Leaf Area Index Estimation of Pergola-Trained Vineyards in Arid Regions Based on UAV RGB and Multispectral Data Using Machine Learning Methods
Round 1
Reviewer 1 Report
Dear authors!
You wrote a very professional paper, and I learned a lot while reading it. There are, however, some points which I request you to address and/or which I would like to bring to your attention. They are listed as "Generic comment" and "specific comments"
My comments have the following format
Line number : your_text -> my_suggested_change
sometimes followed by a comment.
Most of my suggested changes are editorial.
GENERIC
Parameter. This paper confuses “parameter” and “variable”. The two words do not mean the same thing. For instance, you say (Line 15) that “Leaf Area Index (LAI) [is] a valuable parameter” and then (42-43) “that one of the most important variables [is] the leaf area index (LAI)”. LAI is generally a variable. A parameter is a switch, a value that can be modified in a model; a variable is mostly observed. Environmental conditions are usually described by variables (rainfall, sunshine etc). If you fine-tune a plant model by adjusting light interception efficiency, then light interception efficiency is a parameter. Sometimes, there are situations where one can hesitate between parameter and variable, but an observed factor is a variable, not a parameter. Your use of “parameter” in lines 183-184 and 294 is perfect.
Were Vs Are. "Were" is past. "Are" is present. Saying (50-51) “optical methods were based on the direct or diffuse light penetration through the canopy” is wrong. It means: “In the past, optical methods were based on the direct or diffuse light penetration, but this is no longer the current situation”. Therefore, you should say “optical methods are based on measurement of the direct or diffuse light penetration through the canopy.” I know this is difficult for Chinese persons. One can say “vines were sampled every Saturday”, implying that this was done in the past. There are many cases where “were” must be replaced by “are”, including on lines 50, 56, 75, 80, 127, 313, and others. Please correct this.
Number of samples/data point . This comment applies to figure 6 and tables 3 and 4. Please indicate, maybe in the legend, the number of data points/samples the correlations are based on and indicate those that are significant. Indicating “best” R² and RMSE is not very useful.
Language: there are still many weaknesses and the paper should be re-checked, especially the discussion where I have done few language corrections
References: I have not checked any references
SPECIFIC
16: managements such as > management, including
17-19: some Vegetation Indices (VIs) has been successfully used to assess LAI variation, but they were unsuitable to vineyards of different sorts and structures. → some Vegetation Indices (VIs) have been successfully used to assess LAI variations, they are unsuitable for vineyards of different types and structure.
20: study was aimed → study aimed
26: while reached to 0.802 and 0.630 from RGB data → while reaching 0.802 and 0.630, respectively, with RGB data
27: performed better in multispectral data than RGB images → performed better than RGB images with multispectral data.
27: in higher → at higher
29: larger area → large areas
34: were → are
36: was → is
37: the terrible widespread pandemic of COVID-19 → the COVID-19 pandemic
Comment: pandemic implies that it is widespread.
39-40: growing parameters were difficult to monitor and carry out targeted managements → growing conditions are difficult to monitor
44: parameter → factor
47: direct measurements were destructive→ of which the measurement is destructive
50, 56, 75, 80, 127: were → are
55: which based → which is based
61: which they can → which can
63: but they → they
64: Another → In addition
Comment: or “Furthermore”
65: to → with
Comment: or “for”
68-70: rapidly emerging remote sensing platforms and techniques are capable of obtaining field information through analysis of spectral characteristics of crops, complementing the inadequacies of existing ground-based measurement methods
73: products can be used to very large → products that can be used over very large
76: 8m separately. → 8 m, respectively.
78: parameters → variables
79: more higher → higher
82: which becoming → which are becoming
83: is → are
84: or →
Comment: drop “or”
85: better advantages → advantages
86-87: Applying these kinds of data is easy to make decisions for corresponding managements → This kind of data makes management decisions easier
91: However, Machine → Machine
Comment: start a new paragraph with the word “Machine”. Like this
[…] ranging from leaves to landscape and landscape to regional scales.
Machine Learning methods, such as Support Vector Machines (SVM) […]
94: and being → and are being
98: have advantages → have the advantage
100: resultant dried productions
Comment: this is not cleat to me. What is “dried production”? Is it “biomass production” expressed as dry weight? Maybe “dry weight production” would be a good option.
