Next Article in Journal
ELNet: An Efficient and Lightweight Network for Small Object Detection in UAV Imagery
Previous Article in Journal
Crop Classification Using Time-Series Sentinel-1 SAR Data: A Comparison of LSTM, GRU, and TCN with Attention
 
 
Article
Peer-Review Record

Direct Estimation of Forest Aboveground Biomass from UAV LiDAR and RGB Observations in Forest Stands with Various Tree Densities

Remote Sens. 2025, 17(12), 2091; https://doi.org/10.3390/rs17122091
by Kangyu So 1,*, Jenny Chau 1, Sean Rudd 2, Derek T. Robinson 3, Jiaxin Chen 4, Dominic Cyr 5 and Alemu Gonsamo 1
Reviewer 1: Anonymous
Reviewer 2: Anonymous
Remote Sens. 2025, 17(12), 2091; https://doi.org/10.3390/rs17122091
Submission received: 17 May 2025 / Revised: 7 June 2025 / Accepted: 17 June 2025 / Published: 18 June 2025

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors General Comments

This paper presents a novel approach for estimating forest above - ground biomass (AGB) using unmanned aerial vehicle (UAV) - based LiDAR and RGB data combined with deep - learning techniques. The topic is of great significance for forestry management and carbon cycle research. However, there are several issues that need to be addressed to improve the quality and impact of the paper.

 

Specific Comments

L165: "After processing" – What specific processing steps are referred to here? Preprocessing?

 

L171: The ground validation data has a small sample size – only 72 trees, which is insufficient.

L177: Data were collected in both leaf-off and leaf-on seasons, but the manuscript does not analyze whether seasonal differences affect the methodology, results, or accuracy.

L261: The entire Results section is disorganized and lacks subsections. The paragraphs contain excessive numerical data – consider creating a table to compare different data methods or harvested treatments.

L270: The low R² values (e.g., crown diameter R²=0.27) are not discussed. It is recommended to analyze error sources, such as canopy overlap or CHM resolution.

L293: "AGB" does not need to be written out in full again here. The same issue occurs in L314 and L339.

L500: The conclusion emphasizes that "MCWS was particularly effective," but in the results, AGB estimation for some treatments (e.g., 55D) has an R² as low as 0.19. A more balanced summary of performance differences is needed.

Figure 3: Please format the figure according to the journal's guidelines, such as font size and subplot dimensions. The horizontal and vertical axes are barely legible.

Table 3: Add a column indicating the plot numbers for each thinning treatment. What explains the large R² discrepancy between the two methods for 55A?

Author Response

Reviewer 1, Comments 1: (L165) "After processing" – What specific processing steps are referred to here? Preprocessing?
Response 1: We have revised the statement in L165 to clarify that it is a preprocessing step to filter out erroneous points.

Reviewer 1, Comments 2: (L171) The ground validation data has a small sample size – only 72 trees, which is insufficient.
Response 2: We appreciate the reviewer’s concern about sample size. However, we contend that our ground validation dataset (n=72 trees) is statistically sufficient for the following reasons: Representative Coverage: The sampled trees encompass the full range of tree structural variability in the stand, including height (10.70–30.75 m), crown diameter (2.3–10.05 m), and DBH (17.51–52.68 cm), closely aligning with prior field measurements. Effective Sample Units: Each tree represents an independent sample unit, effectively equivalent to 72 plot measurements, enhancing the robustness of our statistical inferences. While larger samples are always desirable, our dataset provides a rigorous basis for analysis, as demonstrated by the consistency with established stand characteristics. 

Reviewer 1, Comments 3: (L177) Data were collected in both leaf-off and leaf-on seasons, but the manuscript does not analyze whether seasonal differences affect the methodology, results, or accuracy.
Response 3: We would like to clarify that the leaf-off data was only used for classification purposes and that the leaf-on data was used to generate height, crown diameter, and AGB estimates. We added an additional explanation in L211 to make this distinction clearer. Analyzing seasonal differences in the results does not yield additional information given that the study area is a mixed forest with deciduous and evergreen tree species. Including leaf-off data, where crowns for deciduous trees will appear smaller compared to leaf-on data, will obviously generate less accurate results. 

