the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Elevating Flash Flood Prediction Accuracy: A Synergistic Approach with PSO and GA Optimization
Abstract. Flash floods are frequent and devastating natural disasters in small mountainous river basins worldwide, causing significant harm to people, infrastructure, and property. Flash flood susceptibility mapping is a crucial tool for damage prevention and reduction. This study is focused on the creation of flash flood susceptibility maps in a mountainous region in northern Vietnam. We enhanced the performance of robust machine learning models, including Support Vector Machines (SVM), Random Forests (RF), and XGBoost (XGB), by applying advanced optimization techniques such as Particle Swarm Optimization (PSO) and Genetic Algorithms (GA). These models were developed based on 14 key factors, including elevation, slope, aspect, curvature, topographic wetness index (TWI), stream power index (SPI), flow accumulation, river density, distance to the river, NDVI, land use/land cover (LULC), rainfall, soil type, geology, and 412 flood inventory points. Nine models were tested, including three standalone ML algorithms (SVM, RF, XGB), three ensemble models optimized with PSO (PSO-SVM, PSO-RF, PSO-XGB), and three optimized with GA (GA-SVM, GA-RF, GA-XGB). The results indicated that ensemble models outperformed standalone ones, with the PSO-XGB, GA-XGB, and GA-RF models exhibiting outstanding performance, achieving accuracy rates of 0.939, 0.927, and 0.933, along with remarkable AUC-ROC scores of 0.957, 0.968, and 0.977, respectively. This innovative study introduces a novel set of associative models, contributing significantly to the advancement of flood prediction techniques. The methodology holds applicability for various regions characterized by similar topographical and climatic attributes. Furthermore, enhancing the precision of flood forecasting contributes to the formulation of mitigation strategies by municipal authorities to mitigate prospective flood-related impacts.
- Preprint
(2987 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on nhess-2024-215', Anonymous Referee #1, 30 May 2025
After a detailed review of the manuscript titled "Elevating Flash Flood Prediction Accuracy: A Synergistic Approach with PSO and GA Optimization," I recommend rejection due to insufficient novelty in the methodology.
The manuscript presents a study on flash flood susceptibility mapping in the Song Ma district, northern Vietnam, employing machine learning models (Support Vector Machines, Random Forests, and Extreme Gradient Boosting) optimized with metaheuristic algorithms (Particle Swarm Optimization and Genetic Algorithms). While the work is well-executed and clearly presented, the approach of integrating PSO and GA with machine learning for flood mapping is not new. This methodology is well-documented in existing literature, and the manuscript does not offer significant innovations to distinguish it from prior studies.
Key Reasons for Rejection:
- Lack of Novelty: The combination of PSO and GA with machine learning models for flood susceptibility mapping is an established technique, not a novel contribution. There are several paper flood mapping using hybrid AI and metaheuristic algorithms such as:
- Bui, D. T., Ngo, P. T. T., Pham, T. D., Jaafari, A., Minh, N. Q., Hoa, P. V., & Samui, P. (2019). A novel hybrid approach based on a swarm intelligence optimized extreme learning machine for flash flood susceptibility mapping. Catena, 179, 184-196.
- Plataridis, K., & Mallios, Z. (2023). Flood susceptibility mapping using hybrid models optimized with Artificial Bee Colony. Journal of Hydrology, 624, 129961.
- Rezaie, F., Panahi, M., Bateni, S. M., Jun, C., Neale, C. M., & Lee, S. (2022). Novel hybrid models by coupling support vector regression (SVR) with meta-heuristic algorithms (WOA and GWO) for flood susceptibility mapping. Natural Hazards, 114(2), 1247-1283.
- Nguyen, H.D. GIS-based hybrid machine learning for flood susceptibility prediction in the Nhat Le–Kien Giang watershed, Vietnam. Earth Sci Inform 15, 2369–2386 (2022). https://doi.org/10.1007/s12145-022-00825-4
- Dodangeh, E., Panahi, M., Rezaie, F., Lee, S., Bui, D. T., Lee, C. W., & Pradhan, B. (2020). Novel hybrid intelligence models for flood-susceptibility prediction: Meta optimization of the GMDH and SVR models with the genetic algorithm and harmony search. Journal of Hydrology, 590, 125423.