103: water → irrigation
107-110: For filling this gap, using UAV multispectral and RGB sensors which flown in conjunction with LAI ground measurements in different vine growth stages, this study constructed several VI-LAI estimation models applying four different ML methods, namely the SVR, RFR, PLSR and GBR→ In order to fill this gap, the present study combined airborne UAV multispectral and RGB sensors with LAI ground measurements at different growth stages to construct several VI-LAI estimation models using four different ML methods, namely the SVR, RFR, PLSR and GBR
116: algorithm which suitable → algorithm suitable
120: how much → to what extent
122: same conditional → similar
128:
Comment: are you referring to vine quality of wine quality? You are also not saying what the grapes are used for: table grapes or wine grapes? The concept of “quality” is very different for the two uses.
130: reaches to 47.8°C and to -28.0°C→ reaches 47.8°C and -28.0°C
Comments: please do specify which years your data refer to (e.g. 1981-2010). It is not clear what temperatures those are. Is this the daily average of the warmest and coldest months, respectively? Is it the average of the daily maximum in July and the daily minimum in January (which is the coldest month in Turpan). Is this the average of the warmest (or coldest) day in the year over a number of years? The information cannot be understood if you don’t clarify this.
131: in winters preventing freezing damages→ during winter to protect them from freezing
134: good qualities, and therefore became a suitable research area for this study. → good quality, and therefore a suitable research subject for this study.
137: hanged → attached
Comment: two changes
147: covering most → covering the main
Comment: saying “most” implies that there are several growing seasons
151: of a good quality of orthomosaic → of good quality orthomosaic
160-161: was installed to UAV, which is able → was installed on the UAV and is able
170: images were resulted → images resulted
180: which developed → which was developed
187: which → as
191-192: quality which taken by VitiCanopy → quality as assessed by VitiCanopy
Comment: the modification suggested above is based on the assumption that VitiCanopy does provide an estimate of measurement quality
198: experimented → studied
Comment: you could also say “sampled”
204: in → from
206: leaves which measured → leaves measured
209: Another, we added → We added
210: was → is
213: correspond → corresponds
238: is proved to be capable of estimating → can estimate
241-243: then calculate the fractions of foliage cover and crown cover by applying the equations as following [60]: → then calculates the fractions of foliage cover and crown cover by applying the following equations [60]:
252: Beer’s Law as following equations [58, 61]:→ Beer’s Law [58, 61]:
249, 261, 306: Where → where
264: following → follows
273: which applicable → which are applicable
264: VIs from → VIs are from
283: In this study, four machine learning → Four machine learning
292: set 30 → set to 30
295-296: using datasets of separately 0.007m, 0.045m GSD data and a mixture of them → using 0.007m and 0.045m GSD datasets separately as well as mixed
299: which can be expressed as following formulas :→ using the standard equations :
312: LAI values, which based → LAI values based
323: except for Green band
Comment: … which is not a surprise!
326: eight of them showed → eight showed
329: showed → show
335: VIs which derived → Vis derived
343:
Comment to figure 7: some relations are obviously non linear, e.g. LAI_viti versus NDVI, SAVI, EVI1 and EVI2 in line (e). The fact should be mentioned in the comments to the figure, and the discussion should attempt to provide an explanation. Also note that your regression lines extend into negative LAI, which is meaningless. You would also achieve much better correlations with a curvilinear regression. I have resampled the best I could (n=180) the points in the left-bottom graph of figure 7. With a shifted power function (3 parameters) I get R²=0.74 (MSE = 0.87, DOF=177) while the linear regression (2 parameters) gets me R²=0.66 (you have R²=0.65, i.e I’m not too far off) (MSE = 0.99, DOF=178). In this case, non-linear seems to be much better than linear. Also see my comment to 542-454: you may have created the non-linearity.
358: models which trained → models trained
366: group → groups
366: not too obvious
Comment: you could say “negligible”
372: showed superior than others → were superior to others
373-375: Among the models based on all VIs, SVR and GBR showed better, PLSR and GBR performed better with only 11 VIs, and there were no stable ML methods performed better in models of using only 3 band data
Comment: this is not clear. Kindly rephrase.
