Review Reports - Integrated Data-Driven Multi-Criteria Analysis and Machine Learning Approaches for Assessment of Flood Susceptibility Mapping

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study aims to perform flood susceptibility mapping. Methods such as Multi-Criteria Analysis and Machine Learning are used for this purpose. The content offers a straightforward approach as it includes steps from the general literature. However, it is weak because it does not present a discussion of similar studies in the literature. The authors are advised to answer the following questions and improve the manuscript.

Figure 1 is composed of 2 parts and on page 5, it is not clear what the vertical axes represent. It is the value of what.
Section 2.2. Which dates of Landsat-8 and Sentinel-2 are used? They should be mentioned in the paper.
Provide a table that shows the resolution and production time, period of each data used as an impact factor. It is not clear when the soil map was produced. There is no information about the LULC map? How is it produced and what is the accuracy?
Figure 3. Aspect should show the directions. What do the numbers in the legend belong to?
Page 19 lines 549-551. References should be provided for this comment, and the similarities should be explained.
The relationship between surface temperature and vegetation should be shown with a scatterplot.
The literature contains many flood susceptibility studies conducted in Pakistan and elsewhere. What are the similarities and differences between this study and others? Were similar data used, and did the methods employed yield similar results? This should be discussed in detail.

Author Response

Reviewer Comments 1

Comments and Suggestions for Authors

Figure 1 is composed of 2 parts, and on page 5, it is not clear what the vertical axes represent. It is the value of what.

Response:

Thank you for this valuable comment. We agree that the vertical axis in the second panel of Figure 1 was not clearly described. In the original figure, the y-axis showed quantities by category, including the number of affected villages, affected population, damaged houses, and affected agricultural area (ha). To avoid ambiguity, we have revised Figure 1 and added Figure 2 separately with its captions to clearly indicate the meaning and units of each variable. The Figure 2 caption and axis labeling have been improved accordingly.

Figure 2. Flood damage statistics for the 2010 and 2022 flood events, including affected villages (number), affected population (people), houses damaged (number), and agricultural area affected (hectare).

Section 2.2. Which dates of Landsat-8 and Sentinel-2 are used? They should be mentioned in the paper.

Response:
Thank you for this suggestion. We agree that the acquisition dates of Landsat-8 and Sentinel-2 imagery should be clearly stated. We have now included the specific image acquisition dates in Section 2.2 of the revised manuscript in Table 1.

Provide a table that shows the resolution and production time, period of each data used as an impact factor. It is not clear when the soil map was produced. There is no information about the LULC map? How is it produced, and what is the accuracy?

Response:

We appreciate this useful suggestion. In response, we have added a new Table 1 summarizing all datasets used as flood conditioning factors, including their spatial resolution, source, production year, and temporal period. We have also clarified the source and production year of the soil map. The reliability of the LULC classification was evaluated using a confusion matrix, and the validation results are presented in the Supplementary Material (Figure S1). Also, we added a detailed description of the LULC preparation method and its accuracy assessment in Section 3.1.2.

Table 1. Different types of datasets are used for FSM.

Data Type	Data Sources	Resolution	Website	Accessed Date
DEM	SRTM	Grid Cell: (30 × 30)	https://earthexplorer.usgs.gov/	2022
LULC	Sentinel-2	Grid Cell: (10 × 10)	https://livingatlas.arcgis.com/landcoverexplorer/	2022
NDVI	Landsat	Grid Cell: (30 × 30)	https://earthexplorer.usgs.gov/	2010 - 2022
LST	Landsat	Grid Cell: (30 × 30)	https://earthexplorer.usgs.gov/	2010 - 2022
Soil	FAO	Grid Cell: (30 × 30)	https://www.fao.org/soils-portal/en/	2022
Stream & river network	Survey of Pakistan	(30 × 30)	http://www.surveyofpakistan.gov.pk/	2022
Climate	Pakistan Meteorological Data (PMD)	Daily basis	https://www.pmd.gov.pk/en/	2022
River flow	WAPDA	Daily basis	https://www.wapda.gov.pk/	2022

Figure 3. Aspect should show the directions. What do the numbers in the legend belong to?

Response:
Thank you for pointing this out. We agree that the aspect is more intuitive when presented in terms of cardinal and intercardinal directions rather than only numeric values. Therefore, we have revised Figure 3 (new Figure 4) so that the aspect classes are expressed as direction categories (e.g., North, Northeast, East, Southeast, South, Southwest, West, Northwest). We have also clarified the legend to explain the meaning of the values.

Figure 4. Factors of flood conditioning that were chosen in the flood susceptibility assessment. Enhanced flood-prone conditions around constructed areas are consistent with a large body of literature that has identified hydrological effects of urbanization, such as decreased infiltration and surface runoff.

