Review History


All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.

Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.

View examples of open peer review.

Summary

  • The initial submission of this article was received on July 7th, 2025 and was peer-reviewed by 3 reviewers and the Academic Editor.
  • The Academic Editor made their initial decision on September 18th, 2025.
  • The first revision was submitted on October 14th, 2025 and was reviewed by 2 reviewers and the Academic Editor.
  • A further revision was submitted on November 21st, 2025 and was reviewed by 1 reviewer and the Academic Editor.
  • The article was Accepted by the Academic Editor on December 2nd, 2025.

Version 0.3 (accepted)

· · Academic Editor

Accept

The authors have addressed the consideration of the reviewers. It can be accepted currently. Congrats!

[# PeerJ Staff Note - this decision was reviewed and approved by Jyotismita Chaki, a PeerJ Section Editor covering this Section #]

Reviewer 1 ·

Basic reporting

The authors have addressed my suggestions. Thank you for your thorough work. I have no further comments.

Experimental design

no comment

Validity of the findings

no comment

Additional comments

no comment

Version 0.2

· · Academic Editor

Major Revisions

The manuscript looks noticeably clearer and better organized, and the reviewers seem to appreciate the effort put into addressing earlier concerns. That said, Reviewer 1 still points out a few sticking points, mostly around how transparent the methods really are. For instance, there is mention of needing a proper quantitative runtime analysis, a clearer rationale for the chosen parameter ranges, and some indication of variability across repeated runs. These gaps may suggest that the approach, while improved, is not entirely airtight. There is also a call for tighter statistical checks, more consistent figures, and the inclusion of repository links. Those additions would likely make the work feel more credible and easier for others to reproduce.

**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.

**Language Note:** When preparing your next revision, please ensure that your manuscript is reviewed either by a colleague who is proficient in English and familiar with the subject matter, or by a professional editing service. PeerJ offers language editing services; if you are interested, you may contact us at [email protected] for pricing details. Kindly include your manuscript number and title in your inquiry. – PeerJ Staff

Reviewer 1 ·

Basic reporting

The authors have made commendable efforts to address my previous concerns. The manuscript is now more clearly structured, and the language has been improved. However, a few issues remain.

Experimental design

The authors have significantly improved the methodological clarity, particularly regarding the ESSA algorithm and the RBF network. However, while the authors mention that the computation time is acceptable, a more quantitative analysis (e.g., average runtime, memory usage, or scalability with increasing data size) would strengthen the practical relevance of the model.
The authors provide optimized hyperparameters in Table 4, but the search space for some parameters (e.g., hidden neurons: [1, 1000]) seems overly broad. A brief justification for these ranges would be helpful.
The authors mention using a single run for reporting results. Including performance variability (e.g., mean ± standard deviation over multiple runs) would provide a better understanding of model robustness.

Validity of the findings

The authors mention plans to include statistical tests in future work. While this is appreciated, even basic statistical significance testing (e.g., t-tests or confidence intervals) on the current results would enhance the credibility of the findings.

Additional comments

While the English has improved, some awkward phrasing and grammatical inconsistencies persist. For example, phrases like “the computation time required remains within an acceptable range” could be more precise.
Although the authors revised several figures, some (e.g., Figure 6) still suffer from inconsistent labeling and legends. Ensure all figures are self-contained, with clearly labeled axes, units, and legends. For example, clarify what “Sampling Point 1#” refers to in the context of the broader dataset.
The authors mention uploading data and code to the publisher’s repository. However, the manuscript should include a direct link or DOI to the repository to ensure transparency and reproducibility.

Reviewer 3 ·

Basic reporting

-

Experimental design

-

Validity of the findings

-

Version 0.1 (original submission)

· · Academic Editor

Major Revisions

**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.

**Language Note:** When preparing your next revision, please ensure that your manuscript is reviewed either by a colleague who is proficient in English and familiar with the subject matter, or by a professional editing service. PeerJ offers language editing services; if you are interested, you may contact us at [email protected] for pricing details. Kindly include your manuscript number and title in your inquiry. – PeerJ Staff

Reviewer 1 ·

Basic reporting

The manuscript presents a hybrid deep learning model designed to predict aquaculture water quality parameters with high precision. It integrates a Self-Attention-enhanced Long Short-Term Memory (SA-LSTM) network optimized via an Enhanced Sparrow Search Algorithm (ESSA) for temporal prediction, and a Radial Basis Function (RBF) neural network for spatial interpolation. The model aims to improve forecasting accuracy for key parameters such as dissolved oxygen and water temperature, using data collected from a pond over 30 days.

The manuscript does not clearly articulate the specific contribution of the work. It is unclear what the novel aspects are beyond combining known techniques (LSTM, RBF, SSA, and attention mechanisms).

