Using High-Resolution UAV Imaging to Measure Canopy Height of Diverse Cover Crops and Predict Biomass
Round 1
Reviewer 1 Report
I think this work is an interesting and valuable exploration of the fine-scale measurement of crop height. It throws up some interesting issues when pondering what is crop height and the effects of scale in addressing the question. I really like the work and encourage its publication. Before that, however, I feel there are a few conceptual considerations that need to be thought through better. Addressing this will lead to some modifications of the analyses and might possibly change some of the conclusions. However, I think this would strengthen the publication considerably.
Most of what follows is based on the assumption that the size of the rising plate used by the authors in the field measurements is 20 x 31 cm in size (0.06 m2) rather than the stated 0.20 x 0.31 cm.
The main issue is that the authors don’t seem to have thought through the issue of scale carefully enough. Most of my comments flow from this one issue. Fundamentally, the authors need to understand and state what the scale is of the phenomenon they are exploring (canopy height). The concept of canopy height (CH) has an intrinsic scale. I would argue it is usually the height of the highest layer of a series of plants. It becomes a slightly less meaningful concept when focussing on one plant, and fairly meaningless when applied at really fine scales (mm to cm).
It then flows from this that measurements need to be done at appropriate scales for the phenomenon being observed. In this study, the field measurements (ruler heights) have a scale set by the size of the rising plate. In this case, the ruler measurements can’t detect changes in CH finer than 20-30 cm. Heights derived from the UAV, on the other hand, can measure differences in height at scales between about 4 and 40 mm.
By the nature of their fine resolution, the UAVs are able to detect any tiny (<20cm) gaps in the canopy (ie between plants) that a rising plate would not detect. Additionally, the UAVs will also detect holes through a plant canopy when foliage cover is under 100% (and such holes should not be considered relevant when assessing CH).
So, because of scale differences, when the authors use ruler heights to assess UAV heights, they aren’t comparing equivalent things. It is this scale-mismatch that makes it inevitable that the authors (and the many studies the authors discuss) found that UAVs return lower average heights than ruler-based methods [line 341] (and this doesn’t mean that ruler-based heights are unreliable as the authors suggest in their discussion). It also means that the authors’ finding that the two sources of height data best match in smooth, dense canopies is also an inevitable outcome (as these are the canopies where the effects of the differences in measurement scale are minimised).
The authors validate their UAV heights by comparing the plot-average UAV heights against the plot-average ruler heights. Because of the differences in scale of these two data sets, this is not a robust test of accuracy. A better test would be to aggregate the UAV data up to the same scale as the ruler measurements (ie ~ 25 x 25 cm) prior to comparing them with the ruler heights. Until such a comparison is done, I don’t think the authors can make any statements about the accuracy of the UAV data using their field data. This provides a much clearer and direct way of assessing accuracy. I found the splitting of plots into mixed and pure, and rough, medium and smooth classes difficult to follow (not helped by the buggy formatting of the in-text citations) and found the relevance of these analyses for assessing accuracy questionable. All these analyses could be circumvented by the above suggested analysis.
Once a proper test of accuracy is done, the authors can then proceed to explore the effects of the different scales of the two UAV sources as functions of species mixtures and canopy heterogeneity – as they have already done in the paper. This however shouldn’t be a discussion about differences in accuracy but about differences in what is being measured at these different scales.
The authors conclude that manual measurement, often entitled “ground truth”, cannot determine CH with sufficient accuracy due to much fewer measurements per sample area [line 539]. This is a ludicrous statement. Firstly, if the field measurements are inadequate, on what grounds do they know that their UAV measurements are adequate? How did they prove their accuracy? Secondly, if their field measurements are inadequate for their validation purposes, then surely this reflects an inadequacy of their field design and measurement strategy.
The authors should note that the fine scale UAV height measurements do not return canopy height (unless the foliage cover of the canopy is 100%). Rather, because they can see through most canopies, UAVs return a height profile. These require further processing to obtain canopy height, which is typically found as some upper percentile height (ie 95th or 99th) over an area. It is important that only UAV data processed in this way are used to represent canopy height, especially in the validation analyses.
A few minor points.
The in-text figure citations are all messed up.
Some methods would benefit from slightly more detail. In particular:
- What was the spacing of ruler measurements throughout the plots? Where only two such measurements taken in each?
