Deep Learning-Based Multi-Source Precipitation Forecasting in Arid Regions Using Different Optimizations: A Case Study from Konya, Turkey

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses an important and timely topic: improving precipitation forecasting in arid regions using LSTM models trained with multi-source datasets (MGM ground-based and NASA/POWER data). The paper is well-structured, the dataset is extensive (39 years), and the methodological comparison between different optimizers (ADAM, RMSProp, SGDM) provides valuable insights. The application to Konya Basin, a water-scarce region, highlights the practical importance of the work. However, the current version suffers from several limitations in methodology, interpretation of results, and depth of discussion. The scientific contribution and novelty need to be better clarified. Therefore, I recommend major revisions before the manuscript can be considered for publication.

Major Comments:

• The manuscript emphasizes LSTM-based precipitation forecasting with different optimizers, but similar approaches have been extensively published in hydrology and climate sciences. The novelty is mainly in the cross-validation with different data sources. The authors should clarify what is genuinely new compared with previous works (e.g., recent studies applying LSTM/GRU with reanalysis and gauge data).
• While the statistical results are clearly presented, there is limited physical explanation of why the model performs better with MGM training than NASA training, or why ADAM outperforms RMSProp and SGDM. A deeper discussion on the hydrometeorological processes in the Konya Basin and how LSTM captures them would significantly strengthen the work.
• The reported R² values (0.99+) and RMSE values (~1–4 mm) suggest extremely high performance. Such results raise concerns about possible overfitting, especially since only monthly univariate precipitation data were used. The authors should provide learning curves, residual analyses, and discuss model generalization capacity more critically.
• The manuscript only considers univariate precipitation forecasting. In hydrometeorology, predictors such as temperature, humidity, circulation indices, and teleconnection patterns (e.g., NAO, ENSO) are often crucial. The absence of such multivariate predictors limits the scientific depth of the study. Future directions should include multivariate inputs to improve robustness.
• Cross-validation between MGM and NASA data is valuable, but the study remains geographically limited to a single basin. The authors should discuss the transferability of the approach to other climatic regions, as well as limitations of using NASA POWER data (e.g., biases in arid regions).
• The manuscript includes Taylor diagrams with conventional ML methods (RF, XGBoost, SVM), but the discussion of these results is superficial. More systematic benchmarking is needed to highlight the added value of LSTM.
• The manuscript is generally readable but contains long, descriptive sentences with limited critical analysis. The discussion section could be more concise and analytical, focusing on interpretation rather than repeating numerical results.

Minor Comments:

• Abstract: Consider shortening and focusing more on main findings and significance.
• Figures: Ensure all figures (e.g., time series, violin plots) have clear axis labels, units, and readable fonts.
• Table 1: The NASA median value (2163) seems inconsistent with MGM (20.9). Please check and correct.
• References: Several references are outdated; more recent works (2021–2024) on AI-based precipitation forecasting should be included.
• Acronyms: Define all acronyms (e.g., MGM, NSE) at first use.
• Conclusion: Current conclusions repeat results. Add more emphasis on practical implications (water management, drought adaptation).

Author Response

Reviewer 1

The manuscript addresses an important and timely topic: improving precipitation forecasting in arid regions using LSTM models trained with multi-source datasets (MGM ground-based and NASA/POWER data). The paper is well-structured, the dataset is extensive (39 years), and the methodological comparison between different optimizers (ADAM, RMSProp, SGDM) provides valuable insights. The application to Konya Basin, a water-scarce region, highlights the practical importance of the work. However, the current version suffers from several limitations in methodology, interpretation of results, and depth of discussion. The scientific contribution and novelty need to be better clarified. Therefore, I recommend major revisions before the manuscript can be considered for publication.

Response to Reviewer 1: I sincerely thank the reviewer for the constructive and valuable comments. I carefully noted your general concerns regarding methodology, interpretation, and the depth of discussion. Following the detailed suggestions from all reviewers, I have thoroughly revised the manuscript to improve clarity, strengthen the scientific contribution, and highlight the novelty of the manuscript. I hope that the revised version addresses your concerns and will be considered suitable for publication.

Major Comments:

• The manuscript emphasizes LSTM-based precipitation forecasting with different optimizers, but similar approaches have been extensively published in hydrology and climate sciences. The novelty is mainly in the cross-validation with different data sources. The authors should clarify what is genuinely new compared with previous works (e.g., recent studies applying LSTM/GRU with reanalysis and gauge data).

