Dry Matter Yield and Nitrogen Content Estimation in Grassland Using Hyperspectral Sensor

Round 1

Reviewer 1 Report

The manuscript “Dry Matter Yield and Nitrogen Content Estimation in Grass-2 land Using Hyperspectral Sensor” (remotesensing-2013915) is a good manuscript and can be developed into one meriting publication. The authors should review what they have written, for clarity to start with; for example, beginning a sentence with ‘2’ is not recommended. The manuscript needs a moderate amount of rewriting and editing.

Is this just another regression model to predict (e.g., nitrogen/protein/biomass) or does it provide new insights in how to sense (remote sense) and predict? A potentially valuable contribution by such a manuscript is to would-be users of their methods or analyses. For example, a student interested in developing predictive models based primarily on reflected wave bands might ask: how to apply random forests; why look at PCA; what are useful preprocessing steps (and how); and what are alternate approaches to identifying important influences on predictions, for example.

I do not want to sound entirely negative. Overall, I like what you did but it is difficult to follow or understand what you did (and why).

Some comments/questions by line

Lines 226-238: explicit names and links - a very good feature that you do not see very often

Line 163: You mention ‘AI’ a few times without explaining. Is this simply a reference to the random forests model?

Was there any RF model tuning to set hyperparameters?

Line 183: you mention setting two hyperparameters, number of estimatoers was set to 100; what was set for max depth?

Maybe explain x-loading analysis

More complete discussion of SHAP would be good – what it is and why use it here

Lines 214-215: Add a sentence or two to explain GA and Phased Regression (PHR)

While one reason for SHAP is to compare feature importance between two different models, there seems to be an implication that the random forests importance only captures positive effect of features. Is this true; you should explain.

Line 291 PCA based approach:

You show three figures and a table, and it would be helpful to expand the discussion of what information is provided by each. For example, what do the peaks or clusters tell you and how are the more important wavelengths identified from the pca

Figures 10 and 11 are missing some wavelength labels

Among all the importance summaries, expand on where they agree or disagree and what does that tell us. Also, there is discussion of wavebands associated with feature values with Figure 12, but not so much with other lists .

 Line 384 mentions ‘oil’ without previous explanation. Is this lipids

Appendixes are a good idea for providing that additional detail for readers - nice

For correction of biovariation, maybe add a sentence or two about now the scatter correction technique works

For centering, maybe another sentence on why better results without adversely affecting the model. Also, is this applicable to both PCA and RF?

Finally, the connection to other research through Appendix B is good. It would be helpful to somehow note with column 1 the top 10 or 20 important wavelengths that you identified in your work

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

This article is trying to develop new more efficient ways to monitor grass composition using inexpensive sensors for estimating dry matter and nitrogen content of pasture.

The article is well written and have good core.

One of the prime concerns is number of previous work in introduction, where more previous work should be written and explained. The second concern is discussion section, where only discussion about obtained result is written but it is not placed in context with another research - only one sentence "Most wavelengths revealed to strongly contribute to the estimation in this paper can be linked to well-known wavelengths in the related papers shown in Table B1"  which is not enough. All the results should be in context with another (state-of-the-art) researches.

Some of the minor corrections:

Line 103 - please add some date (best table) when the grass is cut

Line 213 - missing reference 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

Recommendation: Major Revision

Comments to Author:

Manuscript ID: 2013915

Title: Dry Matter Yield and Nitrogen Content Estimation in Grassland Using Hyperspectral Sensor

Overview and general recommendation:

This manuscript investigated the estimation of DM yields and nitrogen content in grasslands and ways to improve their accuracy. Random Forest Regressor, PCAs, and SHAP analysis selected the feature bands with high accuracy. By combining the plant height measured by an ultrasonic instrument, the estimation accuracy of DM yields and nitrogen content can be improved. Because of this, the current study is on a topic of relevance and general interest to research the fusion of remote sensing and crop models. However, there are still some apparent deficiencies in the results and analysis. I recommend that a Major revision is warranted.

Major comments:

1. Introduction part, the analysis of grassland DM and nitrogen content monitoring is not in-depth enough, and the shortcomings of current research and the fundamental problems to be solved in this study are not pointed out. Such as “the importance of features as only positive values” what is this mean?

2. Table 1, Measured Parameters NC and DM, did you write their measurements backward?

3. Fig. 3, Two years of data can be combined into a single figure to show the changes in DM, plant height, and nitrogen content caused by two years of meteorological differences. At the same time, it is suggested to supplement the charts of critical meteorological conditions for two years.

4. Fig. 6, c), and d) are incorrectly labeled as a) and b) in the title.

5. Table 3. The abbreviations in the table are not annotated

6. Fig. 9, The identification of each indicator vector in the figure is unclear; analyzing the contribution changes of bands is challenging.

7. Fig 13 and 15, These two figures are the data presentation of individual bands in Figures 12 and 14. If there is no significant effect, it is recommended to delete the two diagrams.

8. Section 3.1, the regression models of DM and NC were constructed using Random Forest Regressor, and the accuracy changes after adding plant height were compared. But there is an important question, how are the feature bands screened by PCA and SHAP method related and compared with the bands filtered by r Random Forest Regressor?

9. This manuscript concludes that the green wavelengths, red edge wavelengths, and NIR wavelengths are the important candidate wavebands. However, these bands are consistent with the results of previous studies. Combining plant height to improve the measurement accuracy of DM and NC is the conventional method. What is the innovation of this study?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The manuscript looks good

Author Response

Thanks so much for your comment. The manuscript has been improved by your feedback.

Reviewer 2 Report

The article is much better and I do not have any further remarks.

Author Response

Thanks so much for your comment. The manuscript has been improved by your feedback.

Reviewer 3 Report

I think the author has addressed the questions and suggestions. Well done, thank you.

Author Response

Thanks so much for your comment. The manuscript has been improved by your feedback.