Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Poudel et. al. describes a technique to predicting cover crop biomass using machine learning and UAV-acquired spectral and structural data. While the employment of a neural network with Keras and TensorFlow is admirable, the selection of certain models, optimizers, and activation functions may restrict generality. More information about the dataset, such as the variability in climatic conditions and crop species, might improve the conclusions. Furthermore, comparisons with various machine learning algorithms or other precision agricultural technologies can give a more rigorous assessment. In addition, the following questions and suggestions need explanation.

 

·         The introduction need improvements and provide the objectives of the study at the end of this section.

·         All the figures need improvement, currently these are absurd. The legends’ and axis’ text is very difficult to read.

·         Provide the technical route of the study.

·         The detail of the used optimizers, functions and accuracy indicators is missing.

·         L#234: Please double check the formatting “R2” and also in the whole manuscript.

·         Table3: Highlight the high accuracy results.

·         Provide the results for the DSM as biomass.

·         How height difference for diverse vegetation types was managed?

·         The noise removal of UAV data, which method was used?

 

·         Which VI has performed better? The results are based on the inclusion of the all VIs or some of them? Which combination is better?

Author Response

Comment 1

The introduction need improvements and provide the objectives of the study at the end of this section.

Response 1

The introduction section has been refined in the revised paper to provide stronger context for the study. The objective of the study was to integrate UAV based remote sensing and machine learning for data-driven, and reliable cover crop assessment. Additionally, to identify the model configuration for ensuring reliable biomass estimation. This information is made available in the revised paper. Thanks for this excellent suggestion.

 

Comment 2

All the figures need improvement, currently these are absurd. The legends’ and axis’ text is very difficult to read.

Response 2

The resolution of all figures and legends was somehow impacted when the word document file of the paper was submitted to the system. The revised paper is now submitted as PDF to avoid similar problem from happening again and in addition, the high-resolution figures are submitted as individual image file in the zip file submitted along with the paper submission.

 

Comment 3

Provide the technical route of the study.

Response 3

Technical route of the study in the form of workflow of the cover crop prediction model is included in the revised paper (Figure 2).

 

Comment 4

The detail of the used optimizers, functions and accuracy indicators is missing.

Response 4

Optimizers and activation functions are listed in Table 2 with the specific used in the neural network layers shown in Figure 3 (indicated as dyn). In the revised paper, we also provide clarity about the metric use for performance evaluation of the model in line 179-180. We appreciate this constructive comment.

 

Comment 5

L#234: Please double check the formatting “R2” and also in the whole manuscript.

Response 5

R² formatting has been corrected in the revised paper.

 

Comment 6

Table3: Highlight the high accuracy results.

Response 6

Thank you for this valuable suggestion. High accuracy results are now highlighted in bold in the revised paper (Table 3).

 

Comment 7

Provide the results for the DSM as biomass.

Response 7

Since DSM is intermediate variable, what presented in the revised paper is the products out of this DSM that is the CHM (Modeled canopy height). The results for this variable are provided in Figure 4f and 5f.

 

Comment 8

How height difference for diverse vegetation types was managed?

Response 8

This is a very good point. If the biomass prediction is conducted early in the growing season, this noise of different growth style of the different cover crops species will cause a high degree of uncertainties. For example, Hairy Vetch plants with the growth style of spreading low to the ground will certainly cause consistency when compared to cereal rye and black oats with the rather vertical growth style. In this study, the biomass estimation is conducted toward the end of the north hemisphere winter period (around March). By this time, the hairy vetch plants were very much no longer existence. The main growth only occurred in the early part of winter. In the revised manuscript, this clarity is provided in the material and methods to indicate which species were the main dominant plants at the time of biomass prediction (line 121-124).

 

Comment 9

The noise removal of UAV data, which method was used?

Response 9

The data sanitizing in this study included removal of outliners pixel with the criteria of having NDRE < 0.2 NDRE less than 0.2 is indicative of bare soil condition. This description is now included in the revised paper (line 162-164) along with cited reference (paper #40).

 

Comment 10

Which VI has performed better? The results are based on the inclusion of the all VIs or some of them? Which combination is better?

Response 10

This is a very good question. In the revised paper we included sensitivity analyses of varying combinations of predictors. The results of model run using a total of 31 combinations of these VIs is now included in the revised paper (Table 4). The results of these analyses are provided in Table 4 in the revised paper.

