Retrieving SPAD Values of Summer Maize Using UAV Hyperspectral Data Based on Multiple Machine Learning Algorithm
and Yulong Bao 5Round 1
Reviewer 1 Report
Chlorophyll is one of the important physiological and biochemical parameters of plants, the quick and accurate estimation of its content plays an important role in plant growth monitoring. The author focuses on the current popular UAV imaging technology, combined with 5 spectral processing algorithms, 4 spectral feature screening algorithms and 4 modeling methods to estimate the SPAD value of maize at different growth stages, which has certain scientific significance. However, for publication in this journal, the manuscript will need to undergo substantial revision.
Major comments
1. In Part 2.2.3, the authors indicated that a total of 90 SPAD data were collected for modeling. Are these 90 data the sum of the measured data of three growth periods or the measured amount of data of a single growth period? If it is the sum of three growth periods, what is the distribution of data collection in each period? Please give a detailed explanation.
2. Did the authors construct the SPAD estimation model separately for the three growth periods when constructing the model? Or data from all three growth periods used? If it's the latter, then the spectral reflectance of each growth period will have certain differences due to the growth of crops, will this difference affect the model accuracy? How does the author consider and deal with this difference?
3. Section 3.3 is the evaluation of SPAD value estimation model results by four modeling methods. However, the author spends a lot of space to analyze the modeling process content of PLSR and RF methods in detail, while the model process of XGBoost and DNN is not explained. For the evaluation of the final modeling results, only Figure 11 is presented, but the figure is not explained and compared. In addition, the author pointed out in the abstract and the following paper that the accuracy of the DNN estimation model based on spectral features selected by CART_SPA is the best among all models. However, in figure 11, especially its R2 value is not significantly different from that of the DNN model based on spectral features selected by CART and SPA. Even, there is no significant difference between the R2 of the proposed model and the PLSR model based on the spectral features selected by CART_SPA.
4. In section 3.4, in the visualization of the results, only the fitting results of the four method models to the July 30 maize SPAD value are given, while the model results for the other two periods are given in the discussion in Section 4. I suggest that in this section, visualization results of the optimal model (such as the DNN model based on the spectral features selected by CART_SPA as described by the authors) and the corresponding model based on the full-spectral features are displayed uniformly. In addition, this section compares the visualization results of each modeling method based on the full spectral feature model and the preferred spectral features model. However, in the model evaluation in section 3.3, the description and analysis of the full-spectral-feature-based results for each method is not given. Authors are advised to add relevant content in section 3.3.
5. Section 3.1 gives the comparison and fitting of the ASD and the UAV spectrum. Has the authors analyzed the corresponding SPAD value estimation model based on the ASD spectrum using the same method as the UAV spectrum modeling, to better compare the validity of UAV hyperspectral imaging techniques for maize growth information estimation?
6. The description of SPA in lines 355-360 and 379-384 is a duplicate of the introduction in section 2.3. However, this part of the method description should belong to the materials and methods section.
Minor comments
1. In abstract, the full names of some abbreviations are not uniformly expressed. For example, some initials are capitalized and some are lowercase; Lines 34-35, what is 0.82 in the sentence? R2? Need to give a specific meaning.
2. Inappropriate use of many punctuation marks and spaces, etc. in the abstract and main text. Some sentences are even missing a period in the middle.
3. Throughout the text, SPAD Values and SPAD values are used interspersed, not unified.
4. Section 2.2, data collection, mentions three types of data: UAV, ASD, and SPAD, and the authors give detailed introduction to UAV data collection and SPAD data collection, but lacks an introduction to the ASD data collection process.
5. Section 2.3, about the introduction of the FD algorithm formula, the parameters in the formula given by it cannot correspond to the explained parameters in the following part one by one, which will cause trouble to the reader's reading and understanding.
6. Line 110, "Four regression model methods (PLSR, RF, XGBoost and DNN)" appearing for the first time does not give the full name.
7. Line 135: (0.03 ~ 1.9 mm), (0.003 ~ 0.05 mm), Line 154: 450 ~ 998nm, Lines 325-326: “454 ~ 838nm”, “from 838 to 950 nm”, Line: 410 “PC1～PC3”, et al., the use of "~", "", "～" is inconsistent throughout the text.
8. Lines 148-149: “The data were collected 148 on July 30 (Jointing stage), August 12 (tasseling stage) and August 28 (Milky Maturity 149 Stage) respectively.” The capitalization of the first letter of the reproductive period is inconsistent.
9. In the full text, "S185" and "s185" are used interspersed.
10. Line 160: “(Figure2.a).”, Line 172: “Figure 2(d)”, line 179: “(Figure 2.c)”, Line 326: “(Figure.5a)” Line 329: “Figure 5b” et al. The descriptions of the image names in the main text are not uniform.
11. Change Line 300 “the root mean square error RMSE” to “the root mean square error (RMSE)”?
12. Lines 363-364: “are 450，462，470，474，490，498，502，522，534，582，674，718，786，810，866，363 890，910，926,942å’Œ958nm.” Improper use of punctuation , and even appeared in Chinese!
13. Line 371: “RMSECV and RMSEP” The full text did not find the corresponding full name and did not understand its meaning.
14. Line 394 “3.3 Construction of summer maize SPAD Values estimation model Based on characteristic bands” Word case problem.
15. Line 452-455: “According to the accuracy evaluation results in Figure 11, CARS_SPA is used in the four ml regression models_ The precision of the SPAD Values inversion model con-453 structed by SPA algorithm is the best, so the hyperspectral cars collected on July 30 will 454 be_” Sentence incomplete, where ml maybe "ML"?
