Accuracy Comparison of Estimation on Cotton Leaf and Plant Nitrogen Content Based on UAV Digital Image under Different Nutrition Treatments

Round 1

Reviewer 1 Report (Previous Reviewer 3)

Report Reviewer

 

General brief

The manuscript proposes a rapid and accurate estimation of the foliar nitrogen content (LNC) and plant nitrogen content (PNC) of cotton. This estimation is carried out using RGB images acquired using a low-cost unmanned aerial vehicle (UAV) with a visible light digital camera (RGB) during different growth stages. Several machine learning algorithms were used to develop models to estimate N content and thirty different vegetation indices (VI) in visible light such as:  Random Forest (RF), Support Vector Machine (SVM) and Back Propagation Neural Network (BP).

 

Comments

The abstract was well done. The structure of the introduction is well done; however, it does not explain well why it is important to use different nutrient treatments. In fact, there are no manuscripts evaluating LNC or PNC in other crops treated with different elements. The results are not comprehensive and there are some parts where they are incorrect because they show inadequate values and no units. This condition is related to the lack of explanation of the methodology. The discussions and conclusions are very general and need to be implemented after correction of the previous chapters. Below are some suggestions found throughout the manuscript.

Comments:

 

-Title: The title is very large and confusion, i suggest to modified it.  

 

-Introduction: The introduction needs some slight implementation. For example, the authors may add some results concerning LNC or PNC in the presence of different nutritional treatments, found in the literature.

 

-Row96: The treatment PK-M1 is not described well, application proportions are missing.

 

-Row 94-108: The experimental design requires a figure to understand better.

 

-Row119: The altitude of the flight is very low. Probably, this flight condition is not replicable in others environmental.

 

-Row 124-129: Probably, the photogrammetric process in not well done. Indeed, the reflectance correction missing.

 

-Figure 2: The figure 2 is not cited in the text.

 

-Row 211-216: It is an important error; the unit of measurement is missing.

 

Figure 4: It is not clear what data is being represented, the unit of measurement is missing.

 

 

-Table 5: The description of table 5 is wrong, are the correlations values or other?

 

-Results: all the results presented were not divided for the different N, P, K treatments.

The use of the English language has been done adequately and does not require major corrections. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

The study estimated nitrogen content in cotton from UAV digital image. This study seems interesting and might be useful for crop management. However, following concerns need to be addressed.

 

1.       The leaf nitrogen content (LNC) and plant nitrogen content (PNC) need to be correlated with yield data.

2.       Whether soil analysis has been conducted pre and post-harvest? What is the nitrogen balance?

3.       In the treatments, along with N, PK combinations were planned. However, results in context of PK combinations are not stated/discussed. Whether any effect of PK application on N uptake is there? Further, what is the recommended dose of fertilizer in cotton? On which basis levels of nitrogen and PK were decided for the treatments?

4.       In section 2.2. You have stated each flight activity of 12 min, whereas in Table 1, its 27 min. Please clarify.

5.       Line 143-148 and line 150-154 please add appropriate references for methods.

6.       IN results, while mentioning values add ± SD after the values.

7.       Table 4, the unit needs to be given for the values. Also, statistical analysis needs to be given and stated in text appropriately.

8.       Out of 30 indices, how many indices were significant for explaining the variability of the data? Indices with significant correlation coefficients may be used for further analysis.

9.       Line 239-248, please use symbols ≤ or ≥ at appropriate places.

10.   In the discussion, you have stated use of digital imagery for crop N status. The SPAD meter is commonly use handheld device which gives indirect estimation of chlorophyll content. Based on SPAD values, fertigation application is done in field crops. Compared to that, how image based detection of plant N would be? Estimation of R, G, B values or indices thereof is challenging and computationally demanding.

Okay

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

The authors conducted a study of cotton crops using an unmanned aerial vehicle and an RGB camera.

1. In the introduction, it is necessary to mention studies using other sensors, for example, multispectral cameras, to indicate the features and disadvantages in comparison with studies using RGB cameras.

2. The materials and methods indicate the altitude, flight speed of the drone and the resulting GSD. The authors should justify such a low height, a high speed for such a height and a large GSD that does not correspond to the technical characteristics of the camera. What was the area of aerial photography?

3. The results of the study contain tables and graphs, and the text reflects only the values of them, it is necessary to expand the descriptive part

4. In the discussion, the data obtained should be compared with data from other sensors (multispectral, thermal cameras), flight at low altitude and the effect of shadows and seeding density on the process of photogrammetric data processing should be discussed.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report (New Reviewer)

This paper needs major revision

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report (Previous Reviewer 3)

General brief

Several corrections have been made to the article and I think it has been heavily revised. However, further doubts have arisen from this revision as to the quality of this study. The assumptions seem suitable but there are shortcomings in scientific terms. The abstract and the introduction are well done but can be implemented. Materials and methods have been improved but are confusing Results are poor and there are some parts where they are incorrect. The discussions and conclusions are very general and need an implementation. Below are some suggestions found throughout the manuscript.

