Hyperspectral Imaging for Phenotyping Plant Drought Stress and Nitrogen Interactions Using Multivariate Modeling and Machine Learning Techniques in Wheat

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

Thank you for your valuable effort in this study. I'm very happy to review this manuscript and contribute with some suggestions to improve it.

To be honest, I'm not an expert in this field and cannot make suggestions related to the understanding of the plants and their interactions with nitrogen and water levels. Therefore, my comments are related to the formal presentation of the results and some questions that I couldn't understand well.

Overall, the manuscript is well written. The introduction provides enough background to justify this work. The methodology is well structured, and I don't have comments related to this. However, some improvements need to be considered before accepting this manuscript for publication:

• The abstract needs to be rewritten. Please clarify the main goal of this study and include the main results and the main conclusion.

• The results section should be reviewed completely. For me, it has been impossible to follow the thread of the manuscript. The order and numbering of the figures are incorrect. It's too hard to understand the results because there is no correspondence between the numbers included in the text and the numbers in the figure captions.

• I highly encourage the authors to consider condensing the results a bit, especially section 3.5.1. Please include only the main results and key figures.

• With respect to the discussion, it would be beneficial to include the main implications of this study and discuss the applicability of this methodology on a larger scale.

In the attached PDF file, you can find more comments.

Good luck.

Best regards,

Comments for author File: Comments.pdf

Author Response

                                   Please see the attachment

These are responses to comments made by the reviewer. We appreciate the reviewer for the comments and inputs. The black coloured writings are the reviewers comments while the red are the response to the comments

Reviewer 1

1. The abstract needs to be rewritten. Please clarify the main goal of this study and include the main results and the main conclusion

Response: The abstract has been rewritten spelling out clearly the main goal of the study including the key results and the conclusion as suggested by the reviewer. This can be found on

Page: 1

Line: L12- L38

2. The results section should be reviewed completely. For me, it has been impossible to follow the thread of the manuscript. The order and numbering of the figures are incorrect. It's too hard to understand the results because there is no correspondence between the numbers included in the text and the numbers in the figure Response: The results section has been reviewed thoroughly and corrections has been made on the order and numbering of the figures

Page 12-23,

Line L402- L699

 

3. I highly encourage the authors to consider condensing the results a bit, especially section 3.5.1. Please include only the main results and key figures

Response: As suggested by the reviewers, section 3.5.1 has been condensed highlighting only the main results

Page 19,  Line L586- L619

4. With respect to the discussion, it would be beneficial to include the main implications of this study and discuss the applicability of this methodology on a larger scale

Response: The discussion section has been reviewed highlighting the main implications of the study and the applicability of the proposed methodology on a large scale as suggested by the reviewer.

Page 19,  Line L586- L619

 

In the attached PDF file, you can find more comments

Response:  The comments in the pdf version have been addressed. The Figure labels have been correctly remedied.                                                                                                                                                                       

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

General comments:

The article presents a comprehensive study on using hyperspectral imaging and machine learning models to detect drought stress in wheat under varying nitrogen levels. The study proposes new vegetation indices (VIs) derived from sensitive spectral features selected via a custom-designed ensemble modeling technique. These new VIs, combined with machine learning models (support vector machines, random forest, and deep neural network), demonstrated superior performance in classifying drought stress compared to known VIs, achieving accuracies above 0.94. The combined VIs were also effective in regression models for predicting stomatal conductance and photosynthetic rates, with the random forest regression model performing best. Overall, the study offers valuable insights into improving drought stress detection in crops. I have only minor suggestions for improvement, which are detailed in my comments below.

 

Minor issues:

Introduction:

1. Line 69, please add the full name for VIs (vegetation indices) when you first used it. This is the similar case for other abbreviations like RF in line 107.

2. Line 116, so this study is to use VIs to detect N stress or water stress? It is a little bit confusing since your title is about ‘drought stress’.

 

Methods:

1. You may need to add a full name or explanation for Pn and Gs in description of Figure 1. (photosynthetic rate (Pn) and stomatal conductance (gs)).

2. I think the contents in section 2.4 for the pre-processing of hyperspectral images can be moved to supplementary since they are not the main parts of this study (but more like technical details).