- Can the authors explain, in 1 sentence or so, the basic concept of what structure from motion is?
- Similarly, what is the ‘built-in algorithm’ in Metashape? How does it assess quality?
Line 148: why did nearly all LR images have such ‘low’ image quality? And if this is true, why were they then used in the analyses?
Line 190: isn’t it rather a question of which variable represents the source of truth, against which the other is being assessed (rather than which is the independent variable)?
In the assessments of accuracy, the authors almost exclusively talk in terms of which result has a better R2. This is a metric about the goodness of fit of a model, rather than accuracy per se. Instead, the RMSE values should be the focus of these accuracy discussions (or, even better, the standard errors?).
The logic for why the mustard plots were removed is curious. They are valid results aren’t they? By excluding them because they have higher values and are ‘biasing’ results the authors are effectively saying that they are only interested in building methods that assess crops under ~60 cm – that they don’t want to be able to conclude anything about detecting crops above this height.
The grouping of plots into canopy shapes (smooth, medium, rough) should be done on the UAV data after they have been converted to canopy heights. Currently the authors group them into these classes based on the height profiles and I don’t see the relevance of that to understanding canopy height.
The method used to test which height data source best predicts biomass needs substantial improvement. At the moment, the authors only perform a regression analysis between height and biomass. This doesn’t really reveal anything. A better test would be to use some form of a canopy biomass model that uses CH as an input, and then feed each source of CH into that model and assess the accuracy of each of the model outputs.
The differences between plant height and canopy height are discussed by the authors [line 374]. It is a moot discussion as these two concepts are very similar to each other (the authors themselves couldn’t find an obvious distinction between them in the literature). The real issue that needs discussing is the difference between canopy/plant height and what the UAVs measure, which is a height profile. These are two different things.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The study describes the use of drones to obtain canopy height of cover crops with applications in predicting biomass, and I think the work is interesting, but I have the following questions.
1. It appears that there is a formatting problem with the images in the paper, so please make changes to increase readability.
2. Only two flight altitudes were used in the article to show that the lower the altitude, the higher the accuracy of the obtained canopy height, which I think is not enough. Moreover, although the point cloud density is one of the conditions for the accuracy of canopy height acquisition, light conditions, terrain features, UAV performance, the number and accuracy of control points, etc. are all factors for the image canopy height, so I think the paper should not draw such an arbitrary conclusion.
3. It was shown in the conclusion that in dense and smooth canopies, the UAV method is suitable for obtaining canopy height and, on the contrary, in rough canopies, the accuracy of this method is significantly reduced. I would like to know what criteria define dense and rough and whether there is some measure in the prediction of biomass.
4. In sparse canopy vegetation, I wonder how manual measurements were used to define canopy height and how it is represented?
5. There is an unknown error in line 203, please correct it.
6. The end of line 239 and 249 appear unfinished content.
7. There are many identical pictures in the article, such as Figure 3, Figure 3A, Figure 3B, etc. Please carefully check and make corrections.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The entitled "Using high-resolution UAV imaging to measure canopy height of diverse cover crops and predict biomass” assessed the potential ability of RGB images in extracting the crop height, and its potential for predicting the crop biomass. The manuscript was in good shape, and the methods were logical and the results were reliable. Before the final acceptance, there were small revisions needed to be handled. I would recommend the minor revisions for this manuscript.
Specific comments:
1. Section introduction. The review of crop height extraction is not complete, and there have been many papers on the crop height extraction, and the crop height was widely applied for phenology extraction and yield prediction. Redo this part and add necessary references. I would recommend some for this:
1) Estimating the maize biomass by crop height and narrowband vegetation indices derived from UAV-based hyperspectral images
2. Figure 2, figure 3, figure 4. The number of points should be given clearly. The font should not be covered by the lines. Made necessary modifications.
3. Since the machine learning was widely applied for yield prediction, and many studies have proved the non-linear relationships between the vegetation index, texture index, and crop height. I would suggest the authors add uncertainty and prospect of yield prediction using machine learning methods based on UAV data. I would recommend some reference for this.
1) Object-based image classification of summer crops with machine learning methods
2) Machine Learning-Based Approaches for Predicting SPAD Values of Maize Using Multi-Spectral Images
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The author has solved all my questions.
