Multi-Temporal and Multi-Resolution RGB UAV Surveys for Cost-Efficient Tree Species Mapping in an Afforestation Project
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript evaluates the optimal spatial resolutions for vegetation mapping using a consumer-grade RGB UAV in resource-limited areas. Conducted in Kasho, Pakistan, it found that a 6 cm resolution achieved a good balance between accuracy (R² = 0.92–0.98) and broader coverage (12.3–22.2 hectares), with the highest Benefit-Cost Ratio (BCR: 0.011–0.015). The research highlights the potential of consumer-grade UAVs for cost-efficient and high-precision vegetation mapping in afforestation projects. This study might be interesting to the Remote Sensing readership but still required potential improvements before being published. The main issues are as follows:
- Is there an elevation variation in the study area? If so, how can the UAV-based image resolution be ensured at different elevation levels? Additionally, the manuscript does not provide information on the growth conditions of different tree species in the study area, such as tree height. Are there significant differences in height among different tree species, and would such differences affect the image resolution obtained by the UAV?
- How were the vegetation indices (VIs) of the samples calculated? Were they based on the mean values of Regions of Interest (ROIs) or something else? Please provide detailed methods for calculating the VIs.
- Were the parameters of the SVM used in this study optimized? Please provide the critical SVM parameter values used in this study.
- This study evaluated the performance of UAV-based RGB images with resolutions of 2 cm, 4 cm, and 6 cm for vegetation mapping in afforestation projects. The study found that RGB images with a 6 cm resolution achieved the highest BCR. However, this study did not compare the performance of lower resolutions, such as 8 cm or 10 cm, in vegetation mapping. Readers might inevitably question this conclusion.
- Line 161: The resolution of Figure 3 is too low, please ensure the quality of all Figures in this manuscript.
- Line 228: Change ‘Optima’ to ‘Optimal’ in Figure 4.
- Line 299-301: remove the duplicate sentence of ‘As resolution increases from 2 to 6 cm, mission time decreases significantly.’
- The format of the references should meet the requirements of this journal.
Author Response
Reviewer 1
Question 1:
Is there an elevation variation in the study area? If so, how can the UAV-based image resolution be ensured at different elevation levels? Additionally, the manuscript does not provide information on the growth conditions of different tree species in the study area, such as tree height. Are there significant differences in height among different tree species, and would such differences affect the image resolution obtained by the UAV?
Response1:
Thank you for your questions.
Regarding elevation variation in the study area, we have included additional details on the elevation range, which spans from 306 to 323 meters (lines 107–109). These variations are relatively minor, and the Ground Sampling Distance (GSD) is primarily determined by flight height and camera specifications. To ensure consistent GSD across all plots, we adjusted the UAV flight height to match the elevation of each plot. This approach maintained uniform image resolution throughout the study area.
As for tree height, we observed a range of 1–6 meters (L120). While there are species-specific differences in tree height, these variations do not significantly impact UAV image resolution. The relative height differences (≤5 m) are negligible compared to the operational flight altitude of ~60-80 meters above ground level, as demonstrated in UAV photogrammetry studies where vertical canopy variations had minimal effect on overall GSD consistency (Torres-Sánchez et al., 2015). Furthermore, the generally flat and our flight altitude strategy—adjusting the UAV’s height for each plot to maintain a constant height above ground—ensured uniform resolution across all plots. This method aligns with best practices for GSD stability in terrain-following UAV workflows (Colomina & Molina, 2014; Whitehead et al., 2014).
The primary factors influencing GSD in our study were flight height and camera settings, as supported by photogrammetric principles (Mikhail et al., 2001). We hope this clarification and added references address your concerns
Question 2
How were the vegetation indices (VIs) of the samples calculated? Were they based on the mean values of Regions of Interest (ROIs) or something else? Please provide detailed methods for calculating the VIs.
Response2:
Thank you for your question regarding the calculation of Vegetation Indices (VIs) in our study.
In this study, the VIs were calculated on a pixel-by-pixel basis rather than using the mean values of Regions of Interest (ROIs) . This approach ensures that the indices capture the full spatial variability of vegetation characteristics across the study area, providing a more accurate and detailed assessment. We utilized the Python libraries ArcPy and NumPy to handle the geospatial and computational aspects of the calculations. Specifically, the RGB bands (Red, Green, and Blue) were extracted from composite rasters and converted into NumPy arrays for efficient pixel-wise calculations of VIs using formulas in dicated in the table 1, leveraging arcpy.RasterToNumPyArray() and element-wise operations to calculate the indices. The calculated VIs were then converted back into raster format using the same arcpy.NumPyArrayToRaster() for calculated index values of every pixel) back into a raster format.
