Evaluation of the Use of UAV-Derived Vegetation Indices and Environmental Variables for Grapevine Water Status Monitoring Based on Machine Learning Algorithms and SHAP Analysis
Round 1
Reviewer 1 Report (New Reviewer)
Dear authors,
a very interesting paper indeed.
Pls. find attached some comments that may help to attribute to minor improvements.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report (New Reviewer)
Reviewing Comments for Manuscript remotesensing-2041545
Wei et al. developed a plant water status monitoring method based on the combination of UAV data collection and machine learning algorithms. The authors planted grapes at field scale for two seasons. During grape growth, authors took aerial images with UAV, then computed vegetation parameters from obtained data. In addition, the authors obtained soil/terrain information and meteorological/temporal data. In the end, the authors applied three machine learning algorithms to predict grape stem water potential and compared predicted values with measured values. Overall, the manuscript is well written, the experiment is valid, and the presented data is solid. I recommend accepting the manuscript with minor revisions. Please find my comments as follows:
Comment 1 (Figure 2): Authors presented ‘total daily reference evapotranspiration in Figure 2. However, I cannot identify a source for this data. Is there supporting information for the manuscript? If not, please add a reference/data source for this parameter.
Comment 2 (Section 2.4): Authors mentioned a soil sensor was used to collect apparent electrical conductivity. Besides, the authors also indicated the usage of elevation and slope. Besides these parameters, are there any other common soil parameters used in this study? For example, sand/silt/clay contents also have an impact on plant water uptake. If no other soil variables were used, please provide a justification about why these common soil parameters were not selected for machine learning.
Comment 3 (Figure 5): Authors presented stem water potential with scatters in Figure 5. However, these scatter overlapped with each other on the time scale. Authors may consider using box plots or box plots with scatters to re-visualize this data.
Comment 4 (Section 2): Authors used several sections to introduce variables used in this study. I suggested adding a table that summarized all variables and presented these variables’ short names (used in Figure 6) and definitions. If there is a length limit, please consider adding supporting information.
Comment 5: (Section 3.4): In tables 4 and 5, authors presented model performance from different combinations of predictive variables. And including ancillary variables significantly improve the model performance as compared with only using derived VI. And the SHAP analysis indicated that some ancillary variables have a dominant contribution to the prediction, such as DOY. So, what will be the model performance if only using ancillary variables for the prediction?
Comment 6 (Section 4): Please be concise in the discussion section. Specifically, please remove the context that is not closely related to the results of this study.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
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
Dear authors, the presented manuscript's main topic is the grapevine water status assessment using machine learning algorithms and a combination of vegetation indices, derived from multispectral images acquired by an unmanned aerial vehicle, and environmental and temporal data. This topic is really interesting if we consider climate change conditions and lower water availability for agriculture especially in mediterranean weather conditions. In my opinion this is a well written manuscript, materials and methods are well described and the quality of data presentation is good. However, there are some changes I would like to suggest in this regard to improve this manuscript. The abstract is too long, in the template of this journal, if I am not wrong, they suggest a paragraph of about 200 words maximum. In keywords, I suggest not to use the words that are already in the title. In data acquisition and weather characteristics during the study period (lines 158-167), I would suggest using a table to show and make more clear that information. Conclusion part looks more like an abstract, I would suggest thinking about it as the contributions of this research compared to previous ones. Despite these suggestions, in my opinion this manuscript's major problem is that I cannot find its novel contribution to the study of precision agriculture. In this sense, we can find previous published research about the use of machine learning and plant water stress detection. Therefore, I would like to see more of that research in introduction and clearly establish the difference and contribution between the presented research and previous ones in discussion and conclusion. My overall recommendation is to reconsider the presented manuscript for publication after major revision.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The work is well-structured and there is a high interest, as it presents an application of UAVs to water management of vineyards. It is not very innovative but introduces new climatic area of application while uses coherent chain process of the data, comparable to other areas. It may have an impact about operational use of low-cost technology as it can increase the economic value of vineyard produce, decreasing the heterogeneities in plots and providing more accuracy on crop monitoring. The literature review is adequate and the instruments and mathematical algorithms and models. There is a good discussion of the results. Nevertheless some issue arise when the review was undertaken,
There is a lack of information about the multispectral instrument used and the process of the data. You should indicate the camera model used, and how the reflectance was calculated (using incoming radiation, reference panels, etc…). It is of crucial importance as the physical magnitude must be well calibrated and referenced to extend these calculations to other areas.