376: always performed superior than ones → always outperformed those
377: models which based → models based
378: had inferior performances → performed worse
388: shows discrepancies for different vegetations, structures → varies according to vegetation type, structures
390: researches discussed → papers discuss
396-396: near to Bordeaux → near Bordeaux
396: out it oscillates with the difference of → it changes with
398: which giving rise to much uncertainties → giving rise to many uncertainties
403: attempted to find a variable k, which ranging from 0.29 to 0.56 → experimented with a variable K in the range from 0.29 to 0.56
405: Table 5Table 5 → Table 5
409: not complacently obvious → not obvious
421: photosynthetic vegetations → vegetation types
Comment: all vegetation is usually green, i.e. “photosynthetic”. There are some parasitic plants, mushrooms, moulds etc, but they do not assemble as “vegetation”
425-428: But performances of different ML methods are varied with different datasets. In brief, multispectral sensors would be the best choice for LAI estimations, while RGB sensors also be able to estimate LAI to some extents with its low costs and easy availability. → The performance of different ML methods varies according to datasets: multispectral sensors would be the best choice for LAI estimations, but RGB sensors can also be used as a low cost and easily available alternative.
452-454: Among VIs of multispectral data, DVI is the division of near infrared to red, and
LogR is the logarithm of DVI (Table 2) → Among multispectral data based VIs, DVI is the ratio of near infrared to red, and LogR is the logarithm of DVI (Table 2)
Comment: This may be so, but the log transformation also introduces a non-linearity, which potentially distorts your data
457: and tried → after which we tried
476: values of R 2 and RMSE, which generated → values of R 2 and RMSE, generated
478: datasets, which evaluated according to R² → datasets, based on R²
480: stable → reliably
487-488: many factors and highly collinear → many and highly collinear factors
493: Another, all → In addition, all
495: less → few
501: In sum → In a nutshell
507-508 and 510-511: Red rombatic and black rectangular → Red diamonds and black squares
Comment: I have never come across “rombatic” and I could not find it in any dictionary. The figure is called “lozenge” but usually referred to as “diamonds” in graphs. As to “rectangular”, this is not wrong as a square is a rectangle, but, again, in graphs the shape is mostly called “square”
514: the underlying ability → the ability
Comment: you could also make this a bit stronger by saying “potential” or “usefulness” instead of ability
515: lay a solid foundation → provides solid ground
519: values which estimated → values estimated
527: models which based → models based
530: which R 2 reaches to 0.802 and RMSE of 0.630 → with R² reaching 0.802 and RMSE 0.630
537: However, this study → This study
540: through different growing periods → at different crop stages
Author Response
Dear reviewer,
Many thanks to your comments, and we have tried to address all your comments carefully and improve the manuscript accordingly.
For the resoponses to your comments, please see the attachment.
Thank you again, and best wishes!
Author Response File: 
Author Response.docx
Reviewer 2 Report
The manuscript is interesting, but required extensive editing of English language is making it hard to evaluate.
Example of necessary language improvements (just for abstract):
Line 15-16: rewrite sentence
Line 17: has –> have
Line 18: use present tense, were –> are, no “but”, to –> for
Line 20: no “was aimed to”, use present tense, you should state what are the models used for
Line 25-26: rewrite sentence
Line 27: in –> using
Necessary content improvements:
It is obvious that the article is a result of an extensive research.
But, the manuscript appears not so clear in description of advantages of proposed method(s). Comparison to available methods, e.g. methods developed for “normal” vineyards (applied on pergola-trained vineyard), should point out the accuracy increase of proposed methods.
Author Response
Dear reviewer,
Many thanks to your comments, and we have tried to address all your comments carefully and improve the manuscript accordingly.
For the resoponses to your comments, please see the attachment.
Thank you again, and best wishes!
Author Response File: Author Response.docx
Reviewer 3 Report
In this paper, the authors calibrated the LAI estimation result from mobile phone app, and built machine learning models based on that calibrated LAI and UAV images. The novelty and contribution of the method are limited. A lot of results and conclusions in the manuscript do not have sufficient evidence. The fatal problem is that the authors mixed different sampling conditions (like flight height) together in both data analyzing and model building. What’s more, the discussion section of this article need a thorough reorganize. In its current version, section 4.1 is more like a part of “Results”. Section 4.2 is totally meaningless. Sections 4.3 and 4.4 are confusing. I will list the detailed problems below. Here are some detailed comments.