3. Line 254, the ‘drought stress’ is still a little bit confusing since it sounds like it is ‘water stress’, but actually here the ‘drought stress’ should be the colimitation from both spill water and soil nitrogen, right?

4. Which methods you use belong to machine learning methods, and which belongs to statistical methods? 

5. Line 278, there are two ‘Figure 1’, should be ‘Figure 2’? 

6. Reference for the SHAP algorithm in line 273?

7. Why are there two ‘Input: 30 features Output:top 10 features’ in Figure 2, what are the differences?

8. Please make sure all your equations are left-aligned (e.g. Equation 4).

9. Line 285-286, so here you proposed three new indexes? They are pretty similar, so could you please add some explanations as to why you proposed these three different indexes (different considerations for different indexes)?

10. Please avoid providing the full name and abbreviation multiple times in your manuscript. E.g. Line 301-303, the name for these different machine learning models has been proposed in the introduction part.

11. I suggested using a table to list the characteristics (and also why you choose each of this model in the study) of the machine learning models you used in section 2.8 (since there are some long and repetitive contents for these methods.

12. Line 335, so in this study, you have a strong assumption that the gas exchange (measure as gs and Pn) reflects the drought stress, rather than other stress right? And you use different models to predict gs and Pn using the same inputs which are the VIs (and the VIs is calculated from the processed hyperspectral data). 

13. Line 362, you have used R2 before this line but did not provide the full name, so please double check such similar usage throughout the manuscript.

14. Line 380, the format of the equation for ‘where…’ seems not correct (be left-aligned).

15. Methods part is too long and should be concise and focus on the main story line for this study.

 

Results:

1. Line 427-429. Please add the detailed explanation for these four different experiments designed in the corresponding texts for Figures 2 and 3.

2. I suggest combining the result and discussion part together, especially for contents in 3.1 (there are very interesting interactions between soil water and nitrogen on the measured Pn and gs.

3. I think the way the author presents the result part is a little bit different from the logic present in Figure 1. I think if the author could follow the same logic to present the results, it would be much better for readers to follow (e.g. first show the hyperspectral spectral reflectance and then the extrapolated VIs, and …., just as described in Figure 1).

4. In Figure 6, please capitalize the first letter for ‘Stomatal conductance’ to make it consistent with ‘Photosynthetic rate’.

5. Line 568, so here you consider nitrogen and drought stress respectively?

6. There are no obvious differences between the subfigures of Figure 7, so may consider changing the color scheme or other types of figure to demonstrate the differences/improvements between these model training strategies. 

8. You may also need to provide some other statistical indicator for results in Figure 8, e.g. RMSE. The high R2 is largely contributed by a few very small and large ga measurement points. This is the same case for Figure 9.

9. So the results in Figures 8 and 9 are predicted by using different VIs for gs and Pn. Does the author directly use the processed hyperspectral data to predict gs and Pn?

 

Comments on the Quality of English Language

There are some minor issues for some standard usage of the words, please see my "Comments and Suggestions for Authors".

Author Response

Please see the attachment for response to reviewer's comments

Reviewer 2

These are responses to comments made by the reviewer. We appreciate the reviewer for the comments and inputs. The black coloured writings are the reviewers comments while the red are the response to the comments

Introduction

1. Line 69, please add the full name for VIs (vegetation indices) when you first used it. This is the similar case for other abbreviations like RF in line 107.

Response: The full name of VIs was stated on L69 which is vegetation indices. That is‘’One standard method for analysing HSI data in plant phenotyping is to extract vegetation indices (VIs), which are mathematical combinations of spectral reflectance characteristics of vegetation at different wavelengths’’.

Page: 2 Line: L70- L71

2. Line 116, so this study is to use VIs to detect N stress or water stress. It is a little bit confusing since your title is about ‘drought stress’.

Response: This study utilizes VIs to detect water stress of plants under different N levels. It helps in understanding the dynamic interactions of drought stress and different N levels showing how they affect the reflectance of plant canopy.

Methods

1. You may need to add a full name or explanation for Pn and Gs in description of Figure 1. (photosynthetic rate (Pn) and stomatal conductance (gs)).
2. Response: Full name of Pn and gs has been added to the description of Figure 1 as suggested by the reviewer.

