Predicting Grapevine Physiological Parameters Using Hyperspectral Remote Sensing Integrated with Hybrid Convolutional Neural Network and Ensemble Stacked Regression
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe article discusses the use of hyperspectral remote sensing to predict various physiological parameters of the vine, applying learning models. The methodology used demonstrated the ability to collect detailed and precise data on plant spectral reflectance, which can be correlated with different physiological parameters, such as net assimilation rate, stomatal conductance to water, PSII quantum yield, and transpiration rate. The analysis of this spectral data provided valuable information about plant health and performance, allowing a non-invasive and large-scale assessment of vine physiological characteristics.
The work is clear and detailed in the collection and analysis techniques used. The tables and graphs are clear and well integrated into the text.
My suggestions focus primarily on analysis and discussion. Considering the great potential of the article for replication, as a case study and as an example in the classroom, I suggest some improvements.
1 - In the context of this study, the interpretation of some data and results can be improved. Although the article discusses the use of hyperspectral remote sensing and artificial intelligence models to predict vine physiological parameters, there is a lack of a detailed explanation of the artificial intelligence models used (REGST, CNN and CNN-REGST). This absence can hinder the understanding of readers unfamiliar with the technology about how these models were applied. In addition, the relationship between the spectral data obtained and the results of the physiological parameters could be clearer.
2 - I suggest a review in the results section to include practical implications. Considering how the results obtained can be applied in practice, including in the detection and monitoring of vine diseases, could enrich the discussion and make the results more accessible and relevant to the scientific community and professionals in the field.
3 - In the discussion section, I miss a more detailed explanation about the artificial intelligence models and a clearer connection between the spectral data and the vine physiological parameters. Including a discussion about the current stage of development of the technology reported could strengthen the reader's understanding of the level at which the results obtained are found. An additional small paragraph could resolve this issue.
4 - In the introduction, there was a mention of disease problems in vines, such as Phylloxera. I missed this theme in the discussion. Would it be possible to include it? The authors could help readers understand how these conditions manifest when the vine is contaminated and what the physiological parameters of the plants to be observed would be. This would assist in interpreting the data obtained. An additional small paragraph could resolve this issue.
I hope these suggestions enhance the understanding and application of the study's findings, strengthening its relevance and impact in the field of remote sensing and plant physiology.
Author Response
We have addressed the reviewers suggestions and have uploaded responses in attachment. Please note that changes in response to other reviewers alters the original line numbers.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsThe article addresses a challenge in viticulture using an innovative approach that combines hyperspectral remote sensing with advanced machine learning techniques, to predict grapevine physiological parameters. Overall, the research appears to be an important advancement in the field of viticulture, with practical implications for improving the efficiency and accuracy of vineyard monitoring and management. However, the presentation of the paper requires some improvements. The authors need to address the following issues before it can be accepted.
1. Line 65-66, please show more description for traditional phenotyping procedures. By comparing the differences between traditional and modern methods, readers can better understand the gaps in knowledge.
2. Line 162, Please provide an overview map of the study area, especially for different commercial rootstocks.
3. Line 177, How many samples did you measure? Please show the details of the samples.
4. Line 309, Table 1, If possible, please provide specific parameter information. This will enable readers to assess the replicability and reliability of the findings.
5. Line 402, (r >> 0.5), the Pearson correlation coefficient ‘r’ value for gsw was lower than 0.5, please confirm it.
6. Table 3 should include the number of samples.
7. Line 466-473, From the feature importance score graph, the 490-510nm range may be the first important region, rather than the 500nm wavelength. For the 700-750nm region in Figure 8, I do not think it is an important region, or its importance needs to be reconsidered.
8. Line 521-522, three highly correlated regions were identified (520-570 nm,740-780 nm, and 900-950 nm). This is not consistent with the former description.
9. Line 584-586, ‘While comparing the performance of REGST with or without feature engineering using CNN, the model metric R2 retrieved for each physiological trait were either identical or very close.’. In Table 3, the performance of CNN-REGST was lower than REGST. Please describe the results precisely. The same applies to Line 615-617.
