Influence of Structure and Texture Feature on Retrieval of Ramie Leaf Area Index
Round 1
Reviewer 1 Report
This manuscript (agronomy-2424749) explores how spectral, structure, and texture features from remote sensing images affect leaf area index (LAI) estimation in ramie crops. The findings suggest that multifeature fusion improves LAI estimation accuracy, providing a practical method for large-scale, nondestructive crop monitoring.
Why, besides LAI, were not other indices as important as SLA addressed?
Keywords in alphabetic order;
The legends of the figures and tables need to be better described, with details such that readers do not need to return to the text, and so that all the elements described can be fully understood by comparing the figures/tables with what is present in the legend. Additionally, include the statistical test used and sample size.
Why did not the authors test other models apart from RF and SVM? Why was PLSR chosen over other models? How were the test and validation samples selected for the proposed model?
Figures 2 and 3 should be separate. In Figure 2, the boxplot should contain all the elements analysed (samples). In Figure 3, what does this scale represent? Separate the number values from the side line for better visualization. What kind of correlation (Pearson or Spearman)? What is the number of samples used? Why was not a scale from -1 to +1 chosen?
Figure 4 is poor. This mixture of boxes and graphs should be redone. The scatter plots associated with the equations and models should be grouped, and each of the elements should be described in the figure legend, just as in Table 2. What does CV% represent?
L169. 100 cm; 300 cm (space). Check in all manuscript;
For Figures 5 and 6, 7, 8 and Table 2 and 3; similar comments to those described previously apply.
Why is there a scale from -1.20 to +1.2? Correlation only ranges from -1 to +1! Please review and revise, as the statistical description is wrong. Therefore, rewrite it as'scale'.
The discussion is definitely not suitable for the work. Only two topics are covered, where results are repeated, and only four references are cited in the proposed sections. Revise based on literature. Why is your method better? What are the prospects of using it? How will it benefit the agronomic field? Why is your model better than the preexisting ones? Is your sample size sufficient to truly support your conclusions? How was the minimum sample size selected to ensure the model is ideal and not biased?
The conclusions are repetitions of phrases and ideas from the manuscript. However, it should be a section that addresses the main results and how they apply to the agronomic field, as well as what the future prospects are. It should not be written in bullet-point format. Therefore, please rewrite it completely.
Please check for very old references and replace them with new ones. Review the number of references to improve your discussion.
English needs corrections in some sentences to make them less verbose and more clear and concise.
Author Response
Dear reviewer,
Thank you for your guidance on my article.
I am very happy to receive your comments, which are really precious to me. What I want to say is you are a very responsible reviewer. Your suggestions have taught me a lot of knowledge, and your comments have also encouraged me. Thank you again.
In terms of language, I read through the whole text and made some changes. If you think there is any problem with the language of this article, please feel free to contact me, or I can apply to the journal for language editing.
According to your suggestions, I have made major revisions to the article by using the track changes mode in MS Word. The following are responses to your questions. Do not hesitate to contact me if there have any other questions in my paper.
Q: Why, besides LAI, were not other indices as important as SLA addressed?
A: Physiological and biochemical indexes such as LAI, chlorophyll content, plant height and protein content are important indexes in the growth process of ramie. For ramie, producers mainly harvest its above-ground material because it can be used as animal feed. LAI can t the final yield of ramie. Therefore, we choose it as the research object. Admittedly, other indicators are also important, which we will explore in future studies. Thanks for your questions, I have added appropriate instructions into the article.
Q: Keywords in alphabetic order
A: I have reordered the keywords, thank you.
Q: The legends of the figures and tables need to be better described, with details such that readers do not need to return to the text, and so that all the elements described can be fully understood by comparing the figures/tables with what is present in the legend. Additionally, include the statistical test used and sample size.
A：Thanks for your suggestion. I have modified the chart as required. If you have any other questions, please let me know in time. Thank you very much. I hope I did the right thing.
Q: Why did not the authors test other models apart from RF and SVM? Why was PLSR chosen over other models? How were the test and validation samples selected for the proposed model?
A: In this study, we tested a total of four models (LR, RF, SVR and PLSR), but the final result only showed the best model effect.
We finally decided to select these models after referring to the methods of other studies. Machine learning methods are currently commonly used for crop inversion, which I have added in the article. In addition, neural network is also the mainstream modeling method at present, but in this experiment, the amount of data is not very huge, so PLSR is more appropriate.
A total of 360 samples were obtained and randomly divided into modeling sets and validation sets according to a ratio of 7:3. For each model, the modeling set and the validation set are consistent. I have explained this further in the text.
