A Comparative Analysis of UAV LiDAR and Mobile Laser Scanning for Tree Height and DBH Estimation in a Structurally Complex, Mixed-Species Natural Forest

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript compares the performance of UAV LiDAR and MLS in estimating tree height and DBH in a structurally complex, mixed-species natural forest in the Romanian Carpathians. Field measurements were collected from 110 trees representing three species: Fagus sylvatica, Abies alba, and Picea abies. MLS generated high-density 3D point clouds capturing detailed stem and crown structures, while UAV LiDAR primarily acquired data from the upper canopy. The results indicate that MLS provided highly accurate DBH estimates (R² up to 0.98), while its height estimates were moderately accurate due to occlusion effects. UAV LiDAR consistently underestimated tree height across all species (maximum R² = 0.16), particularly for taller individuals. The authors concluded that the two technologies are complementary: MLS is well-suited for detailed, stem-level inventory, while UAV LiDAR is more effective for canopy structure mapping.

This study addresses an important topic and applies valuable technologies, but its contribution is currently limited by methodological oversimplifications, small sample size, insufficient statistical assessment, and a lack of quantitative analysis in some of the aims.

General Comments:

While the paper addresses a relevant topic and uses advanced technologies, several critical aspects require clarification or more analysis. These improvements strengthen the validity of the conclusions and the broader relevance of the study.

Major comments:

1. Motivation and Relevance of Comparison:

Based on the Introduction, the rationale for comparing two structurally and functionally different technologies (top-down vs. ground-based scanning) is not convincing. The outcome that UAV LiDAR is unsuitable for DBH and MLS struggles with canopy heights is theoretically well-known. Thus, the novelty of the comparison is questionable unless the paper goes beyond known limitations and proposes solutions, corrections, or hybrid approaches. The generalizability of the findings is also limited to all forest types.

2. Species Selection Justification:

The study includes two conifers and one broadleaved species, but no rationale is provided for this specific selection. It would be helpful to explain how these species represent broader structural diversity and what readers can infer from their inclusion. Are they common in other regions? Do they pose different challenges for LiDAR-based measurement?

3. Sample Size and Tree Size Distribution:

The total sample size (110 trees of 3 species) is relatively small to draw reliable conclusions, especially due to the variability in forest structure. Moreover, the manuscript does not present the DBH or height distribution of the sampled trees, making it difficult to assess how well the models or sensors perform within size classes.

4. Plot Design and Spatial Distribution:

All plots appeared to be located close together, possibly within the same stand type. If so, this limits structural variability and weakens the robustness of the comparison. The manuscript should clarify the species composition and heterogeneity of each plot. A table summarizing plot attributes would be useful.

5. Point Cloud Segmentation:

Individual tree segmentation remains one of the most challenging aspects of LiDAR data processing, yet this study oversimplified it. A brief four-step workflow was presented with no citations or discussion of segmentation accuracy. For both UAV and MLS datasets, it's unclear how trees were delineated and matched to field data. This oversimplification could mislead readers into underestimating the complexity of the task.

6. Structural Complexity Not Quantified:

Although the forest is described as “structurally complex,” no quantitative metrics (e.g., canopy height variability or stem density) are provided to support this issue. Including such metrics would help assess the generalizability of the findings and contextualize the occlusion challenges observed.

7. Aim 3 Not Addressed Quantitatively:

The third aim (analyzing the impact of occlusion and canopy layering) is only addressed by qualitative visualizations. No quantitative occlusion metrics are presented. Since this aim is highly case-sensitive, more rigorous and reproducible analysis would be beneficial.

8. Evaluation Metrics and Statistical Rigor:

While RMSE, R², and MBE are standard error metrics, they are not sufficient on their own. Considering that model performances are close, statistical tests (e.g., paired t-tests or ANOVA) may provide stronger support for comparative claims. This is especially important when differences in performance are subtle or species-dependent.

9. Unsupported Discussion Claims:

Several statements in the Discussion section are not directly supported by results. For example:

“MLS provided highly accurate DBH measurements and reasonable height estimates for understory and mid-canopy trees.”

However, the results section does not stratify trees by canopy position. To make such claims, the authors should present performance metrics separately for trees.

10. No Integration of UAV and MLS Data:

While the paper concludes that these technologies are complementary, no attempt is made to integrate them or test a hybrid workflow. Considering that fusion is increasingly seen as the future of forest remote sensing, it is surprising that this study does not include even a basic fusion experiment or discussion of how integration might overcome the limitations observed.

Minor Comments and Figure Issues:

Figure 1: Needs clarification. The map lacks context; which country, forest region, and city is being shown? The arrow is ambiguous. Also, the stated plot size (0.25 ha) does not match the visual dimensions based on the scale bar. Please verify the accuracy.