 

Comments

- In the experimentation missing PK treatment: 0%-100%

- Sensor description missing in the materials and methods.

- The description of image calibration is missing. Furthermore, the use of GCPs is mentioned but it is not written that they were placed in the field, nor is the calibration panel.

- Why was a threshold value of 0.4 used for segmentation?

- Table 5 and figure 3 show the same information. It is necessary to create a single figure that can unify the two pieces of information.

- Table 6 shows that there was a significant effect of the year on the LNC and PNC. This result should be revision.

- In general, the article used the DN values to show the spectral condition. Scientifically, it is not correct to report the values in DN. They must be reported in reflectance values.

- Table 7 and figure 4 show the same information. A single figure must be created that unifies the two pieces of information.

- Figure 6 would appear to be incorrect. Generally, predicted values are shown on the y-axis and measured values on the x-axis.

General brief

Several corrections have been made to the article and I think it has been heavily revised. However, further doubts have arisen from this revision as to the quality of this study. The assumptions seem suitable but there are shortcomings in scientific terms. The abstract and the introduction are well done but can be implemented. Materials and methods have been improved but are confusing Results are poor and there are some parts where they are incorrect. The discussions and conclusions are very general and need an implementation. Below are some suggestions found throughout the manuscript.

 

Comments

- In the experimentation missing PK treatment: 0%-100%

- Sensor description missing in the materials and methods.

- The description of image calibration is missing. Furthermore, the use of GCPs is mentioned but it is not written that they were placed in the field, nor is the calibration panel.

- Why was a threshold value of 0.4 used for segmentation?

- Table 5 and figure 3 show the same information. It is necessary to create a single figure that can unify the two pieces of information.

- Table 6 shows that there was a significant effect of the year on the LNC and PNC. This result should be revision.

- In general, the article used the DN values to show the spectral condition. Scientifically, it is not correct to report the values in DN. They must be reported in reflectance values.

- Table 7 and figure 4 show the same information. A single figure must be created that unifies the two pieces of information.

- Figure 6 would appear to be incorrect. Generally, predicted values are shown on the y-axis and measured values on the x-axis.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

accept 

Results may be presented better

Author Response

thanks for your positive comments

Reviewer 3 Report (New Reviewer)

The authors have done a lot of work to finalize the article. There are still questions about the received GSD regarding the shooting height and the technical characteristics of the camera. The text does not indicate that the authors reduced the GSD during photogrammetric processing.

There is also no discussion about the effect of such a low altitude on the calculation of vegetation indices (the influence of plant shadows and dense sowing).

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 4 Report (New Reviewer)

I appreciated the changes made to the Introduction section. However, I was disappointed by the lack of effort in trying to improve the Conclusion section where basically the authors just pasted a part taken from the Discussions without adding anything else. This section should be completely rewritten. I invite the authors to make this small effort.

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The paper presents the results of research aimed at estimation of leaf nitrogen content and plant nitrogen content of cotton based on RGB UAV images. In general, the such an aim of the study is reached. However, in my opinion some conclusions are not enough supported.

Main issues I would like to point out are as follows:

1. Sampling scheme design should be described in more details.

It is not clear how many samples were taken. You have 48 plots. I understand that 3 cotton plants in each field were selected as samples. But it is not clear if these three samples were than used all together to obtain an average estimate for a field (so each time we have 48 reference measurements) or treated independently (an we have 3x48 reference measurements).

You wrote that 2/3 of samples were randomly used to develop the regression models and 1/3 to validate them. But, it is not clear how the samples were drawn to these sets. Please explain if the same locations of training and validation samples were used in each growing season and each growth stage or the drawing was done independently each time.

2. More details is needed about calculation of predictor variables

You wrote that 48 RGB values is extracted by MATLAB. I understand it means one representative value is derived for each field. Please add more details about the calculation process. For example: What kind of value you consider as appriopriate - is it average value, median or maybe something else? Did you obtain this value based on all pixels within the field or excluded some number of pixels? Did you consider the distribution of pixel values, check for outliers, etc.? Did you take an edge effect into consideration? What was the possible range of RGB values in your images (how many bits were used for pixel values storage)?

3. More details is needed regarding the training of machine learning models.

First of all, you provided the values of hyperparameters you finally applied and wrote (line 152) that these values were set "after training". But, it is not clear how did you established these values in the training phase. Please add more details.

Regarding hyperparameters - what cost (C) value did you use for  SVM model?

Secondly, in my opinion it is not enough and even not appriopriate to put the information about trained models hyperparameters into "Accuracy assessment" section. You should have a special section about machine learning models in "Materials and methods" chapter. You shoud describe the training phase more thoroughly. Did you any preprocessing of predictor values? What approach was apllied to hyperparameters tunning? What measure used to choose the best models? Etc.

