Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article analyzes the problems of increasing soil moisture stress due to climate change. To model and assess soil moisture conditions, it is proposed to use multispectral plant images in two wavelength ranges: the green region (530–570 nm) and the red-edge region (700–710 nm), which are most sensitive to heat stress. The proposed platform utilizes eight different classification models. Experiments were conducted under real-world conditions in a phytotron, analyzing images of leguminous plant leaves at three stages: before, during, and after heat stress.

The article can be published after the following comments are corrected:

- A significant portion of the references are provided artificially without proper analysis of their content: [15–23], [12, 23, 25, 27–30], [25, 34, 35, 39], [3, 12, 13, 30, 42]

- In the review, the authors periodically compare the analyzed method with a traditional approach, but what that traditional approach it is not specified. 

- Well-known mathematical approaches were used, so their analytical description is missing from the article. However, it is recommended to include at least a formal statement of the problem in the article.

-In Figure 4, stage 4 shows the classification models, while stage 5 evaluates the algorithms. The terminology used should be standardized.

- Figure 2 uses an inappropriate visualization method for the results. It's unclear why the range of estimates is provided, rather than the accuracy value. The use of two white rectangles in a single line is particularly misleading.

- The experiments were conducted in a phytotron with high-resolution leaf imaging. For future studies, the authors should consider how image parameters will change during aerial and space photography.

- Remote sensing of the Earth has a higher acquisition rate but lower image resolution. Cloud cover and other factors also limit the use of space imaging. In this regard, the authors correctly conclude that UAVs should be used. However, if arid regions and large areas are being analyzed, the choice of imaging method should be substantiated in more detail. A similar study was conducted in the paper https://doi.org/10.15622/ia.23.4.11

Author Response

Comment 1: A significant portion of the references are provided artificially without proper analysis of their content: [15–23], [12, 23, 25, 27–30], [25, 34, 35, 39], [3, 12, 13, 30, 42]

Our Response: We thank the reviewer for this important observation. We agree that simply listing a large block of citations does not sufficiently demonstrate our engagement with the literature. We have substantially revised the Introduction section. The large citation blocks have been broken down, and the references are now grouped thematically. We have added descriptive text to explain how these specific studies support our statements, categorizing them by application (e.g., disease detection, nutrient analysis, water stress monitoring). This change provides a much clearer context and strengthens our literature review.

Comment 2: In the review, the authors periodically compare the analyzed method with a traditional approach, but what that traditional approach it is not specified.

Our Response: We appreciate the reviewer pointing out this lack of clarity. The term "traditional approach" was indeed too vague. We have now explicitly defined what we mean by "traditional approach" in the Introduction. We state that this term refers to established chemometric techniques, specifically Principal Component Analysis (PCA) and Partial Least-squares Regression (PLSR). We have also ensured this specific terminology is used consistently throughout the manuscript to avoid ambiguity.

Comment 3: Well-known mathematical approaches were used, so their analytical description is missing from the article. However, it is recommended to include at least a formal statement of the problem in the article.

Our Response: We thank the reviewer for this  suggestion. We have added a concise, formal problem statement at the beginning of Section 2.3, "Machine Learning Analysis." This new text mathematically defines our task as a supervised multi-class classification problem by formally describing the Input Space ($X$), Output Space ($Y$), and the Objective (learning a mapping function $f: X \to Y$). This addition frames the problem clearly without adding unnecessary descriptions of the well-known algorithms.

Comment 4: In Figure 4, stage 4 shows the classification models, while stage 5 evaluates the algorithms. The terminology used should be standardized.

Our Response:Thank you for noticing this inconsistency. We agree that the terminology should be standardized for clarity. We have revised the flowchart. In stage 5, the label "Algorithm Evaluation" has been changed to “Model Evaluation”. This ensures consistent and accurate terminology throughout the diagram, as we are evaluating the performance of the trained models.

Comment 5: Figure 2 uses an inappropriate visualization method for the results. It's unclear why the range of estimates is provided, rather than the accuracy value. The use of two white rectangles in a single line is particularly misleading.

