Application of Artificial Intelligence in Hydrological Modeling for Streamflow Prediction in Ungauged Watersheds: A Review

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Editor,

I have reviewed the manuscript entitled "Application of Artificial Intelligence in Hydrological Modeling for Streamflow Prediction in Ungauged Watersheds: A Review." The manuscript addresses a highly relevant and important topic, dealing with a challenging process. It presents recent developments in the literature; however, I have several suggestions to differentiate it further from existing review studies, as outlined below:

1-In the Materials and Methods section, I recommend grouping and detailing the literature using a bibliometric tool (e.g., VOSviewer), especially to highlight changes in recent years.

2-The challenges of data acquisition—such as delays in obtaining data years after measurement—should be addressed. For example, delays caused by the verification and evaluation of institutional data could be mentioned. Additionally, it would be valuable to discuss how modeling is conducted in transboundary rivers. (For example: "Enhancing Long-Term Streamflow Forecasting and Prediction Using Periodicity Data Component: Application of Artificial Intelligence.")

3-Using a GIS program to illustrate results—such as linking the findings in Figure 2 (which I believe was created with CPT and should be revised) and Table 2 to a geographical map or region—would add significant value to the paper.
Example reference: "Hybridized artificial intelligence models with nature-inspired algorithms for river flow modeling: A comprehensive review, assessment, and possible future research directions."

4-In Future Directions section, where recommendations are given in light of the insights gained, would be beneficial. For instance, recent concepts such as nature-based solutions in relation to flow modeling, as well as digital twins and the integration of real data with physical models, are increasingly important in this context.

5-Additional details should be provided on data scale (daily, hourly, monthly, etc.), dataset length and proportion, hybrid algorithms, and LSTM hyperparameter choices.

6-In Table 2, adding more details in the "Research remark" column would strengthen the analysis.

7-The discussion on performance metrics is quite limited. For example, in the case of LSTM, metrics such as RMSE, MARE, and MAE could be explicitly included.

8-Under the Challenges and Future Directions section, the placement of Figure 4 should be reconsidered. Instead of appearing immediately under the heading, it could be positioned after the relevant explanations in the text.

Author Response

Dear Reviewer 1,

Thank you so much for your valuable inputs. These inputs greatly improved our work. Please see attached file for our point by point response

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This review paper by Gacu et al ("Application of Artificial Intelligence in Hydrological Modeling for Streamflow Prediction in Ungauged Watersheds: A Review") examines the application of AI techniques for streamflow prediction in ungauged watersheds, synthesising 111 studies following PRISMA guidelines. While the paper addresses an important and timely topic, several significant gaps and limitations affect its scholarly contribution.

Major limitations

1. The paper's scope and methodology are generally appropriate, covering traditional hydrological models, AI techniques (machine learning and deep learning), and their integration with remote sensing. The authors correctly identify ungauged watersheds as a critical challenge and position AI as a transformative solution. The systematic approach using PRISMA guidelines and the inclusion of 111 studies provides a reasonable foundation. However, the temporal distribution of studies (heavily weighted toward 2021-2025) may introduce recent paper bias at the expense of foundational work.
2. A major limitation is the incomplete coverage of recent advances in the field. The paper misses several key developments that have significantly shaped AI applications in ungauged basin hydrology. Notably absent are discussions of: physics-informed neural networks (PINNs) , which have emerged as a promising approach for incorporating physical constraints into AI models; graph neural networks (GNNs) for basin scale modelling , which are particularly relevant given the network structure of watersheds; foundation models and pre-trained architectures , which represent a paradigm shift toward large-scale, transferable models; and comprehensive uncertainty quantification frameworks , which are crucial for operational applications in ungauged basins. 
3. The technical analysis, while comprehensive in breadth, lacks depth in several areas. The discussion of hybrid physics-AI approaches is superficial and fails to adequately address the growing importance of physics-informed modelling. The treatment of explainable AI (XAI) techniques is minimal, despite their critical importance for building trust in AI models for water resource management. The paper also provides insufficient coverage of advanced uncertainty quantification methods, which are essential for decision-making in data-scarce environments. Additionally, the discussion of graph-based approaches is largely absent, despite their natural applicability to river network modelling.
4. The manuscripts discussion of limitations and future directions, while present, is not comprehensive enough. The authors identify some gaps but miss critical emerging challenges such as model interpretability in operational settings, the need for standardised benchmarking protocols, and the integration of real-time data streams. The paper would benefit from a more thorough examination of reproducibility issues, computational efficiency considerations, and the practical challenges of deploying AI models in resource-limited settings.
5. Several methodological concerns also affect the paper's quality. The evaluation criteria for study inclusion/exclusion could be more rigorous, and the synthesis approach is primarily narrative rather than systematic. The comparative analysis between AI and traditional methods relies heavily on reported performance metrics without adequate consideration of methodological differences between studies. The geographic bias toward developed regions is acknowledged but not adequately addressed in the analysis.
6. Despite these limitations, the paper does make some valuable contributions. It provides a comprehensive overview of traditional and AI-based approaches, offers useful comparative tables and frameworks, and identifies several important research gaps. The integration of remote sensing with AI techniques is well covered, and the paper serves as a useful entry point for researchers new to the field.