Reviewer 1, Comments 4: (L261) The entire Results section is disorganized and lacks subsections. The paragraphs contain excessive numerical data – consider creating a table to compare different data methods or harvested treatments.
Response 4: Thank you very much for this genuine critique. We added subsections to the Results section, dividing it into Unsupervised and Self-supervised Predictions and Delineation, VRH Response to Biomass Estimation, and Biomass Density and Growth. We removed numerical data in the paragraphs that are already in Figure 3 and Table 3, such as overall, unharvested, or harvested observations. We kept numerical data for comparisons not covered in the Figures or Tables, such as moderate vs severe thinning or aggregate vs dispersed, in the paragraphs. Moving these comparisons to an additional table would take up unnecessary space in the manuscript.

Reviewer 1, Comments 5: (L270) The low R² values (e.g., crown diameter R²=0.27) are not discussed. It is recommended to analyze error sources, such as canopy overlap or CHM resolution.
Response 5: Between L395 and L478, we discussed several error sources for low R² values observed in our results. For the example of crown diameter and canopy overlap, we covered this in L396, “First, treetop detection is limited in mixed forests, with omission errors common for smaller crown structures and trees hidden under the canopy (Yun et al., 2021).”, and in L438, “Although growth models often project an inverse relationship between height and crown diameter, MCWS may inaccurately capture crown ratio as the full crown of taller trees are segmented and the canopies of smaller trees are partially hidden.”.

Reviewer 1, Comments 6: (L293) "AGB" does not need to be written out in full again here. The same issue occurs in L314 and L339.
Response 6: We applied the suggested corrections.

Reviewer 1, Comments 7: (L500) The conclusion emphasizes that "MCWS was particularly effective," but in the results, AGB estimation for some treatments (e.g., 55D) has an R² as low as 0.19. A more balanced summary of performance differences is needed.
Response 7: Thank you for this comment. We specified that the method was effective for unharvested (CON R²= 0.80) and severely thinned (33A R²= 0.79, 33D R²= 0.66) forests as the AGB estimation R² for these treatments were high. We revised the summary statement to also discuss the lower R² in moderate thinning treatments (55A and 55D).

Reviewer 1, Comments 8: (Figure 3) Please format the figure according to the journal's guidelines, such as font size and subplot dimensions. The horizontal and vertical axes are barely legible.
Response 8: We applied the suggested corrections.

Reviewer 1, Comments 9: (Table 3) Add a column indicating the plot numbers for each thinning treatment. What explains the large R² discrepancy between the two methods for 55A?
Response 9: We have added the plot numbers for each thinning treatment to Table 3. From L464 to L469, we explained that the large R² discrepancy for 55A between the two methods is because the homogenous clusters of trees left behind by aggregate retention make segmentation difficult for MCWS, but RGB spectral data makes it easier to classify for the self-supervised model. We added an additional statement in L447 to make this explanation clearer.

Reviewer 2 Report

Comments and Suggestions for Authors

The author proposes a forest biomass estimation method based on UAV LiDAR and RGB data, which shows a certain degree of innovation. However, the manuscript still requires Major Revision to meet publication standards. The specific comments are as follows:

  1. Section 2.1: It is recommended to add explanations for aggregate crown retention and dispersed crown retention.

  2. Section 2.2: Regarding how tree height, crown diameter, and DBH are measured on the ground, it is suggested to add a schematic diagram for clearer description.

  3. Section 2.3: How was the threshold of 15-40m determined?

  4. Section 2.3: The description of the method in this section needs to be more specific. What is the principle of the convolutional neural network (CNN) algorithm used here? Based on the current description, it is difficult to understand how individual trees are detected from RGB images.

  5. Table 2: For some tree species, such as Red Maple, the R² is low (only 0.44). The reliability of the fitting requires further discussion.

  6. Table 2: R² values should retain only two decimal places.

  7. The terms R², Ra², Radj² appear in the text and need to be standardized throughout.

  8. Line 264: Is the count of 5122 trees identified by the UAV accurate? Has the identification method (Section 2.3) been validated?

  9. Figure 3: It is recommended to add metrics such as RMSE to more clearly illustrate the accuracy of the algorithm.

  10. Figure 4: The unit for AGB is gC? How was the AGB increment calculated? Was only one phase of UAV data collected?

Author Response

Reviewer 2, Comments 1: (Section 2.1) It is recommended to add explanations for aggregate crown retention and dispersed crown retention.
Response 1: We added more details about the aggregate crown retention and dispersed crown retention treatments in the study area. Basically, aggregated retention mimics natural disturbances, such as windthrow or small wild-fires, by leaving some forest patches intact. In contrast, dispersed retention simulates low-intensity disturbances, allowing for the sporadic survival of certain trees.