Page: 4   Line: L137 -L140

2. I think the contents in section 2.4 for the pre-processing of hyperspectral images can be moved to supplementary since they are not the main parts of this study (but more like technical details).

Response: We humbly disagree with the reviewer on this comment. Section 2.4 highlights the pre-processing steps of the hyperspectral data. Details of how the pre-processing was done is required to help readers know how plant pixel data were obtained before the VIs were extracted. Moreover, this section has been condensed giving references for further readings.

3. Line 254, the ‘drought stress’ is still a little bit confusing since it sounds like it is ‘water stress’, but actually here the ‘drought stress’ should be the colimitation from both spill water and soil nitrogen, right?

Response: This study is based on drought stress which occurs when there is a prolonged period without sufficient rainfall or irrigation, leading to a significant reduction in available soil moisture. Water stress is a broader term that refers to any situation where the water supply to the plant is inadequate for its needs, regardless of the cause. Since the experiment was based on withholding water from plants for a period of time, using drought stress is suitable as compared to water stress. The nitrogen measurement is to understand the interactions of drought stressed plants in variable N conditions.

4. Which methods you use belong to machine learning methods, and which belongs to statistical methods?

Response: The random forest, support vector machines, deep learning networks all fall under machine learning methods while the Partial least square expression, principal components analysis and the ANOVA analysis belongs to the statistical methods.

5. Line 278, there are two ‘Figure 1’, should be ‘Figure 2’?

Response: The error has been corrected. All the figures have been properly labelled.

6. Reference for the SHAP algorithm in line 273?

Response: Reference for the SHAP algorithm has been provided as suggested

Page:8  Line: 278

7. Why are there two ‘Input: 30 features Output: top 10 features’ in Figure 2, what are the differences?.

Response: Figure 2 describes the process of spectral sensitivity selection using ensemble learning approach. The first three models (CFS, chi-square, ReliefF) receives over 200 spectral data and selects the top 10 most sensitive spectral wavelengths. This results in 30 spectral wavelengths(10 from each model) which is used as input to the Boruta SHAP algorithm to extract the top 10 (ranked by the RF-RFE model). Hence the final output from the ensemble model is 10 spectral features (wavelengths).

8. Please make sure all your equations are left-aligned (e.g. Equation 4)..

Response: All the equations have been left-aligned as suggested by the reviewer.

9. Line 285-286, so here you proposed three new indexes? They are pretty similar, so could you please add some explanations as to why you proposed these three different indexes (different considerations for different indexes)?

Response: The three equations were quite distinct: Equations 1, 2 and 3 are simple ratio, normalized difference and absolute difference respectively. The equations are based on general frameworks commonly used in plant spectral phenotyping , and was  adapted for this work. It has been used by different researchers including to propose new indices depending on the sensitive wavebands selected and the phenotypic traits under study.

10. Please avoid providing the full name and abbreviation multiple times in your manuscript. E.g. Line 301-303, the name for these different machine learning models has been proposed in the introduction part.

Response: This has been rectified. The full name with its abbreviations have been provided once and the abbreviations used in the subsequent as suggested.

11. I suggested using a table to list the characteristics (and also why you choose each of this model in the study) of the machine learning models you used in section 2.8 (since there are some long and repetitive contents for these methods

Response: Section 2.8 has been reformatted. As suggested by the reviewer, a table listing the characteristics of the models and why they were chosen for this study has been provided as a supplementary Table S1. 

12. Line 335, so in this study, you have a strong assumption that the gas exchange (measure as gs and Pn) reflects the drought stress, rather than other stress right? And you use different models to predict gs and Pn using the same inputs which are the VIs (and the VIs is calculated from the processed hyperspectral data). 

Response: Yes there is a scientific evidence proving the use of gas exchange measurements such as stomatal conductance and photosynthetic rate in assessing drought stress in plants (reference). Moreover, the above-mentioned gas exchange measurements are also good indicators to measure nitrogen deficiency in plants. Hence, they were first used to analyze drought stress of plants in variable N concentrations. Then using combined VIs (proposed and Known VIs), three machine learning regression models (SVR, RFR and PR) while PLSR was also developed to predict Pn and gs.

Please note: The VIs are calculated form the processed hyperspectral data

13. Line 362, you have used R² before this line but did not provide the full name, so please double check such similar usage throughout the manuscript.