- Ngo, P. T. T., Pham, T. D., Hoang, N. D., Tran, D. A., Amiri, M., Le, T. T., ... & Bui, D. T. (2021). A new hybrid equilibrium optimized SysFor based geospatial data mining for tropical storm-induced flash flood susceptible mapping. Journal of Environmental Management, 280, 111858.
- Kaya, C. M., & Derin, L. (2023). Parameters and methods used in flood susceptibility mapping: a review. Journal of Water and Climate Change, 14(6), 1935-1960.
- Nguyen, H. D. (2022). Flood susceptibility assessment using hybrid machine learning and remote sensing in Quang Tri province, Vietnam. Transactions in GIS, 26(7), 2776-2801.
- Established Literature: Several studies have already explored similar methodologies, including:
- "Flood Mapping with PSO-GA": Utilizes PSO and GA with Support Vector Machines for flood mapping.
- "Metaheuristic Flood Assessment": Applies GA and other metaheuristics with ANFIS for flood zoning.
- "Remote Sensing Flood Mapping": Combines PSO, GA, and Harmony Search with machine learning for flood susceptibility.
These examples highlight that the manuscript’s approach aligns with a well-trodden path in flood prediction research. Although the application to a specific region is detailed, it does not advance the methodological framework beyond what is already known.
Additional Notes:
The manuscript is well-written, with a thorough methodology and robust analysis. However, these qualities do not compensate for the lack of originality, which is a critical factor for publication.
In summary, despite its technical competence, the manuscript does not meet the threshold of novelty required for acceptance. Therefore, I recommend rejection.
Â
Citation: https://doi.org/10.5194/nhess-2024-215-RC1 -
AC1: 'Reply on RC1', Yuei-An Liou, 17 Aug 2025
We sincerely thank the reviewer for the detailed evaluation and for highlighting the importance of novelty in our manuscript. We partially agree with the reviewer’s assessment that the methodology may not stand out as entirely novel compared to some prior studies. However, from our perspective, the study makes significant contributions that add substantial value to the existing literature. Specifically:
- Region-Specific Application in a Data-Scarce Area: Our study is the first to apply an ensemble of PSO- and GA-optimized machine learning models (SVM, RF, XGB) to flash flood susceptibility mapping in a mountainous region of northern Vietnam (Song Ma district). This area, characterized by complex topography and limited flood data, is underrepresented in the literature compared to other regions. By leveraging 412 flood inventory points and 14 tailored predictors (e.g., high-resolution DEM, local rainfall, and soil type), our methodology addresses unique regional challenges, enhancing its practical applicability for disaster management.
- Comprehensive Model Ensemble and Comparative Analysis: We evaluated nine models—three standalone ML models (SVM, RF, XGB), three PSO-optimized models, and three GA-optimized models—demonstrating superior performance (accuracy: 0.927–0.939; AUC-ROC: 0.957–0.977) compared to benchmarks in similar studies.
- Tailored Predictor Integration: The integration of 14 predictors, including region-specific geospatial and climatic factors, combined with a robust flood inventory dataset, represents a tailored adaptation to the study area’s unique characteristics, enhancing model transferability to similar regions.
To address the reviewer’s concern, we have revised the manuscript to better highlight these contributions:
- Expanded the literature review to clarify the research gap and compare our approach with prior studies.
- Added a subsection in the discussion emphasizing the region-specific application and practical significance.
We believe these revisions strengthen the manuscript’s case for its contributions and thank the reviewer for their feedback, which has helped us improve the clarity of our work.
Citation: https://doi.org/10.5194/nhess-2024-215-AC1
-
CC1: 'Comment on nhess-2024-215', Yen-Yi Wu, 25 Jun 2025
This paper showcased how an advanced machine learning approach such as PSO-XGM could help with mapping flood vulnerability. The hybrid approach has become more popular so it is great to see this work show supporting evidence of its robustness.Â
The authors utilized flood inventory points to identify where there are floods. However, there is a spatial scope in floods. One flash flood might involve a large region, while another flood may only swamp a small area. When they randomly picked up "non-flood" locations, how to make sure that the points did not fall into the spatial extent of a flood point? And extend from this question:Â what are the uncertainties and errors this approach may bring in?