This is especially important as you try to compare the performance of several machine learning models and try to rank them in the context of your application.

4. Accuracy assessment issues

You titled the Table 6 as "Accuracy assessment for the estimation of N ...". In my opinion the title shoud be changed. As far as I understand the values in Table 6 are based on the training samples. As you did not use crossvalidation during the training phase (you do not write you did), such results give no reliable information about the general accuracy of trained models. You cannot treat them as "accuracy assessment". But you do. You even put these values into the Abstract as the accuracy of the evaluation you had performed (lines 23-24). It is totally wrong and unacceptable way of usage. You may not claim you achieved such accuracy. You should based here on the results obtained using validation dataset.

In my opinion, you should put the results from validation into similar table. You provide them within the Figure 5, but it would be more convenient for the Reader to have them collected together in a table. By the way, why don't you include SMLR approach into your description of validation results (lines 228-239)?

5. Comparison of maching learning algorithms

One of the issues considered within your study is a comparison of machine learning algorithms performances. Based on information in the current version of your manuscript, I do not think you should compare their results directly. This is because you probably did not assure the equal opportunities for the tested models to perform the best they could. Random Forest performed the best probably because it is the most flexible model among the ones you have tested and do not require any special preprocessing of input predictor variables. For SVM and neural networks, the achieved performances may be deteriorated due to inadequately prepared input data. For both these methods, input data shoud be transformed to a uniform scale of values. In your case you have RGB values which are several orders of magnitude higher then some VI values. To assure the best possible performance of SVM and Neural Network models, all predictors should be centered and scaled. Moreover, for Neural Network you should removed highly correlated variables from your inputs. 

The next question you should ask when the aim is to perform model performances is if the achieved results differ significantly.

6. Additional comments

- lines 65-76: You summarize previous studies about remote sensing applications to cotton N content assessment. Please add more details and references to literature sources

- line 78: SLMR term is used for the first time. Please add its explanation.

- line 118: should be "orthorectified image"

- line 125: Based on information from Table 2, I think it should be "sampled within a day before the UAV campaigns"

- line 151: please add some details about "Data Processing System"

Author Response

1. Sampling scheme design should be described in more details.

It is not clear how many samples were taken. You have 48 plots. I understand that 3 cotton plants in each field were selected as samples. But it is not clear if these three samples were than used all together to obtain an average estimate for a field (so each time we have 48 reference measurements) or treated independently (an we have 3x48 reference measurements). You wrote that 2/3 of samples were randomly used to develop the regression models and 1/3 to validate them. But, it is not clear how the samples were drawn to these sets. Please explain if the same locations of training and validation samples were used in each growing season and each growth stage or the drawing was done independently each time.

Response: We are sorry for our confusing sentences. There are 16 treatments with three replications, totally 48 plots. And in each growth stage, we took 48 RGB images and there were 48 R, G, B values. For plant samples, in each plot we took three plants and put them into one bag as one sample, so there were 48 samples for nitrogen assessment in each growth stages. 

From the samples collected in each growing season, we used cross-validation method to split the dataset into two parts with 2/3 for developing the regression models and 1/3 validating the regression models. The sentences have been revised in manuscript in Line 176-180.

We pooled three years dataset together for selecting training and validation data.

2. More details is needed about calculation of predictor variables

You wrote that 48 RGB values is extracted by MATLAB. I understand it means one representative value is derived for each field. Please add more details about the calculation process. For example: What kind of value you consider as appropriate-is it average value, median or maybe something else? Did you obtain this value based on all pixels within the field or excluded some number of pixels? Did you consider the distribution of pixel values, check for outliers, etc.? Did you take an edge effect into consideration? What was the possible range of RGB values in your images (how many bits were used for pixel values storage)?

Response: The sentences have been revised in Manuscript in Line 122-135. Before we extracted the R, G, B values from images, we did image alignment, geographic reference, mosaicking, generation of dense point clouds, orthorectified image generation and ground control point correction. Before the R, G, B values were extracted from each plot in each growth stage, the image segmentation was processed by 2 g-r-b Super- Green method to obtain the Super-Green images, and then the image was processed by grayscale and enhanced by median filtering. The gray level image was transformed in-to binary image by fixed threshold segmentation (Value was 0.40). the color image with a black background was obtained by color filling, and the R, G, B value in three band were separately set and extracted of each interest region from plot by using MATLAB R2018a (Mathworks Inc., Natick, MA, USA).

After we extracted the RGB value, we used average data for further data analysis. The range of R, G, B in all treatments was 75.4-190, 78-198, 82-195.

The image size was 2517, 696, 3 for pixel, width and height.

3. More details is needed regarding the training of machine learning models.

First of all, you provided the values of hyperparameters you finally applied and wrote (Line 152) that these values were set "after training". But, it is not clear how did you established these values in the training phase. Please add more details.