The structure of the paper does not fully conform to standard scientific article formatting. For example, the contribution of the work is not explicitly stated in the introduction, and the methodology section lacks clarity in its organization.

While the introduction references many related works, it does not clearly define a specific gap in the literature that this study addresses. The novelty and necessity of the proposed model are not sufficiently justified.

Although a table of sample data is provided, the dataset and code necessary for reproducing the results are not made available. This limits the reproducibility of the study.

The results section lacks in-depth analysis. Comparisons with baseline models are presented, but without statistical significance testing or a detailed discussion of why the proposed model performs better.

Experimental design

The research question is implied but not explicitly stated. The manuscript should clearly define what specific problem it aims to solve and how it contributes to the field.

The description of the dataset is vague. While the location and sensor types are mentioned, there is no detailed explanation of data preprocessing, feature selection, or how missing data were handled beyond a brief mention of smoothing.

The manuscript does not provide any pseudocode, algorithmic steps, or computational complexity analysis. This is essential for evaluating the feasibility and scalability of the proposed method.

The absence of code, detailed hyperparameter settings, and data sharing significantly hinders reproducibility. The authors should provide access to the implementation and the dataset.

Validity of the findings

The evaluation lacks statistical rigor. There is no cross-validation reporting.

The manuscript does not explore how the model performs under different conditions, such as varying data quality, sensor noise, or different time windows.

The conclusions are optimistic but not fully substantiated by the presented evidence. The improvements over baseline models are not deeply analyzed or contextualized.

Additional comments

The manuscript introduces a potentially useful hybrid model for spatiotemporal prediction in aquaculture, but the presentation lacks clarity and depth.
The integration of ESSA, SA, LSTM, and RBF is interesting, but the novelty is not well justified.
The figures and tables are helpful but need better captions and integration into the discussion.
The English language is mostly clear, but would benefit from professional proofreading to improve flow and clarity.

Reviewer 2 ·

Basic reporting

- The manuscript is written in clear and professional English. Technical terms (e.g., LSTM, ESSA, RBF) are defined, and mathematical formulations are provided. Minor issues include occasional redundancy (e.g., repetition of “neural neural networks” in the title, line 3) and occasional awkward phrasing that could benefit from light editing.
- The introduction provides a thorough review of relevant studies, including traditional models, neural networks, attention mechanisms, and spatial interpolation approaches. Citations are up-to-date (up to 2025) and well integrated. The knowledge gap, that is, the lack of spatiotemporal predictive frameworks for aquaculture water quality, is clearly identified.
- The article follows standard research structure (Abstract, Introduction, Methods, Results, Discussion, Conclusion). Figures are relevant and well-labeled (e.g., study site, SA-LSTM architecture, RBF structure, prediction comparisons). Tables (sensor parameters, raw data examples, ablation study, prediction metrics) are informative. Figures could benefit from slightly clearer legends and resolution.
- Representative subsets of raw data are included (Table 2), and the authors state that full data are available in supplementary materials. This meets PeerJ’s open data requirements.
- The manuscript is self-contained, presenting both temporal and spatial prediction methodologies, experiments, and results directly linked to the hypotheses.

Experimental design

- The study clearly falls within the scope of computational science and applied AI in aquaculture. It combines ESSA-optimized SA-LSTM with RBF for spatiotemporal prediction, which is original compared to previous time-series-only approaches.
- Well defined: how to improve the accuracy of spatiotemporal prediction of key water quality parameters (DO, temperature, pH, ammonia nitrogen, ORP). The gap, specifically that most existing studies focus on temporal prediction only, is convincingly stated.
- Data collection: 30 days of 10-minute interval measurements from five depths, with reliable sensors (Table 1). Preprocessing methods for missing/outlier data and normalization are clearly described.
Models: LSTM with SA, optimized by ESSA. Comparative methods (LSTM, SA-LSTM, ESSA-SA-LSTM) are tested. Ablation studies and interpolation comparisons (RBF vs. IDW, triangular) are included.
Ethical considerations: Not applicable (no animal/human testing involved).
- Equations, parameters, and computational setup (hardware, software, hyperparameters) are fully described. The workflow diagrams are helpful. With raw data provided, replication is feasible.

Validity of the findings

- The dataset is robust and relevant. Statistical evaluation uses RMSE, MAE, and R², which are appropriate. Improvements (up to about 59% RMSE reduction) are clearly quantified. However, reporting standard deviations or confidence intervals for results would further strengthen robustness.
- Conclusions align directly with the research questions and results. Claims are limited to demonstrated improvements (better spatiotemporal accuracy than baselines). The discussion links results to aquaculture practice (e.g., DO stratification due to photosynthesis).

