An Explainable Machine Learning Framework for Forecasting Lake Water Equivalent Using Satellite Data: A 20-Year Analysis of the Urmia Lake Basin
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript seems to be interesting, but still lacks in some areas. So, my suggestions for the improvement of the manuscript are:
- In the abstract, highlight specific novel contributions of the framework compared to existing models.
- Explain properly the gaps which was addressed by this study. Also, elaborate on the objectives based on these gaps.
- Discuss why explainable models are important.
- Expand on the limitations of the existing explainable AI frameworks used in hydrological modeling.
- A proper explanation is required on handling missing data and the rationale for lag feature selection.
- Discuss whether cross-validation methods were applied during model evaluation.
- Enhance clarity and readability of figures, such as regression plots and SHAP beeswarm plots.
- Compare your model with other methods to show its strengths.
- Share links to data and code, so others can test the study.
Author Response
Reviewer #1
We sincerely thank Reviewer 1 for the thorough and constructive comments, which helped improve the clarity, scientific rigor, and overall quality of our manuscript. Below, we address each point in detail.
1. Highlight Specific Novel Contributions in the Abstract
Comment: Highlight specific novel contributions of the framework compared to existing models in the Abstract.
Response:
We have revised the Abstract to explicitly highlight the main novel contributions: (i) integration of XGBoost-based feature weighting with ensemble learning; (ii) incorporation of SHAP-based explainability directly into the modeling framework, offering a new approach to transparent groundwater forecasting in arid basins; and (iii) comparison with other methods.
2. Clarify Gaps Addressed and Objectives
Comment: Explain properly the gaps addressed by this study and elaborate on the objectives based on these gaps.
Response:
We have revised the Introduction and Literature Review sections to better explain the key gaps in prior studies, including: (i) lack of predictive frameworks for subsurface groundwater indicators like LWE; (ii) reliance on black-box models with limited transparency; and (iii) insufficient integration of explainable AI. Based on these gaps, we defined clear objectives targeting predictive performance and interpretability.
3. Importance of Explainable Models
Comment: Discuss why explainable models are important.
Response:
We added a dedicated discussion in the Introduction and Literature Review emphasizing that explainable models enhance transparency, trust, and actionable decision-making in critical applications such as groundwater management and climate adaptation planning.
4. Limitations of Existing Explainable AI Frameworks
Comment: Expand on the limitations of the existing explainable AI frameworks used in hydrological modeling.
Response:
We now explicitly discuss that existing XAI applications in hydrology often focus only on post-hoc local interpretations, lack global feature insight, and are rarely integrated into model development pipelines. Our study addresses these limitations by embedding explainability earlier in the model training process.
5. Handling Missing Data and Lag Feature Selection
Comment: Provide proper explanation on handling missing data and the rationale for lag feature selection.
Response:
We expanded the Methods section to explain that linear interpolation was used for handling small missing gaps to preserve temporal continuity. Lag features were selected based on hydrological memory principles and confirmed using preliminary autocorrelation analysis, reflecting typical groundwater response times in semi-arid regions. (Line 717)
6. Application of Cross-Validation
Comment: Clarify cross-validation approach.
Response:
We clarified that both Ridge Regression and Random Forest models employed 5-fold cross-validation during hyperparameter tuning and model selection to enhance generalizability and avoid overfitting. (Line 622)
7. Enhance Clarity and Readability of Figures
Comment: Enhance clarity and readability of figures such as regression plots and SHAP plots.
Response:
All figures have been improved with clearer color palettes, appropriate labeling, consistent font sizes, and enhanced layout adjustments for better readability. Specific improvements were made to the regression plot and the SHAP beeswarm plot.
8. Compare Model Strengths Against Other Methods
Comment: Compare your model with other methods to show its strengths.
Response:
A new comparison section and a professional comparison table have been added. Our proposed Random Forest (Weighted + Cross-Validated) model outperformed traditional baselines such as Persistence, Ridge Regression without weighting, simple Random Forest, and standalone XGBoost models.
9. Share Data and Code for Reproducibility
Comment: Share links to data and code, so others can test the study.
