Plot-Level Maize Early Stage Stand Counting and Spacing Detection Using Advanced Deep Learning Algorithms Based on UAV Imagery

Round 1

Reviewer 1 Report

Comments to the paper "Plot leevl maize early-stage stand counting and spacing detection using advanced deep learning algorithms based on UAV imagery".

 

Generic comments:

- Would it make sense just to locate the root points of all of the maize plants, and then extract the relevant information about counts and plant spacing from these locations?

 

Specific comments: 

L12: Change , to .

L62: I have seen use of multispectral cameras (like Micasense RedEdge MX) in several occasions, but nearly no use of hyperspectral cameras in relation to agriculture. Do you mean multispectral cameras instead of hyperspectral cameras?

L100 - L101: I am unsure how the sentence should about the crop spacing be understood.

L114: I would encourage you to share the dataset on zenodo or a similar platform. This will enable others to recreate your results and potentially develop improved methods.

L133: How are maize plants that overlap annotated? With one boundning box for each plant?

L135: Please describe how you have annotated the different features.  Eg. for the big gap describe that a bounding box are drawn such that it contains the following three things: last plant before the gap, the gap and the first plant after the gap. Please also make clear how you discriminate between skips and big gaps.

L136: Do you ensure that all images of plot parcels are oriented similar to the images shown in figure 2?

L156: Please add details about how the models were trained. Did you use transfer learning to adapt an existing model to the training data? If so which existing model was used.

L160: Here you mention that the test data set can be used to prevent potential overfitting. How have you achieved that when you use a fixed number of epochs during training?

L161: Usually the data is divided into the following three data sets: training, validation and test. I would love to see loss curves for the training and test data sets during the training. Please check the description of these data sets and how they are used in this article on wikipedia https://en.wikipedia.org/wiki/Training,_validation,_and_test_data_sets

L188: Please add a reference to where the different metrics are introduced / described in more detail.

L195: The term $p(r)$ in the integral is not described. What is it?

L227: Please highlight the best results in the table and not only just one row. Eg. the best average precision is 0.931 and the best R2 is 0.838. What is your argument for choosing the MAE as the most important metric in assessing the quality of the models?

L229: In the figure, the axis are denoted "predicted value" and "true value". I think it would be more precise to state "precited count" and "true count"

L284: Please highlight models with good (best) performance

L292: Table.4 -> Table 4.

L304: Some elements in the table are truncated to two digits after the period. Please ensure that all values are reported with three digits after the period.

L325: The caption for figure 7 is placed away from the figure.

L348: I would suggest you to redo figure 8 and make it more similar to figure 5. I think it would work better to separate the three different experimental plots so that they appear in three different figures.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript is written with clear understanding of the project addressed. However, there are major concerns that need to be addressed to enhance the quality of the manuscript. My specific comments are as follows:

Introduction:

L36: “….on the expression of a phenotype (e.g., grain yield), known as genotype by environment (G×E) interaction.” Add citation

L61: “UAVs have become an increasingly important tool to acquire high-resolution RGB (visible light spectrum) or hyperspectral (light emitted outside of the visible spectrum)…” Add citation

Based on your objectives, please compare how your study is different from those that have already been published.

Materials and Methods:

Divide Section 2.4: Model development into two sub-sections:

2.4.1. Detecting and Counting Maize Stands

2.4.2 Detecting the Plant-Level Spacing Variabilities (PSV)

Change “2.6 Model Performance Evaluation for Each Task” to “2.5 Evaluation of model performance”

Results and Discussion:

L216: “…the underestimation of overlapping plants was highlighted.” Highlighted in terms of what

Relate your results with existing literatures to support your findings.

Instead of mentioning the results, the authors should justify/explain the findings

 Conclusion:

Revise into 1-2 paragraph only

General comments:

Please check the reference styles and grammar of the manuscript.

 

Minor editing of English language required

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

This study focused on evaluating deep learning object detection models in estimating stand counting for maize breeding plots and quantifying plant-level spacing variabilities. This is a well-conceived study overall but some improvements should be made in the review process, including the addition of relevant references in multiple locations in the manuscript, as well as clarifying the scientific contribution of this study in the subset of previous crop density studies.

First and foremost, the authors should have made an effort to format their manuscript according to journal rules, including the Abstract, which is about 100 words above limit.

Stating that “there is no previous study using UAV imagery and deep learning in detecting the PSV for any crop kind” (lines 73-74) is slightly misleading. While there might not be studies with the exact filter you applied, there are numerous similar studies dealing with crop density overall. I do not agree that studies addressing “plant-level spacing variabilities” are enough distinct from the previously mentioned ones, but I consider that adding either more comprehensive literature review or more specific explanation about the specific importance of plant-level spacing variabilities is mandatory.

What ground sampling distance was obtained from UAV image acquisition in subsection 2.2.? Please add this information. Would your results be more reliable if you created a digital orthophoto to minimize distortion from individual images? Please address this potential limitation of your study in the discussion.

Specific comments:

Lines 57-63: This paragraph should be supported by multiple references.

Line 76: You already introduced this abbreviation in line 66. (This applies to lines 47 and 86 for PSV as well).

Lines 78-80: Besides the reference in which it is first introduced, the statement that it is “the best-known and most popular one-stage object detection algorithm” should be added.

Lines 260 and 266: Again, you already introduced these abbreviations. Your manuscript would benefit from slightly more effort into formatting and checking text.

Figure 8. Please split this figure into three separate figures.

Line 372: This is the first time you mention “UAS”.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors have addressed all the comments. Hence, the paper can be accepted. 

Minor editing of English language required.