Regarding hyperparameters - what cost (C) value did you use for SVM model?

Secondly, in my opinion it is not enough and even not appropriate to put the information about trained models hyperparameters into "Accuracy assessment" section. You should have a special section about machine learning models in "Materials and methods" chapter. You should describe the training phase more thoroughly. Did you any preprocessing of predictor values? What approach was applied to hyperparameters tunning? What measure used to choose the best models? Etc.

This is especially important as you try to compare the performance of several machine learning models and try to rank them in the context of your application.

Response: thanks for your comments, we have added a new section for regression techniques and revised the sentences in material and methods, see Line 167-174.

The regression techniques we used in this study based on previous study, see references Lu et al., 2019; Dong et al., 2021.

[1] Ning Lu, Jie Zhou, Zixu Han, et al. Improved estimation of aboveground biomass in wheat from RGB imagery and point cloud data acquired with a low-cost unmanned aerial vehicle system. Plant Methods, 2019,15:17.

[2]. Zhiqiang Dong, Yang Liu*, Baoxia Ci, Ming Wen, Minghua Li, Xi Lu, Xiaokang Feng, Shuai Wen and Fuyu Ma*. Estimation of nitrate nitrogen content in cotton petioles under drip irrigation based on wavelet neural network approach using spectral indices. Plant Methods, 2021, 17(1):90. DOI:10.1186/s13007-021-00790-x.

4. Accuracy assessment issues

You titled the Table 6 as "Accuracy assessment for the estimation of N ...". In my opinion the title should be changed. As far as I understand the values in Table 6 are based on the training samples. As you did not use cross-validation during the training phase (you do not write you did), such results give no reliable information about the general accuracy of trained models. You cannot treat them as "accuracy assessment". But you do. You even put these values into the Abstract as the accuracy of the evaluation you had performed (lines 23-24). It is totally wrong and unacceptable way of usage. You may not claim you achieved such accuracy. You should base here on the results obtained using validation dataset.

In my opinion, you should put the results from validation into similar table. You provide them within the Figure 5, but it would be more convenient for the Reader to have them collected together in a table. By the way, why don't you include SMLR approach into your description of validation results (lines 228-239)?

Response: Thanks for your comments, we have revised the title of Table 6.

And we also replaced our abstract data using validation dataset.

The results for SMLR we wrote together with other three models.

5. Comparison of matching learning algorithms

One of the issues considered within your study is a comparison of machine learning algorithms performances. Based on information in the current version of your manuscript, I do not think you should compare their results directly. This is because you probably did not assure the equal opportunities for the tested models to perform the best they could. Random Forest performed the best probably because it is the most flexible model among the ones you have tested and do not require any special preprocessing of input predictor variables. For SVM and neural networks, the achieved performances may be deteriorated due to inadequately prepared input data. For both these methods, input data should be transformed to a uniform scale of values. In your case you have RGB values which are several orders of magnitude higher than some VI values. To assure the best possible performance of SVM and Neural Network models, all predictors should be centered and scaled. Moreover, for Neural Network you should remove highly correlated variables from your inputs. 

The next question you should ask when the aim is to perform model performances is if the achieved results differ significantly.

Response: thanks for your comments. We used the RF, SVR and BP in this study based on previous study, and there are different advantages for different techniques. RF regression not only handles a large number of input variables, but also obtains a reasonable prediction accuracy using a small subset of variables. In addition, RF regression is beneficial to overcome the over-fitting problem of simple decision trees.

For SVR, we think it has the ability to handle a small number of training samples, and many studies in remote sensing using SVR to estimate crop biophysical and biochemical parameters. For BP, it has strong non-linear fuzzy approximation ability.

In this study we did not transformed our dataset, but we will consider in the further research work to check whether it will affect the accuracy among different regression techniques.

6. Additional comments

- Lines 65-76: You summarize previous studies about remote sensing applications to cotton N content assessment. Please add more details and references to literature sources

Response: Thanks for reviewer’s comments. We summarized previous studies about remote sensing applications to crop nitrogen content assessment, see references 13,14,15. However, there is little information about whether the cotton N content of leaf or plant level at different growth stages could be detected by RGB sensors based on UAV. So we did not add more details and references.

- Line 78: SLMR term is used for the first time. Please add its explanation.

Response: the full name of SMLR was added in the manuscript.

- Line 118: should be "orthorectified image"

Response: The words has been replaced.

- Line 125: Based on information from Table 2, I think it should be "sampled within a day before the UAV campaigns"

Response: The sentence was revised.

- Line 151: please add some details about "Data Processing System"

Response: "Data Processing System" is a comprehensive but simple-to-use software for data analysis.