Figure 2: Some fitted circles for DBH estimation appear outside the point cloud, and points seem to occur within the tree stem. This raises questions about the fitting algorithm and point classification. How are LiDAR returns being attributed to stem geometry?

Terminology: Use of terms like "sidelap" and "overlap" in flight planning for UAV LiDAR is acceptable but should be clearly explained, as they originate from photogrammetry and may confuse readers unfamiliar with LiDAR-specific applications.

Author Response

Comment 1 (Motivation and relevance of comparison)

Answer 1. We expanded the Introduction to justify comparing UAV LiDAR and MLS specifically in steep, multilayered, mixed-species forests, which remain underrepresented in prior work. We now articulate the gap (most comparisons use plantations or simplified plots) and frame the study around operational constraints (occlusion, terrain, access). We also added a dedicated discussion of hybrid workflows and corrections (Section 4.4) that leverages each platform’s strengths.

Comment 2 (Species selection rationale)

Answer 2. We added a rationale in the Study Area/Methods explaining that Fagus sylvatica, Abies alba, and Picea abies span contrasting crown architectures and stem morphologies that are common across European montane forests. This choice captures structural diversity (broadleaf vs. conifer; horizontally layered vs. conical crowns) and supports broader relevance (Section 2.1).

Comment 3 (Sample size and tree size distribution)

Answer 3. We increased the sample and report distributions. Three additional Picea abies plots were surveyed, yielding 114 trees in mixed plots and 104 trees in single-species plots. We added DBH and height distributions by species (Figure 2) and updated the Study Area and Field Data sections with the new plots and our positioning workflow (QGIS mobile + LiDAR orthophoto). Sample-size limitations are now explicitly acknowledged in the Discussion (Section 4.3).

Comment 4 (Plot design, proximity, and attributes table)

Answer 4. We added Table 1 summarizing key attributes (trees/ha, basal area, quadratic mean diameter, Lorey mean height, species richness, Shannon index; plus SDs of DBH and height). We clarify that POS-1 plots are mixed-species and POS-3 are single-species Picea, that plots are organized as triplets at similar altitude, and that the maximum inter-plot distance is ~2.3 km—yet structure and diversity differ markedly (Methods 2.1; Table 1).

 

Comment 5 (Segmentation workflow and field matching clarity)

Answer 5. We expanded Section 2.5 to detail the VirtSilv segmentation pipeline (trajectory correction, noise filtering, classification, individual-tree delineation, crown refinement) with citation to Niță (2021). Matching to field trees used a top-view orthographic map derived from MLS to mitigate GNSS error; MLS and UAV were co-registered via ICP (CloudCompare). UAV heights were extracted within a 2 m buffer around MLS-validated tree centers relative to the MLS-DTM.

 

Comment 6 (Quantifying “structurally complex”)

Answer 6. We now quantify structural complexity in the Study Area and Table 1 (stem density, basal area, QMD, Lorey height, species richness, Shannon index, SD of DBH and height). These metrics demonstrate higher diversity and size variability in mixed stands versus single-species stands, making “structurally complex” explicit.

 

Comment 7 (Aim 3 not addressed quantitatively—occlusion metrics)

Answer 7. We added a terrain-normalized voxel analysis (Methods 2.X; Results 3.5) that reports, per layer (ground, subcanopy, crown), the union, intersection, MLS-only and UAV-only voxel counts and ratios. Across plots, top-down occlusion (MLS-only) ≈ 0.94–0.97, bottom-up occlusion (UAV-only) ≈ 0.03–0.06, with small intersections. Composite rasters and a metrics CSV are provided; figures show categorical and ratio maps.

 

Comment 8 (Statistical rigor beyond RMSE/R²/MBE)

Answer 8. We added nonparametric tests where assumptions were uncertain: Kruskal–Wallis for plot-wise height errors and Wilcoxon/paired tests where appropriate. As summarized in Results 3.2, MLS DBH percentage errors did not differ significantly between plots (p = 0.134), while MLS height errors differed (p = 0.024). UAV height errors differed strongly between plot types (p < 0.001). These tests support our comparative claims.

 

Comment 9 (Claims about canopy position—provide stratified results)

Answer 9. We stratified results by height-above-ground (HAG) layers (GROUND P0–P30, SUBCANOPY P30–P70, CROWN P70–P100). Layered occlusion maps and metrics quantify where each platform sees volume. The crown shows the highest MLS-only shares; subcanopy and ground remain MLS-dominant with small UAV-only pockets near gaps (Results 3.5; Discussion 4.2).

 

Comment 10 (No integration attempt despite complementary conclusion)

Answer 10. We added Section 4.4 (Practical Implications) describing an integrated workflow: robust ICP co-registration, HAG normalization, shared-grid occlusion diagnostics (0.5–1.0 m voxels), and division of labor (UAV for canopy envelope/coverage; MLS for stems/DBH/subcanopy). We outline trajectory design (loop closures, perimeter + zig-zag) and acquisition timing (outside peak leaf-on where permissible).