Â
Citation: https://doi.org/10.5194/nhess-2024-215-CC1 -
AC2: 'Reply on CC1', Yuei-An Liou, 17 Aug 2025
We sincerely thank the CC1for the positive feedback on the robustness of our hybrid machine learning approach (PSO-XGB) and its contribution to flash flood vulnerability mapping. We appreciate the questions regarding the selection of non-flood locations and the associated uncertainties which have helped us improve the clarity of our methodology.
To address the concern about ensuring that non-flood locations do not fall within the spatial extent of flood points, we clarify that the selection of the 412 non-flood locations was not purely random but involved a careful and systematic process to ensure their validity:
- Historical Data Validation: The non-flood locations were selected from areas within the Song Ma district that have no recorded history of flash floods from 2001 to 2018, based on comprehensive data from the Vietnam Disaster Management Authority (VDMA) and field surveys conducted by the Vietnam Academy for Water Resources (VAWR). These sources provided reliable historical records, ensuring that non-flood points were chosen from regions unaffected by flash floods during this period.
- Geospatial and Topographic Analysis: To further ensure that non-flood points were outside the spatial extent of flood-prone areas, we conducted a detailed geospatial analysis. The selection process incorporated topographic factors such as elevation, slope, and land cover (e.g., non-flood points were prioritized in areas with high elevation, steep slopes, or land cover types less susceptible to flooding, such as dense forest). This analysis was performed using high-resolution Digital Elevation Models (DEMs) and land use/land cover (LULC) data specific to the study area. By cross-referencing these factors with flood inventory data, we minimized the risk of selecting non-flood points within the spatial extent of flood events.
To enhance the clarity of our methodology and address the concerns, we have revised the manuscript as follows:
Revised Section 2.2: We added a detailed explanation of the non-flood point selection process, including the use of historical data, topographic analysis, and buffer zone application to avoid spatial overlap with flood points.
We believe these revisions address the reviewer’s concerns and strengthen the manuscript’s methodological rigor. We are grateful for the reviewer’s insightful comments, which have significantly improved the clarity and transparency of our work.
Citation: https://doi.org/10.5194/nhess-2024-215-AC2
-
AC2: 'Reply on CC1', Yuei-An Liou, 17 Aug 2025
-
RC2: 'Comment on nhess-2024-215', Anonymous Referee #2, 14 Jul 2025
The paper presents a systematic comparison between machine learning models (SVM, RF, and XGBoost) and their optimized counterparts using Particle Swarm Optimization (PSO) and Genetic Algorithms (GA). The authors aim to identify the most effective hybrid configurations for producing high-resolution flash flood susceptibility maps in the Song Ma district of northern Vietnam.
The case study is well documented, and the authors utilize 14 environmental and topographic factors. However, the analysis does not consider any temporal or seasonal variability, which could limit the temporal validity and generalizability of the results over time.
A fundamental shortcoming of the article lies in its lack of reproducibility. The authors do not disclose the specific hyperparameters optimized for each machine learning algorithm, nor the search ranges explored during the optimization process. Most notably, there is no indication of which hyperparameters were selected for tuning in the first place—this omission severely limits the ability of other researchers to replicate or validate the results.
Furthermore, the study would have greatly benefited from a comparative analysis between metaheuristic optimization methods (PSO and GA) and a more conventional technique such as Grid Search, which remains a widely adopted and interpretable approach for hyperparameter optimization in machine learning.
An additional weakness is the absence of any investigation into the sensitivity of model performance to different hyperparameter configurations. Understanding how model accuracy and generalization vary with different hyperparameter settings is crucial, especially when dealing with high-dimensional or non-linear feature spaces.
Crucially, the article also fails to address the issue of overfitting, which is a well-known risk in hyperparameter optimization. Without safeguards such as regularization, cross-validation, or independent validation sets, optimized models may become over-specialized to the training data, reducing their real-world applicability.