Response to Reviewer 1: Thank you for this important remark. I clarified the novelty of the manuscript in the revised version. As highlighted in manuscript, the genuine contribution is the integration of multi-source datasets (MGM ground-based and NASA/POWER satellite), the comparative evaluation of different optimization algorithms (ADAM, RMSProp, SGDM), and the application of cross-validation between independent data sources. This combined framework provides an advancement beyond previous studies that mostly focused on single-source data or standard optimizers.

• While the statistical results are clearly presented, there is limited physical explanation of why the model performs better with MGM training than NASA training, or why ADAM outperforms RMSProp and SGDM. A deeper discussion on the hydrometeorological processes in the Konya Basin and how LSTM captures them would significantly strengthen the work.

Response to Reviewer 1: Thank you for this valuable suggestion. I have expanded the discussion to provide a clearer explanation of the hydrometeorological context of the Konya Basin and why MGM-based training may outperform NASA data. I also added a brief note on the capacity of ADAM to adaptively adjust learning rates, which helps LSTM capture local precipitation variability more effectively compared to RMSProp and SGDM. These additions are included in the revised Discussion section.

“The superior performance of ADAM in this study can be attributed to its adaptive learning rate mechanism, which provides stability and prevents overfitting when dealing with heterogeneous climatic time series, whereas RMSProp and SGDM are more sensitive to local minima.

This can be explained by the fact that MGM station data directly captures the local hydrometeorological dynamics of the Konya Basin (semi-arid continental climate, groundwater-driven feedbacks), which are sometimes smoothed in reanalysis or satellite-based datasets, thereby allowing LSTM to learn more localized precipitation variability.”

• The reported R² values (0.99+) and RMSE values (~1–4 mm) suggest extremely high performance. Such results raise concerns about possible overfitting, especially since only monthly univariate precipitation data were used. The authors should provide learning curves, residual analyses, and discuss model generalization capacity more critically.

Response to Reviewer 1: Thank you for this constructive suggestion. I agree that extremely high R² and low RMSE values may raise concerns about overfitting. To address this, I included several diagnostic analyses in the revised manuscript. Specifically, two-sample t-tests (Table 4) showed no significant bias between observed and modeled precipitation. In addition, cross-validation with independent data sources (MGM vs. NASA; Table 5) confirmed consistent predictive skill, while violin and Taylor diagrams (Figures 8–11) demonstrated stable distributional and correlation patterns. These complementary evaluations indicate that the reported accuracy reflects robust model performance and generalization capacity, rather than overfitting.

• The manuscript only considers univariate precipitation forecasting. In hydrometeorology, predictors such as temperature, humidity, circulation indices, and teleconnection patterns (e.g., NAO, ENSO) are often crucial. The absence of such multivariate predictors limits the scientific depth of the study. Future directions should include multivariate inputs to improve robustness.

Response to Reviewer 1: Thank you for this constructive suggestion. I acknowledge that the current study is limited to univariate precipitation forecasting. As you noted, incorporating multivariate predictors such as temperature, humidity, and large-scale circulation indices (NAO, ENSO) would provide deeper scientific insights. This is an excellent direction for future work.

• Cross-validation between MGM and NASA data is valuable, but the study remains geographically limited to a single basin. The authors should discuss the transferability of the approach to other climatic regions, as well as limitations of using NASA POWER data (e.g., biases in arid regions).

Response to Reviewer 1: Thank you for highlighting this important point. The study was intentionally designed as a basin-scale case study for the Konya Closed Basin to demonstrate the feasibility of integrating multi-source datasets with LSTM. I agree that transferability to other climatic regions and the known limitations of NASA POWER data (e.g., biases in arid environments) should be acknowledged. These aspects are now briefly discussed in the revised Discusion&Conclusions sections and I emphasized that extending the approach to diverse climatic settings will be a priority for future research.

• The manuscript includes Taylor diagrams with conventional ML methods (RF, XGBoost, SVM), but the discussion of these results is superficial. More systematic benchmarking is needed to highlight the added value of LSTM.