Response:
We have uploaded the full dataset and all codes used in this study to GitHub. A new subsection, "Data and Code Availability," has been added, and the repository link is included: https://github.com/habibisara/Urmia-Lake-Climate-Study
Once again, we sincerely appreciate Reviewer 1's valuable feedback and insightful suggestions, which have significantly contributed to enhancing the scientific rigor and overall quality of the manuscript.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript provides a comprehensive assessment of groundwater storage variability in the Urmia Lake Basin using a combination of satellite-based remote sensing, statistical trend analysis, and machine learning modeling. Correlation and regression analyses highlighted sea level pressure and soil moisture as potential climatic influencers of LWE. The findings of the present work highlight the dynamic relationship between climatic factors and groundwater storage in the Urmia Lake Basin. The results provide valuable insights for policymakers to develop climate-adaptive water management strategies to ensure sustainability in arid and semi-arid regions. Here are some comments which may be useful to improve the present work.
(1) Lines 116, 118, 128, 158, 179, [x] → Someone et al. [x]
(2) The authors claim that they develop an explainable machine learning framework for forecasting Lake Water Equivalent (LWE) in the Urmia Lake Basin using a combination of satellite-derived hydrological indicators and climate variables spanning the period from 2003 to 2023. However, the reviewer did not recognize why the machine learning framework is explainable? It is recommended to expand the discussion on how SHAP contributes to interpretability and in what way this framework enhances transparency compared to traditional ML approaches.
(3) Line 214: "As show in Figure 1 It" → Should be corrected to: "As shown in Figure 1, it"
(4) Line 290: "This data was also sourced from [?]" → Please provide the actual data source or citation.
(5) Line 350: The reviewer is confused by how the authors determine the "strongest significant correlation" with LWE. Please clarify the method used to assess significance.
(6) Lines 356–358: The reviewer recommends adding some explanation of the inter-variable relationships.
(7) Table 4: No significant trend in groundwater storage is found for 2003–2013, while a significant downward trend is observed for 2014–2023. However, the entire period (2003–2023) shows a significant upward trend. Please explain why the overall trend appears positive despite the more recent strong decline.
(8) Section 4.5.1: The coefficient of determination (R²) in multiple regression analysis is smaller than 0.4. Is it meaningful to draw conclusions based on this weak model? Could the authors justify the inclusion of this analysis despite low explanatory power?
(9)The conclusion section seems not to summarize the contributions from the machine learning framework in sufficient detail.
(10)Throughout the manuscript: “Lake Water Equivalent” is inconsistently hyphenated as "LakeWater Equivalent" or "Lake Water Equivalent" and abbreviated as both “LWE” and “LWE.” Please unify terminology and formatting.
Author Response
Response to Reviewer 2
We sincerely thank Reviewer 2 for the thoughtful and constructive comments that helped us to improve the clarity, precision, and scientific contribution of our manuscript. Below, we provide detailed point-by-point responses to each comment.
1. Correct Citation Style in Lines 116, 118, 128, 158, 179
Comment: Lines 116, 118, 128, 158, 179: [x] → Someone et al. [x].
Response:
We thank the reviewer for identifying this oversight. We have carefully corrected all citation formatting issues in these lines, ensuring that all references now follow the format "Someone et al. [x]" throughout the manuscript.
2. Expand Explanation of Explainability via SHAP
Comment: Expand the discussion on how SHAP contributes to interpretability and explain how the framework enhances transparency compared to traditional ML approaches.
Response:
We appreciate this important comment. In the revised manuscript, we have expanded the explanation of explainability in Section 4.6. We clearly describe how SHAP (SHapley Additive exPlanations) decomposes model predictions into additive contributions from each input feature based on cooperative game theory. We highlight that SHAP offers both global and local interpretability, allowing transparent understanding of how climatic variables (e.g., temperature, soil moisture, precipitation) influence groundwater storage forecasts. We also emphasize that, unlike traditional black-box ML models, our framework integrates SHAP-based analysis systematically to enhance model transparency, stakeholder trust, and actionable insight. All changes are marked in red.
3. Typographical Correction at Line 214
Comment: Line 214: "As show in Figure 1 It" → Should be corrected to: "As shown in Figure 1, it".
Response:
Thank you for pointing this out. We have corrected the phrase at Line 214 to "As shown in Figure 1, it" as suggested.
4. Missing Data Source at Line 290
Comment: Line 290: "This data was also sourced from [?]" → Please provide the actual data source or citation.
Response:
We appreciate the reviewer catching this omission. We have now specified that the sea level pressure data were sourced from the Iran Meteorological Organization (IMO) dataset and have included the appropriate citation. The correction has been incorporated into Section 3.3 and highlighted in red.