Additional comments

- Some figures (e.g., Figure 6 comparisons) would benefit from clearer legends and slightly higher resolution.
- The novelty is incremental but meaningful for the aquaculture monitoring domain. The integration of ESSA-SA-LSTM with RBF is convincingly presented as advancing beyond prior time-only approaches.

Reviewer 3 ·

Basic reporting

1. Title & language quality:
The manuscript needs careful proofreading. Throughout the text, there are numerous typographical/formatting errors (e.g., “çC” for °C, inconsistent decimal separators such as “9,47”, range symbols like “0>20 (mg·L⁻¹)”, and occasional variable name glitches). Please standardize units/symbols, use a consistent decimal separator, and fix duplication/typos across the paper, figures, and tables.

2. Study site & sensor deployment need clearer description:
Section “Study area and data acquisition” mentions “five floating buoys … at various planes and depths,” and also lists sensors deployed at depths of 0.3–1.9 m. It’s unclear whether each buoy hosts a vertical string of sensors or whether there are five single-depth points. Please provide a precise schematic (with coordinates), total sensor count, depths per mooring, sampling frequency, maintenance/calibration procedures, and a QA/QC summary.

Experimental design

3. Data quality and preprocessing (experimental design):
The imputation (Eq. 1) and outlier handling (Eq. 2) are not fully specified. The outlier rule (“variation range exceeds that of preceding and succeeding moments”) is ambiguous, and the notation (e.g., Xk2) seems inconsistent with the narrative. Please formalize the detection criterion (thresholds, window size), justify the smoothing choice, and report sensitivity analyses showing model robustness to these steps. Include the proportion of data imputed/flagged.

4. Evaluation metrics and units:
The definition of R2 appears incorrect. Please correct Eq. (6) and clarify whether RMSE/MAE were computed on normalized or original scales. For instance, Table 3 reports very small errors with units of mg·L⁻¹; given DO magnitudes (≈8–10 mg·L⁻¹), RMSE ≈ 0.016 mg·L⁻¹ seems implausibly low unless results are still normalized. Report metrics on the original scale for interpretability.

5. Hyperparameter optimization vs fixed settings:
You state that ESSA optimizes learning rate and iteration count, yet the training section fixes Adam at 0.001 and 1,000 epochs. Also, dropout is described as “retention probability 0.8 (equivalent to a dropout rate of 0.3),” which is inconsistent. Please reconcile these discrepancies and provide a table of the final, ESSA-selected hyperparameters (learning rate, epochs/iterations, neurons per layer), along with search bounds used and the best values found.

6. ESSA configuration and reproducibility (experimental design)
The improved SSA (ESSA) is central to your contribution, but key details are missing: population size, proportions of discoverers/joiners/alerters, warning threshold R2, safety threshold Ts, stopping criteria, and computational cost. Please add these, plus variance across multiple runs (e.g., mean±SD over ≥10 seeds) to demonstrate stability and avoid single-run cherry-picking. A concise pseudo-code block and runtime on your stated hardware would help reproducibility.

7. Spatial RBF model specification and validation:
The RBF layer lacks crucial details: basis function type, center selection method, width/shape parameter estimation, training procedure (e.g., k-means + linear solver vs gradient descent), and regularization. More importantly, the spatiotemporal validation should prevent leakage: please use time-blocked splits and leave-location-out (or leave-buoy-out) cross-validation, then report per-variable metrics (not only aggregate numbers) for DO and temperature at the very least.

Validity of the findings

8. Baselines and fairness of comparisons:
For time-series prediction, comparing only to LSTM and SA-LSTM is limited. Please include GRU/Bi-LSTM/TCN and a classical baseline (e.g., ARIMA/SARIMA) with tuned hyperparameters. For spatial interpolation, Kriging is a standard comparator; including Ordinary/Universal Kriging would strengthen claims beyond IDW/triangulation. Add confidence intervals and statistical tests to show that differences are significant.

9. Scope and generalizability:
The dataset covers approximately 30 days from a single pond (March 15–April 15, 2024). This limits seasonality and site variability. Please temper claims and, if possible, evaluate the model on additional months/sites or perform train-on-subset/test-on-held-out-days to mimic deployment. Consider integrating exogenous drivers (solar radiation/irradiance proxies, wind, barometric pressure) to improve generalization, as your introduction emphasizes external influences.

Additional comments

10. Figures, tables, and data/code availability:
Figure 6 subpanels/labels and axis units need to be fully consistent and self-contained; some panels/legends appear mismatched, and the narrative references “Sampling Point 1#” without clarifying selection. Also, please deposit the full raw time series, metadata (sensor models, calibration logs), and complete code (ESSA + SA-LSTM + RBF pipeline) in a public repository, and update the Data Availability statement accordingly (currently it implies “available within the article/supplementary or from the author”). This is essential for verification and reuse.

All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.