 

Comment 11 (Figure 1 map clarity and scale)

Answer 11. Figure 1 was revised to add country/region/city context, a clear north arrow, and labeled arrowheads. Plot footprints were re-scaled to match the scale bar, and captions specify mixed vs. single-species panels.

 

Comment 12 (Figure 2 DBH fitting concerns)

Answer 12. We clarified that DBH is computed from a 10–20 cm slice at 1.3 m above the local DTM, using a robust polygon/circle fit with outlier rejection. Points are first filtered to stem-class returns; misalignments were checked against the MLS top view. Instances where fitted circles appeared offset were due to perspective artifacts; affected panels were corrected/replaced. Details are in Methods 2.4/2.5 and Figure 3 caption.

 

Comment 13 (Define “overlap” and “sidelap”)

Answer 13. We added explicit definitions in Section 2.3: overlap is along-track flight-line overlap; sidelap is across-track (adjacent swaths). Our flights used ~70% overlap and 60% sidelap to stabilize CHMs and point density.

 

Comment 14 (Generalizability and novelty)

Answer 14. The Introduction now frames novelty in the operational context (steep terrain, multilayered canopies, mixed species) and states that our quantitative occlusion metrics and HAG-normalized slicing extend prior comparisons. The Discussion highlights when and how a hybrid workflow is warranted, improving transferability to similar temperate mixed forests.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors, 

I read your study "A Comparative Analysis of UAV LiDAR and Mobile Laser Scanning for Tree Height and DBH Estimation in a Structurally Complex, Mixed-Species Natural Forest" and hereby share my reviewer's comments with you. The motivation of the study is to prove the operational feasibility of UAVLS and MLS in close-to-nature forests of complex characteristics. While most of the comparable studies just work in single-layered open stands, your work contributes significantly to advancments in the field. 

You measured 110 tree diameter and height pairs and compared them to a DJI Zenmuse L1 UAVLS as well as GeoSLAM ZEB Horizon MLS system. While it is already well known that occlusion is the major driver to height estimation in forest ecosystems, you proved high evidence for complex forest ecosystems. 

The graphics in the paper are of especially high quality!

 

Please resolve following major issues, before the paper can be accepted for publication: 

I) l.183: the process of describing individual tree detection could be described in more detail - I understand that region-growing was used for instance segmentation - is this the same as accretion-based growth algorithm? Before this can happen, a set of seed points must be established, which is completely missing from the description. Was there any semantic segmentation done to the point clouds?

II) Slope of the trees
30° slope are a considerable terrain (l.118), and you report most trees standing on steep slopes (l.171). In fact this might also impact tree height estimation: how did you ensure that the tree base point was found according to the standard definition of tree height "starting at the point at the upper part of the slope at the trunk base" - was there any manual correction? Else there might be an offset of several tenth of a meter if falsely taking the lowest part of the slope? 

III) Linear Regression for DBH and height matching
in Figure 3 you provide DBH and height simple linear regressions for the match of reference vs LiDAR measurements. While the graphic provides a detailed insight into the performance of parameter estimation, I have several suggestions with the presentation: Firstly, the axis labels are not matching to the text and should be rather DBH_manual and DBH_automatic or DBH_reference and DBH_LiDAR. Secondly, the regression line should follow (in an ideal case) the 1:1 line (similar to what you did in Figure 4, so I'd recommend to add the 45° diagonal and keep the scaling for both axes x and y constant). Furthermore it is necessary to analyze the coefficient and the intercept of all the regressions. Please provide those as well and perform a test, whether they are significantly different from 0 (for the intercept) and 1 (for the coefficient). 

IV) Interpretation of Occlusion Problematic
Section 3.4 (l.300ff) claims that occlusion was the reason for underestimated tree tops. While this might seem logical in complex forest ecosystems, it might also be caused by wrongly assigned tree points. A general interpretation of occlusion is in my opinion only possible, if the number of filled voxel in the crown detected by MLS is compared to the "real" number of voxel in the crown. It can be assumed, that UAVLS provides more detail for the crown region, so I trust that comparing the filled voxel for both systems would lead to a reliable measure of how much occlusion is present (compare to https://doi.org/10.1016/j.agrformet.2019.01.033 or https://doi.org/10.1016/j.rse.2025.114685). 

