In this study, we present a comprehensive investigation into the classification of tree species from UAV drone laser scanning point clouds, addressing the fundamental challenge of vertical density heterogeneity that characterizes individual-tree point data. Our work is motivated by the observation that conventional deep learning frameworks, when applied to UAV drone point clouds, tend to disproportionately sample dense canopy regions while underrepresenting structurally critical but sparse areas such as stems and lower crowns. To overcome this limitation, we propose TreeConT, a Transformer-based architecture that explicitly incorporates vertical structure awareness into the feature learning pipeline. The core contributions of our work are twofold: first, we introduce a novel progressive down-sampling strategy called Tree-FPS that adaptively balances sampling quotas across vertical strata based on inverse density weighting; second, we enhance the input representation by integrating geometric descriptors with LiDAR intensity data to capture complementary shape and material properties. Through extensive experiments on a large-scale UAV drone dataset comprising 18,749 individual trees across 11 species, we demonstrate that TreeConT achieves superior classification performance, with an overall accuracy of 93.12% and a Kappa coefficient of 0.9207, outperforming strong baselines including PointNet++, DGCNN, PointTransformer, and the baseline PointConT. Our ablation studies confirm that both the Tree-FPS sampling strategy and the multi-feature fusion contribute significantly to the observed improvements, particularly for challenging species pairs with similar crown morphologies.
Our research is situated within the broader context of forest ecology and remote sensing, where accurate species-level classification is essential for biodiversity assessment, carbon stock estimation, and sustainable forest management. UAV drone laser scanning, with its ability to capture fine-scale three-dimensional structural information, has emerged as a powerful tool for individual-tree analysis. However, the inherent vertical density bias in UAV drone point clouds—where canopy regions are densely sampled due to sensor viewing geometry and occlusion effects—poses a significant challenge for deep learning models that rely on uniform sampling assumptions. Traditional farthest point sampling, while effective for general point cloud tasks, exacerbates this bias by preferentially selecting points from high-density regions during progressive down-sampling. Our Tree-FPS strategy directly addresses this issue by introducing a principled mechanism for vertical structure preservation.
The remainder of this article is organized as follows. We first review the relevant literature on tree species classification using UAV drone data, highlighting the limitations of existing approaches. We then present our methodology in detail, including the Tree-FPS sampling algorithm, the geometric and intensity feature extraction, and the overall network architecture. Next, we describe the experimental setup, including the dataset, implementation details, and evaluation metrics. We then present and discuss the experimental results, including comparisons with state-of-the-art methods and ablation studies. Finally, we conclude with a summary of our contributions and directions for future research.
Related Work and Problem Formulation
The classification of tree species from remote sensing data has been extensively studied, with approaches broadly categorized into feature engineering methods and deep learning methods. Feature engineering approaches typically rely on hand-crafted structural parameters or spectral features extracted from point clouds or imagery. For UAV drone laser scanning data, common features include tree height, crown diameter, branching angle, and leaf cluster size. While these features provide interpretable representations, they often fail to capture the full complexity of three-dimensional tree structure, particularly in mixed forests with high structural diversity.
Deep learning methods, in contrast, learn hierarchical features directly from raw data, reducing the reliance on manual feature design. PointNet pioneered the direct processing of unordered point clouds, but its limited capacity for local structure modeling led to the development of more sophisticated architectures. PointNet++ introduced hierarchical feature learning through multi-scale grouping, while DGCNN leveraged graph convolutions to capture neighborhood relationships. More recently, Transformer-based architectures have been applied to point cloud processing, with models like PointTransformer and PCT demonstrating strong performance on various benchmarks.
Despite these advances, existing deep learning methods face two key challenges when applied to UAV drone point clouds of individual trees. First, the vertical density heterogeneity—where canopy regions are densely sampled while stems and lower crowns are sparse—causes sampling bias that propagates through network stages, weakening the representation of structurally important low-density regions. Second, the reliance on spatial coordinates alone limits the network’s ability to discriminate between species with similar crown morphologies but different stem or branch structures. Our work addresses both challenges through targeted architectural innovations.
Methodology
Baseline Architecture: PointConT
Our TreeConT network builds upon the PointConT architecture, which combines Inception-style feature aggregation with content-based attention mechanisms. The baseline PointConT consists of five sequential Inception Feature Aggregator stages, each followed by a down-sampling operation that halves the number of points while doubling the feature dimension. Given an input point cloud $\mathbf{P} \in \mathbb{R}^{N \times 3}$, where $N$ is the number of points, the network first partitions the points into overlapping local patches and maps them from 3D space to a $C$-dimensional feature space. After $M$ stages, the output contains $N/2^m$ points with feature dimension $2^{m-1}C$. A global max pooling layer then aggregates the features into a global vector, which is fed into a linear classifier for species prediction.
The Inception Feature Aggregator module is designed to capture both high-frequency and low-frequency information through parallel branches. The high-frequency branch employs max pooling to extract salient local features, followed by a residual MLP for further enhancement. The low-frequency branch uses average pooling to suppress local noise, followed by content-based attention for global context modeling. The two branches are concatenated and processed by an MLP to produce the output features. This dual-branch design enables the network to simultaneously capture fine-grained geometric details and coarse structural patterns, which is particularly beneficial for tree species classification where both local branch structures and global crown shapes are discriminative.