The selection of models such as the vegetation indices shows an important effort, but I think a good initiative is to use the input reflectance bands into the ML models; as was established in the bibliography most of the Vis studied are not designed to be related to mild and short-term water stress. Only those related more directly with chlorophyll content such as TCARI show to be sensitive. This is suggested because during the discussion of results the authors explain the quantitative analysis on the spectral information included in those indices. Explain why you did not try this scenario.
The algorithms applied show that the temporal variables combined with spectral ones are the more adequate to monitor the water stress (show the higher correlation coefficient), it is logical as the water stress potential is a dynamic variable, with an hourly scale of time. Then, the soil characteristics are more recommendable for large spatial analysis than for temporal studies. I recommend using other spectral variables like radiometric temperature for that purpose. Explain why you did not used as the are available on board of UAVs.
It is important to note that you use as a correlation indicator the coefficient r and r2 as they provide the same information. Please choose one and be coherent along the paper.
Some issues about figures:
i) On figure 5 you should include the rain and irrigation because this parameter has a direct impact on water potential of leaf (following the structure of figure 2). Showing both information together will do more easy the interpretation of the water stress in both seasons.
ii) On figure 2, you show the mean temperature and rainfall, and it should very useful to include the reference evapotranspiration as it represents better than temperature the atmospheric demand. I recommend to set the scale of x axis on one-week in order to have a more legible figure.
As exposed above I recommend accept the paper under minor revision.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Please see the attachment.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
First of all, it is annoying that it is difficult to clearly confirm which part was modified only with the submitted correction comments. It is strongly recommended that you write down the line number correctly or mark the corrected part in the manuscript.
It is strongly recommended that terms be used consistently within the text unless intended.
1. The abbreviation problem has not been solved at all: GWS (Line: 40, 320), DOY (Line: 370, 527, 564), R2 (Line: 294, 339, 348), RMSE (Line: 294, 381), SHAP (Line: 30, 304, 407), ML models (Line 280, 375)
2. TCARI and ExG are redefined in Line 337 and 339 even though they are already defined in Table 1.
3. The same is true of Python term: Line: 263, 267, 282, ….
In addition, the main points mentioned in the previous review remain unresolved.
4. Line 249, 257: Even if the term ‘correlation’ is used in hierarchical clustering, there is no definition for it. If Spearman coefficient is used as before, the problem pointed out in the previous review still exists. Even though this problem is not solved, mentioning it in the text is essential.
5. Line 396-397: Experimental results for SVR are still not found in Tables 4 and 5 in "RFR performances the best for models based on one type of monetary predictor". According to this argument, there should be an SVR model result corresponding to (TCARI or ExG) + (soil/terrain or temporary or weather) in the table, but this cannot be found in the text.
The following are additional comments.
6. As an essential problem, this paper uses nonlinear models RFR and SVR as final models. Nevertheless, the paper provides the R2 value as an evaluation measure. This is not acceptable unless you provide ant validation for using R2 values as an evaluation measure in nonlinear models (because SST = SSE + SSR is not true in nonlinear models). You should only use an evaluation measure that can also be used for nonlinear models, such as RMSE.
[Reference] Andrej-Nikolai Spiess et al., “An evaluation of R2 as an inadequate measure for nonlinear models in pharmacological and biochemical research: a Monte Carlo approach”, BMC Pharmacology, 2010
7. Line 379: You mentioned a combination of 3*2*3*4*1*2*2*2*2*2 can be generated from 8 clusters. But, I obtained 2*1*2*2*3*1*1*1*1 by applying the 0.75 cutoff of line 365 to Figure 6 (left to right order). This should be explained.
8. Line 283-284: Parameters and hyperparameters are different in machine learning model.
Comments for author File: Comments.pdf