Our Response: We agree  that the original Figure 2 was confusing and not the best choice for presenting the results. We have replaced the original stacked bar chart in Figure 2 with a standard grouped bar chart. The new figure now clearly and directly displays the per-class classification accuracy for each of the four best-performing models. We have also updated the figure caption to accurately describe the new visualization. This change removes all ambiguity and makes the performance comparison much more straightforward.

Comments 6 & 7: The experiments were conducted in a phytotron with high-resolution leaf imaging. For future studies, the authors should consider how image parameters will change during aerial and space photography.

Remote sensing of the Earth has a higher acquisition rate but lower image resolution. Cloud cover and other factors also limit the use of space imaging. In this regard, the authors correctly conclude that UAVs should be used. However, if arid regions and large areas are being analyzed, the choice of imaging method should be substantiated in more detail. A similar study was conducted in the paper [https://doi.org/10.15622/ia.23.4.11](https://doi.org/10.15622/ia.23.4.11)

Our Response: We thank the reviewer for these  comments. We have significantly expanded the final paragraph of the Discussion section to address these points. The revised text now explicitly discusses the challenges of translating our lab-based results to field applications, noting the influence of soil background, canopy structure, and atmospheric conditions on spectral data. It also provides a more detailed justification for using UAVs as the next logical step, and acknowledges the limitations of UAVs for monitoring very large areas and discusses the trade-offs of using satellite imagery. Laslty, we integrated the suggested reference (Verkhoturov et al., 2024) to propose future research directions, such as a multi-sensor approach that could combine hyperspectral UAV data with all-weather satellite radar for robust, large-scale monitoring.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper investigates the detection of thermal stress in bean plants using machine learning and hyperspectral data. The authors proposed a well-designed and rigorously executed research framework. The study successfully compared the performance of various machine learning algorithms and precisely identified the key spectral regions most sensitive to thermal stress. The following are some specific suggestions:

1. Random Forest is typically one of the most powerful algorithms for handling tabular and high-dimensional data, with its performance often comparable to or even surpassing that of ANNs. However, in this study, the accuracy of RF was only 65.7%, far below that of DT, SVM, and ANN. This result is highly anomalous. It is recommended that the author(s) add a discussion on why RF performed poorly and provide possible reasons.
2. In the data splitting stage, what was the unit for partitioning the dataset? Was the partitioning done at the plant or pot level? It is suggested to provide more detailed explanations.
3. The study identified the green region (530-570 nm) and the red-edge region (700-710 nm) as key spectral bands. This is a very valuable finding. However, the discussion section could delve deeper. Are these bands a specific response to thermal stress, or a general response of plants to various stresses (such as water stress, nutrient stress)?
4. The author(s) correctly pointed out at the end of the discussion that the study was conducted under controlled laboratory conditions, which is a limitation. This point could be discussed in more detail. When applying this framework to a real-world field environment, what are the specific challenges besides factors like soil and canopy? This would give readers a more comprehensive understanding of the practical application prospects of this technology.
5. It is recommended to revise the caption for Figure 2. A clearer description would be better.
6. In the first sentence of the conclusion section, there is a typo: "...thermal-stressed bean plants.e. Data from...". It should be "...thermal-stressed bean plants. Data from...". Please proofread the entire manuscript carefully to correct such minor spelling or grammatical errors.

Author Response

Comment 1: Random Forest is typically one of the most powerful algorithms for handling tabular and high-dimensional data, with its performance often comparable to or even surpassing that of ANNs. However, in this study, the accuracy of RF was only 65.7%, far below that of DT, SVM, and ANN. This result is highly anomalous. It is recommended that the author(s) add a discussion on why RF performed poorly and provide possible reasons.

Our Response: We thank the reviewer for this  observation. We agree that the underperformance of the Random Forest (RF) algorithm is a counter-intuitive result that warrants discussion. We have added a new paragraph in the Discussion section specifically addressing this point. We propose several potential reasons for this outcome, including the challenges posed by highly correlated features in hyperspectral data, which can dilute the effectiveness of RF's feature subsampling process. We also contrast this with the inherent strengths of the Artificial Neural Network (ANN) in learning complex patterns directly from such data, which helps explain the performance gap we observed.

