Assessing the Leaf Blade Nutrient Status of Pinot Noir Using Hyperspectral Reflectance and Machine Learning Models
Round 1
Reviewer 1 Report
The text provided has been reviewed. The introduction of the manuscript "Assessing the Leaf Blade Nutrient Status of Pinot Noir Using Hyperspectral Reflectance and Machine Learning Models" (remotesensing-2220026) is informative but could benefit from being shortened. Legends in figures and tables should be described more clearly and it is recommended to check the position of Figure 2, the CV% in Table 3, and the mean in Figure 5A. It is also suggested to ensure that all references follow the "Instructions for Authors." The images and schematics are visually appealing and of high quality, but it is recommended to arrange the keywords in alphabetical order, standardize the nomenclature of equipment, reagents, and software with manufacturer, city, state, and country, and correct any English grammar and spelling errors.
Best regards,
Author Response
Reviewer 1
The text provided has been reviewed. The introduction of the manuscript "Assessing the Leaf Blade Nutrient Status of Pinot Noir Using Hyperspectral Reflectance and Machine Learning Models" (remotesensing-2220026) is informative but could benefit from being shortened. Legends in figures and tables should be described more clearly and it is recommended to check the position of Figure 2, the CV% in Table 3, and the mean in Figure 5A. It is also suggested to ensure that all references follow the "Instructions for Authors." The images and schematics are visually appealing and of high quality, but it is recommended to arrange the keywords in alphabetical order, standardize the nomenclature of equipment, reagents, and software with manufacturer, city, state, and country, and correct any English grammar and spelling errors.
Best regards,
Response
The reviewer’s changes have been made in accordance with comments from the other reviewers, which lengthened rather than shortened the final document. English has been reviewed as requested.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Overall, this study certainly has its merits. I like the fact that several preprocessing methods are compared, as well as different feature selection techniques. I think the results of this study are relevant and fit for publication. However, the text itself could use some more work. Especially the discussion and conclusion sections do not feel very polished and need thorough revision.
Main comments:
- There are very numerous typos and language errors. I have already compiled a long list (see below), but the manuscript requires a thorough language revision by a native English speaker.
- The use of permutation variable importance might not be the best choice for hyperspectral data since many of the wavelengths are highly correlated. When one variable is shuffled the model could theoretically still have access to the feature through correlated variable(s). Changing this could result in a clearer view of which regions of the spectrum are good predictors for a specific nutrient. It might be worth trying the workflow given here: https://scikit-learn.org/stable/auto_examples/inspection/plot_permutation_importance_multicollinear.html#sphx-glr-auto-examples-inspection-plot-permutation-importance-multicollinear-py
- Samples were frozen first and then thawed before taking spectral measurements. The reason to do so is not explained well; why not work with fresh material? The freezing and thawing process will undoubtedly have had its effect on the hyperspectral reflectance spectrum. This should be mentioned and its effects discussed in the discussion section. Also, do the authors recommend this method (freezing and thawing) or would they recommend working with fresh leaf material?
- Figure 6: discrepancy between the dots and the regression line; were the observed and predicted values switched? I suggest to doublecheck the code leading to this plot. (Also correct: “Observied” value in Y axis of all 5 subgraphs)
- L 87 ‘chlorophylls’ => the authors either mean chloroplasts or chlorophyll. In fact, neither is correct: only about 1.7% of the N in a plant is contained in chlorophyll, much less than, e.g., N in Photosystem I and in photosystem II. See Berger et al. 2020 (Remote sensing of Environment). However, there is usually a very close relationship between the N content in a plant and the chlorophyll content in a plant.
- Although the data was collected on two different fields, one field did not serve as a test and the other as a training datasets – only this way, the authors would have achieved a truly independent test dataset, which would have made the results much more compelling. Why not? Did the authors test this, and what was the result? Can they explore this in the discussion section.
- In the Vegetation indices, the authors mostly target chlorophyll indices (related to N content). Hence, it is no surprise that the VI-approach only resulted fruitful for N (L324-326). The authors did not select typical pigment indices, e.g. reflecting on anthocyanins (mARI), xantophyll (PRI) or other carotenes. We know that often these pigments are used as ‘screening’ when leaves are low in macronutrients, so including these indices would be interesting. Also, a systematic VI approach (e.g., NDVI-type formulas (A-B)/(A+B) for all (pooled) spectra could result in better predictions.