16. Line 566: “R2” Change the 2 in to superscript.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
There are many grammar errors, punctuation errors, and abbreviation errors in this article, which greatly affect the readability.
There are also many inconsistent wording, for example, SPAD Values vs SPAD values, Maize vs Corn, CARS vs cars; SPA vs spa, S185 vs s185, and so on.
1. Title:
`Suggest to modify the title, for example,
Retrieving SPAD Values of Summer Maize Using UAV Hyperspectral Data Based on Multiple Machine Learning Algorithms
2. Abstract
Line 23: differential spectral reflectance (DSR) logarithmic spectral reflectance (LSR) –> differential spectral reflectance (DSR), logarithmic spectral reflectance (LSR)”.
Line 25: and Then –> and then
Line28: preprocessed by UAV –> obtained by UAV
Line 33: “Adopt” should be deleted.
The abstract should include a detailed description of the quantitative accuracy of the four methods.
3. Introduction
Line 45-46: please reorganize sentences
Line 78: 。> .
Line 85: model, However, -> model. However,
Line 96: The full name should be given when the abbreviation appears for the first time. VIs and SVM.
Line 99: give full name of SPA
4. Materials and Methods
Line 149: Jointing stage, tasseling stage, Milky Maturity Stage -> Jointing stage, Tasseling stage, Milky maturity stage / jointing stage, tasseling stage, milky maturity stage
SPAD-502plus
Line 194: “; “-> “.”
Line 196: give full name of FD
Line 199: give full name of ML
Line 203, in figure 3, Xgboost-> XGBoostï¼› represemtative -> representative; corn-> maize; Train -> Training
Line 211: f peak -> F peak
Line 212-213: How to determine a and b?
Line 245-250: The meaning of bk is introduced in line 249. However, no explanation for b.
Line 302-303: Give explanations of all variables in Equations 3 and 4.
Line 296-301: It is suggested to add absolute error and relative error to the accuracy evaluation index.
5. Results
Line 314-315: delete “below: a. shows the SSR image; b. shows the DSR image; c. shows the RSR image; d. shows the RDSR image; e. shows the LSR image; f. shows the LDSR image.”
Line 319: five-> six
Line 371: what are RMSECV and RMSEP?
Line 375-383: These sentences are more suitable for the method introduction.
Line 402-403: What is the purpose of PCA? PCA is not mentioned in the method part and the technical flowchart
Line 421: What is RSR? RSR is not defined.
Line 425: Figure.6 should be Figure 9.
Line452-457：Please recheck those sentences.
Line 468: Please recheck the figure 12 b and h. These two figure are almost same.
Line 480: Need to supplement references
Line 508: please give more explanations about why the high-frequency information separated by differential transformation is poor in improving the sensitivity of spectrum to chlorophyll content of summer maize in this study.
Line 516, This study does not well support the conclusion that the extraction algorithm of high-frequency information is not suitable for the processing and analysis of vegetation spectral information in the northwest region because it is only concluded in a very small area.
Line 522-523: recheck the sentences.
Line 527-530: XGBoost can resist over fitting. So, in theory, it will perform better than RF. In this study, the results of XGBoost and RF are not significant different. Therefore, it seems possible to infer that RF is not over fitted in this study. However, the authors state that the no significant difference may be due to the over fitting of RF and XGBoost models due to the small number of samples. This explanation is not convincing enough.
Line 532-534: solve what problems?
Line 539-549: It is recommended to rewrite this paragraph.
6. Funding:
Line 575: Is it true that this study is supported by seven projects at the same time?
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Dear authors, thank you for accepting and responding fully and in detail to my comments and suggestions. The manuscript was substantially improved enough to be published in this journal.
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 2 Report
Extensive editing of English language and style are required. Many grammatical errors and inaccurate expressions greatly affect readability.
1. Abstract
1) Why 20 models? Four machine learning regression models, six spectral transformation methods. 4*6=24
2) Is SPAD value estimation model established for each key growth stage of summer maize: July 30 (jointing stage), August 12 (tasseling stage), and August 28 (milky maturity stage)?
3) The abstract is not well written. There are too many introductions to the research steps, but too few descriptions to the conclusions.
4) Lines 32 to 36 need to be rewritten.
2. Introduction
1) Line 47: crop growth -> crop growth monitoring
2) Line 48: he -> The
3. Materials and Method
1) Line 146-150: delete or move to Line 39. Before “As shown in Figure 1,……”
2) Line 158,159,161, Table 1and figure 2: S185 or s185, please be consistent.
3) Line 170: delete 2.2.2
4) Line 178: Please double check 125 spectral curves. five leaves and five points were collected for each leaf. 5*5=25
5) Line 180: off, The -> off. The
6) Line 198-202: Calibration set should be training set? Training set was used for training four ML models.
7) Line 254-256: These sentences are conclusions, not methods. It should not be introduced in the method part.
8) Line 338: what is the difference between n and N? N and n have the same meaning.
9) What is the relationship between wavelength variable selection (2.3) and regression modeling (2.4)? In theory, the optimal wavelength variable is selected through wavelength variable selection algorithm, and then the model is established based on the optimal wavelength variables. However, from Figure 7 to Figure 12, it seems that the authors did not use the optimal wavelength variables selected to build models. Regression algorithms themselves can evaluate and screen the importance of variables.
4 Result
1) Line 346, Whether there are size requirements for ROI?
2) Line 352: Legend font is too small and not clear enough.
3) Section 3.2 is not clear enough; it is recommended to rewrite.
4) Line 482-492: it is recommended to rewrite.
Author Response
Please see the attachment
Author Response File: Author Response.docx