Table 1. RGB UAV Based Vegetation indices calculated for this study.
Indices |
Formula |
References |
|
Red chromatic coordinate (RCC) |
R/(R+G+B) |
[25] |
|
Green chromatic coordinate (GCC) |
G/(R+G+B) |
[25] |
|
Blue chromatic coordinate (BCC) |
B/(R+G+B) |
[25] |
|
Normalized difference index (NDI) |
(RCC-GCC)/(RCC+GCC+0.01) |
[26] |
|
Green leaf index (GLI) |
(2*R-G-B)/(2*R+G+B) |
[31] |
|
Kawashima index (IKAW) |
(R-B)/(R+B) |
[25] |
|
Mean of RGB bands (MRGB) |
(R+G+B)/3 |
[32] |
|
Excess green vegetation index (EXG) |
(2*RCC-GCC-BCC) |
[26] |
|
Visible atmospherically resistance index (VARI) |
(G-R)/(G+R-B) |
[33] |
This approach ensures that the VIs capture the full spatial variability of the data, offering a more accurate and detailed assessment of vegetation characteristics across the study area.
Question 3
Were the parameters of the SVM used in this study optimized? Please provide the critical SVM parameter values used in this study.
Response3:
Thank you for your question regarding the parameters of the Support Vector Machine (SVM) used in our study.
In this study, the SVM classifier was implemented using the default parameters provided by the Scikit-learn library within ArcPy (Line 223-232). Specifically, we used the Radial Basis Function (RBF) kernel, which is a commonly applied default choice for SVM in remote sensing applications. The critical default parameter values used in this study are as follows:
- Kernel: RBF
- C: 1.0 (the default value in Scikit-learn, representing the penalty parameter for misclassification)
- Gamma: 'scale' (the default value in Scikit-learn, which is 1 / (n_features * X.var()) )
Where n_feature=3(RGB)+1(PCA), while X.var() represents the variance of a given independent variable.
These parameters were not optimized through methods such as grid search or random search due to time and computational resource constraints. However, we implemented 5-fold cross-validation (Line 227) to assess the model’s performance and ensure its robustness. After cross-validation, a final SVM model was trained using the entire dataset for the final classification. While parameter optimization could potentially improve classification accuracy, the cross-validation process helped ensure reliable results. As noted by Mountrakis et al. (2011), cross-validation is a widely accepted method for assessing the performance of SVM models in remote sensing applications. They found that cross-validation can provide a robust estimate of model performance, even when parameters are not optimized.Additionally, Gomes et al. (2012) demonstrated that cross-validation can effectively reduce the bias in accuracy assessment, further supporting the reliability of our approach.
Question 4
This study evaluated the performance of UAV-based RGB images with resolutions of 2 cm, 4 cm, and 6 cm for vegetation mapping in afforestation projects. The study found that RGB images with a 6 cm resolution achieved the highest BCR. However, this study did not compare the performance of lower resolutions, such as 8 cm or 10 cm, in vegetation mapping. Readers might inevitably question this conclusion.
Response4:
Thank you for your thoughtful feedback.
Our study evaluated UAV-based RGB imagery at 2–6 cm resolutions, as this range is critical for detecting fine-scale features in early-stage afforestation projects. Testing resolutions beyond 6 cm (e.g., 8–10 cm) was limited by our UAV’s battery capacity, which restricted flight endurance at higher altitudes. Importantly, the afforested trees in our study were young and small (e.g., seedlings and saplings), and resolutions coarser than 6 cm risk oversimplifying their morphological traits, such as leaf structure or stem density, which are vital for accurate species identification and health assessment (Zhou et al., 2023; Torres-Sánchez et al., 2015). Prior research emphasizes that lower resolutions may inadequately capture small vegetation features, reducing classification accuracy in young plantations (Young et al., 2022). While future studies could explore lower resolutions with advanced UAVs, our findings align with evidence that resolutions ≤6 cm better balance detail and coverage for monitoring early growth stages. We have revised the text to clarify this rationale (lines 490–498) and highlighted opportunities to integrate multispectral sensors for scalable monitoring.