Typos and text remarks
40: Remove ‘, than most other crops’
53: I don’t see the direct link between maintaining optimal nutrient levels in a plant and reducing optimal impacts. You could argue that a better understanding of a plant’s needs could lead to a reduction in unnecessary treatments, which could reduce environmental impacts.
59: replace “temporally” with “in time”
61: recommend
62: recommend growers to collect
64: during this period => what do you mean? It would literally mean ‘during the period of tissue analysis in chemical laboratories’, but I guess this is not what’s meant here?
Line 66-68 and line 71-73 are highly repetitive
67 which is time consuming
L73 “This standard measurement is time consuming and has a high labour cost whilst not directly reflecting the nutritional status of the crop.” => revise => what’s the relevance of “whilst not directly reflecting the nutritional status of the crop” in this sense?
86: have made
108 and one micronutrient
135 the aim is to build … and to determine
145: a minimum of information needs to be given here, e.g. on timing (give correct dates) on relevant soil properties
154: Frozen at what temperature? How was this done in the field?
L164: revise sentence
L 169-171: revise sentence
L171 reference panel is
176 each point was measured three times
L172: the spectral does not have a sampling interval: the sensor has a spectral range and a sampling interval.
L 177 Figure 2 figure captions floating, not spaced correctly.
180 and were then averaged
195 wat wavelengths sensitive to target parameters
197 sensitive to leaf macronutrient content
205 of minimum
206 (SD), and of coefficient of variation
226 a scatter plots was constructed to
Figure 3: information is already contained in Table 2, this figure can be moved to the supplementary section, as it is interesting but does not directly add to the story, and there are plenty of figures to be shown.
291 N and P correlated significantly and positively (r=0.58, p<0.01)
299 spectral analysis
311 Prediction accuracy
312 The best results of the machine learning algorithms between the different feature groups
313 based on R² and RMSE. => of the test dataset, right? Mention this clearly. I would strongly suggest to focus primarily on RMSE, less on R².
317 (remove than other models)
318 as no model returns
328 Most biochemical variables
336: Figure 6: The caption should be below the figure
341-344: revise and rephrase
352 shows
355 concentrated
356 1820, and 2150-2180 nm
368 was presented
372 raw reflectance data and their deriviative…
374 in predicting levels of N, K and Ca […]. Levels of P and Mg were poorly represented, …
381 macro- and micronutrient levels
384, with R² above 0.6 for all nutrients
387, and N citrus tree leaf …
390 Although the variability
392 The value range of K concentrations was larger
401 is most suitable for…
402 Two studies have successfully predicted
406 Wei et al[33], who stated
409 wavelengths
426 [55]. The
428 with the
437 good performance for predicting levels of N, K and Ca, but not for P and Mg.
449 it had a low accuracy
455 that the visible
457 wavebands
472 the most sensitive bands
480: which is related to photosynthesis. K plays a key role
486: A previous study that
488 This study also shows
486-489: Needs a reference for this statement
Author Response
Response to reviewer 2.
Overall, this study certainly has its merits. I like the fact that several preprocessing methods are compared, as well as different feature selection techniques. I think the results of this study are relevant and fit for publication. However, the text itself could use some more work. Especially the discussion and conclusion sections do not feel very polished and need thorough revision.
Main comments:
- There are very numerous typos and language errors. I have already compiled a long list (see below), but the manuscript requires a thorough language revision by a native English speaker.
The English has been thoroughly reviewed.