Response to Reviewer 1: Thank you for this valuable suggestion. I initially included Taylor diagrams with conventional ML methods (RF, XGBoost, SVM). However, following the other reviewer’s comments, the focus of the manuscript was streamlined to the comparative assessment of three optimization algorithms (ADAM, RMSProp, SGDM) within the LSTM framework. Accordingly, the later sections of the revised manuscript concentrate on this benchmarking and related discussions, while the conventional ML comparisons were not further elaborated in order to maintain a clear research objective.

• The manuscript is generally readable but contains long, descriptive sentences with limited critical analysis. The discussion section could be more concise and analytical, focusing on interpretation rather than repeating numerical results.

Response to Reviewer 1: Thank you for this constructive remark. The discussion section has been revised to be more concise and analytical, with reduced descriptive sentences and less repetition of numerical results. These revisions were also aligned with the suggestions of the other reviewer.

Minor Comments:

• Abstract: Consider shortening and focusing more on main findings and significance.

Response to Reviewer 1: Thank you for this helpful suggestion. The abstract has been revised to be shorter and more focused on the main findings and significance.

• Figures: Ensure all figures (e.g., time series, violin plots) have clear axis labels, units, and readable fonts.

Response to Reviewer 1: Thank you for this remark. All figures, including time series and violin plots, have been revised to ensure clear axis labels, units, and improved font readability.

• Table 1: The NASA median value (2163) seems inconsistent with MGM (20.9). Please check and correct.

Response to Reviewer 1: Thank you for catching this. You are right—there was a formatting error. I have corrected the NASA median from 2163 to 21.63 in Table 1,

• References: Several references are outdated; more recent works (2021–2024) on AI-based precipitation forecasting should be included.

Response to Reviewer 1: Thank you for this suggestion. Additional recent references (2021–2024) on AI-based precipitation forecasting have been included in the revised manuscript.

• Acronyms: Define all acronyms (e.g., MGM, NSE) at first use.

Response to Reviewer 1: Thank you for this helpful remark. I have carefully checked the manuscript and ensured that all acronyms are defined at their first occurrence (e.g., MGM – Turkish State Meteorological Service, NSE – Nash–Sutcliffe Efficiency).

• Conclusion: Current conclusions repeat results. Add more emphasis on practical implications (water management, drought adaptation).

Response to Reviewer 1: Thank you for this valuable suggestion. I agree that the original conclusions were too result-oriented. In the revised version, I have added a statement to highlight the practical implications of the findings for water management and drought adaptation.

 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The performance of the model is unquestioned – it works great for fitting and simulation. However, the forecasting skill is questionable and the ability of the models to truly predict hydrologic anomalies and extreme events have not been presented. This cannot be really evaluated without testing the model for forecasting at different lag times and forecasting accuracy. The study does not include any investigation to claim a forecasting contribution.

That brings the question to the overall goal and intellectual contribution of the paper, which is not clear at this point. Please clarify and use the last part of the introduction section to declare the research questions, hypotheses (if any), and the objectives of this study.

Comments on the Quality of English Language

Quality of the English language is good. But the writing style could be improved by smaller and more focused paragraphs and using a consistent citation style. I suggest a thorough editing of the whole manuscript to make the technical arguments stronger and clearer.

Author Response

Reviewer 2

The performance of the model is unquestioned – it works great for fitting and simulation. However, the forecasting skill is questionable and the ability of the models to truly predict hydrologic anomalies and extreme events have not been presented. This cannot be really evaluated without testing the model for forecasting at different lag times and forecasting accuracy. The study does not include any investigation to claim a forecasting contribution.

That brings the question to the overall goal and intellectual contribution of the paper, which is not clear at this point. Please clarify and use the last part of the introduction section to declare the research questions, hypotheses (if any), and the objectives of this study.

Respond: Dear reviewer 2, Thank you for your constructive feedback. I would like to note that, based on the comments from all reviewers, the manuscript has been substantially revised and strengthened. Regarding your specific concern, I agree that testing at different lag times and explicitly addressing forecasting skill in terms of extreme events would further enhance the contribution. While such an extension goes beyond the current scope, I have clarified the overall goal and contribution of the paper by explicitly stating the research objectives at the end of the Introduction section.

