KINLI: Time Series Forecasting for Monitoring Poultry Health in Complex Pen Environments

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

A forecasting framework is investigated to monitor poultry health in complex pen environments with challenging characteristics such as sensor noise, timestamp unreliability, and high variance. It demonstrates applicability, practicality, and adaptability through the evaluation of state-of-the-art forecasting algorithms. However, there are still many aspects that need to be improved. The followings are some comments that can be concerned:

1)  The manuscript reads more like an experimental report based on a large number of experiments. The novelty of the work should be highlighted more clearly.

2)     Although the authors argue that the difficulties of preprocessing due to industry-specific data characteristics motivated them to train directly on raw data, the analysis on the potential impact of preprocessing techniques on model performance should be added.

3)  The experimental setup requires more detailed description to ensure reproducibility and clarity.

4)  The work uses large language models (LLMs) without fine-tuning. The authors should discuss whether fine-tuning could improve performance and also reflect on whether the relatively small dataset size is suitable for LLM-based approaches.

5）The authors are advised to carefully proofread and revise the text for accuracy.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Major Comments

• In the Introduction, the authors should explain why turkeys were chosen as the subject of this study.

• The Introduction should also highlight how this research differs from existing work, including a discussion of prior studies and a comparison with them.

• Providing basic descriptive statistics for the KINLI dataset would be helpful for readers. The authors should also mention the data source.

• In Section 2. Materials and Methods, it would be useful to summarize the methodologies in a table. Similarly, in Section 3.2. Considered Algorithms and Models, the methods used should be presented in a table or figure for clarity.

• The statement “With this we will be able to make forecasts at a satisfactory accuracy to help detect problems in the poultry pen ahead of time.” may not be convincing to readers who are not directly engaged in the field. It is unclear whether the reported level of accuracy is practically meaningful in real-world operations. The authors should provide further explanation or comparisons with existing literature.

Minor Comments

• Line 108: ChatGPT [?] → Please clarify.

• Line 257: LLama-3.2-3b [?] → Please clarify.

• Abbreviations such as KE and KINLI should be written in full at first mention.

• Figures 3, 4, 5, and 6 lack y-axis labels.

• In Section 6, subsections are unnecessary.

• In Section 2.2.2. Time Series Forecasting using Foundation Models, one paragraph is excessively long and should be split.

• The paragraph structure in Section 3.2. Considered Algorithms and Models is uncomfortable and needs revision.

• Section 5.1. Issues with LLMs in time series forecasting is also not written in an appropriate academic style and requires restructuring.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

 

The manuscript “KINLI: Time Series Forecasting for Monitoring Poultry Health in Complex Pen Environments” introduces a robust, data-resilient forecasting framework for handling noisy sensor data, missing timestamps, and high-variance features in environments.

 

To further strengthen the manuscript’s clarity, rigor, and scientific relevance, the following points should be addressed:

1. Why did transformer models like FEDformer and Autoformer perform worse? Can dataset characteristics explain this and guide model selection?
2. How were noisy or faulty sensor readings handled (e.g., imputation, denoising, outlier removal)? What fraction of data was corrected or removed, and how did this affect model performance? Were robust training or data augmentation strategies applied?
3. Which error metrics were reported? Please provide MAE, RMSE, and MAPE, and explain thresholds relevant to practical poultry health monitoring.
4. What dataset properties explain why some models succeed while others fail?
5. What were the training times, computational requirements, and inference latencies of the top-performing models? Are they suitable for real-time farm deployment?
6. How well do the models generalize across different barns or farms? Were transferability or adaptation strategies considered?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Please answer the reviewer's information carefully one by one, and the modifications made should be reflected in the response to the review comments. The revisions made should be highlighted in red or blue font in the revised draft. Additionally, it is necessary to include references to relevant literature in the introduction.

Author Response

Comment 1: The manuscript reads more like an experimental report based on a large number of experiments. The novelty of the work should be highlighted more clearly.

Response 1: Several passages have been added to the document to illustrate this. A Concluding Remarks/ Novelty Chapter was added

Comment 2: Although the authors argue that the difficulties of preprocessing due to industry-specific data characteristics motivated them to train directly on raw data, the analysis on the potential impact of preprocessing techniques on model performance should be added.

Response 2: In Section a 5 paragraph for this was added

Comment 3: The experimental setup requires more detailed description to ensure reproducibility and clarity.

Resposne 3: In Section 3 more information was added, see the changes in the red colour. More Inforamtion was added to this, Data Origin, Information about the animals and how the Data was processed to use it for Model Training/ Inference

Comment 4: The work uses large language models (LLMs) without fine-tuning. The authors should discuss whether fine-tuning could improve performance and also reflect on whether the relatively small dataset size is suitable for LLM-based approaches.

Response 4: In Section a 5 paragraph for this was added

Comment 5: The authors are advised to carefully proofread and revise the text for accuracy.

Response 5: The text was proofread. An Issue with the wrong Number of Chapters in the Introduction was fixed

Please Note that Major changes are marked in red. Also a Simple Summary was added

Reviewer 2 Report

Comments and Suggestions for Authors

Major:

The authors state that the conclusion has been “overhauled.” However, from the reviewer’s perspective, there is little noticeable difference. The strength and value of studies of this type lie in demonstrating superior predictive performance and practical effectiveness. Yet, the manuscript lacks an objective discussion of these aspects. If such evidence exists, the authors should explicitly highlight it for the reviewer’s consideration.

Minor : 

• Figures 3, 4, 5, and 6 lack y-axis labels.
• In Section 2.2.2. Time Series Forecasting using Foundation Models, one paragraph is excessively long and should be split
• The paragraph structure in Section 3.2. Considered Algorithms and Models is uncomfortable and needs revision.

Author Response

Comment 1: The authors state that the conclusion has been “overhauled.” However, from the reviewer’s perspective, there is little noticeable difference. The strength and value of studies of this type lie in demonstrating superior predictive performance and practical effectiveness. Yet, the manuscript lacks an objective discussion of these aspects. If such evidence exists, the authors should explicitly highlight it for the reviewer’s consideration.

Response:  Several passages have been added to the document to illustrate this. A Concluding Remarks/ Novelty Chapter was added. A Paragraph about the real world Application was added in Chapter 6.2 and one about the Testing of the Application was added in 6.1

Comment 2: Figures 3, 4, 5, and 6 lack y-axis labels

Response 2: The Labels were added to the figures

 

Comment 3: In Section 2.2.2. Time Series Forecasting using Foundation Models, one paragraph is excessively long and should be split

Response 3: In Section 2.2.2 the one paragraph has been split

Comment 4: The paragraph structure in Section 3.2. Considered Algorithms and Models is uncomfortable and needs revision.

Response 4: In the paragraph spaces has been added to improve readebility

 

Please Note that Major changes are marked in red. Also a Simple Summary was added

Reviewer 3 Report

Comments and Suggestions for Authors

All of my questions have been fully addressed in the author's response.

Author Response

Comment 1: All of my questions have been fully addressed in the author's response.

Answer: All of the reviewer's questions have already been answered.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

It's ok.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have appropriately modified the manuscript according to the reviewer's comments. The revised manuscript is now satisfactory. I would like to thank the authors for their efforts.< !-- notionvc: 85b2720a-1365-48dc-af91-2775a12756a1 -->