the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Flood damage functions for rice: Synthesizing evidence and building data-driven models
Abstract. Floods are a major cause of agricultural losses, yet flood damage models for crops are scarce, often lack validation, uncertainty estimates, and assessments of their performance in new regions. This study introduces CROPDAM-X, a framework for developing and evaluating flood damage models for crops, and applies it to rice. We compile and review 20 damage models from 12 countries, identifying key gaps and limitations. Using empirical survey data from Thailand and Myanmar, we develop a suite of models, including deterministic and probabilistic stage-damage functions, Bayesian regression, and Random Forest, based on key flood characteristics like water depth, duration, and plant growth stage. We assess predictive performance through cross-validation and test how well models trained in one region perform when applied to another. Our results show that model performance depends on complexity and context: Random Forest achieves the highest accuracy, while simpler models offer ease of use in data-scarce settings. The results also demonstrate the potential errors introduced by transferring models spatially, highlighting the need for diverse training data or local calibration. We present the most comprehensive review of flood damage models for rice to date and provide practical guidance on model selection and expected errors when transferring models across regions.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Natural Hazards and Earth System Sciences.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(1866 KB) - Metadata XML
-
Supplement
(3358 KB) - BibTeX
- EndNote
Status: open (until 30 Oct 2025)
- RC1: 'Comment on egusphere-2025-3706', Anonymous Referee #1, 08 Oct 2025 reply
-
RC2: 'Comment on egusphere-2025-3706', Anonymous Referee #2, 09 Oct 2025
reply
This study addresses the issue regarding the impact of floods on rice crops and presents the contribution through model evaluation. However, there are some areas that could be improved for clarity and conciseness, which would enhance the manuscript's overall quality.
Abstract:
Lines 11-12: The term "framework" may not be the most suitable here. The focus of the study seems more on the evaluation and comparison of models rather than the development of a framework. Perhaps rephrasing this part to "model evaluation" or a similar phrase would better capture the essence of the study.
Introduction:
The introduction is quite detailed and comprehensive, but it might benefit from being more concise and directly focused on the specific research gap this study addresses.
Since the study combines empirical data with machine learning techniques, it would be helpful to emphasize how this combination addresses gaps in existing research and highlight the specific problems this study aims to solve.
Table Clarifications:
Table 1: The distinction between "data collection period" and "flood event covered" is not entirely clear. It would be helpful to clarify the difference between these two categories for better reader understanding.
Table 2: The definition of "flood duration" (1-100 days) could be explained in more detail. Is this range based on specific criteria or events? A brief clarification would be beneficial.
Model Development:
This section would benefit from a more detailed explanation of the model development process. Adding a flowchart or a clearer step-by-step description could help readers follow the methodology more easily.
It seems that three models—regression, Bayesian regression, and Random Forest—are being compared. A bit more detail on each model’s methodology and how they were applied to the data would improve understanding.
Table 3: It might be helpful to briefly explain how the Univariable Bayesian Regression, Multivariable Bayesian Regression, and Random Forest models calculate relative yield loss, rather than placing all the details in the supplementary materials. This would improve the readability and flow of the manuscript.
Citation: https://doi.org/10.5194/egusphere-2025-3706-RC2 -
RC3: 'Comment on egusphere-2025-3706', Anonymous Referee #3, 09 Oct 2025
reply
This study introduced a framework named CROPDAM-X for developing and evaluating flood damage functions for crops and applied this framework to rice yield loss estimates in Thailand and Myanmar. This framework also included comprehensive review of the state-of-art flood damage models for rice and provided practical guidance for further applications. Results showed that data-driven models like Random Forest achieved the highest accuracy, while challenges remained when these models were transferred to different areas. Overall, this study is helpful for flood damage estimates in the agriculture sector. However, I still have several concerns and suggestions for improving the current work.
1) Random Forest is just one of the commonly used machine learning models. Could you justify why Random Forest rather than the other machine learning models/algorithms was employed in this study?
2) Line 25: Please provide the data source for estimated losses due to extreme events.
3) Lines 109-110: The proposed framework was named as “CROPDAM-X”, in which “DAM” represents “DAMage”. “DAM” may be interpreted as the hydraulic structure, dams, which is a little confusing.
4) Fig. 1: It is suggested to avoid the acronyms, e.g., MAE, MBE, and CRPS, (which are not explained until in the following sections) in the figure.
5) Line 131: Two variables related flood characteristics were used to develop the models. Is it possible to incorporate the output from existing flood models so that more hydraulic variables can be used for the damage model for crops?
6) Table 3: Are there any references or evidence showing that the minimum damageable flood depth is 2cm? What is h_saturation? Why was the root square of water depth instead of the other transformations used in the linear regression equation?
7) Line 188: What kernel functions are used for the density estimation?
8) Section 3.2.2: The performance comparison of between the ramp functions and the proposed models is based on the median or mean of the evaluation metrics? Line 235: What does it mean that the MAE for RF and BRM covered that of the ramp functions? The IQR or whiskers covered 23%?
9) The caption for Fig. 4 may not be correct.
10) Line 289: Could you quantify how skewed the training data is and how did that affect the model performance?
11) Section 3.5: It would be better to place Section 3.5.1 in the first few sections of this manuscript. Also, please note that there is no one model that fits all, so the ensemble model method might be a better option given various uncertainty sources (see the reference below).
Reference: Huang, T., & Merwade, V. (2023). Uncertainty analysis and quantification in flood insurance rate maps using Bayesian model averaging and hierarchical BMA. Journal of Hydrologic Engineering, 28(2), 04022038.