“Accordingly, the objective of this study is to evaluate the capability of LSTM models with different optimizers to forecast monthly precipitation in the Konya Basin using both ground-based and satellite data. The novelty of this work lies in integrating multi-source datasets (MGM ground-based and NASA/POWER satellite), comparing multiple optimization algorithms (ADAM, RMSProp, SGDM), and performing cross-validation between independent data sources to assess model generalizability. In doing so, the study aims to clarify the practical potential and limitations of deep learning models for operational water management and climate adaptation in arid regions.”

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

See the attached review report file.

Comments for author File: Comments.pdf

Author Response

Reviewer 3

This paper evaluates the use of Long Short-Term Memory (LSTM) neural networks for monthly -based (MGM) and NASA POWER precipitation data, the study compares the performance of three optimization algorithms (ADAM, RMSProp, SGDM). Results show that ADAM consistently provides superior accuracy, with RMSE as low as 1.52 mm and R² above 0.99. A cross-validation design (training with MGM and testing with NASA, and vice versa) further demonstrates robustness and interchangeability of satellite vs. in-situ data. The findings highlight the applicability of LSTM ADAM for sustainable water management in data-scarce arid regions.

The manuscript is well motivated, technically thorough, and addresses a problem of high practical importance. However, some areas would benefit from refinement and clarification.

Respond: Dear reviewer 3, I sincerely thank you for the positive evaluation of the manuscript and for recognizing its motivation, technical depth, and practical importance. I also appreciate your constructive suggestions for refinement and clarification, which have been carefully considered and incorporated into the revised version.

 

Major Comments

• The application of LSTM with different optimizers is a solid exercise but somewhat incremental given prior work. The real novelty lies in the cross-validation with NASA vs. MGM data; this point should be emphasized more strongly in the introduction and conclusions.

Response to Reviewer 3: Thank you for your constructive feedback. the introduction has been revised to better align with the study’s objectives, and the findings have been further discussed with improved interpretation.

(2) The description of hyperparameter settings (epochs, hidden units, learning rates) is detailed, but the rationale behind chosen ranges (100 300 epochs, 10 30 units) is not well justified. Was there any systematic tuning (e.g., grid search)? The discussion of potential overfitting could be expanded, especially since ADAM models achieved nearperfect R² on training and testing.

Response to Reviewer 3: Thank you for this remark. Additional explanations on the rationale for the selected hyperparameter ranges and systematic tuning have been included, and the discussion on potential overfitting has been expanded in the revised manuscript.

(3) The two datasets (MGM vs. NASA POWER) show strong correlation (r = 0.888), but the limitations of satellite data in capturing extreme precipitation in arid regions should be more explicitly acknowledged. It would be helpful to include a bias or error analysis of NASA vs. MGM data before model training, to contextualize generalization results.

Response to Reviewer 3: Thank you for this valuable suggestion. In the revised manuscript, the limitations of satellite data in arid regions have been explicitly acknowledged. Additionally, a bias correction analysis between NASA and MGM precipitation values has been performed and presented with the corresponding equation to better contextualize the generalization results.

(4) The study includes comparisons with conventional ML methods (Random Forest, XGBoost, ANN, SVM), but the presentation (Taylor diagrams, violin plots) could be complemented with a summary table for easier interpretation. Were simpler statistical benchmarks (e.g., ARIMA, persistence models) tested? Including these would strengthen the claim of deep learning superiority.

Response to Reviewer 3: Thank you for this suggestion. In line with the other reviewers’ comments, conventional ML methods (RF, XGBoost, ANN, SVM) and simpler benchmarks were removed as they were not directly aligned with the revised objectives of the study. Instead, the results were strengthened through additional statistical evaluations, including two-sample t-tests, to more rigorously support the findings.

(5) The paper convincingly demonstrates technical success, but the implications for water management in Konya could be more explicitly tied to real-world applications (e.g., irrigation scheduling, drought early warning).

Response to Reviewer 3: Thank you for this important remark. Additional explanations and recommendations have been added in the revised manuscript to more explicitly connect the findings to real-world applications. Furthermore, related implications have been incorporated into the discussion of future directions.

 

Minor Comments

• Figures: Captions for Figures 6 11 should be expanded to fully explain what readers should observe.

Response to Reviewer 3 : Thank you for this remark. The captions for Figures 6–11 have been expanded in the revised manuscript.

• Typographical issues: A few grammatical errors and spacing inconsistencies should be carefully proofread.