Lastly, there is no information provided on the data shuffling or sampling strategy during model training. It remains unclear whether randomization techniques (e.g., data shuffling, stratified sampling, or k-fold cross-validation) were employed to ensure representative and unbiased training/testing splits. This further weakens the methodological transparency and robustness of the study’s claims.
Considering all these aspects, my suggestion is to not recommend the paper for publication in its current form.
Citation: https://doi.org/10.5194/nhess-2024-215-RC2 -
AC3: 'Reply on RC2', Yuei-An Liou, 17 Aug 2025
We sincerely thank RC2 for the detailed and constructive feedback, as well as for recognizing the systematic comparison of machine learning models and the well-documented case study in the Song Ma district. Below, we address each concern raised and outline the revisions made to address them.
- Temporal and Seasonal Variability. We appreciate the referee’s concern regarding temporal and seasonal variability and its impact on the generalizability of our results. In the original manuscript (Section 2.3), we noted that the rainfall dataset, collected from 11 rain gauge stations spanning 53 years (1970–2022), was used to generate a spatial map of rainfall distribution for the peak monsoon months (June, July, and August), which are the primary drivers of flash floods in the Song Ma district. This approach implicitly captures seasonal variability, as the rainfall data reflects the region’s monsoon-driven climate. The flood inventory data (412 points from 2001–2018) further include events across multiple seasons, ensuring a degree of temporal coverage.
- Reproducibility and Hyperparameter Disclosure: We acknowledge the referee’s concern about the lack of transparency in hyperparameter optimization, which is critical for reproducibility. In the original manuscript, we described the use of PSO and GA for hyperparameter tuning but omitted specific details. To address this, we have revised the manuscript by adding a table listing the hyperparameters optimized for each model (SVM: C, gamma; RF: number of trees, max depth, max features; XGBoost: learning rate, max depth, n_estimators) and their search ranges. These hyperparameters were selected based on their established impact on model performance in flood susceptibility studies.
- Comparison with Conventional Optimization (Grid Search): We appreciate the referee’s suggestion to compare PSO and GA with a conventional method like Grid Search. In our study, we conducted a comprehensive comparison of nine models—three standalone ML models (SVM, RF, XGBoost), three PSO-optimized models, and three GA-optimized models—to identify the most effective configurations. This extensive evaluation demonstrated the superior performance of PSO- and GA-optimized models compared to standalone models. While Grid Search is a widely adopted approach, we chose PSO and GA due to their computational efficiency in navigating high-dimensional hyperparameter spaces, which is particularly critical for complex models like XGBoost with multiple interacting parameters. Grid Search, while interpretable, is computationally intensive and less feasible for our study given the large search spaces and resource constraints. To address the reviewer’s concern, we have revised the manuscript in the Discussion section: We included a paragraph justifying the use of PSO and GA, emphasizing their suitability for our study and acknowledging that future work could explore Grid Search for smaller-scale comparisons.
- Sensitivity Analysis of Hyperparameter Configurations: We recognize the importance of understanding how model performance varies with different hyperparameter settings, especially in high-dimensional feature spaces. While we did not conduct a standalone sensitivity analysis, the PSO and GA optimization processes inherently explored a wide range of hyperparameter configurations to identify optimal settings, as evidenced by the high and stable performance of the optimized models (AUC-ROC: 0.957–0.977) on the independent test set (30%, 248 points). This suggests robustness to variations in hyperparameter settings. To address the referee’s concern, we have revised the manuscript by adding a discussion noting that the high test set performance across all optimized models indicates stability against hyperparameter variations, and we acknowledged that future work could include explicit sensitivity analyses for further validation.
- Addressing Overfitting/ Data Shuffling and Sampling Strategy: We thank the referee for raising the concern about overfitting, data shuffling and sampling strategy. As described in Section 2.2, we used a 70:30 train-test split (576 training, 248 testing points) with stratified sampling to maintain an equal proportion of flood and non-flood points. The test set was kept entirely separate to serve as an independent validation set, and the high test set performance (AUC-ROC: 0.957–0.977) suggests minimal overfitting. Besides, the dataset of 824 points (412 flood, 412 non-flood), carefully selected based on historical data (2001–2018) and topographic analysis (elevation, slope, land cover), is sufficiently large and representative for flash flood susceptibility mapping in the Song Ma district. This is comparable to or larger than datasets used in similar studies. The balanced dataset and rigorous train-test split further reduce the risk of overfitting.