Q: Figures 2 and 3 should be separate. In Figure 2, the boxplot should contain all the elements analysed (samples). In Figure 3, what does this scale represent? Separate the number values from the side line for better visualization. What kind of correlation (Pearson or Spearman)? What is the number of samples used? Why was not a scale from -1 to +1 chosen?
A: I determined that Figure 2 contained all the analyzed samples. I divided the samples into two datasets and compared the the data distribution of the two datasets.
Figure 3 has been modified as you suggested. The scale in Figure 3 represents the value of the Pearson correlation coefficient, which I have added in the article. Thank you for your suggestion. The number of samples used is 360, which is all the samples analyzed.
Q: Figure 4 is poor. This mixture of boxes and graphs should be redone. The scatter plots associated with the equations and models should be grouped, and each of the elements should be described in the figure legend, just as in Table 2. What does CV% represent?
A: Thank you very much for your suggestion. Since different datasets and processes are involved here, and different processes need to be compared, so we chose this format for presentation and thought it was appropriate.. We have not thought of a more appropriate way to handle these results, so this part has not been modified. If you or the editorial department have good ideas in the final stage, please contact me in time.
CV represents the coefficient of variation, which can reflect the degree of difference of a series of data.I have already added this in the text, thank you.
Q: L169. 100 cm; 300 cm (space). Check in all manuscript;
A: I'm sorry, but I don't quite understand what you mean, probably because our file L169 shows inconsistency.
Q: For Figures 5 and 6, 7, 8 and Table 2 and 3; similar comments to those described previously apply.
A: Thanks for your suggestion, I have modified it.
Q: Why is there a scale from -1.20 to +1.2? Correlation only ranges from -1 to +1! Please review and revise, as the statistical description is wrong. Therefore, rewrite it as'scale'.
A: There is no problem with the data, but the scale range was set incorrectly when drawing. Thank you for your care, I have revised all the scale range.
Q: The discussion is definitely not suitable for the work. Only two topics are covered, where results are repeated, and only four references are cited in the proposed sections. Revise based on literature. Why is your method better? What are the prospects of using it? How will it benefit the agronomic field? Why is your model better than the preexisting ones? Is your sample size sufficient to truly support your conclusions? How was the minimum sample size selected to ensure the model is ideal and not biased?Please check for very old references and replace them with new ones. Review the number of references to improve your discussion.
A: Your comments are very useful, I have revised the discussion , I hope you are satisfied.
In terms of literature citations, the literature I can collect has been listed in the introduction. Most of the previous studies focused on the data fusion, but did not discuss the limitations of the data, which is the significance of our study. We prove that structural parameters, like spectral indices, have limited application value. Under this limitation, we want to further improve the accuracy of estimation, which is what we want to do next.
In this experiment, the amount of data is 360, which is not a small sample when referring to other studies. However, if we want to build a more stable model, we do need the support of a larger sample, and we will continue to collect it in the next study. Thank you very much for your suggestion, which also gives us a new research method.
Q: The conclusions are repetitions of phrases and ideas from the manuscript. However, it should be a section that addresses the main results and how they apply to the agronomic field, as well as what the future prospects are. It should not be written in bullet-point format. Therefore, please rewrite it completely.
A: Your advice has given me good writing guidance. Thank you very much. I have modified it according to your idea, and I hope it can satisfy you.
Reviewer 2 Report
The manuscript contributes to the Agronomy journal, however is important to add some comments that I already did to the manuscript body.
In line 98, please add a what time the reading were taken (for example 11:00 am to 13:00 pm).
In Figures 4, 7 and 9, please add: The black dashed line represents the 1:1 line.
In the Discussion section please add two o more discussions with other researchers (compare finding) in crops like ramie.
Comments for author File: 
Comments.pdf
Author Response
Dear reviewer,
Thank you for your guidance on my article.
I am very happy to receive your comments, which are really precious to me. What I want to say is you are a very responsible and careful reviewer. Thank you for finding many small problems in my article. At the same time, your comments have also encouraged me. Thank you again.
In terms of language, I read through the whole text and made some changes. If you think there is any problem with the language of this article, please feel free to contact me, or I can apply to the journal for language editing.
According to your suggestions, I I modified the article by using the track changes mode in MS Word. The following are responses to your questions. Do not hesitate to contact me if there have any other questions in my paper.
Q: In line 98, please add a what time the reading were taken (for example 11:00 am to 13:00 pm).
A: Thank you for your guidance, I have added in the article
Q：In Figures 4, 7 and 9, please add: The black dashed line represents the 1:1 line.
A: In these figures, there are many legends. I tried to mark the 1:1 line, but I thought the legend was too complicated and affected the beauty, so I did not modify it. In my opinion, this does not affect the data reading. If you feel it is necessary to make any changes, please let me know
Q: In the Discussion section please add two o more discussions with other researchers (compare finding) in crops like ramie.