Response to Reviewer 3: Thank you for pointing this out. The manuscript has been carefully proofread, and typographical issues such as grammatical errors, spacing inconsistencies, and hyphenation artifacts have been corrected in the revised version.

• References: Literature review is comprehensive, but a few more recent works on hybrid deep learning approaches in precipitation forecasting (e.g., Transformer-LSTM, attention-based methods) could be cited.

Response to Reviewer 3: Thank you for this valuable suggestion. A few recent works on hybrid deep learning approaches in precipitation forecasting, including LSTM and attention-based methods, have been added to the revised manuscript.

Recommendation

The manuscript is technically sound and relevant, but the novelty is modest and the interpretation could be deepened. I recommend major revisions before acceptance, focusing on clarifying methodology choices, emphasizing cross-validation as the key contribution, and expanding discussion of limitations and real-world implications.

Response to Reviewer 3: Thank you for all valuable suggestions.

 

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

While this manuscript tackles an important topic, the writing needs significant improvement, and the scientific rigour does not meet the standards expected. I regret recommending strong rejection in its current form.

Comments

1. The introductory paragraph is not particularly appealing in relation to the title. Also, the lines 39-40 are not very clear. What impact of machine learning does the author refer to here in civil engineering?.
2.  What does the author want to say with Figure 1 in relation to the proposed research?
3.  In lines 94-98, the author mentioned that ML models have superior performance without any reference. 
4. Line 99, "in this chapter"... does not look appropriate here.
5.  The paragraph (lines 99-114) needs more coherence and precision instead of just stating what the aim was, what algorithm was used and so on. More focus should be on 'why'?
6.  Bullet some points as the contribution of the proposed work to the field.
7.  Discussion on "study area" appears like the literature review. It should be very precise and to the point.
8.  Fig 2 needs explanations of legends, graphs and so on as a proper study area map.
9.  All other figures require more self-explanatory captions
10.  Data preparation steps for ML models is missing as different ML models require input in different formats.
11.  Only LSTM is considered; there are many more DL models for forecasting.
12.  The results need to be interpreted in detail.
13. More experimentation is demanded for comprehensive evaluations
14. Adapt discussion
15.  Revise the Figure 5 caption and move this figure to the methodology section
16.  Using AutoML is recommended for optimisation and hyperparameter selection.
17. Adapt the conclusion and future works.
18. Revise funding statement
19. Pay attention to references. Include a recent reference on climatology and precipitation forecasting
20. Organise the paper in a way that makes it more like a scientific work

Comments on the Quality of English Language

English need huge improvements

Author Response

Reviewer 4

 

While this manuscript tackles an important topic, the writing needs significant improvement, and the scientific rigour does not meet the standards expected. I regret recommending strong rejection in its current form.

Respond: Dear reviewer, Thank you for your candid evaluation. Significant revisions have been made to improve the clarity of writing and enhance the scientific rigour of the manuscript. In particular, additional statistical evaluations such as two-sample t-tests and cross-validation analyses have been included, and the discussion has been made more concise and analytical. We sincerely hope that with these revisions the manuscript now meets the standards expected and will be found suitable in its present form.

Comments

1. The introductory paragraph is not particularly appealing in relation to the title. Also, the lines 39-40 are not very clear. What impact of machine learning does the author refer to here in civil engineering?.

Respond: Dear reviewer, Thank you for pointing this out. The introductory paragraph has been revised to be more appealing and directly aligned with the manuscript title. The unclear expression in lines 39–40 regarding the impact of machine learning in civil engineering has also been clarified to specifically emphasize its applications in hydraulic engineering, such as water resources management, flood forecasting, and climate adaptation.

1.  What does the author want to say with Figure 1 in relation to the proposed research?

Respond: Dear reviewer Thank you for this question. Figure 1 was included to illustrate the bibliometric analysis conducted with VOSviewer software. A keyword co-occurrence network was created using the terms “precipitation” and “LSTM” from the Web of Science database, highlighting how research trends have evolved and how deep learning, precipitation forecasting, and climate change are interlinked in the literature. This technique was applied to demonstrate the growing relevance of the proposed research and to position the study within the broader scientific landscape.

