the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
From typhoon rainfall to slope failure: optimizing susceptibility models and dynamic thresholds for landslide warnings in Zixing City, China
Abstract. From typhoon rainfall to slope failure, this study addresses the urgent need for typhoon-adapted hazard warning systems in mountainous regions like Zixing City, China. We develop an integrated framework to optimize dynamic susceptibility models and rainfall thresholds by leveraging machine learning and spatiotemporal rainfall analysis. Using buffer-based negative sampling (0.1–5.0 km) and variable weighting methods (IV, CF, FR), we compare SVM and LightGBM models. The SVM model with FR input at 0.5 km buffer achieved the highest accuracy (AUC=0.913), correctly classifying 86.4 % of landslides in high-risk zones, revealing how typhoon-driven hydrology interacts with slope instability. For rainfall thresholds, the H24-D7 model (24-hour intensity vs. 7-day antecedent rainfall) emerged as optimal (71.8 % accuracy), effectively capturing typhoon-specific triggers like short-term downpours and cumulative soil saturation. Kriging interpolation generated spatially explicit thresholds, identifying granite slopes and road-proximal areas as hotspots for typhoon-induced failures. The final hazard warning system, integrating susceptibility and dynamic thresholds, showed 71.4 % overlap with historical landslides, emphasizing the critical role of typhoon rainfall dynamics in slope failure prediction. This work provides a scalable approach for regions facing typhoon-related landslide risks, prioritizing both spatial heterogeneity and temporal rainfall patterns.
- Preprint
(10434 KB) - Metadata XML
-
Supplement
(957 KB) - BibTeX
- EndNote
Status: open (until 22 Sep 2025)
-
RC1: 'Comment on egusphere-2025-2298', Anonymous Referee #1, 26 Aug 2025
reply
The manuscript “From typhoon rainfall to slope failure: optimising susceptibility models and dynamic thresholds for landslide warnings in Zixing City, China” tackles an important and timely topic with clear potential to contribute to landslide hazard research and early warning practices. The integration of modelling approaches is promising; however, the manuscript in its current form requires substantial revisions. Key issues relate to clarity, consistency, methodological justification, and depth of discussion.
The following are some comments that are intended to be constructive and to help the authors strengthen the manuscript, ensuring it meets NHESS standards and realises its potential contribution.
1. Define all new terms (e.g., IV, CF, FR, SVM, and others) when they first appear in the text, including in the abstract.
2. Why is it necessary to develop a hazard warning system for typhoon-induced landslides in this specific area? Provide a stronger justification. At present, the manuscript only discusses methodological limitations in the introduction. The research gap is unclear, and the rationale for conducting this work specifically in Zixing City is insufficient.
3. The manuscript sometimes uses the term “typhoon-specific hazard monitoring systems” and other times “typhoon rainfall-induced landslide hazard warning system”. It would be better to use consistent terminology throughout. I suggest adopting “typhoon-specific rainfall-induced landslide monitoring systems”, as this best reflects the study’s main objective and reduces confusion for the reader.
4. Provide more information about the study area, including its geographical, geophysical, geological, and hydrological characteristics.
5. Add the units of the factors shown in Figures 2a and 2b.
6. In the text, the authors state that they used 705 landslide points, but Figure 3 (the framework flowchart) refers to 645. Please clarify this inconsistency.
7. There are many machine learning models available for classification tasks. Why did you choose SVM and LightGBM over others? Please justify this choice.
8. Clarify the mechanism for assigning D7 (or other designations) to each landslide point. Specifically, explain how each of the >700 landslide points was linked to one of the 12 rain gauge stations.
9. Provide detailed explanations of all factors with significant results in Table 2. The current explanations are not sufficient.
10. Include the statistical results of the multicollinearity test in the appendix (or supplementary material), and reference them in the main text.
11. Explain how you normalised the resolution of the different factor maps. Since the primary data have different scales, all layers must be resampled to the same resolution to create the susceptibility map.
12. Adjust the font size in Figures 4, 5, and 6. At present, the text appears disproportionately large compared to the maps.
13. Present the AUC values in separate columns for training and testing in Table 3.
14. Avoid the use of unnecessary em dashes (—) throughout the text.
15. Ensure consistency across figures. For example, in Figure 6, landslide points are shown only on the first two maps (SVM and LightGBM), whereas in Figure 7, they are shown on all maps. Standardise this approach.
16. Adjust the sizes of the maps in Figure 8 so that all are presented at the same scale.
17. Why do you describe the final product as a monitoring system? Will it be hosted online for interactive use? If not, it is more accurate to describe it as a hazard zonation map. At times, you also refer to it as a framework. Please avoid such inconsistencies.
18. Consider evaluating the performance of the warning zonation maps (Figures 8d and 8e).
19. In the discussion, you state that the system “can identify regions where slopes are already saturated due to pre-typhoon rainfall and are thus highly susceptible to failure during the typhoon’s high-intensity rainfall phase.” How does it achieve this? Is the system dynamic? The manuscript provides no evidence of using dynamic data; all analyses appear to rely on static datasets. Please clarify.
20. The manuscript lacks a sufficiently scholarly discussion. Strengthen the reasoning behind your findings by incorporating more relevant references.Citation: https://doi.org/10.5194/egusphere-2025-2298-RC1 -
AC1: 'Reply on RC1', Weifeng Xiao, 27 Aug 2025
reply
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-2298/egusphere-2025-2298-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Weifeng Xiao, 27 Aug 2025
reply
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
612 | 30 | 11 | 653 | 24 | 10 | 7 |
- HTML: 612
- PDF: 30
- XML: 11
- Total: 653
- Supplement: 24
- BibTeX: 10
- EndNote: 7
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1