Snowmelt Flood Susceptibility Assessment in Kunlun Mountains Based on the Swin Transformer Deep Learning Method

Round 1

Reviewer 1 Report

This manuscript (remotesensing-2044607) aims to compare several machine learning and deep learning methods (in particular Swin Transformer) for snowmelt flood susceptibility mapping in the Kunlun Mountains region, China from 1999 to 2020. Although it is an easy-to-follow manuscript, it is not entirely new to use these common machine learning methods in flood susceptibility mapping. Another very serious concern is that some related studies have been neglected. A further and detailed literature review must be conducted. Also, the current results of this study can hardly be reviewed because of those problems about data and methodology. Therefore, at least a “Major Revision” is required. My suggestions and comments are presented as follows:

- 1. Both the Abstract and the Introduction Section are weak because the authors did not clearly raise a real important scientific question or gap related to flood susceptibility mapping. Therefore, potential readers can hardly identify the need that the authors should have to provide a new solution from an international perspective. What I have learned from the introduction is that the authors applied some established models to a snow-covered mountain in China. Note that those machine learning and deep learning techniques are not new methods in geohazard susceptibility mapping.

- 2. In Line 64~67: the authors mentioned that: "Nevertheless, the assessment of snowmelt flood susceptibility in large regions is still difficult to achieve because snowmelt hydrological models tend to focus on long-term runoff simulation, and the models' spatial parameterization is complicated". Please explain in more details why the assessment of snowmelt flood susceptibility in large regions is still difficult.

- 3. In Line 67~70: the authors mentioned that: "due to the sparse data and the fact that the snowmelt flood formation process is more complex than the physical process of rainfall runoff, traditional data-driven methods are difficult to apply in quantifying the relationship between indicators and flooding". Please explain in more details why traditional data-driven methods are difficult to apply. What is the complex element? What is the real difference between rainfall flooding and snowmelt flooding? Did this manuscript effectively deal with these problems?

- 4. Actually, Machine learning (ML) and deep learning (DL) methods are not counterpart. Deep learning is part of a broader family of machine learning methods based on artificial neural networks with representation learning.

- 5. The authors need to answer clearly why we must build the flood susceptibility models in larger regions? Actually, the influencing factors and spatial patterns will be very different across different sub-regions because of the existence of significant spatial heterogeneity.

- 6. The authors need to answer clearly why the Swin Transformer (Swin-T) method was selected. As far as I known, this method is not the current state-of-the-art deep learning method.

- 7. Section 2.1. Study Area, Figure 1: please provide the snow-covered regions in this study area because it is very important for snowmelt flooding.

- 8. Another serious concern is that the authors must look further into the latest research in this field. In fact, the literature review is far from enough. In particular, the maximum entropy (MAXENT) algorithm has been successfully used in flood susceptibility mapping. However, this well-accepted technique is totally ignored in the manuscript, and the following articles should be cited. The Introduction section is meant to set the context for your research work and highlight how it contributes to the knowledge in this field and builds on previous similar studies.

Predicting future urban waterlogging-prone areas by coupling the maximum entropy and FLUS model. Sustainable Cities and Society, 2022, 80: 103812

Utilizing User-Generated Content and GIS for Flood Susceptibility Modeling in Mountainous Areas: A Case Study of Jian City in China. Sustainability, 2021.

Hazard and vulnerability in urban flood risk mapping: Machine learning techniques and considering the role of urban districts. International Journal of Disaster Risk Reduction, 2020.

- 9. In Line 164~167: please explain clearly why the flood inventory was created using data from flood events between 1999 and 2020. Were the flood inventories observed in 23 years ago still active? It should be better to remain only those frequently-happened flooding events.

- 10. Section 2.3. Flood Conditioning Factor: the selection criterion of the influencing factors of snowmelt flooding is still not clearly explained. For example, snowmelt flooding will be caused not only by local precipitation (at/over each grid cell), but also by flood water from high precipitation area in the upstream regions.

- 11. In Table 1, the authors failed to provide many specific details of the input data, such as the pre-processing processes, date, and accuracies. I suggest the authors to add this information in this table. In particular, did the authors utilize the long-term multi-temporal spatial data given that the flooding data are from 1999 to 2020? How to deal with the data with different spatial resolution?

- 12. Please explain why the support vector machine (SVM) and random forest (RF), and two deep learning models, deep neural network (DNN) and convolutional neural network (CNN) were selected for comparison. For example, why not use the other more common MAXENT, or artificial neural network models (ANN)?

- 13. It seems that this manuscript did not consider the serious problem of the multicollinearity of different flood conditioning factors.

- 14. In Table 2, the authors need to clearly explain the detailed determination procedures of all these hyper-parameters.

- 15. The accuracy values for both training and testing samples should be presented. The current results of this study can hardly be reviewed because of those problems about data and methodology.

- 16. It is suggested that the flood susceptibility hotspot regions should be identified by using the hotspot analysis tool, which can provide meaningful and interesting results.

- 17. What about the snowmelt flood susceptibility maps obtained from the other machine learning methods?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In Figure 1, please remove the nine-dash line from the map because of its illegality.

It should explain why the Kunlun Mountains region was selected as the study area.

Why the ReliefF method is used to assess the effective contribution of various factors for classification?

Why was the Jackknife test method chosen to test the sensitivity to 15 melt flood factors?

Are there any challenges using the Swin-T-based approach?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors and Editor,

Thanks for your kind invitation to review the manuscript entitled “Snowmelt flood susceptibility assessment in Kunlun Mountains based on Swin Transformer deep learning method” by Ruibiao Yang, Guoxiong Zheng, Ping Hu, Ying Liu, Wenqiang Xu and Anming Bao.

This paper carries out an assessment of the susceptibility to meltwater flooding with deep learning techniques. I recommend a Major Revision. Here are my suggestions:

1. The article has a strong regional study component. The writing should be oriented to a generic approach (applicable to more places in the world) supported by research in a study area.

2. The Introduction section is very poor. Just looking up " Role of Snowmelt in Floods" in the Journal Water turns up a number of interesting papers not covered here. It is necessary to frame the work in a broader form. This section has to improve substantially. Bibliographic references are scarce, especially in the international area. It is necessary to provide more literature.

3. Also the introduction fails to provide a suitable perceptive on the novelty and importance of the study, while this information needs to be stated clearly in the text. The introduction does not incorporate the main methodological concept adopted in this manuscript.

4. It would be better to highlight major difficulties and challenges and your original achievements to overcome them in a clearer way in the abstract and introduction.

5. The methodology limitation should be mentioned.

6. The Discussion is loose. Should summarize the manuscript's main finding(s) in the context of the broader scientific literature and address any limitations of the study or results that conflict with other published work.

With kind regards

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thank you for incorporating my previous comments and suggestions.

Reviewer 3 Report

Dear Authors and Editor,

The authors have put diligent efforts in responding to reviewers' comments and they are revised the paper accordingly. The paper is ready. I am happy to recommend acceptance of the paper.

Fine document, congratulations to the authors.

Best regards,