 

Minor concerns: 

Please check your references! Reading the introduction I found several citations that do not support the statement that was made in the citing sentence (for example 7, 8 does not regard canopy-dwelling organisms, 22 does not suport MLS data, 28 does not refer to tree height measurements)

l.42: DBH might be a rough proxy for stand age, but is not considered to be directly relatable to individual tree age, especially when considering varying forest productivities and also thinning regimes. 

l.53: I would rather guess that the two limiting factors in collecting data across large areas fails due to a) amount of labour and b) experience and skill of measuring person

l.79: there is one citation in a incorrect format (Schöps et al 2015)

l.141: if the height measurements were conducted twice, were there two independent operators, or was one person measuring? did the measurement happen from the same location or was it varied?

l.173: you used loops for capturing the quarter hectare plots - can you please also provide the three trajectories in one additional figure to better understand the walking pattern?

l.191: Figure 2 nicely demonstrates the dbh and height estimation, especially with the noise inherent in the point cloud. Did the DBH estimation rely on circles, arcs or flexible splines, and was there any noise reduction performed before fitting? This question also connects to the major issue I) of describing the methods more in detail. 

l.216: there is still sample text from the results section in the manuscript, which has to be removed. 

l.236: Figure 4 uses three very similar colors for the various species (yellow, orange and red) and it is very difficult for me to tell them apart. Please use safe colors from a more contrasting scale (e.g. Viridis) and also change the "x" point sign to a species specific symbol (square, triangle, etc.) to also discriminate the species in the black and white version. Also the bars at the histogram below might follow the new color-coding of the species. 

l.271: the UAV derived heights were found to underestimate true tree height because of limitations in the raster-based CHM - why did you still decide for CHM raster height estimation and did not use direct height measurement by determining the highest point of the isolated tree point cloud?

l.279: the plot visualizations might be turned, to align the squares with the borders of the surrounding box (to avoid the black space at the edges). 

l.386: can you comment on the importance of scanning trajectory planning? spatial misalignments are the major cause of corrupt point-clouds, so do you think a certain walking protocol might improve future data acquisition campaigns?

 

Author Response

Comment 1 (Tree detection description: region-growing/accretion, seeds, semantics)

Answer 1. We expanded Methods (Section 2.4/2.5) to describe the accretion (region-growing) procedure: seeds are initialized at local apex candidates detected on a smoothed HAG surface (non-maximum suppression over a 3D neighborhood). Regions then grow vertically and radially under connectivity and curvature constraints, merging only when crown overlap and vertical continuity criteria are met. We clarify that our pipeline is geometric/structural, not semantic DL; points are classed via rule-based ground/off-ground and stem/crown filters within the VirtSilv workflow, with citation to Niță (2021).

 

Comment 2 (Slope and tree-base definition; manual corrections)

Answer 2. We now state that the tree base is defined at the intersection of the stem and the local DTM (HAG≈0) sampled under the stem axis. On slopes, base height is taken from the upslope/downslope median within a 0.25–0.5 m footprint to avoid bias. We note that ambiguous cases (buttresses, deadwood contact) were flagged and corrected manually using the MLS orthographic view; the number of corrections is reported. 

 

Comment 3 (Regression presentation: labels, 1:1 line, equal scaling, tests)

Answer 3. The regression panels were revised (now Figure 4): axes are relabeled DBH_reference vs DBH_MLS and H_reference vs H_platform; a 1:1 line is added; axes use equal scaling; we report slope, intercept, 95% CIs, and two-sided tests of slope=1 and intercept=0. RMSE, MAE, MBE, and R2  are shown in the insets.

 

Comment 4 (Occlusion by voxel count: MLS vs UAV crowns)

Answer 4. We added quantitative occlusion metrics per crown layer (HAG P70–P100): union, intersection, MLS-only, UAV-only voxel counts and ratios. For example, POS3_1 crown: MLS-only ≈ 0.970, UAV-only ≈ 0.030; POS3_2 crown: MLS-only ≈ 0.969, UAV-only ≈ 0.031. These are reported in Results 3.5 and summarized in the composite panels.

 

Comment 5 (References accuracy)

Answer 5. We reviewed and corrected citations for accuracy and relevance throughout, updated several placements, and ensured consistency between in-text citations and reference list (including year, venue, and DOI where applicable).

Comment 6 (DBH–age statement)

Answer 6. The Introduction was revised to clarify that DBH is not a direct measure of age; it is a proxy for size and vigor influenced by site conditions, competition, and history, while still critical for allometry and biomass estimation.

 

Comment 7 (Limiting factors wording)

Answer 7. We changed the phrasing at l.53 to emphasize labour and measurement skill as key constraints of field methods, in line with the reviewer’s suggestion.

Comment 8 (Schöps et al., 2015 citation format)

Answer 8. The citation was corrected for spelling/diacritics (Schöps), year, and venue; the reference list entry was updated accordingly.