10. More discussion of relevant findings needs to be provided.
11. Figure S1 and Figure S3 are the same, please check them.
Addressing these issues will significantly enhance the clarity, reliability, and overall quality of the paper, rendering it more suitable for publication.
Author Response
We have addressed the reviewers suggestions and have uploaded responses in attachment. Please note that changes in response to other reviewers alters the original line numbers.
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsUnderstanding the physiological status is critical for accurate prediction of long-term variations of carbon, water, and energy fluxes in plants. In the current study, the authors tried to retrieve several important physiological parameters (net assimilation rate, stomatal conductance to water vapor, quantum yield of PSII and transpiration rate) from UAV hyperspectral data with different machine learning (ML) approaches.
I have several comments/suggestions before a positive recommendation for acceptance.
1. L. 12. “This study investigated rootstock influence on scion photosynthetic parameter and the prediction potential of canopy level measurement using aerial hyperspectral imaging”. However, according to the title, the paper should focus on how to retrieve grapevine physiological parameters from hyper-spectral remote sensing data (also consistent with the content presented in L. 148-154.). If the authors intend to highlight the impacts of rootstock on photosynthetic parameters, they should at least evaluate the effectiveness of ML methods in predicting physiological parameters across different rootstocks.
2. L. 20-21. Provide more information about the results of different models.
3. L. 43-66. Reduce the length of these paragraphs. Only present the rootstock genotype has impacts on physiological processes of grapevine.
4. L. 182. ‘a LI-COR (Li-6800) portable photosynthesis system Li-6800 (LI-COR Biosciences,’. Remove one ‘LI-COR (Li-6800)’ from this sentence.
5. L. 186-188. How many vine trees were selected for measurements each time? How many leaves were measured for each canopy? These are important because these leaf scale ground-truth data are used for UAV scale model development.
6. L. 212-213. If the image has a spatial resolution of 1.87 cm per pixel, more than one pixel should be extracted for each canopy. Please clarify how these spectra were processed to match the leaf scale ground-truth data described in section 2.2.
7. Many parts of Figure 2, 3 and 4 are repeated, please merge them.
8. “Results” section: Present the performances of the modeling algorithms across different rootstock genotypes.
9. “Discussion” section: It has been demonstrated in many previous studies, both experimentally and theoretically, that the performance of multiple regression models can be tremendously improved if only informative variables are included in the model. As the you have identified the important spectral features in section 3.4, please discuss the effectiveness of ML models only with important wavelengths.
10. Supplementary Material. The labels for X, Y axes of Figure S1, S2 and S3 are inconsistent with the figure caption texts.
11. Figure S4 caption (L. 22:): ‘rat’ to ‘rate’.
12. Use different markers to represent different genotypes in Figure S1, S2, S3 and S4.
Author Response
We have addressed the reviewer's comments. The attached document provides response to all the comments. Please note that document lines have changed with the various revisions.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have addressed most of my comments.
I think the authors can improve Figures S1-S4 with their current results.
Although all genotypes were used together for statistical analyses, each leaf sample's genotype can be easily identified. Thus, for each single figure (such as Figure S1 PLSR), you can use 'o' to represent M/1103P, '+' to represent M/3309C, 'x' to represent M/5C..., instead of using the same dot for all different genotypes. This will provide the readers with more intuitive results.
Author Response
"I think the authors can improve Figures S1-S4 with their current results.
Although all genotypes were used together for statistical analyses, each leaf sample's genotype can be easily identified. Thus, for each single figure (such as Figure S1 PLSR), you can use 'o' to represent M/1103P, '+' to represent M/3309C, 'x' to represent M/5C..., instead of using the same dot for all different genotypes. This will provide the readers with more intuitive results."
We have accommodated this request and used different symbols for the genotypes. Please note that this required reanalysis to reproduce the 32 graphs in the same manner. Please note that CNN initializes weights randomly each time so there were very slight changes in CNN and CNN-REGST model in net assimilation (A) and stomatal conductance (gsw), identified by track changes in Table 3 i. and ii, respectively. This did not change our results statements. We inserted the slightly changed value in line 592 of this revision and made one small correction on line 421. We used track changes, after accepting all the previous submission edits, so the new small changes can be seen.