We believe these revisions address the referee’s concerns while enhancing the manuscript’s transparency and robustness. We are deeply grateful for the insightful feedback, which has significantly improved the clarity and quality of our work. We kindly request the opportunity to revise and resubmit the manuscript to incorporate these changes.
Citation: https://doi.org/10.5194/nhess-2024-215-AC3
-
AC3: 'Reply on RC2', Yuei-An Liou, 17 Aug 2025
Status: closed
-
RC1: 'Comment on nhess-2024-215', Anonymous Referee #1, 30 May 2025
After a detailed review of the manuscript titled "Elevating Flash Flood Prediction Accuracy: A Synergistic Approach with PSO and GA Optimization," I recommend rejection due to insufficient novelty in the methodology.
The manuscript presents a study on flash flood susceptibility mapping in the Song Ma district, northern Vietnam, employing machine learning models (Support Vector Machines, Random Forests, and Extreme Gradient Boosting) optimized with metaheuristic algorithms (Particle Swarm Optimization and Genetic Algorithms). While the work is well-executed and clearly presented, the approach of integrating PSO and GA with machine learning for flood mapping is not new. This methodology is well-documented in existing literature, and the manuscript does not offer significant innovations to distinguish it from prior studies.
Key Reasons for Rejection:
- Lack of Novelty: The combination of PSO and GA with machine learning models for flood susceptibility mapping is an established technique, not a novel contribution. There are several paper flood mapping using hybrid AI and metaheuristic algorithms such as:
- Bui, D. T., Ngo, P. T. T., Pham, T. D., Jaafari, A., Minh, N. Q., Hoa, P. V., & Samui, P. (2019). A novel hybrid approach based on a swarm intelligence optimized extreme learning machine for flash flood susceptibility mapping. Catena, 179, 184-196.
- Plataridis, K., & Mallios, Z. (2023). Flood susceptibility mapping using hybrid models optimized with Artificial Bee Colony. Journal of Hydrology, 624, 129961.
- Rezaie, F., Panahi, M., Bateni, S. M., Jun, C., Neale, C. M., & Lee, S. (2022). Novel hybrid models by coupling support vector regression (SVR) with meta-heuristic algorithms (WOA and GWO) for flood susceptibility mapping. Natural Hazards, 114(2), 1247-1283.
- Nguyen, H.D. GIS-based hybrid machine learning for flood susceptibility prediction in the Nhat Le–Kien Giang watershed, Vietnam. Earth Sci Inform 15, 2369–2386 (2022). https://doi.org/10.1007/s12145-022-00825-4
- Dodangeh, E., Panahi, M., Rezaie, F., Lee, S., Bui, D. T., Lee, C. W., & Pradhan, B. (2020). Novel hybrid intelligence models for flood-susceptibility prediction: Meta optimization of the GMDH and SVR models with the genetic algorithm and harmony search. Journal of Hydrology, 590, 125423.
- Ngo, P. T. T., Pham, T. D., Hoang, N. D., Tran, D. A., Amiri, M., Le, T. T., ... & Bui, D. T. (2021). A new hybrid equilibrium optimized SysFor based geospatial data mining for tropical storm-induced flash flood susceptible mapping. Journal of Environmental Management, 280, 111858.
- Kaya, C. M., & Derin, L. (2023). Parameters and methods used in flood susceptibility mapping: a review. Journal of Water and Climate Change, 14(6), 1935-1960.
- Nguyen, H. D. (2022). Flood susceptibility assessment using hybrid machine learning and remote sensing in Quang Tri province, Vietnam. Transactions in GIS, 26(7), 2776-2801.
- Established Literature: Several studies have already explored similar methodologies, including:
- "Flood Mapping with PSO-GA": Utilizes PSO and GA with Support Vector Machines for flood mapping.