Response: The full name for R² has been provided on line 310 and other similar areas in the manuscript as required

Page: 9  Line: L 310

14. Line 380, the format of the equation for ‘where…’ seems not correct (be left-aligned).

Response: The position of the equation on L380 has been changed. It has been left aligned as suggested by the reviewer. This is now on line L376

Page:11 Line: L376

15. Methods part is too long and should be concise and focus on the main story line for this study.

Response: The Method section has been reviewed. Section 2.8 have been reviewed and part put in a tabular form as supplementary Table. This has ensured the section is comprehensive yet concise. 

Results

1. Line 427-429. Please add the detailed explanation for these four different experiments designed in the corresponding texts for Figures 2 and 3.

Response: Detailed explanation of the four treatments have been added to Figure 2 and 3 now Figure 3 and 4 respectively as suggested by the reviewer.

Page:13

Line: L39-442 and L446-448

2. I suggest combining the result and discussion part together, especially for contents in 3.1 (there are very interesting interactions between soil water and nitrogen on the measured Pn and gs.

Response: From the journal structure and regulations, the results and discussion section are to be distinct. Combining these two sections might go against what the journal instructs

3. I think the way the author presents the result part is a little bit different from the logic present in Figure 1. I think if the author could follow the same logic to present the results, it would be much better for readers to follow (e.g. first show the hyperspectral spectral reflectance and then the extrapolated VIs, and …., just as described in Figure 1).

Response: We understand the results part is a bit confusing and a note of description of the first two sections (which are deviations of the workflow of Figure 1 has been provided). It is important to understand the effect of drought stress on the different gas exchange measurements which are used as ground truth measurement in subsequent analysis. It is also important to understand the effect of drought and variable N on the whole spectra of each treatments before the extraction of the individual VIs. Hence section 3.1 and 3.2.

Page: 12 Line: L402 - L411

4. In Figure 6, please capitalize the first letter for ‘Stomatal conductance’ to make it consistent with ‘Photosynthetic rate’.

Response: The first letter for ‘Stomatal Conductance’ has been capitalized as suggested by the reviewer.

Page: 16 Line: 521

5. Line 568, so here you consider nitrogen and drought stress respectively?

Response: Line 568 considers drought stress in particular in two nitrogen levels (high and low nitrogen). Here the key goal is to analyze drought stress in variable nitrogen conditions.

6. There are no obvious differences between the subfigures of Figure 7, so may consider changing the color scheme or other types of figure to demonstrate the differences/improvements between these model training strategies. 

Response:

There are differences between the subgroups in Figure 7. Each subgroup represents an item/trait of study. For instance, Figure 7a, analyses the ML model trained using only the known VIs whereas Figure 7b is for the performance analysis of ML for proposed indices. The confusion matrices are displayed as heatmaps, with the classification accuracy increasing with the depth of colour. Hence the deeper the diagonal heatmaps, the better the performance of the models. Same colour were used for uniformity.

7. You may also need to provide some other statistical indicator for results in Figure 8, e.g. RMSE. The high R2 is largely contributed by a few very small and large ga measurement points. This is the same case for Figure 9.

Response: Other metrics for the measurements of the regression models in Figure 8 and 9 now Figure 10 and 11 have already been provided as Table 5.

Page: 22 Line: L688 – L702

8. So the results in Figures 8 and 9 are predicted by using different VIs for gs and Pn. Does the author directly use the processed hyperspectral data to predict gs and Pn?

Response: The results in Figure 8 and 9 which is now Figure 10 and 11 respectively are predicted using different VIs for gs and Pn. The VIs are of three parts: the known VIs, proposed VIs and combined VIs(from combining known and proposed VIs). They were all derived from the processed hyperspectral data  

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript presents the results of an experiment in which two levels of nitrogen and two levels of drought are tested to study the spectral response of the wheat crop and to develop indices for drought detection. It is suggested to incorporate wheat and nitrogen in the title or justify why they are not considered.

The introduction is well structured, shows the background of similar works, and presents the problem. However, it could be strengthened by deepening the importance of the study of N and its possible interactions with other variables.

Some aspects of the methodology are missing. It is suggested to specify how the irrigation water supply was made in the water stress treatments, e.g., it is mentioned that the water supply was cut off, but it is unclear if it was gradual, constant, and for how long.