Comment 9 (Height measurement replication: operators and positions)

Answer 9. We now state that a single operator acquired two replicate height measurements per tree from two azimuths at a fixed horizontal distance (maintained with the Vertex), averaging the two. Where line-of-sight was obstructed, the azimuth changed while holding the distance constant; this is noted in Section 2.2.

 

Comment 10 (Trajectory visualization)

Answer 10. We added a trajectory figure  (Figure 2 ).

 

Comment 11 (DBH fitting method and noise reduction)

Answer 11. Section 2.4 clarifies that DBH is computed from a 10–20 cm slice at 1.3 m HAG after statistical outlier removal (radius-based + z-score) and stem-class filtering. We fit a robust circle model using iteratively reweighted least squares (Huber) initialized by a RANSAC pre-fit; we report the equivalent circular diameter and provide quality flags (inlier ratio, residual RMSE). Irregular sections are handled by trimming and re-fitting.

 

Comment 12 (Remove leftover text)

Answer 12. The stray placeholder text at l.216 was removed.

 

Comment 13 (Colorblind-safe palette and symbols)

Answer 13. All plots now use colorblind-safe palettes (e.g., Viridis/Okabe-Ito). Species use distinct markers (e.g., circle = Picea, square = Fagus, triangle = Abies) and the same hues across panels; histograms follow the same scheme.

 

Comment 14 (CHM vs direct point-cloud height)

Answer 14. We justify the CHM approach for its operational reproducibility and wall-to-wall coverage. After co-registration with MLS, we also extract direct point-cloud maxima within 2 m buffers for sensitivity analysis; both approaches show the same bias trend under occlusion, with the point-based method slightly reducing smoothing artifacts. This rationale and comparison are added to Sections 2.3 and 3.3.

Comment 15 (Rotate plots to align with bounding box)

Answer 15. The plots are North oriented and we kept them as it is

Comment 16 (Walking protocol importance for MLS accuracy)

Answer 16. We added explicit discussion (Sections 4.3–4.4) on walking protocol: loop closures to bound drift, perimeter + zig-zag passes to maximize angular diversity around crowns, and speed control. We link these choices to observed improvements in apex detectability and reduced SLAM drift, and we provide trajectory guidelines as practical recommendations.

Reviewer 3 Report

Comments and Suggestions for Authors

General Comments:
This manuscript presents a comparative assessment of UAV-based LiDAR and Mobile Laser Scanning (MLS) for estimating tree height and diameter at breast height (DBH) in a structurally complex, high conservation value forest in Romania. The topic is timely and of potential practical value, particularly considering the increasing interest in remote sensing-based forest monitoring in difficult terrain. However, the current version of the manuscript has several significant issues that compromise the reliability and reproducibility of the findings. I outline below major concerns that should be addressed in a substantially revised version of the manuscript.

Major Comments:
1. Lack of explanation regarding data preprocessing and sample matching
The manuscript does not describe essential preprocessing steps such as outlier removal, DBH thresholding, handling of missing data, or the matching criteria for identifying the same trees across the two platforms (UAV vs. MLS). Given that UAV and MLS systems may detect different sets of trees due to occlusion or resolution differences, it is crucial to clarify how tree-level correspondence was established. Without this information, accuracy metrics such as RMSE and MAE cannot be interpreted reliably.

2. Concerning residual patterns in Figure 4 
The residual plots in Figure 4 show extreme prediction errors in both DBH (up to ±40 cm) and height (±10 m), suggesting possible issues with field measurement accuracy, object matching, or sensor limitations. Such substantial residuals warrant further investigation and discussion. However, the manuscript does not acknowledge or attempt to explain these large deviations.

3. Inadequate description of modeling approach and software environment
The manuscript vaguely refers to “regression analysis” but does not provide details about the type of model used (e.g., linear, nonlinear), variables included, or statistical settings. It also lacks information on the computational tools or software used. This significantly hinders reproducibility and transparency of the analysis.

4. Absence of summary tables for key results
All numerical results are presented only in figures, with no summary tables comparing performance metrics across platforms. A concise table showing RMSE, MAE, and bias by platform and species would help readers better interpret and compare the results.

5. Potential use of automated writing tools
While not conclusive, the manuscript’s writing style—characterized by short repetitive paragraphs, generic transitions, and stylistic uniformity—suggests the possibility that large language model (LLM)-based tools may have been used during drafting. MDPI requires authors to clearly disclose the use of AI-assisted writing tools in the acknowledgments or disclosure sections. If such tools were used, the authors should transparently specify the tool and its extent of use. If not, some parts of the text would benefit from more editorial refinement to enhance natural flow and reduce redundancy.

6. Limited references and lack of recent literature
The reference list is relatively short given the breadth of the topic. Important recent studies on UAV and MLS-based forest inventory (especially from 2022 onward) are missing. The paper would benefit from a more comprehensive literature review and broader contextualization of its contribution.

