Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript proposes a yield prediction strategy that aligns meteorological variables with management phases (intervals between agronomic operations) and integrates this representation into a hybrid CNN–LSTM model with an attention mechanism. Incorporating agronomic domain knowledge into feature engineering is well motivated and may have tangible practical value for planning and sustainability.

Aspects requiring revision to strengthen the rigor and utility of the work:

1. Since the core contribution hinges on the definition of management phases, these phases should be described using explicit and reproducible rules: how the date of each operation is identified, how missing events are handled, how repeated events are treated, how atypical cycles are managed, and whether any exclusion criteria are applied. Providing a flowchart and/or pseudocode for the phase-segmentation procedure would substantially improve methodological transparency.
2. The manuscript should state explicitly that phases are defined by management operations and do not necessarily correspond to phenological stages. A brief discussion of when these constructs align—and when they diverge—would facilitate interpretation and help delineate the method’s generalizability to other crops and production systems.
3. The current approach relies on phase-level aggregation (e.g., means/cumulative sums). The authors should justify this choice and/or extend the representation with robust statistics that capture extremes and variability (e.g., maxima/minima, percentiles, number of days exceeding a threshold, duration of rainfall/temperature spells). In addition, incorporating phase length as a covariate may help capture event-duration effects on yield.
4. To ensure replicability, the manuscript should detail: the missing-data imputation method, outlier criteria, the exact normalization/standardization procedure (and whether parameters are fit using the training set only), and the characteristics of the meteorological data source (number of stations and approximate distance to production areas). A table reporting the percentage of missing values by variable and phase would add clarity.
5. Regarding the experimental design and hyperparameter configuration, it is appreciated that training settings are harmonized across deep-learning models. To strengthen robustness, the authors should report (or provide in an appendix): (i) the hyperparameter search strategy (grid, random, manual), (ii) the ranges explored for each model, and (iii) the validation-based selection criterion. Model comparisons can be sensitive to architectural choices (e.g., number of layers, units, kernel sizes, dropout, learning rate).
6. The use of attention provides an opportunity for additional scientific value. The authors are encouraged to include an analysis of attention weights by phase and by variable—either globally (averages) or conditionally (e.g., stratified by climate regimes)—to identify which phases/variables drive predictions. This would better connect the approach to agronomic interpretation and management decision-making.
7. The study is conducted on a single crop and region. The manuscript should explicitly discuss limitations and conditions for transferring the approach. If feasible, adding a temporal generalization evaluation (e.g., training on earlier years and testing on later years) would help demonstrate performance under distribution shift.
8. The authors should specify the level of yield aggregation (plot/field/farm/cycle) and ensure unit consistency (kg/ha). If records differ in cultivated area or management practices, the manuscript should explain how such heterogeneity is controlled for or normalized.
9. With respect to manuscript structure and writing, the English is generally understandable, but clarity would benefit from improved concision and consistent terminology (e.g., “management phase” vs. “production phase”). A focused language edit is recommended, particularly in the Introduction, Methods, and Discussion, to reduce ambiguity and improve flow.
10. Engagement with recent scholarship is adequate, but could be strengthened by citing more recent work on temporal models with attention/Transformers and multimodal approaches (e.g., weather + soil + sensors/remote sensing). This would sharpen the positioning of the contribution relative to contemporary alternatives.
11. The manuscript mentions waste reduction and supply-chain planning. It is suggested to add paragraph providing concrete examples of operational decisions enabled by the prediction (e.g., harvest scheduling, contracting, logistics, input planning) and, where possible, a qualitative or quantitative indication of potential benefits.
12. The authors are encouraged to include a worked prediction example and informative visualizations. Specifically: (i) select 1–3 representative production cycles (e.g., high/medium/low yield) and show predicted versus observed values; (ii) provide an observed-vs.-predicted (parity) plot with a 1:1 line and global metrics (RMSE/MAPE/R²); and (iii) add a residual plot (residual vs. predicted or residual vs. observed) to assess bias (under-/overestimation at extremes) and heteroscedasticity. If attention is used, it would also be valuable to visualize phase-level attention weights for the illustrative case to connect performance with agronomic interpretability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposes a hybrid deep learning model (CNN–LSTM–Attention) that incorporates weather data aligned with production management phases to improve crop yield prediction. The empirical results on an eight-year Bok choy dataset suggest improved predictive performance compared with several baseline models.