Reviewer 2 Report

Comments and Suggestions for Authors

This study predicts biomass in cover crops using vegetation indices, digital elevation models, and deep learning techniques. Although it is an interesting work, the methodology and fieldwork are more concise than they should be. It could have been replicated in other seasons or differentiated by crop types.

The importance of cover crops must be better justified. Cover crops help prevent soil erosion, regulate moisture, attract pollinating insects, facilitate weed and pest control, and serve as mulch or a source of green manure and organic matter. Additionally, they can be used for grazing or forage. They are not typically used for human consumption nor are they the species that yield the highest productivity for animal consumption. If their primary purpose is not to generate yields but to protect the soil without depleting nutrients, how can the estimation of biomass in them be better justified?

The state of the art for this type of study must be further explored in the introduction section.
The resolution of the images needs to be improved.
Why are geographic coordinates being used instead of projected ones? You must specify the reference system.
The methods must be explained in greater detail. Currently, the figures are unclear.

I recommend adding a flowchart to better illustrate the methodology.

Line 163: Where can I see this? “Plant height estimate (PLH) at NERS showed dual peaks with the dominant one skewed toward the high values, whereas at CRS the peak was toward the low values.”
In the dataset preparation, did you further explore the data? Were there any atypical data points?
The explanation about histogram comparisons could be improved. Perhaps a single illustrative figure would suffice. Figures 3, 4, 5, and 6 could be consolidated into one.

Line 208: Can you clarify the training process?
Figures 7 and 8 do not provide clear or useful information.
Where did you explicitly explain the percentages for training, validation, and testing?
All figures must be improved.

The discussion is interesting but requires a deeper comparison with other studies that have used deep learning methods.

In general, the methodology needs to be further elaborated, and the quality of the figures must be significantly improved.

Author Response

Comment 1

This study predicts biomass in cover crops using vegetation indices, digital elevation models, and deep learning techniques. Although it is an interesting work, the methodology and fieldwork are more concise than they should be. It could have been replicated in other seasons or differentiated by crop types.

Response 1

We are continuing this work in the on-going season with the goal of teasing out the prediction by species, exactly as suggested. It is, however, important to document early benchmarks of this effort to establish at least understanding which combination of optimizer and activation functions and what combination of predictors are the best for the overall biomass prediction. Results from this present paper is critical for the upcoming work and given the fast changing of machine learning topic, it is important to document results as recent as possible. We incorporate the statement about the future investigation direction in the revised paper (line 395-396). Thanks for the great suggestion!

 

Comment 2

The importance of cover crops must be better justified. Cover crops help prevent soil erosion, regulate moisture, attract pollinating insects, facilitate weed and pest control, and serve as mulch or a source of green manure and organic matter. Additionally, they can be used for grazing or forage. They are not typically used for human consumption nor are they the species that yield the highest productivity for animal consumption. If their primary purpose is not to generate yields but to protect the soil without depleting nutrients, how can the estimation of biomass in them be better justified?

Response 2

Thank you for this useful feedback. The importance of cover crops for agronomic and ecosystem services has now been incorporated in the revised paper in with supporting literature. The prediction is not solely aimed at estimating yield and productivity. Instead, the discussion highlights how biomass estimation can quantify cover crops' ability to provide greater canopy coverage, prevent soil erosion, facilitate weed and pest control, and serve as mulch or a source of green manure. This is further supported by studies that focus on higher biomass production and the importance of its prediction

 

Comment 3

The state of the art for this type of study must be further explored in the introduction section.

Response 3

The sentence highlighting the potential of UAV-based remote sensing and machine learning for a reliable data-driven approach to crop assessment is included in the revised paper.

 

Comment 4

The resolution of the images needs to be improved.

Response 4

The resolution of all figures and legends was somehow impacted when the word document file of the paper was submitted to the system. The revised paper is now submitted as PDF to avoid similar problem from happening again and in addition, the high-resolution figures are submitted as individual image file in the zip file submitted along with the paper submission.

 

Comment 5

Why are geographic coordinates being used instead of projected ones? You must specify the reference system.

Response 5

Geographic (EPSG 4326) was used as the coordinate reference system for describing the study locations in the material and method text. This is a common practice in literature to simply describe the general location of the study sites and this is justified given the purpose and nature of such coordinates data [1]. To improve clarity of the paper and to help readers with study orientation, in the revised paper, the information about coordinate reference system used (EPSG 32615, UTM zone 15, WGS 84 ensemble datum) is provided as well as reference grid data in Figure 1 both as Easting/Northings and Latitude/Longitude.