Author Response

Comment 1 (Data preprocessing and tree matching across platforms)

Answer 1. We added a dedicated subsection in Methods (Sections 2.4–2.5) detailing the full preprocessing pipeline and matching criteria. 

Comment 2 (Large residuals in Figure 4)

Answer 2. We investigated the extreme residuals and now report causes and mitigations. DBH outliers (up to ±40 cm) were mistakes that were corrected mo height outliers (±10 m) arose from UAV occlusion in layered crowns or mis-identified apices in dense MLS crowns. We comment in the text

Comment 3 (Modeling approach and software environment)

Answer 3. We now specify models and tools. For platform–field comparisons we used ordinary least squares (OLS) with equal-axis plots, and as sensitivity we added Deming regression (errors-in-variables; variance ratio from empirical residual variances). We report slope, intercept, 95% CIs, and two-sided tests of slope = 1 and intercept = 0, alongside R2 , RMSE, MAE, and MBE. Analyses were run in Python 3.11 using statsmodels 0.14, scikit-learn 1.4, numpy/pandas, with figures in matplotlib; point-cloud processing used CloudCompare (ICP) and LAStools as stated. Package versions are listed in the methods

Comment 4 (Provide summary tables for key results)

Answer 4. We added concise tablenTable 2 (main text): RMSE/MAE/MBE and slope/intercept (with CIs) by platform × species.

Comment 5 (Potential use of automated writing tools)

Answer 5. We revised the prose for clarity and removed repetitive phrasing. We also added an AI-use disclosure in the Acknowledgments stating that language assistance tools were used for editing (not for data analysis or interpretation) and that the authors take full responsibility for the content, in accordance with MDPI guidance.

Comment 6 (Expand recent literature and context)

Answer 6. We expanded the literature review (Introduction and Discussion) to include recent work on: (i) UAV–MLS comparative inventory performance, (ii) CHM bias and penetration limits under leaf-on conditions, (iii) SLAM robustness and trajectory design in forests, and (iv) HAG-normalized cross-platform analyses and data fusion. These additions are integrated where we discuss motivation, methods (occlusion/HAG), and practical recommendations. The new citations are marked in the revised manuscript and reference list.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The corrections and responses to the comments are appreciated. However, several comments from the first round of review remain unaddressed and should be resolved before publication.

Innovation and motivation

It is explained that: “While it is well established that UAV LiDAR generally outperforms ground-based approaches for capturing upper canopy structure, and that MLS excels in detailed stem mapping [27–29], most comparisons have been limited to simplified or single-layered forest types where occlusion and structural complexity are minimal.” While UAV LiDAR has been shown to be effective in capturing the top layers of canopy in simplified or single-layered forests, as mentioned by the authors, it is not clearly evident that this hypothesis holds true for more complex stands. For readers, this result is already known, as noted previously, and the explanations provided are not convincing.

Species selection rationale

It was explained that: “Their differing crown architectures, branching densities, and bark reflectance properties influence LiDAR signal interaction, making them ideal for testing sensor performance under varied structural conditions.” The authors refered to Calders et al. (2020); however, this paper did not provide information explaining why the differences of the three species make them ideal for such comparison. Please doublecheck the reference. Furthermore, if these species are only important at the national scale, the study might have limited relevance or interest at the global scale.

Sample Size and Tree Size Distribution

Increasing the number of observations and including plots showing the size distribution is appreciated; however, the number of observations appear to be very low in some height and DBH classes (as low as one observation). This significantly impacts the reliability of the findings.

Point Cloud Segmentation:

There is no much change in the segmentation section, and it still is a black box that requires clarification. The explanations in the Methodology should provide enough details to allow readers to reproduce the results.

Author Response

Comment:
Innovation and motivation

It is explained that: “While it is well established that UAV LiDAR generally outperforms ground-based approaches for capturing upper canopy structure, and that MLS excels in detailed stem mapping [27–29], most comparisons have been limited to simplified or single-layered forest types where occlusion and structural complexity are minimal.” While UAV LiDAR has been shown to be effective in capturing the top layers of canopy in simplified or single-layered forests, as mentioned by the authors, it is not clearly evident that this hypothesis holds true for more complex stands. For readers, this result is already known, as noted previously, and the explanations provided are not convincing.

Answer:
We appreciate the reviewer’s concern regarding the need to better highlight the novelty and motivation for our study. In section 4.1 the supporting references on the challenges of UAV and MLS performance forests (Peng et al., 2020; Caroti et al., 2015; Liu et al., 2018; Neuville et al., 2021; Gao et al., 2022) are strengthening the narrative. Specifically, we now explicitly emphasize that while previous studies have shown UAV LiDAR efficiency for canopy mapping and MLS accuracy for stem detection, the majority of comparisons were performed in plantations or simplified, single-layered stands with minimal vertical occlusion. In contrast, our work provides a quantitative evaluation in multi-layered, uneven-aged, high conservation value forests under leaf-on conditions and steep terrain, conditions rarely examined in prior research. We have also clarified that the voxel-based occlusion analysis presented here offers new quantitative insights into how canopy closure and vertical layering govern platform complementarity in operational settings, extending beyond what is available in the current literature.