Comment 2: In the data splitting stage, what was the unit for partitioning the dataset? Was the partitioning done at the plant or pot level? It is suggested to provide more detailed explanations.

Our Response: We thank the reviewer for raising this critical point. We apologize that our original description was not sufficiently clear. We have completely revised the relevant paragraph in Section 2.3, "Machine Learning Analysis," to provide a detailed explanation of our data splitting strategy. We now explicitly state that the partitioning was performed at the pot level to prevent any data leakage and to ensure a robust evaluation of the models' generalization capabilities. We specify that the 36 pots were randomly split into training (29 pots) and testing (7 pots) groups, ensuring that the models were tested on entirely new plants not seen during training.

Comment 3: The study identified the green region (530-570 nm) and the red-edge region (700-710 nm) as key spectral bands. This is a very valuable finding. However, the discussion section could delve deeper. Are these bands a specific response to thermal stress, or a general response of plants to various stresses (such as water stress, nutrient stress)?

Our Response: We agree that this point deserves a more in-depth discussion. We have added a new, detailed paragraph to the Discussion section to address this question. We acknowledge that the physiological responses in the green and red-edge regions (i.e., chlorophyll degradation and changes in leaf structure) are indeed general stress indicators. However, we argue that the exceptional accuracy of our framework suggests that the machine learning models were able to capture the unique patterns, magnitudes, and dynamics within these spectral regions that are characteristic of thermal stress. We also propose a future controlled experiment to definitively isolate the unique spectral signature of thermal stress from other common stressors.

Comment 4: The author(s) correctly pointed out at the end of the discussion that the study was conducted under controlled laboratory conditions, which is a limitation. This point could be discussed in more detail. When applying this framework to a real-world field environment, what are the specific challenges besides factors like soil and canopy? This would give readers a more comprehensive understanding of the practical application prospects of this technology.

Our Response: We thank the reviewer for encouraging us to expand on this  limitation. We have  expanded the final paragraphs of the Discussion section to provide a more detailed account of the challenges in transitioning this technology from the lab to the field. Beyond the influences of soil and canopy, we now discuss specific obstacles such as: (1) variability in illumination (sun angle, clouds) and atmospheric effects; (2) the complexity of the 3D canopy structure (leaf angles, wind); and (3) the presence of confounding biological factors, where multiple stressors may occur simultaneously.

Comment 5: It is recommended to revise the caption for Figure 2. A clearer description would be better.

Our Response: We agree. The original caption was unclear, especially in light of the confusing chart format. In conjunction with redesigning Figure 2 into a clearer grouped bar chart (as also suggested by Reviewer 1), we have completely rewritten the caption. The new caption is now more descriptive and accurately explains what the figure represents: the per-class classification accuracy for our best-performing models.

Comment 6: In the first sentence of the conclusion section, there is a typo: "...thermal-stressed bean plants.e. Data from...". It should be "...thermal-stressed bean plants. Data from...". Please proofread the entire manuscript carefully to correct such minor spelling or grammatical errors.

Our Response: We thank the reviewer for catching this error. We have corrected the specified typo in the Conclusion section. Furthermore, we have performed a thorough proofreading of the entire manuscript to identify and correct any remaining minor spelling or grammatical errors.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Firstly, I must commend the authors for their excellent work in responding to the first round of review comments. The quality of the manuscript has been significantly improved, especially in terms of methodological transparency (clarifying the division of data based on flower pots), discussion of anomalous results (in-depth analysis of the reasons for the poor performance of random forests), and an in-depth discussion of the findings and limitations of the study. The authors' revisions were careful, thorough and constructive, which made the manuscript much more scientifically rigorous and readable. My remaining comments are aimed at further polishing the depth of the article's exposition and enhancing the impact of its conclusions.