12) Fig. 5 is a stacked bar chart, which makes it difficult to compare the difference in terms of the share of each category.
13) Line 409: Please explain the metrics values are referred to the median or the mean.Citation: https://doi.org/10.5194/egusphere-2025-3706-RC3 -
RC4: 'Comment on egusphere-2025-3706', Anonymous Referee #4, 22 Oct 2025
reply
The authors of the article propose a method for comparing models aimed at estimating variations in rice yields based on survey data following flood events in several case studies. The scientific stakes are high, as rice is a crop that is heavily impacted in Asia. While the structure of the article is clear, further clarification of the stated objectives and real contributions of the article is needed as well as a discussion.1. IntroductionThe authors would benefit from presenting the issues in terms of model performance and the issues in terms of modelling flood-related agricultural damage separately. A literature review on the issues of modelling flood-related agricultural damage already exists (Bremond et al, 2013). Why not focus on the specific modelling of yield variations in rice and target its specific dimensions?In my view, the scientific contribution of the article lies in the comparison of different modelling approaches and their performances. This part needs to be consolidated in the state of the art.In the Research contributions section, the objectives stated are not really those presented in the article:1. an inventory of flood damage models for agriculture → This is an inventory of models for rice2. the article does propose a four-step methodological framework, but it is only applied to rice cultivation.The name of the CROPDAM-X methodological approach seems to me to echo the floodam.agri method (https://floodam.org/floodam.agri/). I encourage the authors to look at the specific features of their approach compared to this existing one.2. MethodologyThe experimental design does not seem clear enough to me to judge the relevance of the analyses carried out subsequently. What data were used for modelling ? On which case studies? How many simulations? It seems to me that a diagram or table should clarify these aspects.Similarly, the presentation of the survey data on which all the performance analyses are based is not detailed enough to understand the relevance of the model transferability: What data was collected in the surveys? Yields? Variation in yield? If so, what was the reference yield for the different case studies? When did the floods occur in the case studies? ...3. Results and discussionIt would be appropriate to present the model outputs for the various case studies before comparing their performance.The results presented must be consistent with the stated objectives. It is unclear whether the objective is to compare performance or to inventory and discuss the performances of the various approaches. This needs to be clarified.A section discussing the results is missing.4. ConclusionsThe conclusion needs to be improved. However, this requires first clarifying the main focus of the article.Citation: https://doi.org/
10.5194/egusphere-2025-3706-RC4
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 1,109 | 48 | 15 | 1,172 | 50 | 15 | 17 |
- HTML: 1,109
- PDF: 48
- XML: 15
- Total: 1,172
- Supplement: 50
- BibTeX: 15
- EndNote: 17
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The study models rice yield loss using different approaches. It uses generalized and localized splits of the data. It probes deterministic, probabilistic, Bayesian regression, and random forest regression types of models. Based on my evaluation, I have serious doubts about the claim that the authors made regarding the results. When you test models making different splits, it is wrong to choose the best-performing model as the best. You need to focus on the models that have been evaluated to reduce bias through independent evaluation. In this case, the authors chose the generalized model as the best, but in my opinion, their selection is biased because it uses all the data. Instead, they should focus on the localized split (CRV). The models called "transferred" are the ones that give an independent evaluation.
I also recommend to the authors to evaluate the models spatially and to use, at least, the locations to investigate how dependent on the locations the models are. This could supply insights into other variables that could be used in the future to improve the models.
Majors.
Methodology
The methodology is briefly explained. It should be extended for a better comprehension of the reader.
Regarding model evaluation, from my perspective, the only useful evaluations are the LOOCV and the CRV. The first lets you assess model stability by analyzing the variability of the 10 folds you receive. If it is shrunk, it is stable; if it is sparse, it is unstable. These results should be discussed. In the second evaluation, you provide an independent and useful analysis by using data from other locations and validating it against unknown locations. This is the real truth of how the model works.
Results
Regarding the CRV, you can also use ten folds; however, I do not understand why only one dot is shown in Fig. 3. The results presented in this figure, in my opinion, do not support your conclusions. The best-performing models, based on the CRPS, are BRM of SDF (prob). You cannot assign equal weight to the distribution of the ten folds for just one point evaluation (it is probably biased). You should run a ten-fold evaluation for all the models and compare them fairly.
Generalized models perform better because of overfitting. Again, the independent evaluation (CRV) tells you how the model could perform in unknown conditions.
Section 3.5.1 seems irrelevant to this study. It is not clear how it supports the results obtained.
Conclusion
The conclusion is excessively detailed; it should provide a concise summary of the key findings. The first paragraph corresponds to the introduction. I have serious doubts about the conclusion regarding which figure the authors are referring to (40,000 ha). extract from the results. It is not strongly supported by the results shown in the manuscript.
Minors
L145: A comma is used to separate thousands. Please clarify which figure you are referring to when mentioning 40,00 ha.
Table 3: What is the value of h saturation? The table only displays the value of h min.
Table 3: Could you please clarify where exactly the supplementary information is?
Fig. S1 should be in the manuscript.
Fig S2 should be in the manuscript. It allows the reader to have a clear view of the variables used in the models.
Table 3: The models should be briefly described in the manuscript rather than just shown in the supplementary.
L173-174: What is the purpose of running a model with LNO?
Fig. 4. These are boxplots, not violin plots.
Fig. 6 is described before Fig. 5.