Advancing Tree Species Classification with Multi-Temporal UAV Imagery, GEOBIA, and Machine Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper discusses and analyzes tree detection in UAV images using four machine learning algorithms. Overall, the paper is well written and the experiments are sufficient. There are some issues that need to be addressed carefully.

1. Currently, a large amount of research has been conducted on tree detection in remote sensing images based on deep learning, but this paper does not mention it.
2. Some recent research literature on tree detection using UAV imagery should be added to the introduction section.
3. The paper needs to design some tree features for image classification. Deep learning algorithms can autonomously learn tree features. What are the advantages of traditional machine learning algorithms over deep learning? The paper should clarify this.
4. The research area in the paper is relatively simple overall. Could other more complex areas be added to test and validate the proposed method, thereby further verifying the reliability of the paper's analysis results?

Author Response

-Comment 01- Currently, a large amount of research has been conducted on tree detection in remote sensing images based on deep learning, but this paper does not mention it.

Response: We thank the reviewer for highlighting the extensive research on deep learning (DL) for tree detection in remote sensing, which our paper did not initially address. We agree this is a valuable addition for context. We have revised the Introduction to acknowledge this body of work, citing recent reviews and studies on DL methods like CNNs for UAV/remote sensing classification. We also justify our focus on traditional ML for its interpretability and suitability to our smaller, phenologically focused dataset, complementing DL approaches. These citations have been added to the reference list.

 

-Comment 02- Some recent research literature on tree detection using UAV imagery should be added to the introduction section.

Response: Thank for pointing out the need to include recent literature on tree detection using UAV imagery. We have revised the Introduction to acknowledge this, citing recent studies on UAV base tree classification. This addition provides context for our multi-temporal approach and has been integrated after the UAV advantages discussion. The new references have been added to the bibliography.

 

-Comment 03- The paper needs to design some tree features for image classification. Deep learning algorithms can autonomously learn tree features. What are the advantages of traditional machine learning algorithms over deep learning? The paper should clarify this.

Response: We agree that DL's autonomous feature learning is a strength, but traditional ML provides advantages in interpretability, computational efficiency, and performance on smaller datasets like ours. We have revised the Introduction to clarify this, emphasizing our designed tree features (spectral, textural, geometrical, canopy height via GEOBIA) and explicitly stating ML's benefits over DL, while noting how our approach complements these strengths for phenological UAV data.

 

-Comment 04- The research area in the paper is relatively simple overall. Could other more complex areas be added to test and validate the proposed method, thereby further verifying the reliability of the paper's analysis results?

Response: Thank the reviewer for noting the relative simplicity of our study area and suggesting validation in more complex environments to enhance reliability. While the botanical garden provided a dense, phenologically diverse, we agree broader testing is valuable. We have revised the Conclusions section to explicitly propose future applications in complex areas. This addition strengthens the discussion of generalizability without altering our current findings.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper presents a novel framework for tree species classification using multi-temporal UAV imagery, GEOBIA (Geographic Object-Based Image Analysis), and machine learning. The study was conducted in the botanical garden of the University of the Punjab, utilizing high-resolution (2.3 cm/pixel) imagery captured across four seasons. Four machine learning models Random Forest (RF), Extra Trees (ET), XGBoost, and Support Vector Machine (SVM) were evaluated, with RF achieving the highest overall accuracy (86%). The results highlight the importance of multi-temporal data and features like canopy height and texture in improving classification accuracy. This approach holds promise for applications in forest management and biodiversity conservation.

(Lines 300-330, Results Section): The paper reports impressive classification accuracy (86% OA with RF) but fails to rigorously address *class imbalance, a critical issue given the uneven sample sizes (e.g., *Jatropha integerrima = 39 samples vs. Melia azadarach = 15 samples). SMOTE oversampling is mentioned (Line 255), but no metrics (e.g., per-class F1-scores, precision-recall curves) are provided to prove its effectiveness. The confusion matrices (Figure 8) hint at misclassification biases (e.g., Millettia ovalifolia performs poorly), but the paper lacks a quantitative imbalance analysis* (e.g., Matthews Correlation Coefficient per class).  I wpould suggest authors to add Class-Wise Metrics include a table of precision/recall/F1 for each species to expose bias (e.g., is RF’s 86% OA driven by overfitting majority classes?). 

Discuss Limitations: Explicitly state how imbalance affects generalizability (e.g., rare species may be misclassified in real-world forests). 