Question 5
Line 161: The resolution of Figure 3 is too low, please ensure the quality of all Figures in this manuscript.
Response5:
Thank you for bringing up the resolution issue.
We have improved the resolution of all figures in the manuscript, specifically Figures 1 (Line 119) and 3 (Line 164). Additionally, following Reviewer 2's suggestion to remove Figure 2, the numbering has been adjusted, and Figure 3 is now renumbered as Figure 2.
Question 6
Line 228: Change ‘Optima’ to ‘Optimal’ in Figure 4.
Response6:
Thank you for pointing out the typo.
We have corrected it and replaced "Optima" with "Optimal" in Figure 4(previous), but now it is Figure 3.
Question 7
Line 299-301: remove the duplicate sentence of ‘As resolution increases from 2 to 6 cm, mission time decreases significantly.’
Response7:
Thank you for your thorough review.
We have removed the duplicate sentence, "As resolution increases from 2 to 6 cm, mission time decreases significantly," and tracked the change accordingly with new line number (322).
References
Colomina, I., & Molina, P. (2014). Unmanned aerial systems for photogrammetry and remote sensing: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 92, 79–97.
Gomes, T.A.; Prudêncio, R.B.; Soares, C.; Rossi, A.L.; Carvalho, A. Combining meta-learning and search techniques to select parameters for support vector machines. Neurocomputing 2012, 75, 3–13.
Mikhail, E. M., Bethel, J. S., & McGlone, J. C. (2001). Introduction to modern photogrammetry. John Wiley & Sons.
Mountrakis, G.; Im, J.; Ogole, C. Support vector machines in remote sensing: A review. ISPRS Journal of Photogrammetry and Remote Sensing 2011, 66, 236–247.
Torres-Sánchez, J., López-Granados, F., De Castro, A. I., & Peña-Barragán, J. M. (2015). Configuration and specifications of an unmanned aerial vehicle (UAV) for early site-specific weed management. PLoS ONE, 10(6), e0128967. https://doi.org/10.1371/journal.pone.0128967
Whitehead, K., Hugenholtz, C. H., Myshak, S., et al. (2014). Remote sensing of the environment with small unmanned aircraft systems (UASs), part 1: A review of progress and challenges. Journal of Unmanned Vehicle Systems, 2(3), 69–85.
Young, D.J.N., Koontz, M.J. and Weeks, J.M., 2022. Optimizing aerial imagery collection and processing parameters for drone-based individual tree mapping in structurally complex conifer forests. Ecological Applications, [online] Available at: https://doi.org/10.1111/2041-210X.13860 https://doi.org/10.1111/2041-210X.13860.
Zhou, X., Wang, H., Chen, C., Nagy, G., Jancso, T. and Huang, H., 2023. Detection of Growth Change of Young Forest Based on UAV RGB Images at Single-Tree Level. Forests, 14(1), p.141. Available at: https://doi.org/10.3390/f14010141 https://doi.org/10.3390/f14010141.
Author Response File: Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsL22-40: Explain better in the abstract what exactly is being mapped and what is meant by accuracy. The title and abstract only mention “vegetation mapping”, which is rather unspecific. What exactly is being mapped? Just vegetation vs. everything else (bare earth, built-up areas, water, etc.) or are you mapping different vegetation types (grassland, shrubs, trees) or trees versus other vegetation or something else? This should be specified in the abstract and possibly also in the title.
L46: Vegetation mapping by itself does not assess carbon sequestration. Please rephrase these sentences. Forest type mapping, or better tree detection including tree height and tree species, are a basis for assessing carbon sequestration potential.
L140 (Figure 2): This figure does not seem relevant for understanding the paper or replicating the work. The illustration of the flight planning grid app does not add necessary information to the research paper. Consider removing it.
L142-144: The information about Dronelink is also not necessary. For understanding and replicating the work it is irrelevant which flight planning software was used. There are many options which provide the same result. Consider taking it out. The flight parameters on the other hand, in particular front and side overlap are important and should stay in the paper.
L151: March is late winter (not late fall).
120 – Chapter 2.2 Data Collection: It would be helpful to mention here how the time series, leaf-off and leaf-on data (orthomosaics) were matched to each other in terms of their geolocation. It seems you have not used differential GPS correction (PPK/RTK) because consumer-grade UAVs do not provide this option. Did you use ground control plates? – Later in the text (L235) you mention Ground Control Points. Please include them here in the methods and explain if and how they were marked to remain at the same location between flights.