1.  In lines 94-98, the author mentioned that ML models have superior performance without any reference. 

Response to Reviewer : Thank you for pointing this out. The statement in lines 94–98 regarding the superior performance of ML models has been revised and supported with appropriate references. The revised text now reads:

“In recent years, there has been a marked shift in researchers' focus on advanced data-driven methods in hydrology. As shown in Figure 1, keywords such as deep learning, LSTM, precipitation, and climate change occupy central positions in the research landscape. This trend reflects a growing preference for deep learning, LSTM, and various machine learning approaches, which have been widely adopted due to their superior predictive capabilities compared to traditional statistical methods [12–15].”

 

1. Line 99, "in this chapter"... does not look appropriate here.

Response to Reviewer : Thank you for noticing this. The phrase “in this chapter” in line 99 has been revised to “in this study” to ensure appropriateness in the context.

1.  The paragraph (lines 99-114) needs more coherence and precision instead of just stating what the aim was, what algorithm was used and so on. More focus should be on 'why'?

Response to Reviewer: Thank you for this valuable suggestion. The paragraph in lines 99–114 has been revised for better coherence and precision. In line with this comment and the other reviewers’ suggestions, the section was improved to emphasize the rationale (‘why’) behind using multi-source data, LSTM, and different optimization algorithms, rather than only describing the aim and methods.

 

1.  Bullet some points as the contribution of the proposed work to the field.

Response to Reviewer : Thank you for this constructive comment. The paragraph in lines 99–114 has been revised for improved coherence and precision. In addition, the contributions of the proposed work have been explicitly highlighted in bullet points to emphasize novelty and relevance.

“In this study, monthly precipitation data from Konya, one of Turkey’s driest regions, were collected for the period 1984–2022 from both ground-based meteorological stations and NASA’s POWER database. The main objective was to assess the capability of deep learning methods, specifically Long Short-Term Memory (LSTM) networks, in forecasting monthly precipitation under arid conditions. For model development, the dataset was divided into training (80%) and testing (20%) subsets. Three widely used optimization algorithms—Adaptive Moment Estimation (ADAM), Root Mean Square Propagation (RMSProp), and Stochastic Gradient Descent with Momentum (SGDM)—were compared using standard statistical metrics to evaluate forecasting accuracy and model robustness.

To further examine generalizability, a cross-validation framework was implemented by training on one data source (MGM ground-based or NASA/POWER satellite) and testing on the other. This approach enabled the evaluation of consistency and reliability across independent datasets, thereby assessing their potential interchangeability.

The novelty of this work lies in:

• Integrating multi-source datasets (MGM and NASA/POWER) for precipitation forecasting in a semi-arid region,
• Benchmarking multiple optimization algorithms within the LSTM framework,
• Conducting cross-validation between independent data sources to test model robustness and transferability.

By addressing these aspects, the study clarifies both the practical potential and the limitations of deep learning models for operational water management and climate adaptation in arid regions.”

 

1.  Discussion on "study area" appears like the literature review. It should be very precise and to the point.

Response to Reviewer : Thank you for this observation. The “Study Area” section was intentionally kept somewhat detailed in order to provide sufficient background, as the Konya Closed Basin is a hydrologically and environmentally sensitive region frequently discussed in the literature. While some contextual information may resemble a review, it was considered necessary to help readers unfamiliar with the area understand its significance.

1.  Fig 2 needs explanations of legends, graphs and so on as a proper study area map.

Response to Reviewer : Thank you for this remark. Explanations of the legends, color gradients, and topographic context have been added to Figure 2, and the average elevation of the basin (~1000 m) is now indicated to make it a proper study area map.

1.  All other figures require more self-explanatory captions

Response to Reviewer : Thank you for this helpful suggestion. The captions of all figures have been revised to be more self-explanatory, including clear descriptions of data sources, units, and graphical elements.

1.  Data preparation steps for ML models is missing as different ML models require input in different formats.

Response to Reviewer : Thank you for this remark. In line with the other reviewers’ suggestions, the manuscript was revised to focus exclusively on the LSTM framework rather than multiple ML models. Accordingly, the data preparation steps and methodological explanations have been streamlined and presented specifically for the LSTM-based experiments.

1.  Only LSTM is considered; there are many more DL models for forecasting.

Response to Reviewer : Thank you for this observation. Following the other reviewers’ suggestions, the scope of the manuscript was narrowed to focus exclusively on LSTM, and the analysis and explanations were revised accordingly.