- "Metaheuristic Flood Assessment": Applies GA and other metaheuristics with ANFIS for flood zoning.
- "Remote Sensing Flood Mapping": Combines PSO, GA, and Harmony Search with machine learning for flood susceptibility.
These examples highlight that the manuscript’s approach aligns with a well-trodden path in flood prediction research. Although the application to a specific region is detailed, it does not advance the methodological framework beyond what is already known.
Additional Notes:
The manuscript is well-written, with a thorough methodology and robust analysis. However, these qualities do not compensate for the lack of originality, which is a critical factor for publication.
In summary, despite its technical competence, the manuscript does not meet the threshold of novelty required for acceptance. Therefore, I recommend rejection.
Â
Citation: https://doi.org/10.5194/nhess-2024-215-RC1 -
AC1: 'Reply on RC1', Yuei-An Liou, 17 Aug 2025
We sincerely thank the reviewer for the detailed evaluation and for highlighting the importance of novelty in our manuscript. We partially agree with the reviewer’s assessment that the methodology may not stand out as entirely novel compared to some prior studies. However, from our perspective, the study makes significant contributions that add substantial value to the existing literature. Specifically:
- Region-Specific Application in a Data-Scarce Area: Our study is the first to apply an ensemble of PSO- and GA-optimized machine learning models (SVM, RF, XGB) to flash flood susceptibility mapping in a mountainous region of northern Vietnam (Song Ma district). This area, characterized by complex topography and limited flood data, is underrepresented in the literature compared to other regions. By leveraging 412 flood inventory points and 14 tailored predictors (e.g., high-resolution DEM, local rainfall, and soil type), our methodology addresses unique regional challenges, enhancing its practical applicability for disaster management.
- Comprehensive Model Ensemble and Comparative Analysis: We evaluated nine models—three standalone ML models (SVM, RF, XGB), three PSO-optimized models, and three GA-optimized models—demonstrating superior performance (accuracy: 0.927–0.939; AUC-ROC: 0.957–0.977) compared to benchmarks in similar studies.
- Tailored Predictor Integration: The integration of 14 predictors, including region-specific geospatial and climatic factors, combined with a robust flood inventory dataset, represents a tailored adaptation to the study area’s unique characteristics, enhancing model transferability to similar regions.
To address the reviewer’s concern, we have revised the manuscript to better highlight these contributions:
- Expanded the literature review to clarify the research gap and compare our approach with prior studies.
- Added a subsection in the discussion emphasizing the region-specific application and practical significance.
We believe these revisions strengthen the manuscript’s case for its contributions and thank the reviewer for their feedback, which has helped us improve the clarity of our work.
Citation: https://doi.org/10.5194/nhess-2024-215-AC1
-
CC1: 'Comment on nhess-2024-215', Yen-Yi Wu, 25 Jun 2025
This paper showcased how an advanced machine learning approach such as PSO-XGM could help with mapping flood vulnerability. The hybrid approach has become more popular so it is great to see this work show supporting evidence of its robustness.Â
The authors utilized flood inventory points to identify where there are floods. However, there is a spatial scope in floods. One flash flood might involve a large region, while another flood may only swamp a small area. When they randomly picked up "non-flood" locations, how to make sure that the points did not fall into the spatial extent of a flood point? And extend from this question:Â what are the uncertainties and errors this approach may bring in?
Â
Citation: https://doi.org/10.5194/nhess-2024-215-CC1 -
AC2: 'Reply on CC1', Yuei-An Liou, 17 Aug 2025
We sincerely thank the CC1for the positive feedback on the robustness of our hybrid machine learning approach (PSO-XGB) and its contribution to flash flood vulnerability mapping. We appreciate the questions regarding the selection of non-flood locations and the associated uncertainties which have helped us improve the clarity of our methodology.