In addition, it is recommended that the reasons for using only two levels of nitrogen and two levels of water stress be clarified. Alternatively, justify and support why more levels were not included. Including more levels would allow a better understanding of the water-nitrogen stress interaction.

It is recommended to describe in more detail the equipment used for hyperspectral imaging, especially the characteristics of the hyperspectral camera (radiometric, spectral, and spatial resolution), and to detail if the camera has its power source or if natural solar light was used. It is suggested that a photograph of the equipment used and a sample of the images acquired be added since they are the primary data sources for the work.

Any author suggested the normalized difference ratio between 910 and 950 nm wavelengths. Is there any reference to support this range?

L243-245. There is a lack of information on the application of the segmented images. Were they used to generate the values of VIs, or what was their final application?

L278. Revise the numbering of the figures. Figure 2 instead of Figure 1

L268, this sentence was already mentioned in L265-266.

L285-L287. It is suggested to support why these three types of equations were used.

L362-366. The description of the R2, RMSE, and MAE threshold values is missing, making it unclear when the model is valid or invalid.

The titles of all figures and tables need to be revised as they do not match what is described in the text; e.g., L391 says Figure 3 but should be Figure 2. Examples like the above are in several places in the manuscript.

Figures 2 and 3 show results from a one-way ANOVA analysis, but the methodology does not describe this process. This process should be explained where appropriate.

L435-455. Terms such as high, low, visibly different, show peaks, high, …., are frequently used to describe spectral characteristics. Still, the quantities in reflectance units are not mentioned, which can lead to confusion and subjective interpretation.

In Figure 4, the differences could be better represented by normalizing the reflectance values. In addition, the authors should evaluate the possibility of including a graph for each treatment (WWLN, WWHN, DSHN, and DSLN) that consists of the spectral reflectance at 0 DADS, 6 DADS, and 15 DADS to facilitate the visualization of the differences throughout the phenological cycle. 

L506-508. It is unclear to which wavelength pairs of the color map shown in Figure 7

The gas exchange analysis appears to be part of another experiment that was not necessarily drought-related. For example, section 3.3 discusses the effects of drought on gs and Pn gas exchange changes but does not note that those results are used posteriorly in estimating plant stress affect detection from hyperspectral images. Thus, this information remains only secondary or ancillary.

L554 and L556. The value of r seems to be incorrect. It should be -0.78 and -0.67, respectively.

Table E-1 and L557-556. It is suggested that a unique name or code be assigned to each index in Table E-1 to match the text describing "Figure 6. Correlations between the proposed indices and the Pn and gs measurements" to facilitate the reader's interpretation.

The discussion section is weak because it only describes the results again but does not discuss at length issues such as the reasons for the changes in spectral characteristics that may be due to a direct effect of drought, an effect of nitrogen concentration, or a combined effect of both factors.

The discussion does not delve into the analysis of the VIs used, and the reasons why the selected indices were the best to estimate the effect of drought is absent. In addition, it is suggested that the discussion and analysis of the spectral regions that showed better results in the correlation with drought and nitrogen be extended; this could help to understand the processes occurring in the plant.

A discussion on the ML models tested to estimate drought's effect is missing. A comparison of the models used could be included to explain the reasons for these results and broaden the knowledge of the analysis processes of this type of data. 

Author Response

Please see the attachment for response to reviewers comments

These are responses to comments made by the reviewer. We appreciate the reviewer for the comments and inputs. The black coloured writings are the reviewers comments while the red are the response to the comments

Reviewer 3

1. The manuscript presents the results of an experiment in which two levels of nitrogen and two levels of drought are tested to study the spectral response of the wheat crop and to develop indices for drought detection. It is suggested to incorporate wheat and nitrogen in the title or justify why they are not considered.

Response: We agree the reviewer that wheat and nitrogen terms should be incorporated in the title since the study was done on wheat and variable nitrogen levels. .. The title is now changed to ‘Hyperspectral Imaging for Phenotyping Plant Drought Stress Using Multivariate Modelling and Machine learning Techniques.