Reviewer 2 Report

 

General comment

The manuscript presents very interesting and useful estimation of N content in leafs and plants from UAV Digital Image. The rapid estimation of N content in plant is essential for the decisions of cotton fertilization, mainly in early stages of the growth.  The authors successfully developed regression models between LNC and PNC contents in different growth cotton stages and UAV images. I believe that the final algorithm will be very useful for the precise nitrogen fertilization of cotton and I believe also other crops as it will substitute  the slow and expensive laboratory analysis . The manuscript is written in a good English, the results are clear and promissing. Possibly in conclusion, some information or sentence about the future real UAV use for the plant fertilization should be appropriate.

 

Questions:

Had authors some unfertilized control? Did authors try to have a low nitrogen dose in order to be able to notice also N insufficiency in leaves or in cotton plants? Eventually, is the dose 195,5 kg N/ha so low to see eventual problems with nitrogen nutrition?

 

Specific comments

Lines 85-86:Is the nutrient content of soils the average of three years or were there differences in nutrient content measured only in the first year? In addition a short explanation of used methods for N and mainly K determination should be useful – similarly as Olsen P.

The P2O5 and K2O contents should be changed and recalculated in P and K content.

Lines 89-90: Please indicate the size of plots

Lines 96: Phosphorus…. not phosphorus  (new sentence)

Line 163: divide R2of

Author Response

1. Had authors some unfertilized control? Did authors try to have a low nitrogen dose in order to be able to notice also N insufficiency in leaves or in cotton plants? Eventually, is the dose 195.5 kg N ha^-1 so low to see eventual problems with nitrogen nutrition?

Response: Thanks for reviewer’s comment.

There is no unfertilized control in this study. We set four different nitrogen application treatment, and 195.5 kg ha^-1 was the lowest nitrogen dose. According to our previous research work, the 195.5 kg ha^-1 nitrogen dose significantly affected cotton growth, nitrogen accumulation and yield production. 

[1]. Yan Chen, Ming Wen, Yang Lu, Minghua Li, Xi Lu, Jie Yuan, Yang Liu*, and Fuyu Ma. Establishment of a critical nitrogen dilution curve for drip-irrigated cotton under reduced nitrogen application rates. Journal of Plant Nutrition, 2022,12(45): 1786-1798. DOI: 10.1080/01904167.2022.2027973

[2]. Ming Wen, Wenqing Zhao, Wenxuan Guo, Xiaojun Wang, Penbing Li, Jing Cui, Yang Liu* and Fuyu Ma*. Coupling effects of reduced nitrogen, phosphorus and potassium on drip-irrigated cotton growth and yield formation in Northern Xinjiang. Archives of Agronomy and Soil Science, 2021, 68(9):1239-1250. DOI: 10.1080/03650340.2021.1881776

[3]. Yang Liu, Ming Wen, Minghua Li, Wenqing Zhao, Penbing Li, Yan Jiang, Jing Cui, and Fuyu Ma*. Effects of reduced nitrogen application rate on drip-irrigated cotton dry matter accumulation and yield under different phosphorus and potassium managements. Agronomy Journal, 2021,113(3):2524-2533. DOI: 10.1002/agj2.20625

[4]. Zhiqiang Dong, Yang Liu*, Baoxia Ci, Ming Wen, Minghua Li, Xi Lu, Xiaokang Feng, Shuai Wen and Fuyu Ma*. Estimation of nitrate nitrogen content in cotton petioles under drip irrigation based on wavelet neural network approach using spectral indices. Plant Methods, 2021, 17(1):90. DOI:10.1186/s13007-021-00790-x.

2. Line 85-86: Is the nutrient content of soils the average of three years or were there differences in nutrient content measured only in the first year? In addition a short explanation of used methods for N and mainly K determination should be useful-similarly as Olsen P.

Response: The nutrient content of soils is the average of three years, and the information was added in Line 89-90.

The short information of used methods for N was added in Line 144-146.

As there is no K and P content used in this manuscript, so we did not add the determination of P and K in this study.

3. The P₂O₅ and K₂O contents should be changed and recalculated in P and K content.

Response: The P₂O₅ and K₂O contents have been changed and recalculated in P and K content, and the P and K content were added in Line 99.

4. Line 89-90: Please indicate the size of plots.

Response: The size of each plot was mentioned in Line 97-98.

5. Line 96: Phosphorus… not phosphorus (new sentence)

Response: phosphorus has been replaced by Phosphorus in Line 98.

6. Line 163: divide R²of

Response: space was added between R² and of in Line 173.

Reviewer 3 Report

Report Reviewer

 

Summary brief

The manuscript proposed a rapid and accurate estimation of leaf nitrogen content (LNC) and plant nitrogen content (PNC) of cotton in a non-destructive manner using RGB images of cotton fields. Images were acquired using a low-cost unmanned aerial vehicle (UAV) with a visible light digital camera during different growth stages. The correlation between N content and visible light vegetation indices (VIs) was analysed using different machine learning algorithms such as:  Random Forest (RF), Support Vector Machine (SVM) and Back Propagation Neural Network (BP) to develop models for estimating N content.