A：I have added and revised the discussion, and thank you for your suggestions.
Reviewer 3 Report
1. Introduction
There are some traditional LAI measurement methods which are non-destructive and objective.
More review is required.
You mentioned PROSPECT, however, you ignored PROSAIL. Why? This series is canopy scale RTMs and you can obtain LAI values by inversion.
Machine Learning techniques have also been used for estimation of Leaf Area Index and then you used some of them.
You should have provided information about them.
Figure 1.
Scale bar is missing.
2.2.1. Ground LAI acquisition
You used the indirect method.
2.2.2. UAV multispectral data acquisition and processing
Did you use GCPs for bundle adjustment?
Please provide the reasons why two calibration plates with reflectance of 5% and 30% were used for radiometric correction.
You said nine texture eigenvalues were calculated. How many pixels did you use for the calculation?
2.3. Model construction and evaluation
You used Linear regression (LR), random forest (RF), supports vector regression (SVR) and 129 partial least squares regression (PLSR).
Which hyperparameters did you optimize and how did you optimize them?
3.1.2. LAI inversion performance in different CC dataset
Why did you only show the results of SVR in Figure 4?
3.2.2. LAI inversion performance in different CC dataset
Why did you only show the results of PLSR in Figure 7?
3.3. Effect of TF on ramie LAI inversion
Why did you only show the results of SVM in Figure 9?
3.4. Multi-feature fusion to improve LAI estimation accuracy
Why didn't conduct any sensitivity analysis for evaluating the importance of features?
4. Discussion
There are few comparisons with the previous studies.
You said it is difficult to achieve precise estimation of LAI using spectral information. Some studies have used spectral information for evaluating vegetation structures. Could you clarify the reasons why vegetation structures had to be obtained tradition methods in this study?
Texture eigenvalues could be changed by the window size. Can you say that the appropriate sizes were chosen?
Author Response
Dear reviewer,
Thank you for your guidance on my article.
I am very happy to receive your comments, which are really precious to me. What I want to say is you are a very responsible and professional reviewer. Your suggestions have taught me a lot of knowledge. Thank you again.
In terms of language, I read through the whole text and made some changes. If you think there is any problem with the language of this article, please feel free to contact me, or I can apply to the journal for language editing.
According to your suggestions, I have made major revisions to the article by using the track changes mode in MS Word. And I hope my modifications can make you satisfied.The following are responses to your questions. Do not hesitate to contact me if there have any other questions in my paper.
- Introduction
Q: There are some traditional LAI measurement methods which are non-destructive and objective. More review is required.
A: Thank you for your suggestions, and I have revised and supplemented some contents. The traditional ramie LAI measurement needs to be sampled and then taken back to the laboratory for measurement with a scale. In our previous phenotype collection process, the measurement of ramie LAI was a very painful thing, so we proposed to use UAV remote sensing.
Q: You mentioned PROSPECT, however, you ignored PROSAIL. Why? This series is canopy scale RTMs and you can obtain LAI values by inversion.
A: Thank you very much for your question. We are concerned that these models of PROSAIL and PROSPECT have been used in many studies, and they show better effects. We would like to be able to use RTMs in future studies, but at this time we don't have any more data to support this study. Your suggestion also gives us a new research direction, and then we will learn this method and apply it. Thanks.
Q: Machine Learning techniques have also been used for estimation of Leaf Area Index and then you used some of them. You should have provided information about them.
A: Thanks for your suggestion, I have added this part to my article.
- Materials and method
Q: Figure 1. Scale bar is missing.
A: Thanks for your suggestion, I have modified it according to your requirements.
Q: 2.2.1. Ground LAI acquisition. You used the indirect method.
A: Yes, we measured LAI by indirect method. In the introduction, I added this content.
When we need to make frequent measurements on large-scale materials, it is difficult to measure LAI by using instruments. If the UAV remote sensing can get the same accuracy as the instrument measurement, it will have more application value.
Q: 2.2.2. UAV multispectral data acquisition and processing. Did you use GCPs for bundle adjustment?
A: In the experimental area, we set up six ground control points for geographical location correction. And the DJI Phantom 4 pro has higher RTK accuracy.
Q: Please provide the reasons why two calibration plates with reflectance of 5% and 30% were used for radiometric correction.
A: We think that the calibration plates with uniform reflectivity distribution may obtain better correction effect, so we choose these two reflectance plates with large reflectivity span. At the time of purchase, we also consulted the DJI official and also referred to their recommendations. If you want to be more accurate, you can set several calibration plates, and choose a reflectivity with a certain span.
Thank you for your question, and if you have any other suggestions, I'd be happy to learn more from you.