Page:1 Line: L113- 121

2. The introduction is well structured, shows the background of similar works, and presents the problem. However, it could be strengthened by deepening the importance of the study of N and its possible interactions with other variables

Response: The introduction section has been reviewed to include the effects of interactions of the nitrogen and drought stress on plants as well as the importance of the study. This has improved the introduction section of this paper.

Page:3 Line: L105- L113

3. Some aspects of the methodology are missing. It is suggested to specify how the irrigation water supply was made in the water stress treatments, e.g., it is mentioned that the water supply was cut off, but it is unclear if it was gradual, constant, and for how long

Response: Further notes have been added to the methodology section to include how the drought was imposed and how the moisture content of the  the well-watered plants were maintained.

Line: L172-L173 and L178-L179

4. In addition, it is recommended that the reasons for using only two levels of nitrogen and two levels of water stress be clarified. Alternatively, justify and support why more levels were not included. Including more levels would allow a better understanding of the water-nitrogen stress interaction.

Response: This study is preliminary research aimed at using non-invasive methods to understand the interactions of drought and variable N levels and their effect on spectral reflectance of plants. Although there are several studies on drought or plant N level analysis, studies involving the use of non-destructive image-based approaches to understand multiple stress dynamics in plants is at the infancy stage. Hence it was prudent in considering the two extremes of the multiple stresses: drought and well-watered (for drought stress) and high and low N (for nitrogen deficiency). This is the basis for further studies which may consider other levels of nitrogen and water.

5. It is recommended to describe in more detail the equipment used for hyperspectral imaging, especially the characteristics of the hyperspectral camera (radiometric, spectral, and spatial resolution), and to detail if the camera has its power source or if natural solar light was used. It is suggested that a photograph of the equipment used and a sample of the images acquired be added since they are the primary data sources for the work.

Response: Further details of the camera specifications have been given. References on the use of this type of camera have been provided to aid readers understand how this type of camera operates and its purpose of use in plant phenotyping studies. Sample of the pseudo image of the original and hyperspectral data is shown on supplementary Figure S1 on page 33, Line 1058-1063

Page: 6  Line: L216 -L225

6. Any author suggested the normalized difference ratio between 910 and 950 nm wavelengths. Is there any reference to support this range?

Response: Several studies have utilized spectral differences combined with Otsu thresholding to segment hyperspectral data. In most cases, the segmentation approaches attempt a pixel level classification based on the spectral signature of each pixel. From, the study of Okyere et al, 2023 and Williams et al, 2017, there is a large difference between spectral signatures of plant and non-plant pixels at the NIR regions of the spectrum. Hence careful selection of normalized difference spectra combined with Otsu thresholding could result in efficient hypercube segmentation. For this study, It was observed that the normalized spectral differences between 910 nm and 950nm provided vast differences between plant and non-plant pixels, hence was used for the segmentation. The results of the segmentation is shown in supplementary Figure S1.

7. L243-245. There is a lack of information on the application of the segmented images. Were they used to generate the values of VIs, or what was their final application?

Response: Yes, the VIs were extracted from the segmented images. The VIs (both known and proposed) were combined with machine learning models to identify drought stress plants (through classification) and predict gas exchange measurements (g_s and P_n) through regression analysis

8. Revise the numbering of the figures. Figure 2 instead of Figure 1

Response: The numbering of the Figures have been revised. All Figures have been properly labelled to reflect what is in the test.

Page: 9 Line: L305

9. Revise the numbering of the figures. Figure 2 instead of Figure 1

Response: The numbering of the Figures have been revised. All Figures have been properly labelled to reflect what is in the test.

10. L268, this sentence was already mentioned in L265-266

Response: The sentence in L268 has been removed since it had been mentioned already on L265-266

Page : 7 Line: L265

11. L285-L287. It is suggested to support why these three types of equations were use

Response: The equations for the proposed indices were selected because they help to minimize the effect of varying light conditions including sunlight intensity, angle of sunlight etc on the plant reflectance measurements. Additionally, the indices resulting from these equations reduce the effects of atmospheric conditions such as haze, aerosols and scattering on plant reflectance. The reason for using these equations are highlighted on page 9

Page: 9 Line: L311-315

12. The description of the R2, RMSE, and MAE threshold values is missing, making it unclear when the model is valid or invalid.

Response: The R², RMSE and MAE has been clearly explained in the text showing the threshold used in this study.

