Individual Tree Aboveground Biomass Estimation Based on UAV Stereo Images in a Eucalyptus Plantation

Round 1

Reviewer 1 Report

Dear Authors 

The manuscript entitled “Individual Tree Aboveground Biomass Estimation Based on UAV Stereo Images in Eucalyptus Plantation” This research three algorithms were used to estimate the AGB of individual eucalyptus trees with different age, and the effects of different feature variables are explored to achieve accurate estimation of individual tree biomass. However, the authors must address a major concern before the manuscript may be considered for publication. The following are my observations on these topics.

• Authors should mention the full form in the paper initial appearance.

Abstract:

• In the abstract, authors should keep the material flowing. It's also a good idea to highlight the most noteworthy findings and novelties.

• There is no issue statement or research gap.

Introduction:

• The introduction portion should be strengthened. The existing research gap and why the study is necessary.

• Rewrite the line 134 to 145

Materials and methods

2.1. Study Area

• Add climatic information related to the study area.

• Rearrange the formula according to the journal format. 

Result and Analysis

• Explanation of the result should be rewritten concisely.

• Rewrite this section “Individual Tree Segmentation Results of Eucalyptus Plantation”.

• All the figures must be correct the visibility of the figure is not good.

Conclusions:

• Rewrite the conclusion portion and possible to short the conclusion section.

The English requires slight revision for spell check and grammar 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

See attached.

Comments for author File: Comments.docx

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript “Individual tree aboveground biomass estimation based on 2 UAV stereo images in eucalyptus plantation” explored the possibility using cost-efficient data acquisition by UAV for estimation of biomass of individual eucalyptus trees. The study comprehensively tested the performances of various features derived from UAV data as predictors for biomass estimation and model algorithms. However, the methods used in this study have some problems and need to be clearly described.

1.     How the features derived from UAV data were linked to individual trees? Calculated based on the segmented tree boundary? For spectral or vegetation indices, mean, median, or other kind of statistical value?  

2.     Features selection based on correlation coefficients may be good for linear regression, but not for non-linear or non-parametric algorithms, such as Random Forest algorithm and CatBoost. Random Forest has ability of ranking variable importance, why not use it?

3.     I assumed all variables (except for the phenological features) were derived from UAV images acquired in July, the same time as field survey. Because Eucalyptus grow very fast, after three month or a half year, the biomass of individual tree changed much. Thus, the phenological features cannot be linked to the individual tree biomass of July.

4.     Tree age certainly is one important factors determining the biomass of trees. Due to the fast-growing characteristics of eucalyptus, based on our study, age in years as a predictive variable is too coarse, while age in months is better. Besides, in this study, age only has 3 values (2, 3, 4), not sufficient for describing the variation in biomass among the ages.   

5.     Estimation accuracy of canopy area from UAV is relatively low. Using inaccurate values as references to estimate biomass must lead low estimation accuracy.

6.     There were various scenarios in combination of variables and models, but were not clearly described in method section, but presented in result section, causing great confusions. Suggest make a flowchart illustrating each step from data acquisition, feature extraction, model building, results, and according to the flowchart, describe method step by step.

7.     In model construction section, do not only introduce the algorithms, should describe how the method is used in this study. For example, what kind of input variables, how to set the parameters in the algorithms?

8.     Because all features were derived from UAV data, UAV is the only data source. Thus it cannot be called as multi-source data.

 Other suggestions:

1.     Maps were not in good quality. For example, Figure 1, the coordinates only showed on the top of upper left map, missing north arrow. Which map is the scale bar  accociated? Figure caption needs more detailed.

2.     There many figures showing the estimation results (figure 8,9,10, 11, 12, 13), consider to combine them in a table.

 

General, English writting is fine, but needs carefully checks of grammar errors. In additon, avoid using same words or phrases repeatly in a paragraph, pay more attension on the context between sentences and paragraphs.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Overall

Overall this study represents an interesting assessment of parameters and methodologies to estimate Eucalyptus AGB. The introduction is good but the results are poorly analysed and presented, and should be better discussed.

One of the main flaws, in my opinion, is the consistent use of AGB RMSE to compare model performances. Without a scale RMSE has little meaning. You shouldn’t use absolute values of RMSE values to compare situations with different AGB, due to tree age or density.

Some more statistical analyses to see whether differences are significant or not are also needed.  And, the discussion should address all of the results presented (e.g. seasonal changes are not discussed).

Once the results section is corrected/improved, using a better measure than RMSE for comparisons, and testing whether the observed differences are significant, the discussion section has to be rewritten accordingly. This article needs major changes to be acceptable.

 

Main issues

L. 59 The RMSE is not very informative, as its value depends on the range of data values. An error of 106m³/ha would be big for a total value of 200 m³/ha, yet small for a total of 1000 m³/ha. I propose using another measure, like NRMSE or a relative RMSE.

L. 59-70 The whole comparison of studies using RMSE is flawed if the trees are of different age or density. As differences in this error maybe due to differences in tree biomass per hectare.

L.98 The relative RMSE is very high. More than a third. I would not conclude that they show a great potential for the methodology.

Table 1 – add SD to provide a measure of variability

Figures 3 to 5. These actually represent the “Pairwise correlations between spectral characteristics and AGB”, as also correlations between vegetation indices are shown. However, I see no need to present all of these correlation results if the only ones effectively used and discussed are the correlations with AGB. I propose to present only these correlations and eliminate the correlations between the parameters, and to present the correlations in a table, which is more informative than the coloured graphs. Later, when a few parameters are selected as the most relevant ones (TH, age, CA etc.) you can and should check the correlation between them.

Please sort the parameters (in the new table) to facilitate interpretation. For instance, the results of Figure 4 should be sorted either by band (RGB) or by texture parameter.

