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

No further comments.

Author Response

Dear Reviewer,

Thanks so much for your second round of reviews.

Reviewer 3 Report

The author provided a response or made revisions to the comments from my initial review, and I am relatively satisfied with the modifications. In the response to the first comment, the author acknowledged the limitations of the proposed method in identifying dams (as distinct from bridges and roads, as dams act as barriers), and suggested adding corresponding discussions in the main text. Additionally, the author's response to the second comment also recommended presenting it in the main text.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

The author has responded to my review comments one by one and revised the manuscript accordingly. I have no further comments. In my opinion, the article meets the requirements for public publication.

Author Response

Dear Reviewer,

Thanks so much for your second round of reviews.