Comment: Species selection rationale

It was explained that: “Their differing crown architectures, branching densities, and bark reflectance properties influence LiDAR signal interaction, making them ideal for testing sensor performance under varied structural conditions.” The authors refered to Calders et al. (2020); however, this paper did not provide information explaining why the differences of the three species make them ideal for such comparison. Please doublecheck the reference. Furthermore, if these species are only important at the national scale, the study might have limited relevance or interest at the global scale.

Answer: We thank the reviewer for this observation. We have kept Calders et al. (2020) because it specifically addresses the use of LiDAR in structurally complex ecosystems, which aligns with the focus of our study. However, we have revised the paragraph in Section 2.1 to clarify both the structural contrasts among the three selected species and the broader ecological relevance beyond the national scale. In addition, we have added several references to support the discussion on species morphology, crown architecture, and LiDAR interactions (e.g., Pretzsch et al., 2014; Paluch et al., 2021; LaRue et al., 2022; Assmann et al., 2025).

Comment: Sample Size and Tree Size Distribution

Increasing the number of observations and including plots showing the size distribution is appreciated; however, the number of observations appear to be very low in some height and DBH classes (as low as one observation). This significantly impacts the reliability of the findings.

Answer:
We thank the reviewer for this observation. In the revised manuscript, we have retained the expanded dataset and the plots showing DBH and height distributions but also acknowledge that some extreme size classes remain underrepresented. To address this concern, we added a sentence in the Sources of Uncertainty and Limitations section explicitly noting that low sample sizes in some DBH and height classes may reduce the statistical robustness of the results in these ranges. We also highlight that future studies should include additional plots, multi-season acquisitions, and broader structural gradients to better represent rarely occurring size classes and improve statistical reliability.

Comment: Point Cloud Segmentation:

There is no much change in the segmentation section, and it still is a black box that requires clarification. The explanations in the Methodology should provide enough details to allow readers to reproduce the results.

Answer:
We thank the reviewer for highlighting the need for greater clarity in the segmentation workflow. In response, we have adapted the text in the Methods section to describe each processing stage in more detail. 

We have also modified Figure 3 to visually illustrate the entire VirtSilv pipeline—from raw MLS scans to individual tree-level digital twins—so that readers can clearly follow each step of the process. Together, these textual revisions and the updated figure ensure that the segmentation and modeling workflow is now fully transparent and reproducible using a python workflow or even a cloudcompare procedure.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors,

after carefully reviewing your revised manuscript and the changes you have made, I would like to thank you for thoroughly addressing all of my comments and for making significant improvements to the respective sections. The revisions have greatly enhanced the clarity and coherence of your work.

In particular, the results are now much better supported by the methodological workflow you propose, which strengthens the overall validity of your findings. 

Once again, I would like to commend you for the outstanding quality of the graphics. The additional figures further enhance the clarity and completeness of the study.

Your findings make a valuable contribution to advancing our understanding of the benefits of merged point clouds, and I am confident that this work will be of interest to the research community.

Congratulations on that well presented study!

Author Response

Dear Reviewer,

We sincerely thank you for your kind words and for acknowledging the revisions we made to the manuscript. We are delighted that the changes have improved the clarity, coherence, and overall quality of the study, and that the methodological workflow and graphical elements now better support the results and conclusions.

We greatly appreciate your recognition of our efforts to address all comments thoroughly. Your constructive feedback has been instrumental in strengthening the manuscript and ensuring it makes a meaningful contribution to understanding the benefits of integrating UAV and MLS point clouds for forest structure characterization.

Thank you once again for your valuable time and support throughout the review process.

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript compares the performance of UAV LiDAR and Mobile Laser Scanning (MLS) for estimating tree height and DBH in structurally complex natural forests. The study addresses a relevant and timely question, and the design and analyses are generally sound. However, there are several critical issues that must be addressed before the manuscript can be considered for publication.

Comments

1. Missing Figure 10
Sections 3.5 and 4 refer repeatedly to Figure 10 (Occlusion mapping and voxel statistics), yet the figure itself is not included in the manuscript. As occlusion analysis is one of the central contributions of this study, the absence of this figure represents a serious deficiency. The figure must be provided, clearly visualizing UAV-only, MLS-only, and overlapping voxel volumes, along with density statistics.