- The use of permutation variable importance might not be the best choice for hyperspectral data since many of the wavelengths are highly correlated. When one variable is shuffled the model could theoretically still have access to the feature through correlated variable(s). Changing this could result in a clearer view of which regions of the spectrum are good predictors for a specific nutrient. It might be worth trying the workflow given here: https://scikit-learn.org/stable/auto_examples/inspection/plot_permutation_importance_multicollinear.html#sphx-glr-auto-examples-inspection-plot-permutation-importance-multicollinear-py
Reply: Line 243-250, added hierarchical clustering in the method section. Using predictor variables from hierarchical clustering have improved some model performance (see Table 4). However, for some of the previously best-performing models (SVR-Pearson correlation selected derivative variables), the predictive performance of the models using hierarchical clustering was reduced, based on R2, RMSE and MAE. The multicollinearity in hyperspectral data proposed by the reviewer does affect the interpretation of the machine learning model. In line 463-468, discussion about this problem has been added.
- Samples were frozen first and then thawed before taking spectral measurements. The reason to do so is not explained well; why not work with fresh material? The freezing and thawing process will undoubtedly have had its effect on the hyperspectral reflectance spectrum. This should be mentioned and its effects discussed in the discussion section. Also, do the authors recommend this method (freezing and thawing) or would they recommend working with fresh leaf material?
Reply: Line 534-540, added discussion and proposed further study advice. Line 169 adds more information about why we didn’t use fresh material.
- Figure 6: discrepancy between the dots and the regression line; were the observed and predicted values switched? I suggest to doublecheck the code leading to this plot. (Also correct: “Observied” value in Y axis of all 5 subgraphs)
Double check the code and correct “observed”
- L 87 ‘chlorophylls’ => the authors either mean chloroplasts or chlorophyll. In fact, neither is correct: only about 1.7% of the N in a plant is contained in chlorophyll, much less than, e.g., N in Photosystem I and in photosystem II. See Berger et al. 2020 (Remote sensing of Environment). However, there is usually a very close relationship between the N content in a plant and the chlorophyll content in a plant.
Deleted that sentence
- Although the data was collected on two different fields, one field did not serve as a test and the other as a training datasets – only this way, the authors would have achieved a truly independent test dataset, which would have made the results much more compelling. Why not? Did the authors test this, and what was the result? Can they explore this in the discussion section.
Reply: This has been tested, the result didn’t return a good model performance. But as the reviewer said using a truly independent test dataset does make the result more compelling. Line 390-395 adds more information about this.
- In the Vegetation indices, the authors mostly target chlorophyll indices (related to N content). Hence, it is no surprise that the VI-approach only resulted fruitful for N (L324-326). The authors did not select typical pigment indices, e.g. reflecting on anthocyanins (mARI), xantophyll (PRI) or other carotenes. We know that often these pigments are used as ‘screening’ when leaves are low in macronutrients, so including these indices would be interesting. Also, a systematic VI approach (e.g., NDVI-type formulas (A-B)/(A+B) for all (pooled) spectra could result in better predictions.
Reply: Table 1 adds more VI with relevant anthocyanin and carotenes. However, this didn’t improve the predictive performance. In Table 1, N_870_1450 and N_1645_1715 was normalised indices used in a previous study (reference 24). They showed a relationship between Potassium and Phosphorus levels in wheat. Line 435-438 suggests further study can use more combinations of VI.
Typos and text have been addressed, thank you for pointing many of them out.
Typos and text remarks
Author Response File: Author Response.docx
Reviewer 3 Report
This is an interesting study and the authors have collected sufficient dataset for methodology. The paper is generally well written and structured. I have provided some remarks on the file attached to be corrected.
Comments for author File: 
Comments.pdf
Author Response
Reply reviewer 3
Line 130, This has been mentioned in line 44, 45. Please see the reference Ashley, R. Grapevine Nutrition-an Australian Perspective. Foster’s Wine Estates Americas 2011, 1000. Ashley stated that in grapevine, Magnesium is a component of chlorophyll, thus contributes to carbohydrate production in leaves through photosynthesis. The magnesium symptoms can cause the leaf margin to become yellow or the basal leaves to become red. Calcium plays a role in the structure of the vine and may be associated with bunch stem necrosis.
Line 169. Add more information about why after two days of sampling. Lines 534-540, mentions the drawback of this sampling strategy in the discussion.
We noticed that veraison was highlighted in your pdf. Review. Veraison is used frequently as period of grape berry ripening. The authors have seen its use non-italicised if the reviewer thinks it should be italics?
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The version I received has changes, however the keywords were not rearranged in alphabetical order!