 

General concept comments

Unfortunately, the article has some important conceptual and methodological errors. The introduction is well done but does not explain well the knowledge useful for understanding the article. The methodology applied is not well explained and has several errors. The results are not complete and do not lead to valid scientific results because the methodology is not adequate. The discussions are also not exhaustive. However, a PDF with observations made directly on the text has been attached for further details.

Comments for author File: Comments.pdf

Author Response

1. The article has some important conceptual and methodological errors. The introduction is well done but does not explain well the knowledge useful for understanding the article. The methodology applied is not well explained and has several errors. The results are not complete and do not lead to valid scientific results because the methodology is not adequate. The discussion are also not exhaustive. However, a PDF with observations made directly on the text has been attached for further details.

Response: Thanks for reviewer’s comments. The manuscript has been revised.

2. Line 89：How large was each plot?

Response: the size of each plot was mentioned in Line 97-98.

3. Line 90-97: Various fertilizer treatments have been described, as the title suggests, however there is nothing in the results and discussion.

Response: The various fertilizer treatments were set in this study because of the soil environment is not only composed of N major element, it also contains phosphorus and potassium two major elements. The different NPK fertigation managements affect cotton N absorption and utilization at different growth stages. While the cotton LNC or PNC estimation by using digital cameras under different fertigation conditions is poorly understood.

The manuscript aims to explorer whether the color indices from RGB images can be used to estimate the LNC or PNC of cotton at different growth stages under different fertigation conditions. The results mainly focused on the changes of LNC, PNC, color indices and model establishment under different fertigation conditions, so there is no detailed information about P and K elements, as well as fertigation managements in Results and Discussion.

4. Line 106: GPS or GNSS?

Response: We have double checked DJI Mavic Pro has both GPS and GLONASS, and in this study we set as GPS.

5. Line 114: The flight altitude is very low. In the literature, this parameter is much higher.

Response: The flight altitude generally ranges from 10 to 50 m, some studies set even higher for bigger experimental area. Based on our previous research results (manuscript is ready to submit, data not shown), we set the altitude as 10 m with a ground resolution of 1.3 cm in this study, which could meet our requirements for field image acquisition and data analysis.

6. Line 139-145: - Why did you choose to select these indices?

- Explain more about how you divided an RGB image into individual bands by having a single sensor.

Response: There were many studies using RGB images derived from UAV to monitor crop growth status, the indices we choose in this study were based on those references. Additionally, nitrogen content could affect crop canopy chlorophyll content, which also affect visible light absorption. So we choose those indices to find optimal indices for cotton nitrogen nutrition diagnosis.

References:

[1] Jiale Jiang, Weidi Cai, Hengbiao Zheng, et al. Using digital cameras on an unmanned aerial vehicle to derive optimum color vegetation indices for leaf nitrogen concentration monitoring in winter wheat. Remote sensing, 2019, 11,2667.

[2] Songyang Li, Fei Yuan, Syed Tahir Ata-UI-Karim, et al. Combining color indices and textures of UAV-based digital imagery for rice LAI estimation. Remote Sensing, 2019, 11,1763.

[3] Ning Lu, Jie Zhou, Zixu Han, et al. Improved estimation of aboveground biomass in wheat from RGB imagery and point cloud data acquired with a low-cost unmanned aerial vehicle system. Plant Methods, 2019,15:17.

For how divided an RGB image into individual bands by having a single sensor.

The detailed methodology for data processing we have added in the Manuscript, see Line 122-128.

7. Line 187: What is the unit of measurement?

Response: there is no unit for R, G, B value.

8. Line 189: Explain better what is this value.

Response: Sorry for mistake, we have deleted the unit. There is no unit for R, G, B value.

9. Line 191: What is the unit of measurement?

Response: there is no unit for R, G, B value.

10. Line 250: This is a very important error. It is not possible to use certain inappropriate references.

Response: The references we cited in this sentence were [30] Blackmer, T. M.; Schepers, J. S.; Varvel, G. E.; Walter‐Shea, E. A. Nitrogen deficiency detection using reflected shortwave radiation from irrigated corn canopies. Agronomy Journal. 1996, 88, 1-5. [31] Zhao, D.; Reddy, K. R.; Kakani, V. G.; Reddy, V. R. Nitrogen deficiency effects on plant growth, leaf photosynthesis, and hyperspectral reflectance properties of sorghum. European Journal of Agronomy. 2005, 22, 391-403.

In [30], the author mentioned that “Chlorophyll meter technology has potential for rapidly assessing crop N availability by measuring transmittance of light through a leaf about 650 and 940 nm. These meters can identify N deficiency of corn because an N deficiency reduces chlorophyll content of leaves, which in turn increases the amount of light transmitted through a leaf”, so we considered to cite in this manuscript.