5. Clarification on Strongest Significant Correlation
Comment: Clarify how the strongest significant correlation with LWE was determined.
Response:
Thank you for this request. We have added clarification in Section 4.2. The "strongest significant correlation" was determined based on two criteria: (i) the absolute magnitude of the Spearman correlation coefficient (|ρ|), and (ii) the associated p-value (p < 0.05). Among the climatic variables, sea level pressure had the highest absolute correlation with LWE (ρ = -0.527, p = 0.014), satisfying both criteria. This clarification is now provided in the manuscript.
6. Add Explanation of Inter-variable Relationships
Comment: Add explanation regarding inter-variable relationships (Lines 356–358).
Response:
We agree with the reviewer’s suggestion and have added detailed discussion in Section 4.2. We now explain that temperature and sea level pressure are positively correlated, while soil moisture is negatively correlated with both. These relationships reflect climatic dynamics typical of semi-arid regions, where higher temperatures and pressures reduce soil moisture through increased evaporation and decreased infiltration. These new discussions are highlighted in red in the revised manuscript.
7. Explain Trend Results Across Periods (Table 4)
Comment: Explain why the overall trend (2003–2023) appears positive despite a recent strong decline.
Response:
Thank you for this important comment. We have updated the trend analysis discussion in Section 4.4. We clarify that although the 2014–2023 period shows a significant recovery in groundwater storage, the sharp decline during 2003–2013 was stronger in magnitude, leading to an overall negative trend across the full 2003–2023 period. The apparent contradiction arises because early-period declines outweigh late-period recoveries when evaluated cumulatively. We also specify that this segmentation was performed to capture phase-specific hydrological behaviors, which a full-period analysis might otherwise mask.
8. Justify Inclusion of Multiple Regression Despite Low R²
Comment: Justify the inclusion of the multiple regression model despite a low R² value.
Response:
We appreciate the reviewer’s concern. In Section 4.5.1, we now explain that while the R² value (0.363) is modest, this is typical for environmental studies dealing with complex hydrological systems influenced by numerous unobserved factors. We explicitly state that the multiple regression analysis was included to: (i) explore direct linear relationships among climate variables and LWE, (ii) demonstrate the limitations of linear models in this context, and (iii) motivate the need for more advanced forecasting approaches. Additionally, we point out that the weak linear performance highlights the value of non-linear models like Random Forest and XGBoost, which performed significantly better in subsequent analyses.
9. Expand Conclusion to Summarize Machine Learning Framework Contributions
Comment: Expand the Conclusion to better summarize the machine learning contributions.
Response:
Thank you for this suggestion. We have revised the Conclusion to clearly summarize the contributions of the proposed machine learning framework. We now emphasize that the framework combines XGBoost-based feature re-weighting, Random Forest modeling, lag-based memory features, seasonal encoding, and SHAP-based explainability. We also highlight the framework’s strong predictive accuracy (Random Forest: RMSE = 3.27, R² = 0.89) and its ability to provide interpretable forecasts for sustainable groundwater management. These revisions are highlighted in red.
10. Unify Terminology for "Lake Water Equivalent" and "LWE"
Comment: Unify the terminology and abbreviation of "Lake Water Equivalent" (LWE).
Response:
We thank the reviewer for noting this inconsistency. We have carefully reviewed the entire manuscript and unified the terminology. "Lake Water Equivalent" is now consistently used, and it is abbreviated as "LWE" throughout the text, tables, figures, and captions.
Once again, we sincerely appreciate Reviewer 2's thoughtful and constructive comments, which have helped us significantly enhance the clarity and quality of the manuscript.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsReview comments for Manuscript Number : water-3612910-peer-review-v1
Name of Journal : Water
Name of Publisher : MDPI
Tile of Manuscript :An Explainable Machine Learning Framework for Forecasting
Lake Water Equivalent Using Satellite Data: A 20-Year Analysis
of the Urmia Lake Basin
Manuscript Number : water-3612910-peer-review-v1
Following are minor review comments:
1.
The objectives of study to be clearly stated based on last portion of abstract.
2.
In prediction of water storage in lake, author to explain why rain fall and silt entering lake is not considered.
3.
Key words to be organized in alphabetical order
4.
Please explain how your research is superior as compared to study on “Water clarity mapping of global lakes using a novel hybrid deep-learning-based recurrent model with Landsat OLI image”.
5.
Please explain how input parameters are different as compared to study on “An optical mechanism-based deep learning approach for deriving water trophic state of China's lakes from Landsat images.”.