Q: You said nine texture eigenvalues were calculated. How many pixels did you use for the calculation?
A: Thank you for your professional question. We use software to extract texture eigenvalues. In the software, AOI with fixed size was drawn. AOI is drawn inside each plot, leaving a certain edge. Therefore, we can ensure that the acquisition of image data in each plot is unified. However, we do not know the specific number of pixels occupied by AOI, I am very sorry.
Q: 2.3. Model construction and evaluation. You used Linear regression (LR), random forest (RF), supports vector regression (SVR) and 129 partial least squares regression (PLSR). Which hyperparameters did you optimize and how did you optimize them?
A: In the modeling process, we directly call the model in python sklearn, we use GridSearch method to find the best parameters of different models, and use RFE to filter the feature values related to the target variable.
In this experiment, the same method was adopted for the construction of all datasets. In addition, the study paid more attention to the influence of different canopy feature levels on LAI measurement, and explored whether structural features were as limited as spectral features. So in other ways, we didn't optimize much. I hope my explanation is satisfactory to you.
Q: Why did you only show the results of SVR in Figure 4? Why did you only show the results of PLSR in Figure 7? Why did you only show the results of SVM in Figure 9?
A: Thank you for your question. In this study, we used four ML methods for modeling, but different ML models have different performance for different datasets, so we chose the best one. SVR performed better in 3.1.2 when coverage data and spectral data were used to estimate LAI, and PLSR performed better in 3.2.2 when plant height data and spectral data were used to estimate LAI. We believe that this does not affect the ultimate purpose of the study, because we are conducting a comparison experiment and the same model is used between the control groups.
Of course, according to your suggestion, we also spent more content to explain the machine learning method, and also made some additions in the discussion section.
Q: 3.4. Multi-feature fusion to improve LAI estimation accuracy. Why didn't conduct any sensitivity analysis for evaluating the importance of features?
A:Thank you very much for your advice. In this experiment, the final accuracy improvement has shown that data fusion can effectively improve accuracy, so we did not carry out sensitivity analysis. We also want to know whether these variables have limitations, such as saturation of spectral data and limited value of structural data for LAI estimation in the later period. These questions are answered in the findings above.
But, thank you very much. You have provided us with a good idea. Sensitivity analysis of variables is a good way to explain the contribution of different variables to LAI estimation. If you can, you can very well guide how to do this, because I have read less literature about this method. Thank you very much.
- Discussion
Q: There are few comparisons with the previous studies.
A: Your comments are very useful, I have revised the discussion , I hope you are satisfied.
In terms of literature citations, the literature I can collect has been listed in the introduction. Most of the previous studies focused on the data fusion, but did not discuss the limitations of the data, which is the significance of our study.
We prove that structural parameters, like spectral indices, have limited application value. Under this limitation, we want to further improve the accuracy of estimation, which is what we want to do next.
Q: You said it is difficult to achieve precise estimation of LAI using spectral information. Some studies have used spectral information for evaluating vegetation structures. Could you clarify the reasons why vegetation structures had to be obtained tradition methods in this study?
A: Thank you for your responsible and conscientious. What I mean is that spectral information has some limitations. When the spectral index tends to be saturated, the effect of only using spectral information to estimate LAI may not be good. Many studies have proved that the fusion of spectral information and other information can improve the accuracy of LAI estimation, and the same results were obtained in this experiment.
However, previous studies only explored the effect of data fusion, but did not further analyze whether structural features also had limitations. Therefore, we divided different datasets for verification. After we proved that the structural features also have limitations, it may provide support for future studies to further improve the accuracy of LAI estimation.I hope my explanation is satisfactory to you.
Q: Texture eigenvalues could be changed by the window size. Can you say that the appropriate sizes were chosen?
A: I'm sorry we didn't consider the window size in this study. However, the window size of texture eigenvalues extraction in each plot is the same, so there should be no problem with the research results in the controlled experiment.
Thank you very much for your suggestion. We will remember this detail in the future research, and it also provides us with a good research idea. Thank you here.
Round 2
Reviewer 1 Report
I appreciate the authors' responses. Minor modifications are required. Please remove the lines around the figures, revise the captions once more, add references to the three discussion topics, and perform an English language check. Best.
Revise English.
Author Response
Dear reviewer,
I am very glad to receive your reply so soon.
I have modified this paper as required, please refer to cover letter for details.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors added some elements to enrich manuscript.
Author Response
Dear reviewer,
I am very glad to receive your reply so soon. I also feel that the content and structure of the article have been enriched after the first revision. Thank you again for your suggestion.
In view of some small problems still exist in the article, I have made modifications.
I have reviewed the whole text again and modified some English expressions. Moreover, in the final review, I will ask the journal for help, hoping that they can show more accurate English language.