The technical implementation is competent, and the dataset is valuable. However, the manuscript currently reads primarily as an applied machine learning performance study rather than a contribution to sustainability science. The sustainability implications are asserted rather than demonstrated, methodological rigor can be strengthened, and several conceptual and validation gaps limit the scholarly contribution. Substantial revisions are required to elevate the work beyond incremental model comparison.

 

Strengths

• The focus on aligning weather variables with agronomic production phases is logically motivated and agriculturally plausible.
• The use of 1,714 production cycles across eight years enhances practical relevance.
• CNN, LSTM, and attention components are clearly presented, including equations and hyperparameters.
• Multiple alternative models are evaluated under consistent training settings.
• Reported gains in RMSE, MAPE, and R² are numerically substantial.

Major Concerns

• The manuscript repeatedly claims that improved prediction promotes sustainability (e.g., balancing supply and demand, reducing waste), yet it provides no sustainability metrics, simulations, or impact analyses. Prediction accuracy alone does not automatically translate into sustainability gains. There is no quantification of: i) Reduced input use (water, fertilizer), ii) Reduced post-harvest waste, iii) Economic stability improvements, iv) Emissions reduction, v) Risk mitigation under climate variability.
  

  The contribution must extend beyond model accuracy. The authors must either demonstrate measurable sustainability outcomes or explicitly frame the paper as a predictive modeling contribution and more carefully justify its sustainability relevance.

• The dataset is confined to: One crop (Bok choy), One cooperative, One region, One meteorological station. This raises concerns regarding: Climatic variability, Cross-crop transferability, Regional robustness, Structural changes over time.

  The manuscript does not test: Year-wise stability, Performance under extreme weather years, Cross-plot generalization (37 plots are mentioned but not analyzed separately).

  Without such analyses, claims of robustness remain unverified.

• Results are reported from a single chronological split. There is: No multi-seed repetition, No variance reporting, No confidence intervals, No statistical significance testing, No ablation beyond phase removal.

  Deep learning models can vary significantly across runs. Reporting single-point metrics risks overstating stability. This is a serious methodological weakness.

• The attention mechanism is described mathematically, but: No attention weight visualization is provided. No agronomic interpretation of attention scores is presented. No feature importance analysis is conducted.

  Thus, the claim that the model dynamically weights phase importance remains theoretical rather than empirically demonstrated. For agricultural sustainability research, interpretability is not optional — it is essential for decision-making credibility.

• While CNN, LSTM, CNN-LSTM, LSTM-AM, and XGBoost are included, the study omits: Transformer-based time series models, Temporal Convolutional Networks (TCN), Strong tabular models (e.g., CatBoost, LightGBM), Statistical baselines (e.g., ARIMA with engineered features).

  Given that the paper acknowledges more advanced architectures in future work, excluding them weakens the “state-of-the-art” claim.

• Several technical decisions lack justification: i) Averaging weather within phases instead of using distributions or extrema. ii) Mean imputation for gaps >3 days. iii) Winsorization without sensitivity analysis. iv) Use of MAE loss but evaluation emphasizing RMSE. v) Additionally, there is inconsistency between activation functions in equations (tanh) and hyperparameter table (ReLU).

  These issues suggest insufficient methodological depth.

Minor Issues

• The Contribution section largely repeats the Introduction. Condense and sharpen the narrative.
• The “58% reduction” claim should clarify baseline context and avoid marketing-style phrasing.
• Tables contain typographical encoding errors (e.g., broken words).
• Cross-study RMSE comparisons are not directly meaningful because yield scales differ. This limitation should be explicitly acknowledged.
• English clarity: Remove repeated sentences and tighten technical explanations.

Suggestions

• Add multi-run experiments (≥5 seeds) with mean ± standard deviation.
• Provide attention weight visualization and agronomic interpretation.
• Include year-wise performance analysis to assess temporal robustness.
• Strengthen baselines with at least one transformer or TCN model.
• Clarify preprocessing sensitivity via ablation experiments.
• Explicitly quantify sustainability implications through scenario modeling.
• Resolve activation inconsistencies and improve technical precision.
• Streamline narrative and reduce redundancy.