Manuscript can be submitted after including above minor changes for further review.
Comments for author File: Comments.pdf
Minor correction for English language required.
Author Response
Response to Reviewer 3
We sincerely thank Reviewer 3 for the valuable and constructive comments that helped us to further enhance the clarity, depth, and scientific quality of our manuscript. Below, we provide detailed point-by-point responses to each comment.
1. Clearly State the Study Objectives in Abstract
Comment: The objectives of study to be clearly stated based on last portion of abstract.
Response:
We sincerely thank the reviewer for this comment. We have revised the abstract to explicitly state the study objectives in its final portion. The updated abstract now clearly outlines the three main objectives of the study: (i) forecasting Lake Water Equivalent (LWE) in the Urmia Lake Basin using an ensemble-based machine learning framework; (ii) enhancing predictive modeling through XGBoost-guided feature weighting; and (iii) improving model transparency and interpretation using SHAP-based explainability techniques. The objectives have been structured clearly to improve the logical flow and clarity for the reader.
2. Clarify Exclusion of Rainfall and Silt Considerations
Comment: In prediction of water storage in lake, author to explain why rain fall and silt entering lake is not considered.
Response:
Thank you for this important comment. Rainfall has been considered indirectly in our study through precipitation data, which was included as one of the key climatic variables influencing LWE. However, sediment (silt) inflow into the lake was not explicitly considered due to two main reasons: (i) publicly available, consistent satellite-based datasets covering long-term silt/sediment measurements for the Urmia Lake Basin over 2003–2023 are lacking; and (ii) the GRACE-based Lake Water Equivalent (LWE) data primarily captures mass changes related to water storage rather than sediment deposition. Furthermore, sedimentation typically affects bathymetry and water quality rather than significantly altering total water mass at the scales monitored by GRACE satellites.
3. Arrange Keywords Alphabetically
Comment: Key words to be organized in alphabetical order.
Response:
Thank you for this correction. The keywords have been reorganized in alphabetical order in the revised manuscript as follows: Groundwater Storage; Machine Learning; Mann-Kendall Test; Remote Sensing; Time-series Analysis; Urmia Lake.
4. Explain Superiority Over "Water Clarity Mapping" Study
Comment: Please explain how your research is superior as compared to study on “Water clarity mapping of global lakes using a novel hybrid deep-learning-based recurrent model with Landsat OLI image.”
Response:
We appreciate the reviewer’s suggestion to clarify this point. A brief discussion has now been added to the literature review section. In comparison to the study titled “Water clarity mapping of global lakes using a novel hybrid deep-learning-based recurrent model with Landsat OLI image,” our research advances the field in two main ways: (i) while the referenced study focuses on optical clarity and surface properties of global lakes using deep learning applied to optical imagery, our work specifically targets subsurface water storage dynamics (Lake Water Equivalent) using GRACE-based gravimetric satellite observations and climatic time series; (ii) our framework integrates ensemble learning with SHAP-based explainability to provide transparent predictions, while the referenced study primarily emphasizes predictive performance without a detailed explainability framework. Therefore, our study contributes a new dimension to groundwater storage forecasting by combining predictive accuracy with explainability, essential for climate-adaptive water resource management.
5. Differentiate Input Parameters from "Water Trophic State" Study
Comment: Please explain how input parameters are different as compared to study on “An optical mechanism-based deep learning approach for deriving water trophic state of China's lakes from Landsat images.”
Response:
Thank you for this comment. We have now addressed this in the revised manuscript. The input parameters in our study differ fundamentally from those used in “An optical mechanism-based deep learning approach for deriving water trophic state of China's lakes.” The referenced study focuses on optical properties (e.g., reflectance values) derived from Landsat imagery to assess surface water quality (trophic state). In contrast, our study utilizes satellite-derived climatic indicators (e.g., precipitation, soil moisture, sea level pressure, temperature) and gravimetric-based LWE data from GRACE satellites, which are related to hydrological storage and groundwater dynamics. Thus, while the referenced study emphasizes surface water quality estimation based on optical characteristics, our study addresses groundwater storage variability using a climate-driven and physically based modeling approach.
Once again, we sincerely appreciate Reviewer 3's valuable feedback and insightful suggestions, which have contributed to improving the overall scientific rigor and presentation of the manuscript.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have implemented all the suggested changes in the manuscript. Now, it is suitable for publication in the journal.