In [31], the author mentioned that “Nitrogen treatments mainly affected spectral reflectance in both red and near-infrared (NIR) regions with the increase in reflectance in the red region and a decrease in reflectance in the NIR region for N-deficient corn canopy”, and “reflectances near 550 and 710 nm were better for detecting corn plant N deficiencies compared with reflectances at other wavelengths” so we considered to cite in this manuscript.

11. Line 256-257: There are not the references.

Response: the references were added in Line 267.

Round 2

Reviewer 1 Report

Dear Authors,

I regret it, but I must recommend that your paper should be rejected because in my opinion your research were not conducted correctly and the applied methodology has serious flaws. The particular reasons are as follows:

One of the aims of your research is to compare the performances of several machine learning models for particular task of nitrogen content estimation in cotton plants. To fairly compare the performance of algorithms, they have to be tuned thoroughly and the data have to be prepared for them in an appriopriate way. If not, they may underperform and the comparison may lead to wrong conclusions. In your case, you did not prepare the data in the way the algorithms need.

For SVM and neural networks, the achieved performances may be deteriorated due to inadequately prepared input data. For both these methods, input data should be transformed to a uniform scale of values. In your case you have RGB values which are several orders of magnitude higher than some VI values. To assure the best possible performance of SVM and Neural Network models, all predictors should be centered and scaled. Moreover, for Neural Network you should remove highly correlated variables from your inputs.

Secondly, although I asked for it, you also did not provide details on how the hyperparameter values were selected. So, it is not clear if the optimal values for your dataset were used. Again, it is crucial if we want to conclude about the aplicability of particular algorithms.

 

You also wrote you applied the crossvalidation approach. However, from your text it seems you did a holdout approach. My impression is, you just divided the dataset into training and validation sets, did the training based on trainig data using set hyperparameter values and validate the performance of the prediction using validation dataset. It may be considered as enough, if you want to assess the value of prediction in your particular model. But it is not appriopriate if you want to compare the performances of algorithms for a particular purpose in general.

Cross-validation is a resampling method that uses different portions of the data to test and train a model on different iterations. 

My additional comment on data preparation description:

I understand that you have used an approach based on Excess Green Index as used in Woebbecke et al. (1995).  Please use in text the formula (2g – r - b) for clarity (I mean put the formula in brackets). Refer to source publication of this approach (eg. Woebbecke et al. (1995)). I would also suggest to add a figure to illustrate the approach (eg. Similar to Fig 2 in Wang et al. 2020.)

References:

1. Le Wang , Ming Wen , Pengbing Li , Minghua Li , Zhiqiang Dong , Jing Cui ,Yang Liu & Fuyu Ma (2020): Growth and yield responses of drip-irrigated cotton to two differentmethods of simulated hail damages, Archives of Agronomy and Soil Science

2. D. Woebbecke, G. Meyer, K. VonBargen, D. Mortensen Color indices for weed identification under various soil, residue, and lighting conditions. Trans. ASAE, 38 (1) (1995), pp. 271-281

 

 

Please add the sentence from your response “After we extracted the RGB value, we used average data for further data analysis.” to the text about data collection (eg. In line 136)

 

Author Response

Reviewer 1

Dear Authors,

1. I regret it, but I must recommend that your paper should be rejected because in my opinion your research were not conducted correctly and the applied methodology has serious flaws. The particular reasons are as follows:

One of the aims of your research is to compare the performances of several machine learning models for particular task of nitrogen content estimation in cotton plants. To fairly compare the performance of algorithms, they have to be tuned thoroughly and the data have to be prepared for them in an appriopriate way. If not, they may underperform and the comparison may lead to wrong conclusions. In your case, you did not prepare the data in the way the algorithms need.

For SVM and neural networks, the achieved performances may be deteriorated due to inadequately prepared input data. For both these methods, input data should be transformed to a uniform scale of values. In your case you have RGB values which are several orders of magnitude higher than some VI values. To assure the best possible performance of SVM and Neural Network models, all predictors should be centered and scaled. Moreover, for Neural Network you should remove highly correlated variables from your inputs.

Secondly, although I asked for it, you also did not provide details on how the hyperparameter values were selected. So, it is not clear if the optimal values for your dataset were used. Again, it is crucial if we want to conclude about the applicability of particular algorithms.

 You also wrote you applied the crossvalidation approach. However, from your text it seems you did a holdout approach. My impression is, you just divided the dataset into training and validation sets, did the training based on training data using set hyperparameter values and validate the performance of the prediction using validation dataset. It may be considered as enough, if you want to assess the value of prediction in your particular model. But it is not appropriate if you want to compare the performances of algorithms for a particular purpose in general.

Cross-validation is a resampling method that uses different portions of the data to test and train a model on different iterations. 