Comments on the Quality of English Language

The manuscript would benefit from professional language editing. There are grammatical inconsistencies, repeated phrases, awkward transitions, and redundancy. Technical explanations would gain clarity through concision and improved sentence structure.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript entitled “A Hybrid Deep Learning Model for Crop Yield Prediction Taking Weather Data Associated with Production Management” presented a combination of three models including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Attention Mechanisms (AM). The current manuscript needs to address the following items as :

1- Contribution

The contributions of the paper are not clearly defined. I strongly recommend that the authors list the main contributions in bullet-point format at the end of the Introduction section to make it easy and understand the novelty of the work. Also, one of the questions arise through reading the manuscript is:How the proposed model outperforms existing approaches.

2- Related Work

The current manuscript lacks comprehensive previous work  discussion of related work. While Table 7 is discussed later in the manuscript, it would be more appropriate to include this comparison in the Related Work section rather than in the Discussion. Also, the table should also include the benchmark data set as well for the objective and limitations.

3- Dataset

Although the authors describe the dataset features, the manuscript does not provide a repository link, or clear description of data accessibility.

4- Table and figures

The current manuscript did not clearly describe how the input/output are involved in the prosed models. Also, figure 1 did not present how the dataset is generated through the models.

5- Proposed Model Architecture

The authors describe CNN, LSTM, and Attention Mechanism in each subsection. However, the manuscript does not clearly explain how these components are integrated into one framework. For example, which stages involve before implementing those models. Also, what about the dimension complexity perform on these models to enhance the outcome.

Additionally, the proposed algorithm lacks a formal description of model input, outputs as well as processing for the proposed model.

6- Discussion

The authors have addressed CNN, LSTM, and Attention mechanisms. However, the authors did not clearly present the hyperparameter or optimize the outcome for each of these models.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript makes an original and potentially significant contribution by integrating meteorological information aligned with management phases into a hybrid CNN-LSTM-AM architecture. Compared with the previous version, the revised manuscript demonstrates substantial improvement, both in the robustness of its predictive results and in the updating and articulation of the discussion. It is also evident that the authors have addressed several important comments raised during the first round of review, including the conceptual distinction between management phases and phenological stages, the incorporation of recent literature, the analysis of attention weights, and the inclusion of evaluation figures together with illustrative cases.

Nevertheless, before the manuscript can be considered for acceptance, I recommend a minor revision aimed at strengthening its formal consistency and presentation. In particular, Figure 1 should be improved in terms of visual quality, as its current resolution is insufficient for adequate readability. In addition, reorganizing it into a vertical rather than a horizontal layout would enhance its clarity. For Figure 2, it would be advisable to explicitly include the labels (a) and (b) in each subfigure to facilitate identification and cross-reference within the text. Finally, a final revision of style and grammar is recommended in order to correct any remaining linguistic inconsistencies and improve the overall fluency of the manuscript.

Overall, I believe that the manuscript has sufficient merit to warrant publication, provided that these final adjustments are adequately addressed.

Author Response

Response to Reviewer 1 Comments

1. Summary

Thank you very much for taking the time to review this manuscript and for your positive evaluation of our work. We sincerely appreciate the constructive feedback provided during both rounds of review, which has significantly contributed to the improvement of the manuscript. Please find the detailed responses below and the corresponding revisions/corrections highlighted in track changes in the re-submitted files.

2. Questions for General Evaluation

Reviewer's Evaluation	Response	Revisions
Is the content succinctly described and contextualized? — Yes	Thank you.	No revision needed.
Are the research design, questions, hypotheses and methods clearly stated? — Yes	Thank you.	No revision needed.
Are the arguments and discussion of findings coherent, balanced and compelling? — Yes	Thank you.	No revision needed.
For empirical research, are the results clearly presented? — Yes	Thank you.	No revision needed.
Is the article adequately referenced? — Yes	Thank you.	No revision needed.
Are the conclusions thoroughly supported by the results? — Yes	Thank you.	No revision needed.

