the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 13 Jun 2025
Landslide Hazard Microzonation Using a Hybrid Integrated Approach to Reduce Disaster Risk: A Case Study of Jecheon, South Korea
Abstract. Effective landslide prevention and mitigation necessitate the development of reliable landslide susceptibility maps. However, previous studies have primarily focused on assessing the overall performance of predicted susceptibility rather than examining the spatial characteristics of the predicted Landslide Susceptibility Index (LSI). This study aims to evaluate the efficacy of predicted LSIs derived from widely used statistical models while considering the spatial characteristics of landslides. To achieve this goal, four commonly used LSI models, namely logistic regression (LR), certainty factor (CF), frequency ratio (FR), and information value (IV), were utilized to map landslide susceptibility in Jecheon, South Korea. The models were developed using 112 landslide inventories and taking into account topography, hydrogeology, soils, forests, and lithological heterogeneities. Subsequently, the predicted LSIs were compared with the 1D topography profiles of recent debris events delineated from the high-resolution aerial and drone imagery. The distribution of anticipated LSIs along the landslide source area to the landslide runout and deposit zones was found to be inconsistent with the landslide characteristics. Nevertheless, the overall accuracy of the FR, IV, CF, and LR models demonstrated the strong predictive capabilities of these models. To address this spatial inconsistency issue, we proposed a hybrid integrated approach to achieve higher accuracy than the individual LSI models. Subsequently, a landslide hazard microzonation map was prepared and validated based on the in-situ observations and inventory data. It was observed that 94.6 % of landslide inventory occurrences fell within severe to high susceptibility zones. Precision results, such as an area under the curve of 0.906, mean square error of 0.25, mean absolute error of 0.08, root mean square error of 0.28, and a precision of 88.3 %, suggest that the hybrid integrated approach is more useful for landcover planning and mitigating landslide-induced disaster risks compare to individual LSI models.
- Preprint
(12869 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on egusphere-2025-1169', Anonymous Referee #1, 27 Jun 2025
This study utilizes 112 landslide inventory points in Jecheon, South Korea, to generate landslide susceptibility maps using four commonly applied models: Frequency Ratio (FR), Certainty Factor (CF), Logistic Regression (LR), and Information Value (IV). The spatial consistency of the predicted Landslide Susceptibility Index (LSI) with observed landslide profiles was assessed using high-resolution aerial and drone imagery. To address observed spatial inconsistencies, a hybrid integrated model was proposed, resulting in a microzonation hazard map that reportedly shows improved predictive performance, with high precision metrics such as an AUC of 0.906. However, the manuscript has several critical flaws. First, the landslide distribution map is of low quality and the number of samples is insufficient to represent the actual pattern of landslide development in the region, especially in the context of rainfall-induced clustered events. The limited sample size also raises concerns about severe overfitting, which can lead to high accuracy on training data but poor generalization in real-world applications. Additionally, the manuscript lacks meaningful and insightful scientific discussion. The spatial inconsistency between predicted susceptibility and actual landslide characteristics is acknowledged but not adequately analyzed or explained. There is also a lack of in-depth analysis of model limitations, data uncertainties, and regional applicability, which significantly weakens the scientific value of the work. Based on these issues, I recommend rejection.
Citation: https://doi.org/10.5194/egusphere-2025-1169-RC1 - AC1: 'Reply on RC1', Sang-Guk Yum, 19 Aug 2025
-
RC2: 'Comment on egusphere-2025-1169', Anonymous Referee #2, 30 Jun 2025
The manuscript has some serious problems in methodology and data.
For the methodology, the used model such as frequency ratio (FR), certainty factor (CF), logistic regression (LR), and information value (IV), is not novel and new. Moreover, the as frequency ratio (FR), certainty factor (CF), and information value (IV) is similar and almost same models based on probability. I recommend the new models such as machine learning models. Also, the relationships between landslide and input factors should be analyzed for the reason using the FR. Then only related factors should be used for the analysis. The authors just describe the result of FR and used all factors which is related to landslide or not.
For the data, the authors have used the 112 landslides The number of landslides is too small to apply the modes. The Also, the study does not explicitly describe using a separate validation dataset (e.g., splitting the 112 landslides into training/testing subsets). Performance metrics like AUC were apparently computed on the same inventory used for modeling, which could lead to overfitting concerns. A randomized train-test split or cross-validation would increase confidence that the models generalize beyond the known inventory.
There is no “Discussion section. The section should be included with “Conclusion” section.
Citation: https://doi.org/10.5194/egusphere-2025-1169-RC2 - AC2: 'Reply on RC2', Sang-Guk Yum, 19 Aug 2025
Status: closed
-
RC1: 'Comment on egusphere-2025-1169', Anonymous Referee #1, 27 Jun 2025
This study utilizes 112 landslide inventory points in Jecheon, South Korea, to generate landslide susceptibility maps using four commonly applied models: Frequency Ratio (FR), Certainty Factor (CF), Logistic Regression (LR), and Information Value (IV). The spatial consistency of the predicted Landslide Susceptibility Index (LSI) with observed landslide profiles was assessed using high-resolution aerial and drone imagery. To address observed spatial inconsistencies, a hybrid integrated model was proposed, resulting in a microzonation hazard map that reportedly shows improved predictive performance, with high precision metrics such as an AUC of 0.906. However, the manuscript has several critical flaws. First, the landslide distribution map is of low quality and the number of samples is insufficient to represent the actual pattern of landslide development in the region, especially in the context of rainfall-induced clustered events. The limited sample size also raises concerns about severe overfitting, which can lead to high accuracy on training data but poor generalization in real-world applications. Additionally, the manuscript lacks meaningful and insightful scientific discussion. The spatial inconsistency between predicted susceptibility and actual landslide characteristics is acknowledged but not adequately analyzed or explained. There is also a lack of in-depth analysis of model limitations, data uncertainties, and regional applicability, which significantly weakens the scientific value of the work. Based on these issues, I recommend rejection.
Citation: https://doi.org/10.5194/egusphere-2025-1169-RC1 - AC1: 'Reply on RC1', Sang-Guk Yum, 19 Aug 2025
-
RC2: 'Comment on egusphere-2025-1169', Anonymous Referee #2, 30 Jun 2025
The manuscript has some serious problems in methodology and data.
For the methodology, the used model such as frequency ratio (FR), certainty factor (CF), logistic regression (LR), and information value (IV), is not novel and new. Moreover, the as frequency ratio (FR), certainty factor (CF), and information value (IV) is similar and almost same models based on probability. I recommend the new models such as machine learning models. Also, the relationships between landslide and input factors should be analyzed for the reason using the FR. Then only related factors should be used for the analysis. The authors just describe the result of FR and used all factors which is related to landslide or not.
For the data, the authors have used the 112 landslides The number of landslides is too small to apply the modes. The Also, the study does not explicitly describe using a separate validation dataset (e.g., splitting the 112 landslides into training/testing subsets). Performance metrics like AUC were apparently computed on the same inventory used for modeling, which could lead to overfitting concerns. A randomized train-test split or cross-validation would increase confidence that the models generalize beyond the known inventory.
There is no “Discussion section. The section should be included with “Conclusion” section.
Citation: https://doi.org/10.5194/egusphere-2025-1169-RC2 - AC2: 'Reply on RC2', Sang-Guk Yum, 19 Aug 2025
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 269 | 79 | 19 | 367 | 11 | 26 |
- HTML: 269
- PDF: 79
- XML: 19
- Total: 367
- BibTeX: 11
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1