1.  The results need to be interpreted in detail.

Response to Reviewer : Thank you for this suggestion. The results have been interpreted in greater detail in the revised manuscript. In addition, the discussion and conclusion sections were further revised and expanded in line with the other reviewers’ comments.

1. More experimentation is demanded for comprehensive evaluations

Response to Reviewer: Thank you for this comment. Additional statistical evaluations, including two-sample t-tests, have been incorporated into the revised manuscript to strengthen the comprehensiveness of the experimental assessment.

1. Adapt discussion

Response to Reviewer : Thank you for this suggestion. The discussion section has been adapted and revised accordingly in the new version of the manuscript.

1.  Revise the Figure 5 caption and move this figure to the methodology section

Response to Reviewer : Thank you for this suggestion. The caption of Figure 5 has been revised, and the figure has been relocated to the Methodology section in the revised manuscript.

1.  Using AutoML is recommended for optimisation and hyperparameter selection.

Response to Reviewer : Thank you for this helpful suggestion. AutoML-based trials were performed for optimisation and hyperparameter selection. The results consistently indicated that the most suitable range was achieved with 300 iterations and 30 hidden units, as all of the best-performing configurations fell within this interval.

1. Adapt the conclusion and future works.

Response to Reviewer : Thank you for this suggestion. The conclusion and future works sections have been revised and adapted accordingly in the updated manuscript.

1. Revise funding statement

Response to Reviewer : Thank you for this remark. The funding statement has been revised in the updated manuscript.

1. Pay attention to references. Include a recent reference on climatology and precipitation forecasting

Response to Reviewer : Thank you for this suggestion. Recent references on climatology and precipitation forecasting have been added to the revised manuscript.

1. Organise the paper in a way that makes it more like a scientific work

Response to Reviewer : Thank you for all important remark. The manuscript has been revised and reorganized in line with your comment and the other reviewers’ suggestions to improve its overall scientific structure and presentation.

 

Best

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This study explores the application of a deep learning-based LSTM network for monthly precipitation prediction in the arid Konya region of Turkey. It integrates ground observation data (MGM) and remote sensing data (NASA POWER), and compares the performance of three optimization algorithms (ADAM, RMSProp, and SGDM). The research methodology is rigorous, utilizing diverse data sources. The results are comprehensively evaluated using various statistical metrics (RMSE, MAE, R², NSE), and the model's generalization ability is demonstrated through cross-validation. The study holds certain application value and academic contribution in the fields of water resource management and precipitation prediction in arid regions, particularly regarding data source independence and the superior performance of the LSTM-ADAM combination. However, the article still has room for improvement, especially in the depth of result analysis, discussion of limitations, and clarity of future research directions. Overall, the paper is suitable for publication after appropriate revisions.

 

1. Add a theoretical analysis of the superior performance of the ADAM optimizer in the "Discussion" section, incorporating its adaptive learning rate mechanism, and explain why it performs better in precipitation time series forecasting.
2. Provide a detailed explanation in the "Discussion" or "Conclusions" section regarding potential biases in NASA POWER data and their impact on model results, and discuss how these biases could be further corrected in future research.
3. Supplement the discussion on the limitations of the model, particularly regarding the applicability to different climatic regions and data representativeness issues, and propose specific strategies to address these limitations.
4. Include a direct comparison with existing LSTM-based precipitation forecasting studies in the Introduction or Discussion section, highlighting the innovations of this study in terms of methodology, data, or application scenarios.

 

Author Response

Reviewer 1

1-Add a theoretical analysis of the superior performance of the ADAM optimizer in the "Discussion" section, incorporating its adaptive learning rate mechanism, and explain why it performs better in precipitation time series forecasting.

Response to Reviewer 1: Thank you for this valuable suggestion. Theoretical information related to the superior performance of the ADAM optimizer has been highlighted in red in the Discussion section (lines 436–443). The added text is as follows:

“A comparative analysis of various deep learning training algorithms revealed that the ADAM optimizer consistently outperformed its counterparts in terms of predictive accuracy. This observation aligns with existing literature, which emphasizes ADAM’s ability to accelerate convergence while enhancing model accuracy [37]. The superior performance of ADAM in this study can be attributed to its adaptive learning rate mechanism, which provides stability and prevents overfitting when dealing with heterogeneous climatic time series, whereas RMSProp and SGDM are more sensitive to local minima.”