Reference

1. Van Sinckle, J. Basic GIS Coordinates. Third Edition. CRC Press. Talor & Francis Group, New York.

 

Comment 6

The methods must be explained in greater detail. Currently, the figures are unclear.

Response 6

The problem with resolution of figures has been fixed in the revised paper. Clarity on the methods has also increased with the inclusion of flow chart of the overall model.

 

Comment 7

I recommend adding a flowchart to better illustrate the methodology.

Response 7

Thank you for such excellent feedback. We have added this flow chart to improve clarity about the approach used in this paper (Figure 2).

 

Comment 8

Line 163: Where can I see this? “Plant height estimate (PLH) at NERS showed dual peaks with the dominant one skewed toward the high values, whereas at CRS the peak was toward the low values.”

Response 8

Thank you for noticing the lack of reference in the sentence. In the revised manuscript the reference has been added, this dual peak with the dominant one skewed toward the high values in plant heigh estimates at NERS location is shown in Figure 3f, whereas the description of CRS peak for PLH is in Figure 4f. Likewise, other statements describing histograms of these variables have been revised in the revised paper by inclusion of the reference figure panels.

 

Comment 9

In the dataset preparation, did you further explore the data? Were there any atypical data points?

Response 9

The data sanitizing in this study included removal of outliners pixel with the criteria of having NDRE < 0.2. NDRE less than 0.2 is indicative of bare soil condition. The relevant literature was cited in this regard.

 

Comment 10

The explanation about histogram comparisons could be improved. Perhaps a single illustrative figure would suffice. Figures 3, 4, 5, and 6 could be consolidated into one.

Response 10

Thank you for the suggestion. However, it was not feasible to combine these figures. Instead, given the data showed similarity between the extracted data for the prediction versus the whole pixels histogram of the remote sensing data, we decided to remove the previous Figure 5 and 6 in order to reduce the length of this paper. This is also in consideration of the newly added results data in Table 4, which are have provide more insight about the findings than these removed Figures.

 

Comment 11

Line 208: Can you clarify the training process?

Response 11

The training process involved monitoring the mean absolute error (mae) in each epoch and improving the performance of the prediction by reducing mae.

 

Comment 12

Figures 7 and 8 do not provide clear or useful information.

Response 12

Figures 7 and 8 provide a descriptive explanation of what’s going on during the training. Given the different optimizers and activation functions influence the training progress, it is important to document these conditions. Therefore, we retain these figures in the revised paper. Instead, we remove previously known as Figures  5 and 6 in order to shorten the paper. In such conditions, it is still reasonable to retain Figure 7 and 8, now assigned as Figure 6 and 7 in the revised paper.

 

Comment 13

Where did you explicitly explain the percentages for training, validation, and testing?

Response 13

The revised paper provides clarity on the training-validation-split of 50-25-25 (line 180). It was also cited the reason for using this split setup.

 

Comment 14

All figures must be improved.

Response 14

All figures have been improved in the revised paper.

 

Comment 15

The discussion is interesting but requires a deeper comparison with other studies that have used deep learning methods.

Response 15

Thank you for acknowledging the interesting nature of the discussion. We provide additional highlights in the discussion section regarding the present study as compared to other studies utilizing a deep learning approach (line 359-380). We believe this great suggestion has helped improve the discussion section of the revised paper.

 

Comment 16

In general, the methodology needs to be further elaborated, and the quality of the figures must be significantly improved.

Response 16

The quality of the figures has been improved in the revised paper. Likewise, the methodology has been improved with additional clarifying sentences and inclusion of big picture workflow.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Please reduce the number of keywords.

Double check the figures' quality. 

 

Author Response

Comment 1

Please reduce the number of keywords.

Double check the figures' quality.

 

Response 1

As suggested, keywords have been reduced. We have also double check the quality of the figures as embedded in paper word document. We noticed earlier, the figures in the supplied pdf version had lower resolution. Therefore, we no longer supply this pdf file as we were not able to configure the production of pdf file retaining the original resolution of the figures. Instead, along with the word document submission (similar to what we did last time, to supply the word document we are also providing the figures as stand along files (png) compressed in a zip file. Similar to the practice we did during round 1 submission, we again compress the word document of the paper in zip file, hoping the preserve the size and formatting better, particularly to avoid resolution distortions in the figures, which occurred during pre-round1 review submission.

Reviewer 2 Report

Comments and Suggestions for Authors

Good job.

Author Response

Comment 1

Good job.

 

Response 1

Thank you for your acknowledgement. Once again, we want to thank you for your thorough feedback helping us to improve this paper.