Evaluation of Point Hyperspectral Reflectance and Multivariate Regression Models for Grapevine Water Status Estimation

Round 1

Reviewer 1 Report

The manuscript sought to evaluate various techniques and algorithms in characterising water stress of grape-vines using hyperspectral data. The paper is well written and easy to understand. There is relatively little that is new in this manuscript save for the fact that is being conducted on grapevines. This lead to the next observation which begs the question “so what” to be asked after reading the manuscript. This question simple speaks to the implications of the results. It can be observed that the discussion seems to be focused more on the techniques rather than implications of the results which adds more in terms of articles but not knowledge. These techniques have been widely implemented in different crops irrigate and rainfed. It is in this regard that I suggest that the authors add a section on the implications of the results towards the viticulture fraternity as this will give even more strength to the article. This section could discuss the spectral bands and how it links to the water stress in grapes.

General comments

It is not clear how and how many times was the instrument was calibrated during the acquisition of spectra

Table 5 requires to be improved as it is confusing.

Author Response

Reviewer 1:

1. In response the main comment: Vegetation indices (VI) were also computed and compared with the modeling pipelines proposed in this study. This enabled the results of this study to be compared with previous work in the field where plant water status was estimated using hyperspectral reflectance. It displays the advantage of the proposed pipelines in this study. The relationship between the important spectral bands identified in this study and grapevine water status was explored and described in section 4.2. A few sentences have been added to the conclusions to reinforce the importance of this work to the viticulture industry.
2. In response to general comment 1: Additional explanation has been added to line 196.
3. In response to general comment 2: Table 5 has been modified for clarity and is now Table 7, as Reviewer 2’s comments required the addition of two tables.

Reviewer 2 Report

The study is interesting and the analysed issue definitely should be analysed. Nevertheless, the text needs some improvement.

In the introduction there are no examples of the previous studies that use remote sensing indices to retrieve water content, please add this information.  Apart from that, the introduction is quite clear. The aim is well defined. The used datasets are defined as data groups – please consider naming them as the datasets.

The method section is quite long. Please consider shortening some of the parts. The same refers to the results.

In the discussion part the authors indicated the ranges that are important when estimating the stem water potential. In my opinion, it is not fully clear that these ranges can be useful in estimating the stem water potential. Could ranges be related to the vegetation condition affected by factors other than changes in water content? The authors are describing this issue in the discussion part, but this has been not fully explained.

Detailed comments are in the attached file.

Comments for author File: Comments.pdf

Author Response

Reviewer 2:

1. In response to the main comment: 11 plant water status related vegetation indices (VI) were computed and described in section 2.5.4. For all models using VI as an input variable, the best performance in terms of R² was 0.41. This proves the advantage of the modeling pipelines proposed in this study.
2. In response to specific point 1: The same plants were used for collecting both reflectance and stem water potential data. Additional explanation has been added to line 203.
3. In response to specific point 2: Normalized difference indices (NDI) were calculated and used to fit the models. The results of modeling accuracy (R² and RMSE) using simple ratio indices (SI) as inputs, compared to those using NDI as inputs, were almost the same. This can be attributed to the high correlation between SIs and NDIs calculated from the same bands, as the coefficients of correlation between them were higher than 0.99 for all the SI-NDI pairs using the same bands. Some of the NDI denominators were zero, so the outputs of those NDIs were flagged as error. Therefore, SI was selected as inputs for fitting the models.
4. In response to specific point 3: The formula (1) has been revised.
5. In response to specific point 4: A table (Table 1) has been added to clarify the number of surveyed canopies on each measurement date.
6. In response to specific point 5: The definition of VIP is stated in section 2.8.1, where partial least squares regression is introduced.
7. In response to specific point 6: The heatmaps in Figure 12 indicate variable importance for each variable which was used as input for those modeling pipelines that provided the best performance on the test set. Variable importance within the subset was computed by partial least squares regression (PLSR), random forest regression (RFR), and support vector regression (SVR). Correlation coefficients were used to select variables, so were not presented in the heatmap.

Reviewer 3 Report

   The manuscript by Wei et al. is devoted to development of methods of the grapevine water stress estimation on basis of hyperspectral measurements. The work is potentially interesting; however, there are remarks and question.

   The main comment: results by authors should be strongly integrated into the broad context of remote sensing of the water stress. Is the work in accordance with other works in the field? What is more effective: classical water indices (WI and NDWI) or approaches proposed by authors?

   Specific points:

1. Sections 2.3 and 2.4: It is not fully clear: were same plants used for measurements of water potentials and reflectance? Or were averaged values compared?
2. Section 2.5.3: Why were simple ratios used? Difference reflectance indices seems to be more stable.
3. Equation (1) seems to include error. Maybe, yi – y(mean) should be replaced on (yi-Y(mean))^2.
4. Lines 257-258: It is not clear: were 85 samples measured in each time of measurements? Or were the samples measured on basis of all time points (e.g., for all experimental time)?
5. Line 490: What was “variable importance in projection”? It should be clarified.
6. Why correlation coefficients are not used in heatmaps (Figure 12)?

Author Response

The comments of Reviewer 3, have been addressed with responses to reviewer one and two as they were very similar to those expressed by Reviewer 2.

Round 2

Reviewer 2 Report

Thank you for the corrections to the manuscript. Most of my suggestions were answered.

Author Response

We are pleased the reviewer was satisfied by our corrections,