Minor issues

The English is generally good but needs modest improvement for publication. Areas requiring improvement (suggestions only):

1. Minor Grammar and Style Issues

• Some awkward phrasing that could be streamlined
• Occasional redundancy in expression
• Minor preposition and article usage inconsistencies
• Some sentences are overly complex and could be simplified

2. Specific Examples from the Text:

• Line 48-49: "are commonly used for the said purposes that typically rely on..." (awkward phrasing)
• Lines 182-183: "[insert years if applicable, e.g., 2000 and 2025]" (incomplete editing)

3. Flow and Readability:

• Some paragraphs contain very long sentences that could be broken down
• Occasional repetitive phrasing
• Some transitions between paragraphs could be smoother 

Recommendation: Major Revision

The paper addresses an important topic and provides a useful foundation, but significant revisions are needed to address the gaps in coverage of recent advances, particularly in physics-informed neural networks, graph neural networks, uncertainty quantification, and explainable AI. The technical depth should be enhanced, the discussion of hybrid approaches should be expanded, and a more systematic approach to literature synthesis would strengthen the contribution. With these revisions, the paper could serve as a valuable resource for the hydrological modelling community.

Author Response

Dear Reviewer 2

Thank you so much for your valuable inputs to improve our work. Please see attached file for our point by point response to your comments. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Editor,

I have carefully reviewed all the authors’ responses and confirm that they have addressed the reviewers’ comments thoroughly and implemented the required major revisions. In my opinion, the manuscript has reached an acceptable standard and can be considered for publication.

Best regards

Author Response

Dear Reviewer 1,

Thank you so much for your valuable inputs. It made our paper better. its all appreciated

Reviewer 2 Report

Comments and Suggestions for Authors

I appreciate the authors’ efforts to revise their manuscript in response to my comments. The revised version shows expanded coverage of recent advances (PINNs, GNNs, foundation models, explainable AI, uncertainty quantification) and some additional discussion of limitations, future directions, and methodological clarifications. However, several of my major concerns remain insufficiently addressed:

1. Depth of Technical Analysis

While the authors have added sections that mention PINNs, GNNs, foundation models, and UQ frameworks, these remain surface level descriptions rather than a critical synthesis of how these approaches reshape hydrological modelling.

My original concern was not just about acknowledging these advances but providing a rigorous comparative discussion of their strengths, weaknesses, and applicability to ungauged basins. The revisions read more like extended summaries than substantive engagement.

2. Systematic vs. Narrative Synthesis

The authors reiterate that their approach is “primarily narrative.” Despite claims of clarifying inclusion/exclusion criteria and using bibliometric tools, the synthesis still lacks a systematic comparative framework. This remains a weakness, as the strength of a PRISMA-based review lies in structured, transparent synthesis rather than narrative overview. The revisions fall short of addressing this methodological gap.

3. Comparative Analysis of AI vs. Traditional Models

The authors state they have considered methodological differences across studies (e.g., input data, scale), but the revisions do not appear to include a systematic framework or tabulated comparison. Without such structure, the comparative discussion continues to rely heavily on reported performance metrics without critically contextualising differences.

4. Geographic Bias

The authors acknowledge underrepresentation of developing regions but address it only by suggesting “future research.” This does not substantively reduce the bias in the present review. A more meaningful revision would involve actively expanding coverage of some studies from underrepresented regions (even if fewer in number) and critically analysing the implications of this imbalance.

5. Future Directions and Limitations

The expanded discussion includes additional concepts (e.g., digital twins, cloud platforms, nature-based solutions). While useful, this section remains somewhat of an “unstructured listing of concepts” rather than a structured roadmap. Critical challenges such as reproducibility, benchmarking protocols, and operational deployment barriers are still discussed only briefly, without sufficient analysis of their implications.

The revisions improve the breadth of coverage and readability but do not fully resolve the central issues of technical depth, systematic synthesis, and methodological rigor. In its current form, the manuscript provides a broader overview but still lacks the critical, systematic engagement expected of a high-quality review article.

Recommendation – Major Revision

While the revisions to date represent progress, substantial further work is still required to meet the expectations of a systematic and critical review. If these issues are comprehensively addressed, however, the paper has the potential to make a meaningful and valuable contribution to the literature on AI applications in hydrology.

Author Response

Dear Reviewer 2

Thank you for your valuable inputs to our paper. We have realized some things that would greatly improve our work. Please see attached file for our point by point response to your comments. Thank you so much, everything is appreciated

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have responded rigorously and constructively, addressing all major concerns with substantial revision. Their integration of systematic elements (bibliometric analysis, comparative tables, and structured roadmaps) and more critical engagement satisfies the requirements of a high-quality technical review in the field of AI and hydrology.

The review could further benefit from (if feasible):

• Additional explicit framework or flowchart for literature inclusion/exclusion (greater reproducibility).
• Deeper theory-driven synthesis (e.g., integrating critical interpretive synthesis practices for theory building or reconceptualisation).
• Even more thorough benchmarking of reproducibility and transparency protocols.

I recommend that the paper be accepted after minor revision, mainly for polish or very minor additions. The substantive methodological and analytical requests have been commendably satisfied, with improvements in clarity, critical engagement, and transparency that now meet or nearly meet high standards for the genre. The paper is ready for acceptance, provided these minor enhancements are made if feasible.

Author Response

Dear Reviewer 2

Thank you so much for your additional inputs to improve our work. Please see attached point by point response.

Author Response File: Author Response.pdf