2-Provide a detailed explanation in the "Discussion" or "Conclusions" section regarding potential biases in NASA POWER data and their impact on model results, and discuss how these biases could be further corrected in future research.

Response to Reviewer 1: Thank you for this valuable suggestion. A statement discussing potential biases in NASA POWER data and their implications has been added and highlighted in red in the Conclusions section (lines 530–531) as follows:

“Moreover, possible biases in NASA POWER data, particularly in arid zones, should be considered when transferring this approach beyond the study area.”

 

3-Supplement the discussion on the limitations of the model, particularly regarding the applicability to different climatic regions and data representativeness issues, and propose specific strategies to address these limitations.

 

Response to Reviewer 1: Thank you for this valuable suggestion. The discussion on model limitations and future strategies has been expanded and highlighted in red in the Conclusions section (lines 529–545). These additions emphasize the applicability of the proposed framework to different climatic regions, acknowledge potential data representativeness issues, and outline future research directions involving multivariate datasets to enhance model accuracy and generalization.

 

4-Include a direct comparison with existing LSTM-based precipitation forecasting studies in the Introduction or Discussion section, highlighting the innovations of this study in terms of methodology, data, or application scenarios.

Response to Reviewer 1: Thank you for this valuable suggestion. A comparative discussion with existing LSTM-based precipitation forecasting studies and the key innovations of this research have been highlighted in red in the Introduction section (lines 89–103 and 117–123). These additions emphasize the distinct methodological contributions, including the integration of multi-source datasets, the benchmarking of multiple optimization algorithms, and the cross-validation approach using independent data sources.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript has been improved as per reviewers' comments. However, the introductory paragraph is now too long. Please break it into two parts and make them meaningful. Also, is Figure 1 necessary. It is unclear and confusing - is it really adding any value. I suggest removal.

Author Response

Respond: Dear Reviewer 2, Thank you for your constructive feedback. The introductory paragraph has been divided into two meaningful parts to improve the logical flow and readability, as suggested. Regarding Figure 1, it has been retained intentionally because it supports the methodological rationale of the study by illustrating the interconnections among key research themes (precipitation, LSTM, deep learning, and climate change). This visualization helps to contextualize the selected approach and highlight the research gap addressed in this work. Considering its relevance to both the methodology and previous reviewer suggestions, keeping Figure 1 was deemed more appropriate.

Reviewer 3 Report

Comments and Suggestions for Authors

The revised manuscript satisfactorily addresses the concerns raised in the previous review. The methodological justifications, expanded discussion, and clearer framing of novelty have significantly improved the work. I recommend acceptance for publication.

Author Response

Respond: Dear Reviewer 3, Sincere thanks for the positive evaluation and constructive feedback throughout the review process. The comments and recommendations have been highly valuable in improving the overall quality and clarity of the manuscript.

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have attempted to address the reviewers' comments. I thank the authors for putting in some effort.

However, the methodology presented in the manuscript is not adequately described or justified.

For example, Figure 2 contains multiple training and testing sections, but it is unclear which one is used at each stage and how. If the authors are training the model on one dataset and testing it on another, what is the impact of domain shift and differences in data distribution?

Additionally, it is challenging to discern how the objectives outlined in the manuscript are achieved through the methodology and results. Authors are referred to some references on forecasting, such as a)https://www.mdpi.com/2227-7390/8/9/1441 b)https://www.sciencedirect.com/science/article/pii/S0022169425005554

Based on the above concerns and the lack of sufficient clarification and justification, it is still recommended to reject the manuscript.

Author Response

Respond: Dear Reviewer 4, Thank you for your constructive feedback and careful evaluation. The methodological framework has been substantially clarified and expanded in the revised manuscript. The workflow diagram (Figure 5) has been updated to clearly show the sequence of training and testing phases, including the cross-validation structure using MGM and NASA data sources. In the Methods section (lines 266–268 and 399–425), a detailed explanation is provided on how the model was trained on one dataset and tested on another to evaluate data-source independence and domain shift effects.
Furthermore, the Discussion section (lines 447–456) elaborates on how potential differences in data distribution between ground-based and remote-sensing datasets were addressed through the integrated design of the model.