(Lines 220-240): The paper combines spectral, textural, and geometric features but doesn’t evaluate multicollinearity (e.g., Pearson correlation heatmap). Redundant features (e.g., GLCM homogeneity vs. GLDV angular moment) may inflate model complexity without improving accuracy. The authors might run a feature importance analysis with SHAP values or permutation importance to identify truly discriminative features. 

(Lines 270-290): RF’s hyperparameters (e.g., n_estimators=100, max_features="sqrt") are standard but not optimized. A grid search or Bayesian optimization could push accuracy further. 

UAV Data Limitations (Lines 120-140): The 2.3 cm/pixel resolution is excellent, but no discussion on *occlusion effects* (e.g., overlapping canopies in dense areas) or radiometric consistency across seasons (e.g., sunlight angle changes). 

 

 

 

 

Lines 1-10 (Title and Authors): The title is clear and descriptive, but "Adevancing" should be corrected to "Advancing" (typo). The list of authors and affiliations is well-organized.

Lines 20-30 (Abstract): The abstract comprehensively covers the objectives, methods, and key findings. It would be beneficial to briefly mention the study's limitations, such as sample size or geographical constraints.

Lines 100-110 (Study Area): The description of the study area and selection criteria is adequate, but the map (Figure 1) lacks clarity and detail.

Lines 150-160 (Reference Data): Table 2 effectively summarizes species distribution, but more details on the sampling method (e.g., number of points per species) would be helpful.

Lines 200-210 (Tree Crown Extraction): The segmentation parameters (scale, compactness, shape) are well-explained, but their impact on the results could be further discussed.

Lines 250-260 (Machine Learning Models): The descriptions of the models (RF, ET, XGBoost, SVM) are sufficient, but the performance comparison in the results section (Lines 300-310) could be enhanced with clearer visualizations.

Lines 350-360 (Accuracy Assessment): Accuracy metrics (OA, F1-score, Kappa) are appropriately used, but individual confusion matrices for each model would improve clarity.

Lines 400-410 (Discussion): The discussion of results is robust, but addressing practical challenges (e.g., computational costs or cloud cover effects) would add depth.

Lines 450-460 (Conclusion): The conclusion is comprehensive, but suggestions for future research (e.g., integrating LiDAR data) could be expanded.

Lines 500-510 (References): The reference list is thorough and relevant but including more recent sources (2024-2025) would strengthen the paper.

 

Author Response

-(Lines 300-330, Results Section): The paper reports impressive classification accuracy (86% OA with RF) but fails to rigorously address *class imbalance, a critical issue given the uneven sample sizes (e.g., *Jatropha integerrima = 39 samples vs. Melia azadarach = 15 samples). SMOTE oversampling is mentioned (Line 255), but no metrics (e.g., per-class F1-scores, precision-recall curves) are provided to prove its effectiveness. The confusion matrices (Figure 8) hint at misclassification biases (e.g., Millettia ovalifolia performs poorly), but the paper lacks a quantitative imbalance analysis* (e.g., Matthews Correlation Coefficient per class).  I would suggest authors to add Class-Wise Metrics include a table of precision/recall/F1 for each species to expose bias (e.g., is RF’s 86% OA driven by overfitting majority classes?).

Response: We appreciate the reviewer's feedback on the need for a more rigorous treatment of class imbalance in the results section (lines 300-330). While our overall accuracy (OA) of 86% with RF is encouraging, we recognize the uneven sample sizes as a potential source of bias. To demonstrate SMOTE's effectiveness in mitigating this, we have added Table 6 with per-class precision, recall, and F1-scores for the RF model on multi-temporal. These metrics show balanced performance overall, with no evidence of major overfitting to majority classes though minority classes. We have also expanded the text to quantitatively discuss this, enhancing the analysis without altering core findings.

 

-Discuss Limitations: Explicitly state how imbalance affects generalizability (e.g., rare species may be misclassified in real-world forests).

Response: Thank for highlighting the need to discuss class imbalance's impact on generalizability. We have expanded the limitations section in the Conclusions to explicitly state this, noting that imbalance may lead to misclassifications in diverse real-world forests. This addition provides a more rigorous acknowledgment of the issue.

 

-(Lines 220-240): The paper combines spectral, textural, and geometric features but doesn’t evaluate multicollinearity (e.g., Pearson correlation heatmap). Redundant features (e.g., GLCM homogeneity vs. GLDV angular moment) may inflate model complexity without improving accuracy. The authors might run a feature importance analysis with SHAP values or permutation importance to identify truly discriminative features.