3. Point-by-point Response to Comments and Suggestions for Authors

Comment 1: "Figure 1 should be improved in terms of visual quality, as its current resolution is insufficient for adequate readability. In addition, reorganizing it into a vertical rather than a horizontal layout would enhance its clarity."

Response 1: Thank you for pointing this out. We agree with this comment. Accordingly, we have redesigned Figure 1 to adopt a vertical layout, which better illustrates the sequential data flow and dimensional transitions across the CNN, LSTM, and Attention modules. The figure resolution has also been significantly improved to ensure readability. The updated Figure 1 can be found in Section 3.2.3 of the revised manuscript.

Comment 2: "For Figure 2, it would be advisable to explicitly include the labels (a) and (b) in each subfigure to facilitate identification and cross-reference within the text."

Response 2: We agree with this suggestion. We have added explicit labels "(a)" and "(b)" directly within each subfigure of Figure 2, corresponding to the Observed vs. Predicted yield parity plot and the Residual plot, respectively. This revision facilitates easier identification and cross-reference within the text. The updated Figure 2 can be found in Section 5 (Discussion) of the revised manuscript.

Comment 3: "A final revision of style and grammar is recommended in order to correct any remaining linguistic inconsistencies and improve the overall fluency of the manuscript."

Response 3: Thank you for this suggestion. We have carefully proofread the entire manuscript and corrected several linguistic inconsistencies, including: (1) fixing typographical errors where "R²" was incorrectly rendered as "R?" in the Discussion section (Section 5, paragraphs 1 and 5); (2) improving sentence fluency and grammatical consistency throughout the manuscript. All corrections are highlighted in the tracked changes of the revised manuscript.

4. Response to Comments on the Quality of English Language

Point 1: The reviewer indicated that "The English could be improved to more clearly express the research."

Response 1: We have conducted a thorough revision of the English language throughout the manuscript to improve clarity and fluency. Specific corrections include fixing typographical errors (e.g., "R?" corrected to "R²" in multiple locations in the Discussion) and refining sentence structures for improved readability. All changes are visible in the tracked changes of the resubmitted manuscript.

5. Additional Clarifications

We would like to express our sincere gratitude to the reviewer for the careful and constructive evaluation. The suggestions have been instrumental in improving the overall quality and presentation of our manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Here is a refined version with one single issue clearly stated:

The manuscript has improved considerably compared to the previous version, particularly in terms of methodological rigor, clarity, and overall completeness. The inclusion of multi-seed experiments, preprocessing ablation analysis, and attention-based interpretability substantially strengthens the technical contribution. The revised version is clearer, better structured, and more consistent throughout.

However, one issue remains that should be addressed prior to final acceptance. Specifically, some statements, particularly those referring to “substantial improvement” and “sustainability benefits”, are stronger than what is directly supported by the presented results. As the sustainability implications are based on scenario analysis rather than empirical validation, the corresponding claims should be moderated to better reflect the evidence.

Author Response

Response to Reviewer 2 Comments

1. Summary

Thank you very much for taking the time to review this manuscript and for your thorough and constructive feedback. We greatly appreciate the recognition of the improvements made in the revised version, particularly regarding methodological rigor, clarity, and the inclusion of multi-seed experiments and attention-based interpretability. Please find the detailed responses below and the corresponding revisions/corrections highlighted in track changes in the re-submitted files.

2. Questions for General Evaluation

Reviewer's Evaluation	Response	Revisions
Is the content succinctly described and contextualized? — Can be improved	Please see the point-by-point response below.	Revised. See Response 1 below.
Are the research design, questions, hypotheses and methods clearly stated? — Yes	Thank you.	No revision needed.
Are the arguments and discussion of findings coherent, balanced and compelling? — Can be improved	Please see the point-by-point response below.	Revised. See Response 1 below.
For empirical research, are the results clearly presented? — Yes	Thank you.	No revision needed.
Is the article adequately referenced? — Yes	Thank you.	No revision needed.
Are the conclusions thoroughly supported by the results? — Can be improved	Please see the point-by-point response below.	Revised. See Response 1 below.