1. It is very important that the authors do a good job of explaining why Random Forests (RF) perform poorly in the discussion section. In order to further enhance the depth of the discussion, it is suggested that the authors could give more thought and elaboration to the following question: Why does ANN outperform other models that perform equally well? For example, what are the advantages of ANN over Decision Trees (DTs) and CN2 Rule Induction, which also perform well?
2. In addition to the direct recommendation to use UAVs for validation, a more practical intermediate step would be to design and test a new, simplified vegetation index based on the two key bands identified in this study, 530-570 nm and 700-710 nm. This index may be more sensitive to heat stress than existing generic indices such as NDVI. This would be the perfect bridge between sophisticated hyperspectral analyses and future low-cost, multispectral portable sensors or UAV applications. Presenting this specific next research direction at the end of the "Conclusion" or "Discussion" will greatly enhance the practical guidance value of this paper.
3. To further enrich the discussion and contextualize the findings within recent advancements in intelligent agriculture, the authors might consider incorporating the following relevant studies:

[Zhu, H., Lin, C., Dong, Z., Xu, J. L., & He, Y. (2025). Early Yield Prediction of Oilseed Rape Using UAV-Based Hyperspectral Imaging Combined with Machine Learning Algorithms. Agriculture, 15(10), 1100.]

[Zhu, H., Lin, C., Liu, G., Wang, D., Qin, S., Li, A., ... & He, Y. (2024). Intelligent agriculture: Deep learning in UAV-based remote sensing imagery for crop diseases and pests detection. Frontiers in Plant Science, 15, 1435016.]

Author Response

Thank you very much for your  feedback. 

We have carefully considered and addressed your remaining comments. Below is a point-by-point response detailing the changes made in this second revision.

Comment 1: It is very important that the authors do a good job of explaining why Random Forests (RF) perform poorly in the discussion section. In order to further enhance the depth of the discussion, it is suggested that the authors could give more thought and elaboration to the following question: Why does ANN outperform other models that perform equally well? For example, what are the advantages of ANN over Decision Trees (DTs) and CN2 Rule Induction, which also perform well?

Our Response: Thank you for this suggestion. We have expanded the Discussion section to include a more detailed explanation of why the Artificial Neural Network (ANN) might have outperformed the also well-performing CN2 Rule Induction and Decision Tree (DT) models. The revised text (integrated into Paragraph 2 of the Discussion) elaborates on the ANN's inherent strengths in modeling complex non-linear relationships and performing implicit feature weighting across numerous correlated bands, which is particularly advantageous for hyperspectral data, potentially allowing it to capture the subtle signature of thermal stress with slightly higher fidelity than the rule- or tree-based approaches. We also refined the subsequent paragraph (Paragraph 3) comparing ANN and RF, focusing on hyperparameter sensitivity for RF and the baseline comparison context.

Comment 2: In addition to the direct recommendation to use UAVs for validation, a more practical intermediate step would be to design and test a new, simplified vegetation index based on the two key bands identified in this study, 530-570 nm and 700-710 nm... Presenting this specific next research direction at the end of the "Conclusion" or "Discussion" will greatly enhance the practical guidance value of this paper.

Our Response: We agree that proposing a targeted vegetation index  enhances the translational value of our findings. We have added a specific recommendation to the Conclusion section. The added text explicitly proposes the formulation and field-testing of a novel vegetation index based on the identified green (530–570 nm) and red-edge (700–710 nm) regions as a key next step. We highlight its potential as a cost-effective tool and a bridge between hyperspectral analysis and more accessible multispectral sensors.

Comment 3: To further enrich the discussion and contextualize the findings within recent advancements in intelligent agriculture, the authors might consider incorporating the following relevant studies: [Zhu et al., 2025; Zhu et al., 2024]

Our Response: Thank you for suggesting these recent studies. We have incorporated both suggested citations ([Zhu et al., 2025] and [Zhu et al., 2024]) into the final paragraph of the Discussion section. 

Thank you once again for your time and expertise.

Sincerely,
Lucas Prado Osco (on behalf of all authors)
University of Western São Paulo
Presidente Prudente, SP, Brazil
lucasosco@unoeste.br