Page 19 lines 549-551. References should be provided for this comment, and the similarities should be explained.

Response:

Thank you for this observation. We agree that the statement requires supporting references and a clearer explanation of the similarity. In the revised manuscript, we added comparative studies showing that AHP and FAHP often produce similar overall susceptibility patterns, while FAHP provides better treatment of uncertainty and ambiguity in expert-based weighting. We also clarified that the slightly improved performance of FAHP in our study is consistent with previous work reporting modest but meaningful gains of FAHP over conventional AHP in flood susceptibility applications, especially in spatially complex settings.

The relationship between surface temperature and vegetation should be shown with a scatterplot.

Response:
Thank you for this valuable suggestion. We agree that a scatterplot can better illustrate the relationship between land surface temperature and vegetation. Accordingly, we have added a scatterplot showing the relationship between land surface temperature and normalized difference vegetation index values, along with a brief interpretation of the observed trend in the (4.4. Relationship of LST and NDVI) Discussion section.

The literature contains many flood susceptibility studies conducted in Pakistan and elsewhere. What are the similarities and differences between this study and others? Were similar data used, and did the methods employed yield similar results? This should be discussed in detail.

Response:
We thank the reviewer for this important comment. We agree that the manuscript needed a more comprehensive comparison with prior flood-susceptibility studies. In the revised manuscript, we have significantly expanded the Discussion section to compare our study with relevant work conducted in Pakistan and other regions. Specifically, we discussed similarities and differences in:

The selection of flood conditioning factors
The datasets used
The applied methods
The resulting flood susceptibility patterns and model performance.

We also highlighted the unique contributions of the present study, particularly in the study area, the integrated use of Multi-Criteria Analysis and Machine Learning approaches, and the interpretation of factor importance.

Reviewer 2 Report

Comments and Suggestions for Authors

None

Comments for author File: Comments.pdf

Author Response

Review Comments 2

This study conducts a flood susceptibility assessment for the Mohmand Dam Catchment in the Swat River Basin, Pakistan, by integrating two multi-criteria decision analysis methods (AHP and FAHP) and five machine learning models (Random Forest, Support Vector Machine, Logistic Regression, K-Nearest Neighbors, and Multi-Layer Perceptron) based on fourteen topographical, hydrological, environmental, meteorological, and anthropogenic factors. Remote sensing data were extracted using Google Earth Engine, key influencing factors were identified through SHAP analysis, and the predictive performance of different models was compared. The study provides valuable references for regional flood management and is recommended for publication after the following details are clarified:

In Figure 3, the color differentiation for some subfigures is low; different soil types have similar colors in the legend, making them difficult to distinguish.

Response:
Thank you for this helpful observation. We agree that the color contrast in the soil map was not sufficiently clear in the original version. To improve readability and visual interpretation, we revised Figure 3 (new Figure 4) by using a more distinguishable color scheme for the soil classes and by improving the legend presentation. This revision makes it easier to differentiate between adjacent soil categories.

All datasets were resampled to a shared spatial resolution during the preprocessing stage; the impact of this process on the accuracy of the original data could be further clarified.

Response:
Thank you for this important comment. We agree that the implications of resampling should be clarified. In the revised manuscript, we explicitly state that all datasets were resampled to a common spatial resolution to ensure spatial consistency and compatibility in the multi-source analysis. We also added a clarification that resampling may introduce some degree of smoothing or local information loss, particularly for coarser or categorical datasets, but this preprocessing step was necessary to maintain pixel-wise comparability among all conditioning factors. We further note that the shared resolution was selected as a practical compromise between data consistency and preservation of spatial detail.

In the analysis of the spatial distribution of flood-conditioning factors, the analysis for some factors is somewhat one-dimensional.

Response:
Thank you for this valuable suggestion. We agree that the interpretation of some flood-conditioning factors in the previous version was brief. In the revised manuscript, we expanded the discussion of the spatial distribution of these factors by linking them more explicitly to flood-generation processes in the study area. In particular, we improved the interpretation of topographic, hydrological, land-cover, and anthropogenic factors by explaining how their spatial variation may influence runoff concentration, infiltration capacity, drainage efficiency, and exposure to flood hazards.

The study is based on the Mohmand Dam Catchment, with model training and validation relying on historical flood points from this area. The transferability of this method to other geographical or climatic regions could be further discussed in the conclusion.

Response:
Thank you for this important comment. We agree that the transferability of the framework should be discussed more clearly. In the revised Conclusion section, we now explain that the proposed methodology is transferable in structure because it integrates widely used geospatial predictors, multi-criteria decision analysis, and machine learning models. However, we also clarify that direct model transfer to other regions may be limited by differences in topography, hydrology, land use, climate regime, data quality, and flood-generation mechanisms. Therefore, while the workflow itself is adaptable, factor weighting, model calibration, and validation should be region-specific.