Response: thanks for your comments. About the data scale you mentioned in comments, we did not write clearly in the Results Part. We totally used 30 VIs in this study, while after we did correlation analysis, there are twenty three of the thirty VIs showed positive or negative correlations. The bn, GBRI, ExB, g-b and INT, r-b were the most strongly and weakly correlated to LNC, respectively (GBRI and g-b: r = 0.80, p-value < 0.01; bn and ExB: r = -0.80, p-value < 0.01; INT: r = 0.05, p-value > 0.05; r-b: r = -0.05, p-value > 0.05). For the PNC, twenty eight of the thirty VIs showed positive or negative correlations. The highest correlations with PNC were found for bn (r = 0.79, p-value < 0.01), with the lowest for r-b (r = 0.01, p-value > 0.05). Generally, the correlations of VIs with PNC were weaker than those with LNC.

Among the twenty three of thirty VIs in LNC and twenty eight of thirty VIs in PNC, the input parameters for model development was that the correlation coefficient between VIs and LNC(PNC) more than 0.50. So for model development, we did not use R, G, B original values as the correlation coefficient was not over 0.50. Further, for the sensitive VIs we chose, the data was calculated based on the formula, so all VIs we chose was in a small scale, and it can meet the algorithms need.

For detailed information about input VIs (the input parameters for model development were that the correlation coefficient between VIs and LNC(PNC) more than 0.50) we have added in Line 192-194.

For data split, honestly speaking, the software we used is DPS (Data Processing System), when we use the RF, BP, SVR model, the default split method was cross-validation.

2. My additional comment on data preparation description:

I understand that you have used an approach based on Excess Green Index as used in Woebbecke et al. (1995).  Please use in text the formula (2g – r - b) for clarity (I mean put the formula in brackets). Refer to source publication of this approach (eg. Woebbecke et al. (1995)). I would also suggest to add a figure to illustrate the approach (eg. Similar to Fig 2 in Wang et al. 2020.)

References:

1. Le Wang , Ming Wen , Pengbing Li , Minghua Li , Zhiqiang Dong , Jing Cui ,Yang Liu & Fuyu Ma (2020): Growth and yield responses of drip-irrigated cotton to two differentmethods of simulated hail damages, Archives of Agronomy and Soil Science
2. D. Woebbecke, G. Meyer, K. VonBargen, D. Mortensen Color indices for weed identification under various soil, residue, and lighting conditions. Trans. ASAE, 38 (1) (1995), pp. 271-281

Response: the formula has put in the brackets. And the figure has been added according to reviewer’s comments as new Figure 2.

3. Please add the sentence from your response “After we extracted the RGB value, we used average data for further data analysis.” to the text about data collection (eg. In line 136)

Response: we have added the sentences “After we extracted the RGB value, we used average data for further data analysis.” in Data Collection part, see Line 110-111.

Author Response File: Author Response.docx

Reviewer 3 Report

Report Revision 2

 

The manuscript was modified in several parts, increasing its scientific value. However, there are some problems with the design and the results of the experiment. In general, materials and methos have been edited well, description and understanding have improved. However, the problem in this section is the very low flight altitude. In the previous revision the authors did not indicate in Table 3 whether the reference is the author of the index or the user who used it.

In the Result section, figure 5 should be larger and the red or blue dots are not described. From three years of data the results showed few and not very satisfactory. I recommend implementing this part, even adding images. Furthermore, the effect of different fertilizations on the spectral and vegetative response of the plant should be highlighted, for example by carrying out a multi-way ANOVA analysis to verify if there were any interactions.

Author Response

The manuscript was modified in several parts, increasing its scientific value. However, there are some problems with the design and the results of the experiment. In general, materials and methos have been edited well, description and understanding have improved. However, the problem in this section is the very low flight altitude. In the previous revision the authors did not indicate in Table 3 whether the reference is the author of the index or the user who used it.

Response: thanks for reviewer’s comments. The references in Table 3 were the user who used it for research, not the author of the index.

According to our field scale and data quality, the flight altitude for 10 meters could meet our research needs.

2. In the Result section, figure 5 should be larger and the red or blue dots are not described. From three years of data the results showed few and not very satisfactory. I recommend implementing this part, even adding images. Furthermore, the effect of different fertilizations on the spectral and vegetative response of the plant should be highlighted, for example by carrying out a multi-way ANOVA analysis to verify if there were any interactions.

Response: thanks for reviewer’s comments. we have described the red and blue dots in the Figure 5 below the figure. This Figure 5 is for model validation, and we have both 48 data in LNC or PNC for 3 years.

For multi-way ANOVA analysis, we are sorry for not getting the reviewer’s point. Does it mean that the multi-way ANOVA analysis for sensitive VIs we choose for model development or total 30 VIs with different fertigation treatment during 2018-2020. Or the multi-way ANOVA analysis for LNC and PNC in different fertigation treatment during 2018-2020.

Sorry for that, so we did not update this apart analysis.

Author Response File: Author Response.docx