Page 3, line 105. “LAI estimation were focused mostly on Vertical Shoot Positioning trained vineyards, and relatively few studies have been conducted for pergola-trained vineyards.” What’s the difference? Why do we need a novel study on pergola-trained vineyards? Can we directly use the Vertical shoot positioning trained vineyards model?
Page 2, line 119. “using VIS derived from UAV based RGB and multispectral data applying several ML methods”. This description has ambiguity. It seems more like you are going to build a model involving both RGB and multispectral data. In fact, you just want to do some comparisons of the two sensors.
Page 4, line 149. “images were captured from flight heights of 17m and 100m”. Why? Please explain your purpose clearly in this part because you never mentioned the flight height of UAV before.
Page 4, line 161. “green (520-600nm) , red (630-690), and near infrared (760-900nm)” the spectral bands are pretty broad. Are they proper for quantitative remote sensing? You need some references to support this sensor.
Page 5, table 1. Why is there a “91m” in the flight height column? You never mentioned it in the context of this article.
Page 5, figure 2. I suggest to add pictures captured from different height and growth period of vine in this figure.
Page 7, figure 4. This diagram is not strict. You put two drones in the sky, but you claimed that the two sensors were installed together and shoot images simultaneously. This is important because once the two sensor shoot from different angles, the BRDF will carry new problems to your study. Another suggestion is that the second column of the small table in the lower right corner of this figure is better to be moved to the rightmost.
Page 8, table 2. Some indices, such as GtoR and RmG were only listed in the RGB camera region. However, they could also be applied to your multispectral sensor. Why don’t use these two indices in multispectral sensor?
Also in this table, you wrote a lot of “Nir”s, please correct them.
Page 9. Line 302. Why do you use R square to assess the performance of the models? This index is always used to evaluate the model fitting effect.
Page 9, figure 5. The model error is unneglectable. Considering that VitiCanopy calculated LAI is treated as ground truth in the following parts of the manuscript., the error transfer between VI-VitiCanopy LAI and VitiCanopy-ture LAI have to be discussed.
Page 10, figure 6. What caused the obvious difference of RGB indices from different heights? This is not reasonable compared with the multispectral sensor’s result. Please provide the full pictures of the two sensors at different heights.
Page 11, figure 7. You compared the estimation results of 0.007 and 0.045m GSD data. However, the samples involved in the two models are different. The comparison is not reasonable.
Page 12, line 360 and line 363. Similar with the previous problem. You cannot compare two statistical models using different sample sets.
Page 14, section 4.1. This section should be result. Because this is a necessary step of your method. To make this part more like a part of discussion, you can do a stepwise simulation of k and LAI to show the stability of the calibration.
Page 15, line 465. “Both in multispectral and RGB data, feature selection procedures improved model performance comparing to suing of all features (VIs) or only 3 band data”. If so, why don’t do the selection in your method? And you can’t just claim it without any data and results.
Page 16, line 476. “show the maximum and minimum values of R2 and RMSE”. Why show maximum and minimum values? Max and min values cannot tell the statistical distribution of the datasets. Stdv or quantiles are more proper here.
Page 16, line 492. “better stability in the mixture of different GSD data than using single one”. This is obvious because mixture dataset has more samples.
Page 16, line 494. “lower stability and accuracy in lower dimensional datasets”. You cannot ascribe the lower stability and accuracy to the so said “dimensional”. In fact, the 3 bands cannot reflect enough sensitive information in you figure 6. That’s the main reason of the bad performance.
Page 16, line 496. “dimension and spatial resolution are the main factors that affect the final performance of ML models.” This is not reasonable. How can dimension and spatial resolution be the main reason? How about the indices you calculated before? What exactly do you want to express here?
Author Response
Dear reviewer,
Many thanks to your comments, and we have tried to address all your comments carefully and improve the manuscript accordingly.
For the resoponses to your comments, please see the attachment.
Thank you again, and best wishes!
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The manuscript is significantly improved especially regarding quality and novelty of English language as well as the quality of presentation.
Reviewer 3 Report
Authors have considerably revised the current manuscript and most of the prior concerns have been addressed. I think the current version is sufficiently improved.