2. Conclusions Too Brief
The conclusion (Section 5) is overly short and does not sufficiently reflect the study’s objectives and findings. Simply stating that “MLS performs best for DBH and UAV for canopy” underrepresents the contribution of the work. The conclusions should explicitly tie back to the three stated objectives (accuracy and bias quantification, species/structural effects, and occlusion limitations), and highlight broader implications for biomass and carbon estimation, conservation monitoring, and forest management. Practical recommendations for hybrid UAV–MLS workflows and directions for future research (e.g., multi-season data, improved SLAM robustness, automated data fusion) should also be included.

3. Excessive Self-Citation
The reference list contains at least seven papers authored or co-authored by the corresponding author (Niță). This constitutes a relatively high proportion of the total citations and creates the impression of over-reliance on self-citation. Please retain only the most essential self-references and incorporate additional independent, international studies to provide a more balanced scholarly context.

4. Absence of UAV DBH Estimation
The manuscript states that UAV LiDAR cannot be used for DBH estimation. While this is broadly true, there have been recent attempts at UAV-LiDAR-based stem detection. Relevant literature should be acknowledged, and the distinction between those approaches and the present study should be clearly articulated.

5. Lack of Statistical Significance Testing
Currently, only descriptive metrics such as R² and RMSE are reported. To substantiate claims of performance differences between platforms, statistical significance testing (e.g., paired t-test, Wilcoxon signed-rank test) should be conducted. This would considerably strengthen the persuasiveness of the results.

Author Response

General note:
We thank the reviewers for their constructive comments. In response, we have revised the manuscript to address all concerns. Below we provide a point-by-point response, indicating where changes were made and what new analyses, figures, or clarifications were added.

1. Missing Figure 10

Comment: Sections 3.5 and 4 refer repeatedly to Figure 10 (Occlusion mapping and voxel statistics), yet the figure itself is not included in the manuscript.

Response:
We apologize for this oversight, the figure was in the tracked changes manuscript pdf format. Now Figure 10 has now been included in the revised manuscript as accepted change. It clearly visualizes UAV-only, MLS-only, and overlapping voxel volumes across vertical forest layers, along with point density statistics, as described in Sections 3.5 and 4. The figure caption and text were updated for clarity.

2. Conclusions Too Brief

Comment: The conclusion was too short and did not reflect the study’s objectives or broader implications.

Response:
We have completely rewritten Section 5 (Conclusions) to explicitly tie back to the three objectives—accuracy and bias quantification, species/structural effects, and occlusion limitations. The revised conclusion now: Summarizes the main findings for UAV and MLS performance; Highlights implications for biomass and carbon estimation, habitat monitoring, and forest management;
Provides practical recommendations for hybrid UAV–MLS workflows; Suggests directions for future research, including multi-season acquisitions, improved SLAM robustness, and automated UAV–MLS data fusion.

3. Excessive Self-Citation

Comment: The reference list contained too many self-citations.

Response:
We have reduced the self-citation rate from 14% to under 10% by removing non-essential self-references and adding several independent international studies on UAV LiDAR, MLS, and forest structure mapping. The reference list is now more balanced and reflects a broader body of work.

4. Absence of UAV DBH Estimation

Comment: UAV LiDAR–based stem detection has been attempted in other studies and should be acknowledged.

Response:
We now explicitly acknowledge recent studies on UAV-based DBH estimation in Section 2.3 (e.g., Ye, N.; Mason, E.; Xu, C.; Morgenroth, J. Estimating Individual Tree DBH and Biomass of Durable Eucalyptus Using UAV LiDAR. Ecol Inform 2025, 89, 103169, doi:10.1016/J.ECOINF.2025.103169.     Moreira, B.M.; Goyanes, G.; Pina, P.; Vassilev, O.; Heleno, S. Assessment of the Influence of Survey Design and Processing Choices on the Accuracy of Tree Diameter at Breast Height (DBH) Measurements Using UAV-Based Photogrammetry. Drones 2021, Vol. 5, Page 43 2021, 5, 43, doi:10.3390/DRONES5020043.). We clarify that while UAV LiDAR stem detection is possible under optimized flight conditions and with specialized algorithms, our study focused on operational, leaf-on conditions in dense, multi-layered forests where UAV point densities and canopy occlusion strongly limit DBH estimation. In contrast, MLS provided inventory-grade DBH accuracy under these realistic conditions.

5. Lack of Statistical Significance Testing

Comment: The manuscript reports only descriptive statistics (R², RMSE) without significance testing.

Response:
We clarify in Section 3.2 that Kruskal–Wallis tests were used to compare MLS and UAV errors across plots, with p-values now explicitly reported in the Results. The tests confirmed significant differences in height errors between plot types (p < 0.05), while DBH errors showed no significant differences (p > 0.1). Together with R² and RMSE, these results sufficiently demonstrate performance differences without requiring redundant tests.