To address the concern about ensuring that non-flood locations do not fall within the spatial extent of flood points, we clarify that the selection of the 412 non-flood locations was not purely random but involved a careful and systematic process to ensure their validity:
- Historical Data Validation: The non-flood locations were selected from areas within the Song Ma district that have no recorded history of flash floods from 2001 to 2018, based on comprehensive data from the Vietnam Disaster Management Authority (VDMA) and field surveys conducted by the Vietnam Academy for Water Resources (VAWR). These sources provided reliable historical records, ensuring that non-flood points were chosen from regions unaffected by flash floods during this period.
- Geospatial and Topographic Analysis: To further ensure that non-flood points were outside the spatial extent of flood-prone areas, we conducted a detailed geospatial analysis. The selection process incorporated topographic factors such as elevation, slope, and land cover (e.g., non-flood points were prioritized in areas with high elevation, steep slopes, or land cover types less susceptible to flooding, such as dense forest). This analysis was performed using high-resolution Digital Elevation Models (DEMs) and land use/land cover (LULC) data specific to the study area. By cross-referencing these factors with flood inventory data, we minimized the risk of selecting non-flood points within the spatial extent of flood events.
To enhance the clarity of our methodology and address the concerns, we have revised the manuscript as follows:
Revised Section 2.2: We added a detailed explanation of the non-flood point selection process, including the use of historical data, topographic analysis, and buffer zone application to avoid spatial overlap with flood points.
We believe these revisions address the reviewer’s concerns and strengthen the manuscript’s methodological rigor. We are grateful for the reviewer’s insightful comments, which have significantly improved the clarity and transparency of our work.
Citation: https://doi.org/10.5194/nhess-2024-215-AC2
-
AC2: 'Reply on CC1', Yuei-An Liou, 17 Aug 2025
-
RC2: 'Comment on nhess-2024-215', Anonymous Referee #2, 14 Jul 2025
The paper presents a systematic comparison between machine learning models (SVM, RF, and XGBoost) and their optimized counterparts using Particle Swarm Optimization (PSO) and Genetic Algorithms (GA). The authors aim to identify the most effective hybrid configurations for producing high-resolution flash flood susceptibility maps in the Song Ma district of northern Vietnam.
The case study is well documented, and the authors utilize 14 environmental and topographic factors. However, the analysis does not consider any temporal or seasonal variability, which could limit the temporal validity and generalizability of the results over time.
A fundamental shortcoming of the article lies in its lack of reproducibility. The authors do not disclose the specific hyperparameters optimized for each machine learning algorithm, nor the search ranges explored during the optimization process. Most notably, there is no indication of which hyperparameters were selected for tuning in the first place—this omission severely limits the ability of other researchers to replicate or validate the results.
Furthermore, the study would have greatly benefited from a comparative analysis between metaheuristic optimization methods (PSO and GA) and a more conventional technique such as Grid Search, which remains a widely adopted and interpretable approach for hyperparameter optimization in machine learning.
An additional weakness is the absence of any investigation into the sensitivity of model performance to different hyperparameter configurations. Understanding how model accuracy and generalization vary with different hyperparameter settings is crucial, especially when dealing with high-dimensional or non-linear feature spaces.
Crucially, the article also fails to address the issue of overfitting, which is a well-known risk in hyperparameter optimization. Without safeguards such as regularization, cross-validation, or independent validation sets, optimized models may become over-specialized to the training data, reducing their real-world applicability.
Lastly, there is no information provided on the data shuffling or sampling strategy during model training. It remains unclear whether randomization techniques (e.g., data shuffling, stratified sampling, or k-fold cross-validation) were employed to ensure representative and unbiased training/testing splits. This further weakens the methodological transparency and robustness of the study’s claims.
Considering all these aspects, my suggestion is to not recommend the paper for publication in its current form.
Citation: https://doi.org/10.5194/nhess-2024-215-RC2 -
AC3: 'Reply on RC2', Yuei-An Liou, 17 Aug 2025
We sincerely thank RC2 for the detailed and constructive feedback, as well as for recognizing the systematic comparison of machine learning models and the well-documented case study in the Song Ma district. Below, we address each concern raised and outline the revisions made to address them.