The methods employed in this study have seen new developments and applications in recent years; it is recommended that relevant recent literature be supplemented to reflect

Response:
Thank you for this useful suggestion. We agree that the manuscript should better reflect recent developments in flood susceptibility assessment. In the revised manuscript, we added recent literature on the integration of machine learning, multi-criteria decision analysis, remote sensing, and explainable AI approaches such as SHAP. These additions help position the present study within the latest methodological developments and demonstrate how recent research increasingly emphasizes both predictive accuracy and interpretability in flood susceptibility mapping. Recent examples include SHAP-based machine learning frameworks for urban flood susceptibility, integrated AHP–remote sensing–machine learning studies in Pakistan, and broader comparative or explainable modeling approaches for flood susceptibility mapping.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have updated the text with new figures, references, and answers to questions. However, two points given below require clarification.

1. It appears that Dynamic World data was used for the LULC analysis. Information regarding the overall accuracy of this data can be provided. However, the accuracy analysis performed by the authors is unacceptable because the accuracy of a class cannot be tested with only a few pixels; it is not statistically significant.

2. How the points selected for the plots shown in Figure 11. NDVI and LST eventually move in opposite directions (negative correlation) because high NDVI (dense vegetation) leads to high evapotranspiration, which lowers LST. In Figure 4, i(NDVI) and j(LST) show that, but not your plots. Have you used these two images for different years?

Author Response

It appears that Dynamic World data was used for the LULC analysis. Information regarding the overall accuracy of this data can be provided. However, the accuracy analysis performed by the authors is unacceptable because the accuracy of a class cannot be tested with only a few pixels; it is not statistically significant.

Response:

Thank you for this important comment. We agree that the initial accuracy assessment based on only a few pixels per class was not statistically robust. Because the study area is a mountainous catchment, it was not feasible to collect validation samples uniformly from the entire catchment or to obtain a large number of ground reference points. Moreover, the dataset used in this study has been reported as highly accurate in previously published and cited studies, which supports its reliability for the present analysis. In response, we revised the manuscript by removing the limited class-wise validation and by clarifying that the study used Dynamic World as an existing LULC product rather than producing a new classification. We now cite the published validation of Dynamic World and relevant independent evaluation studies, which show that the dataset has reasonable overall performance at a global scale but variable class-wise and regional accuracy, especially in heterogeneous landscapes (Brown et al., 2022; Venter et al., 2022; Xu et al., 2024). We also added a limitation statement noting that a statistically rigorous local accuracy assessment would require a sufficiently large and stratified reference sample and should be addressed in future work.

Brown, C. F., Brumby, S. P., Guzder-Williams, B., Birch, T., Hyde, S. B., Mazzariello, J., Czerwinski, W., Pasquarella, V. J., Haertel, R., Ilyushchenko, S., Schwehr, K., Weisse, M., Stolle, F., Hanson, C., Guinan, O., Moore, R., Tait, A. M., Do, T., Perez-Hoyos, A., … Bholanath, R. (2022). Dynamic World, near real-time global 10 m land use land cover mapping. Scientific Data, 9, 251. https://doi.org/10.1038/s41597-022-01307-4

Venter, Z. S., Sydenham, M. A. K., Meijer, J. R., & others. (2022). Global 10 m land use land cover datasets: A comparison of Dynamic World, WorldCover and Esri Land Cover. Remote Sensing, 14(16), 4101. https://doi.org/10.3390/rs14164101

Xu, P., Tsendbazar, N.-E., Herold, M., de Bruin, S., Koopmans, M., Birch, T., Carter, S., Fritz, S., Lesiv, M., Mazur, E., Pickens, A., Potapov, P., Stolle, F., Tyukavina, A., Van De Kerchove, R., & Zanaga, D. (2024). Comparative validation of recent 10 m-resolution global land cover maps. Remote Sensing of Environment, 315, 114316. https://doi.org/10.1016/j.rse.2024.114316

How the points selected for the plots shown in Figure 11. NDVI and LST eventually move in opposite directions (negative correlation) because high NDVI (dense vegetation) leads to high evapotranspiration, which lowers LST. In Figure 4, i(NDVI) and j(LST) show that, but not your plots. Have you used these two images for different years?

Response:

Thank you for this important observation. In Figure 4i,j shows the spatial distribution of NDVI and LST for a representative year, whereas Figure 11 presents the temporal relationship between monthly mean NDVI and monthly mean LST for different years. Since Figure 4 reflects spatial variability within a single year and Figure 11 reflects temporal variability based on yearly averaged values, the two figures are not directly comparable. The temporal pattern visible in the maps may not necessarily appear in the same way in the annual mean value analysis.