Response: We thank the reviewer for noting the potential for multicollinearity and redundancy in features. To rigorously address this, we have revised the section 2.8 to include SHAP analysis, with Fig 9.

 

-(Lines 270-290): RF’s hyperparameters (e.g., n_estimators=100, max_features="sqrt") are standard but not optimized. A grid search or Bayesian optimization could push accuracy further.

Response: We thank the reviewer for suggesting hyperparameter optimization in section 2.6 (lines 270-290). In our analysis, we used PyCaret's grid search (tune_model) to optimize RF parameters, resulting in the reported values. We have revised the section to explicitly describe this process and its impact.

 

-UAV Data Limitations (Lines 120-140): The 2.3 cm/pixel resolution is excellent, but no discussion on *occlusion effects* (e.g., overlapping canopies in dense areas) or radiometric consistency across seasons (e.g., sunlight angle changes).

Response: Thank for complimenting the 2.3 cm/pixel resolution and for pointing out the need to discuss UAV data limitations in section 2.3 (lines 120-140). We have revised the paragraph to explicitly address occlusion effects from overlapping canopies in dense areas and potential radiometric inconsistencies due to seasonal sunlight angle variations, despite consistent acquisition protocols. This enhances methodological transparency without altering our core findings.

 

-Lines 1-10 (Title and Authors): The title is clear and descriptive, but "Adevancing" should be corrected to "Advancing" (typo). The list of authors and affiliations is well-organized.

Response: We have corrected this typo error

 

-Lines 20-30 (Abstract): The abstract comprehensively covers the objectives, methods, and key findings. It would be beneficial to briefly mention the study's limitations, such as sample size or geographical constraints.

Response: We have revised the abstract to include some key constraints, such as the moderate sample size and geographical focus on a controlled botanical garden environment, which may limit generalizability to natural forests.

 

-Lines 100-110 (Study Area): The description of the study area and selection criteria is adequate, but the map (Figure 1) lacks clarity and detail.

Response: We have updated the map, the revised figure has been incorporated, improving visual representation without altering the text description

 

-Lines 150-160 (Reference Data): Table 2 effectively summarizes species distribution, but more details on the sampling method (e.g., number of points per species) would be helpful.

Response: Our field survey was a comprehensive inventory of identifiable trees across the garden's sections, rather than a random stratified sample. We have revised section 2.4 to describe it as such, detailing the systematic approach (including use of canopy height and tree diameter for guiding identification and excluding very small trees during crown extraction).

 

-Lines 200-210 (Tree Crown Extraction): The segmentation parameters (scale, compactness, shape) are well-explained, but their impact on the results could be further discussed.

Response: We have revised the paragraph to discuss their impact on results.

 

-Lines 250-260 (Machine Learning Models): The descriptions of the models (RF, ET, XGBoost, SVM) are sufficient, but the performance comparison in the results section (Lines 300-310) could be enhanced with clearer visualizations.

Response: We have updated Figure 7 to include a comparing OA, F1-score, and kappa across models on multi-temporal data, providing clearer at-a-glance insights into relative performances (e.g., RF's lead).

 

-Lines 350-360 (Accuracy Assessment): Accuracy metrics (OA, F1-score, Kappa) are appropriately used, but individual confusion matrices for each model would improve clarity.

Response: We thank the reviewer for affirming the appropriate use of accuracy metrics (OA, F1-score, kappa) in section 2.8 (lines 350-360). Figure 8 already provides separate matrices for each model (RF, ET, XGBoost, SVM) on multi-temporal data.

 

-Lines 400-410 (Discussion): The discussion of results is robust, but addressing practical challenges (e.g., computational costs or cloud cover effects) would add depth.

Response: We have revised the Discussion section to explicitly discuss computational costs and cloud cover effects, integrating these into the limitations for a more comprehensive analysis.

 

-Lines 450-460 (Conclusion): The conclusion is comprehensive, but suggestions for future research (e.g., integrating LiDAR data) could be expanded.

Response: We have revised the section to elaborate on integrating LiDAR, while adding complementary suggestions like hyperspectral sensors, enhancing the forward-looking aspect without introducing new findings

 

-Lines 500-510 (References): The reference list is thorough and relevant but including more recent sources (2024-2025) would strengthen the paper.

Response: We thank the reviewer for affirming the thoroughness and relevance of our reference list and suggesting more recent sources. We have strengthened it by adding some latest studies on UAV-based tree species classification, added to the bibliography, enhancing the paper strength without altering existing content

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have well addressed all my comments. I have no more concerns.