Line 3.2: Resolution Impact: This paragraph explains an obvious fact for anybody who has done drone surveys. It simply states that coarser resolution results allow greater aerial coverage. This can be stated just for clarity in a single sentence, since it is common knowledge. Also in Figure 5 it is not clear what “time” refers to. I assume it is the flight time. However, there is more to it, since higher resolution flight require not only longer flights but also produce higher data volumes and in particular increase the processing time. If you include all these factor together then the paragraph would become more relevant and could stay in the paper.
Author Response
Reviewer 2
Question 1
L22-40: Explain better in the abstract what exactly is being mapped and what is meant by accuracy. The title and abstract only mention “vegetation mapping”, which is rather unspecific. What exactly is being mapped? Just vegetation vs. everything else (bare earth, built-up areas, water, etc.) or are you mapping different vegetation types (grassland, shrubs, trees) or trees versus other vegetation or something else? This should be specified in the abstract and possibly also in the title.
Response1:
Thank you for your constructive feedback.
In our study, vegetation mapping refers explicitly to the mapping vegetation at the species level (e.g., species such as Acacias.spp and Eucalyptus.spp) and also included 2 non-vegetation classes, including bare land and water. We included both vegetated and non-vegetated ground points model training and validation as classifier training needs both the presence and absence of a given vegetation class, in this case, a given tree species. These non-vegetation classes were also included to ensure robust discrimination.
Regarding the term "accuracy," it refers to the overall classification accuracy, using R2. While R2 is traditionally used in regression analysis, it can be adapted for classification tasks by treating the predicted class probabilities as continuous values and comparing them to the actual class labels. This approach allows us to quantify the proportion of variance in the class labels explained by the model's predictions.
We have revised both the title and the abstract to reflect this specificity. The new title is “Multit-temporal and Mulit-resolution RGB UAV Surveys for Cost Efficient Tree Species Mapping in Afforestation Project” and in the abstract, we have added vegetation classes and non-vegetation classes for clarification.
Question 2:
L46: Vegetation mapping by itself does not assess carbon sequestration. Please rephrase these sentences. Forest type mapping, or better tree detection including tree height and tree species, are a basis for assessing carbon sequestration potential.
Response2:
Thank you for highlighting this critical distinction.
We agree that vegetation mapping alone does not directly quantify carbon sequestration. In response to your feedback, we have revised Lines 46–50 to the new text that states “As it furnishes fundamental data essential for estimating carbon sequestration—encompassing various vegetation types, canopy height, structure, etc.—it concurrently acts as a significant tool for assessing carbon sequestration potential [2], essential for climate change mitigation and achieving the Sustainable Development Goals (SDGs).”
Question 3
L140 (Figure 2): This figure does not seem relevant for understanding the paper or replicating the work. The illustration of the flight planning grid app does not add necessary information to the research paper. Consider removing it.
Response3:
Thank you for your valuable suggestion.
We have taken your feedback into consideration and have removed Figure 2 (Line 143) to ensure the paper remains focused and relevant.
Question 4
L142-144: The information about Dronelink is also not necessary. For understanding and replicating the work it is irrelevant which flight planning software was used. There are many options which provide the same result. Consider taking it out. The flight parameters on the other hand, in particular front and side overlap are important and should stay in the paper.
Response4:
Thank you for your recommendation.
We have removed the unnecessary information about Dronelink (Lines 145-148), while ensuring that the important flight parameters, particularly the front and side overlap, are retained in the paper (Lines 148-151).
Question 5
L151: March is late winter (not late fall).
Response5:
Thank you for your thorough review.
We have corrected the term to accurately reflect that March is late winter, as suggested (Line 153).
Question 6
120 – Chapter 2.2 Data Collection: It would be helpful to mention here how the time series, leaf-off and leaf-on data (orthomosaics) were matched to each other in terms of their geolocation. It seems you have not used differential GPS correction (PPK/RTK) because consumer-grade UAVs do not provide this option. Did you use ground control plates? – Later in the text (L235) you mention Ground Control Points. Please include them here in the methods and explain if and how they were marked to remain at the same location between flights.
Response6:
Thank you for your insightful query.