Page: 10-11 Line: L372 -L381

13. The titles of all figures and tables need to be revised as they do not match what is described in the text; e.g., L391 says Figure 3 but should be Figure 2. Examples like the above are in several places in the manuscript.

Response: The Figures and tables have been revised and properly labelled. The labelling in the figure now reflects what are in the test

14. Figures 2 and 3 show results from a one-way ANOVA analysis, but the methodology does not describe this process. This process should be explained where appropriate.

Response: The ANOVA analysis was used for the quantitative assessment of the differences in the four treatments for gas measurements analysis. A summary of the processes and tools for ANOVA analysis have been stipulated.

Page: 5 Line: L203-209

15. L435-455. Terms such as high, low, visibly different, show peaks, high, …., are frequently used to describe spectral characteristics. Still, the quantities in reflectance units are not mentioned, which can lead to confusion and subjective interpretation.

Response: L435-455 have been reviewed to include the absolute spectral reflectance values with their units as suggested by the reviewer. This will help readers in understanding and interpretation of the graphs in this section

Page: 13-14 Line: L457- L479

16. In Figure 4, the differences could be better represented by normalizing the reflectance values. In addition, the authors should evaluate the possibility of including a graph for each treatment (WWLN, WWHN, DSHN, and DSLN) that consists of the spectral reflectance at 0 DADS, 6 DADS, and 15 DADS to facilitate the visualization of the differences throughout the phenological cycle

Response: The spectral reflectance of the treatments presented are already normalized between 0 and 1. Normalizing the spectra ensures that reflectance data is accurate and consistent allowing for more reliable comparison and interpretation. The graph already depicts the spectral reflectance from a wide range of spectrum (400-1000 nm). Plotting another graph for each treatment will be quite difficult unless the specific spectrum for which the comparison is made is known.

17. L554 and L556. The value of r seems to be incorrect. It should be -0.78 and -0.67, respectively.

Response: The value of r has been corrected

18. Table E-1 and L557-556. It is suggested that a unique name or code be assigned to each index in Table E-1 to match the text describing "Figure 6. Correlations between the proposed indices and the Pn and gs measurements" to facilitate the reader's interpretation.

Response: Unique name has been assigned in Table E-1 to match the text describing it as well as Figure 6 as suggested by the reviewer.

 Page: 31 -32

19. The discussion section is weak because it only describes the results again but does not discuss at length issues such as the reasons for the changes in spectral characteristics that may be due to a direct effect of drought, an effect of nitrogen concentration, or a combined effect of both factors.

Response: The discussion part had explanation of the behaviour of the spectral characteristics that may be due to the drought and nitrogen deficiency. This part has been improved to include further explanations citing other research in relation to this finding.

20. The discussion does not delve into the analysis of the VIs used, and the reasons why the selected indices were the best to estimate the effect of drought is absent. In addition, it is suggested that the discussion and analysis of the spectral regions that showed better results in the correlation with drought and nitrogen be extended; this could help to understand the processes occurring in the plant.

Response: The discussion section on line highlights the possible reason for selecting the 10 known Vis. Also, the wavelengths selected from the ensemble learning models have been highlighted with possible reasons for the selection of those wavelengths.

21. A discussion on the ML models tested to estimate drought's effect is missing. A comparison of the models used could be included to explain the reasons for these results and broaden the knowledge of the analysis processes of this type of data.

Response: Discussion of the ML models tested is available in the discussion section. Line 769 -783 compares the ML methods used

 

References

Williams, D., Britten, A., McCallum, S., Jones, H., Aitkenhead, M., Karley, A., et al.

(2017). A method for automatic segmentation and splitting of hyperspectral images of raspberry plants collected in field conditions. Plant Methods 13 (1), 1–12. doi: 10.1186/ s13007-017-0226-y

 

Okyere FG, Cudjoe D, Sadeghi-Tehran P, Virlet N, Riche AB, Castle M, Greche L, Simms D, Mhada M, Mohareb F and Hawkesford MJ (2023) Modeling the spatial-spectral characteristics of plants for nutrient status identification using hyperspectral data and deep learning methods. Front. Plant Sci. 14:1209500. doi: 10.3389/fpls.2023.1209500 COPYRIGHT

 

Author Response File: Author Response.pdf