L. 312/313 “Among the three forest ages, Eucalyptus plantations with three years of age achieved the highest F of 0.93.” – any idea why? Because more samples were analysed? Less variability between samples?

L.330 “Pearson correlation between 18 spectral features and biomass was analysed” Pearson correlation shows a linear relationship applicable to normally distributed data without outliers. Did spectral feature-AGB plots confirm these assumptions?

L. 392/396 You write “Additionally, there are some gaps between eucalyptus tree crowns, which are not completely connected, but the segmented individual tree crowns are completely connected in most cases, and the difference may lead to a certain difference between the extracted individual tree CA and the measured CA.” If I understand this sentence right, this would imply that the Extracted CA would tend to be larger than the CA measured in the field. Figure 7a suggests the contrary! This should be discussed. It would be useful to have the regression formulas in the graphs to see the slope and intercept values.

L. 402 which were the top six spectral characteristics selected? The ones with the highest absolute correlation? Are the presented results the mean value for the 6 parameters? What was the variability (SD)? Please test and indicate if the differences are significant!

Figures 8-11 and 13. Again, RMSE is not a good measure. Use a relative RMSE or NRMSE. Why use 3D figures? Please indicate SD in the figures and if the differences were significant. Also, the y-axis should be comparable and start at “0”. The differences in Figure 11a seem enormous but are indeed very small.

Were these results tested per tree-age (patch) and consistent per tree age?

Sections 3.2.1 to 3.2.3 could be merged in a single section with a single results table.

L. 619-623 You write “In this study, the incorporation of forest age greatly enhances the accuracy of the Eucalyptus AGB model, with a maximum increase in R2 of 0.45. This phenomenon can be attributed to the fast growth rate of Eucalyptus, leading to substantial variations in individual tree biomass at different ages within a year.” First of all, the gain in R2 is in comparison to the use of spectral signatures. If you compared it to a model with TH, the gain would be much lower (as suggested by Figure 15). So, age is very relevant when other relevant variables are missing. Secondly, if there are “substantial variations in individual tree biomass at different ages” (and I don’t know what you men with “within a year”), age does not always represent AGB well. This is therefore not an argument why age is a good estimator.

L.672 Forest age is not a variable extracted from UAV-acquired data.

L.674-677 The differences in accuracy should be discussed

The seasonal AGB variation presented in Figure 15 is not discussed at all.

 

Editorial details

Overall, insert space before reference brackets []

Figure 1. the latitude axis has no labels.

L. 180 add the version of the Pix4D software

L. 215 should read “formulas (2) to (4).”

L. 226/227 should read “formulas (5) to (7).”

Table 3 – add reference(s) and description of parameters (N, P)

L. 256 explain the abbreviation CA (canopy area is only mentioned in Figure 7, L. 384)

L. 282/283 eliminate spaces before and after “RandomForestRegressor”

L. 292 eliminate spaces before and after “CatBoostRegressor”

L. 297 should read “referenced as (8) and (9).”

Table 4 help the reader reminding him what the parameters mean (TP, FN, …). Although previously mentioned, I find it useful to repeat the description of abbreviations when they first appear in a new section (here, in the Results section).

Figure 2 – rewrite legend in a more concise way, e.g.: “Figure 2. Individual tree segmentation results of Eucalyptus plots with different forest ages. Results for 2-year old (a-c), 3-year old (d-f) and 4-year old (g-i) forests, with orthophoto mosaic, point data derived from the stereo images, and individual tree segmentation based on point data, from left to right.”

I presume the images in the left column are orthomosaics of several overlapping photos.

Shouldn’t the red box be transparent, so that the reader can see the segmentation result?

I can’t see the red triangles of the “real  point”, just “#”.

Furthermore, the font-size in the figures should be adapted according to the legend font-size, and the symbol sizes are also far too big.

Also, please add scales to the aerial imagery.

L. 316 “Eucalyptus trees are found to be undersegmentation” … undersegmented …

Figure 6 – I suspect the figures were not correctly converted to PDF. The boundaries should be transparent and thin-lined so that the tree crown is visible, and I can’t identify real points and detected points in the image. Furthermore, the font-size in the figures should be adapted according to the legend font-size, and the symbol sizes are also far too big.

Figure 7a – CA (m²), not TH per square meter

Figure 7b – TH (m), not TH per meter, or per square meter

L.384/385 Is canopy area = crown area? Better use a coherent terminology

Figure 12. please use different colours from those used for the methods in Figures 8-11 and 13.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors has addressed and did sufficient revision on all the comments raised during first  review.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Glad to see this improved version of your research, well done!

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

 

Thank authors for their great efforts to revise the manuscript and for taking reviewers’ comments and suggestions seriously. The revised manuscript was much improved.

As the research indicates that age is the most important factor determining the AGB of eucalyptus tree. Therefore, it is better to sample different ages of individual trees, 5 values is really insufficient for constructing a reliable model. If possible, collect more samples at various ages to improve the validity of models.

Some questions about the results:

1. Figure 7 shows that the scatter points between the extracted CA and the measure CA is closer to 1:1 line (R2 0.91), better than those with TH, but R2 for TH (0.94) is greater than for CA (0.91). Please check them carefully.

2.  In Figure 8, correlation coefficients between r, g, and b (mean values of spectral bands) and AGB are very different from those between MEA-R,  MEA-G, and MEA-B (also means of spectral bands), even the direction. Can you explain it?

3.  Suggest to avoid using ‘And’ or ‘On…’ to begin sentences or paragraphs.  

 

Other suggestions can be found in attached PDF file.

 

 

Comments for author File: Comments.pdf

Some sentences need revision for readibility.

Author Response

Please see the attachment.

Author Response File: Author Response.docx