To ensure precise geolocation alignment between leaf-off and leaf-on orthomosaics, we employed the following protocol:
- Ground Control Points (GCPs):
- At the center of each plot, permanent, high-visibility Ground Control Points (GCPs) were installed at across the study area. Additionally, the center of each tree within the plot was also designated as a permanent GCP. These GCPs were surveyed once before the initial flight and remained undisturbed throughout the study.
- GCP coordinates were recorded using a Garmin GPSMAP 64s (horizontal accuracy: ~3 m) and later refined during photogrammetric processing. Even though, horizontal error may seem big, however, since the relative alignment between leaf-on and leaf-off orthomosaics was prioritized by meticulously pinpointing the same GCPs in all overlapping images from both campaigns, ensuring pixel-level alignment of invariant features (e.g., plot center, non-vegetation structures). Discrepancies in absolute geolocation (due to GPS limitations) were secondary to maintaining sub-pixel relative alignment between multi-temporal datasets.
- Image Georeferencing Workflow:
- For both leaf-off and leaf-on datasets, GCPs were manually identified in all overlapping UAV images using ArcMap. This enabled bundle adjustment to minimize geospatial errors (<0.5 m RMSE).
- Post-processing image-to-image alignment in ArcMap (nearest neighbor resampling) further synchronized the orthomosaics, leveraging shared GCPs and invariant landscape features (e.g., roads, big targeted tree species) to ensure sub-meter overlap accuracy.
- Flight Consistency:
- Identical flight plans (altitude, overlap: 80% front/side, grid pattern) were replicated for both campaigns, ensuring consistent image acquisition geometry.
This approach, though limited by consumer-grade UAV, aligns with best practices for multi-temporal UAV studies lacking PPK/RTK (James et al., 2020). We have updated Section 2.2 to explicitly detail GCP deployment and alignment workflows (Lines 172–191).
Question 7
Line 3.2: Resolution Impact: This paragraph explains an obvious fact for anybody who has done drone surveys. It simply states that coarser resolution results allow greater aerial coverage. This can be stated just for clarity in a single sentence, since it is common knowledge. Also, in Figure 5 it is not clear what “time” refers to. I assume it is the flight time. However, there is more to it, since higher resolution flight require not only longer flights but also produce higher data volumes and in particular increase the processing time. If you include all these factors together then the paragraph would become more relevant and could stay in the paper.
Response7:
Thank you for your insightful feedback.
We have clarified the term "time" in Figure 5(re-numbered Figure 4) by expanding the analysis to include total time, which now accounts for flight time, orthomosaic generation time, and classification time for each resolution. We have also replaced the previous violin plots with error bar graphs, which now clearly illustrate the relationship between resolution, total time, and area coverage, while also highlighting the impact of higher resolutions on data volume and processing time. We have made the description more direct and concise. We believe these changes enhance the relevance of the paragraph and offer a more complete picture of the factors involved in drone surveys at different resolutions.
References
James, M. R., Robson, S., d’Oleire-Oltmanns, S., & Niethammer, U. (2020). Optimising UAV topographic surveys processed with structure-from-motion: Ground control quality, quantity and bundle adjustment. Geomorphology, 280, 51–66. https://doi.org/10.1016/j.geomorph.2016.11.021
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsUllah et al. uses a consumer-grade RGB unmanned aerial vehicle (UAV) to evaluate optimal spatial and temporal resolutions (leaf-off and leaf-on) for precise, economically viable vegetation mapping. The results showed that 6 cm resolution achieved reliable accuracy (R² = 0.92–0.98) with broader coverage (12.3–22.2 hectares), while 2 cm and 4 cm resolutions offered higher accuracy (R² = 0.96–0.99) but limited coverage (4.8–14.2 hectares). The authors claims that 6 cm resolution covers a large region, so the viability of consumer-grade RGB UAVs for large-scale vegetation mapping at a cost-efficient 6 cm resolution.
Generally, it is well-written.
Some comments:
Table 3 title: Descriptive analysis of classification metrics across resolutions and seasonal conditions for Plot 1, 2, and 3. It had better be Plots considering the plot number to be plural here.
The authors could add another column to the table: Plot number, and put plot 1, plot 2, and plot 3 there. Now there are three tables under the same title, and it is kind of confusing.
Author Response
Please see the attachments. Thank you.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe revised manuscript addressed my concerns.
Author Response
Thank you for your comments