3. Point-by-point Response to Comments and Suggestions for Authors

Comment 1: "Some statements, particularly those referring to 'substantial improvement' and 'sustainability benefits', are stronger than what is directly supported by the presented results. As the sustainability implications are based on scenario analysis rather than empirical validation, the corresponding claims should be moderated to better reflect the evidence."

Response 1: We agree with this comment and thank the reviewer for the careful reading. We have moderated the relevant claims in the following specific locations:

(1) In the Abstract, the phrase "substantial reduction" has been revised to "notable reduction" to more accurately reflect the nature of the improvement.

Revised text: "...representing a notable reduction in prediction error (58% lower RMSE) compared to XGBoost."

(2) In the Discussion (Section 5), the sustainability benefits paragraph has been revised. The phrase "may generate sustainability benefits" has been changed to "could potentially contribute to more sustainable practices," and an explicit caveat about the need for empirical validation has been added.

Revised text: "...Such improvement could potentially contribute to more sustainable practices by reducing the mismatch between harvested volume and contracted demand..."

Revised text: "Nevertheless, they suggest that accurate yield forecasting has the potential to support sustainability-oriented agricultural management, though empirical validation of these sustainability outcomes is needed in future studies."

(3) The case-specific scenario paragraph (Section 5) has been revised to more explicitly emphasize the scenario-based nature of the projections: "It should be emphasized that these values are derived from scenario-based projections rather than direct empirical observations, and therefore should be interpreted with caution. Future field validation studies are recommended to substantiate these estimated sustainability benefits."

All changes are marked in tracked changes in the revised manuscript.

4. Response to Comments on the Quality of English Language

Point 1: The reviewer indicated that "The English is fine and does not require any improvement."

Response 1: Thank you for the positive evaluation of our English language quality. We have nonetheless conducted a final proofread and corrected minor typographical errors (e.g., "R?" → "R²") to ensure consistency throughout the manuscript.

5. Additional Clarifications

We sincerely appreciate the reviewer's thorough and constructive evaluation. The suggestion to moderate sustainability-related claims has improved the scientific rigor and balanced tone of the Discussion section.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The updated manuscript addresses all the reviewers comments, so the current manuscript is clear, and the quality of the work has improved. 

Author Response

Response to Reviewer 3 Comments

1. Summary

Thank you very much for taking the time to review this manuscript and for confirming that the updated manuscript addresses all prior reviewer comments. We are grateful for the reviewer's thorough evaluation and positive assessment of the manuscript quality.

2. Questions for General Evaluation

Reviewer's Evaluation	Response	Revisions
Is the content succinctly described and contextualized? — Yes	Thank you.	No revision needed.
Are the research design, questions, hypotheses and methods clearly stated? — Yes	Thank you.	No revision needed.
Are the arguments and discussion of findings coherent, balanced and compelling? — Yes	Thank you.	No revision needed.
For empirical research, are the results clearly presented? — Yes	Thank you.	No revision needed.
Is the article adequately referenced? — Yes	Thank you.	No revision needed.
Are the conclusions thoroughly supported by the results? — Yes	Thank you.	No revision needed.

3. Point-by-point Response to Comments and Suggestions for Authors

Comment 1: "The updated manuscript addresses all the reviewers comments, so the current manuscript is clear, and the quality of the work has improved."

Response 1: Thank you for the positive assessment. We greatly appreciate the reviewer's confirmation that the previous round of revisions has been satisfactorily addressed. In this round, we have made additional minor revisions based on feedback from other reviewers, including: (1) improving the visual quality and layout of Figure 1, (2) adding explicit (a) and (b) labels to Figure 2 subfigures, (3) moderating sustainability-related claims in the Discussion to better reflect the scenario-based nature of the analysis, and (4) correcting minor typographical and grammatical errors. All changes are highlighted in the tracked changes of the resubmitted manuscript.

4. Response to Comments on the Quality of English Language

Point 1: The reviewer indicated that "The English is fine and does not require any improvement."

Response 1: Thank you. We have nonetheless performed a final proofread and corrected minor typographical errors to ensure consistency throughout the manuscript.

5. Additional Clarifications

We sincerely appreciate the reviewer's positive evaluation and the time invested in reviewing our manuscript. The constructive feedback throughout the review process has been instrumental in improving the overall quality of this work.

Author Response File: Author Response.pdf