Reviewer 2, Comments 2: (Section 2.2) Regarding how tree height, crown diameter, and DBH are measured on the ground, it is suggested to add a schematic diagram for clearer description.
Response 2: Thank you for this suggestion. We added a schematic diagram to explain our process for measuring tree height and crown diameter. We also included an equation to explain the calculation behind tree height using measurements from the clinometer. 

Reviewer 2, Comments 3: (Section 2.3) How was the threshold of 15-40m determined?
Response 3: We determined the threshold of 15-40m to filter out erroneous points and exclude non-mature trees. Previous surveys of the study area have a mean tree height below 30m, so any point higher than 40m would be an error in the point cloud than an actual tree. The minimum threshold of 15m excludes saplings and juvenile trees, which are detected less reliably on UAV observations and are not represented well in the Tallo database for biomass relationships.

Reviewer 2, Comments 4: (Section 2.3) The description of the method in this section needs to be more specific. What is the principle of the convolutional neural network (CNN) algorithm used here? Based on the current description, it is difficult to understand how individual trees are detected from RGB images.
Response 4: Thank you for this genuine critique. The model is DeepForest, which uses deep learning object detection networks to detect ecological objects in UAV imagery. We discussed some characteristics of the convolutional neural network, such as its RetinaNet one-stage detector and ResNet-50 classification backbone, and the principle behind these choices. We revised our description of the convolutional neural network to include the model name and an explanation of the delineation process to provide further depth and specificity.

Reviewer 2, Comments 5: (Table 2) For some tree species, such as Red Maple, the R² is low (only 0.44). The reliability of the fitting requires further discussion.
Response 5: Thank you for the suggestion. Yes, the R² for some tree species is low, which we attribute to the Tallo database being an amalgamation of measurements from individual studies across Canada, involving different growing environments and possibly different measurement methods. For example, data for Red Maple was collected in various distinct ecoregions, which may have influenced the growth rate of the trunk and crown. We expanded our coverage of database limitations in Section 4.4 to include this discussion. 

Reviewer 2, Comments 6: (Table 2) R² values should retain only two decimal places.
Response 6: We applied the suggested corrections.

Reviewer 2, Comments 7: The terms R², Ra², Radj² appear in the text and need to be standardized throughout.
Response 7: We applied the suggested corrections.

Reviewer 2, Comments 8: (Line 264) Is the count of 5122 trees identified by the UAV accurate? Has the identification method (Section 2.3) been validated?
Response 8: Although we do not have an exact in situ count of all trees in the area to validate our UAV-based estimation, 5122 is within the reasonable range of tree count based on previous surveys of the study area. According to Table 1, which provides the density of red pine trees for each treatment stand, there are an estimated 3300-3500 red pines in the study area. Considering the stand is a mixed forest, the additional 1600-1800 trees can be attributed to the Maple, Oak, and Eastern White Pine populations present. We also validated the detection rate of our methods with ground measurements, with the unsupervised LiDAR algorithm and self-supervised RGB model yielding a detection rate of 93.06% (L292) and 75.44% (L303) respectively.

Reviewer 2, Comments 9: (Figure 3) It is recommended to add metrics such as RMSE to more clearly illustrate the accuracy of the algorithm.
Response 9: Thank you for the suggestion. We have updated Figure 3 to include RMSE.

Reviewer 2, Comments 10: (Figure 4) The unit for AGB is gC? How was the AGB increment calculated? Was only one phase of UAV data collected?
Response 10: We apologize for this confusion, the unit for AGB should be kg in this case. We described our calculation process for AGB increment in L352. We input our tree height estimates from our UAV data and annual DBH growth from a previous DBH inventory survey of the study area into our allometric equations to calculate the annual AGB increment. We added a description of our calculation process to the caption for Figure 4 and the results text in Section 3.3. UAV data was only collected within one year (2023), and we acknowledge that tree height change is an important component of AGB increment. Considering that the majority of trees involved in our AGB estimates are mature, the annual height change is likely minuscule. Nevertheless, we added this detail to our limitations in Section 4.4.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

 The manuscript has been revised as required.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript has been thoroughly revised in accordance with my prior feedback, and I am confident it is now ready for publication.

Back to TopTop