- Temporal and Seasonal Variability. We appreciate the referee’s concern regarding temporal and seasonal variability and its impact on the generalizability of our results. In the original manuscript (Section 2.3), we noted that the rainfall dataset, collected from 11 rain gauge stations spanning 53 years (1970–2022), was used to generate a spatial map of rainfall distribution for the peak monsoon months (June, July, and August), which are the primary drivers of flash floods in the Song Ma district. This approach implicitly captures seasonal variability, as the rainfall data reflects the region’s monsoon-driven climate. The flood inventory data (412 points from 2001–2018) further include events across multiple seasons, ensuring a degree of temporal coverage.
- Reproducibility and Hyperparameter Disclosure: We acknowledge the referee’s concern about the lack of transparency in hyperparameter optimization, which is critical for reproducibility. In the original manuscript, we described the use of PSO and GA for hyperparameter tuning but omitted specific details. To address this, we have revised the manuscript by adding a table listing the hyperparameters optimized for each model (SVM: C, gamma; RF: number of trees, max depth, max features; XGBoost: learning rate, max depth, n_estimators) and their search ranges. These hyperparameters were selected based on their established impact on model performance in flood susceptibility studies.
- Comparison with Conventional Optimization (Grid Search): We appreciate the referee’s suggestion to compare PSO and GA with a conventional method like Grid Search. In our study, we conducted a comprehensive comparison of nine models—three standalone ML models (SVM, RF, XGBoost), three PSO-optimized models, and three GA-optimized models—to identify the most effective configurations. This extensive evaluation demonstrated the superior performance of PSO- and GA-optimized models compared to standalone models. While Grid Search is a widely adopted approach, we chose PSO and GA due to their computational efficiency in navigating high-dimensional hyperparameter spaces, which is particularly critical for complex models like XGBoost with multiple interacting parameters. Grid Search, while interpretable, is computationally intensive and less feasible for our study given the large search spaces and resource constraints. To address the reviewer’s concern, we have revised the manuscript in the Discussion section: We included a paragraph justifying the use of PSO and GA, emphasizing their suitability for our study and acknowledging that future work could explore Grid Search for smaller-scale comparisons.
- Sensitivity Analysis of Hyperparameter Configurations: We recognize the importance of understanding how model performance varies with different hyperparameter settings, especially in high-dimensional feature spaces. While we did not conduct a standalone sensitivity analysis, the PSO and GA optimization processes inherently explored a wide range of hyperparameter configurations to identify optimal settings, as evidenced by the high and stable performance of the optimized models (AUC-ROC: 0.957–0.977) on the independent test set (30%, 248 points). This suggests robustness to variations in hyperparameter settings. To address the referee’s concern, we have revised the manuscript by adding a discussion noting that the high test set performance across all optimized models indicates stability against hyperparameter variations, and we acknowledged that future work could include explicit sensitivity analyses for further validation.
- Addressing Overfitting/ Data Shuffling and Sampling Strategy: We thank the referee for raising the concern about overfitting, data shuffling and sampling strategy. As described in Section 2.2, we used a 70:30 train-test split (576 training, 248 testing points) with stratified sampling to maintain an equal proportion of flood and non-flood points. The test set was kept entirely separate to serve as an independent validation set, and the high test set performance (AUC-ROC: 0.957–0.977) suggests minimal overfitting. Besides, the dataset of 824 points (412 flood, 412 non-flood), carefully selected based on historical data (2001–2018) and topographic analysis (elevation, slope, land cover), is sufficiently large and representative for flash flood susceptibility mapping in the Song Ma district. This is comparable to or larger than datasets used in similar studies. The balanced dataset and rigorous train-test split further reduce the risk of overfitting.
We believe these revisions address the referee’s concerns while enhancing the manuscript’s transparency and robustness. We are deeply grateful for the insightful feedback, which has significantly improved the clarity and quality of our work. We kindly request the opportunity to revise and resubmit the manuscript to incorporate these changes.
Citation: https://doi.org/10.5194/nhess-2024-215-AC3
-
AC3: 'Reply on RC2', Yuei-An Liou, 17 Aug 2025
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
355 | 118 | 21 | 494 | 29 | 40 |
- HTML: 355
- PDF: 118
- XML: 21
- Total: 494
- BibTeX: 29
- EndNote: 40
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1