The content-based attention mechanism, a key component of PointConT, performs clustering in feature space to group semantically similar points before applying self-attention within each cluster. This reduces computational complexity while enhancing local semantic consistency. Given an input feature $\mathbf{X} \in \mathbb{R}^{S \times D}$, where $S$ is the sequence length and $D$ is the feature dimension, the module computes Query, Key, and Value matrices through linear transformations:
$$
\mathbf{Q} = \mathbf{X}\mathbf{W}_Q, \quad \mathbf{K} = \mathbf{X}\mathbf{W}_K, \quad \mathbf{V} = \mathbf{X}\mathbf{W}_V
$$
where $\mathbf{W}_Q$, $\mathbf{W}_K$, and $\mathbf{W}_V$ are learnable weight matrices. The points are then partitioned into $L$ clusters using binary clustering, and self-attention is computed independently within each cluster:
$$
\mathbf{Y}_i = \text{Softmax}\left(\frac{\mathbf{Q}_i \mathbf{K}_i^\top}{\sqrt{D}}\right) \odot \mathbf{V}_i
$$
The cluster outputs are concatenated to form the final attention result. This mechanism provides a good balance between computational efficiency and representational power, making it suitable for large-scale point cloud processing.
Tree-FPS: Vertical Structure-Aware Sampling
The central contribution of our work is the Tree-FPS sampling strategy, which replaces the standard farthest point sampling in the progressive down-sampling stages of the network. Standard FPS iteratively selects points that are farthest from the already selected set, resulting in approximately uniform spatial coverage. While this works well for general point cloud tasks, it introduces significant bias for UAV drone point clouds of trees, where canopy regions have much higher point density than stem regions. As a result, the down-sampled point set becomes dominated by canopy points, and the stem structure—which is critical for species discrimination—is progressively weakened across network stages.
Tree-FPS addresses this issue through a three-step process: vertical stratification, adaptive quota allocation, and stratified local sampling. Given an input point cloud $\mathbf{P} = \{p_m = (x_m, y_m, z_m) | m = 1, 2, \ldots, n\}$, we first compute the vertical range $[z_{\min}, z_{\max}]$ and divide it into $k$ equal-height intervals with layer thickness $\Delta h = (z_{\max} – z_{\min}) / k$. The number of layers $k$ is a hyperparameter that balances structural resolution and sampling stability. Based on our empirical analysis, we set $k = 50$ as the default, which provides sufficient vertical granularity while ensuring that each layer contains enough points for local FPS.
For each height interval $i$, we count the number of points $n_i$ as a measure of relative point density. To allocate sampling quotas, we introduce an inverse-density weighting scheme that assigns higher weights to sparser layers:
$$
w_i = \lambda \cdot \frac{1}{(n_i + \epsilon)^\alpha} + (1 – \lambda)
$$
where $\epsilon$ is a small stability constant to avoid numerical issues when $n_i = 0$, $\alpha$ is a decay parameter that controls the strength of density compensation, and $\lambda \in [0, 1]$ is a mixing factor that ensures a minimum sampling quota for high-density layers. The weight function is designed to be robust to extreme density variations: when $n_i$ is very small, the weight approaches $\lambda/\epsilon^\alpha + (1-\lambda)$, which is bounded; when $n_i$ is very large, the weight approaches $(1-\lambda)$, which is also bounded. This prevents the sampling from becoming too aggressive in either direction.
The initial sampling quota for each layer is then computed as:
$$
N_i^{(1)} = \left\lfloor N \times \frac{w_i}{\sum_{j=1}^{k} w_j} \right\rfloor, \quad \text{with } N_i^{(1)} \leq n_i
$$
where $N$ is the total desired number of sampled points. We then perform local FPS within each layer to select $N_i^{(1)}$ points. If the total number of selected points is less than $N$, we repeat the process on the remaining points:
$$
\mathbf{N} = \bigcup_{l=1}^{L} \mathbf{N}^{(l)}
$$
where $\mathbf{N}^{(l)}$ is the set of points selected in the $l$-th iteration, and the process continues until $|\mathbf{N}| = N$. This iterative refinement ensures that the final sampled set accurately reflects the target distribution.
The key insight behind Tree-FPS is that low-density regions, such as stems and lower branches, contain information that is disproportionately important for species classification relative to their point count. By allocating more samples to these regions, we ensure that the network maintains access to critical structural cues throughout the feature abstraction process. The inverse-density weighting scheme is designed to be adaptive: when density differences are extreme, the compensation is stronger; when density is more uniform, the weights become more balanced, and the sampling approaches standard FPS.
Geometric and Intensity Feature Fusion
In addition to the sampling strategy, we enhance the input representation by fusing spatial coordinates with geometric descriptors and LiDAR intensity. This multi-feature approach provides complementary information that improves the network’s ability to discriminate between species with similar gross morphology but different fine-scale structure or material properties.
For each point in the point cloud, we compute three geometric descriptors based on the covariance matrix of its local neighborhood. Given a set of neighboring points $\mathcal{P} = \{p_1, p_2, \ldots, p_k\}$, we compute the covariance matrix:
$$
\mathbf{C} = \frac{1}{k} \sum_{i=1}^{k} (p_i – \bar{p})(p_i – \bar{p})^\top
$$
where $\bar{p}$ is the centroid of the neighborhood. Eigenvalue decomposition of $\mathbf{C}$ yields three eigenvalues $\lambda_1 \geq \lambda_2 \geq \lambda_3$ and corresponding eigenvectors $v_1, v_2, v_3$. From these eigenvalues, we derive three geometric descriptors:
$$
\text{DA}_1 = \frac{\lambda_1}{\sum_{i=1}^{3} \lambda_i} \quad \text{(Pointness)}
$$
$$
\text{DA}_2 = \frac{\lambda_1 – \lambda_2}{\lambda_1} \quad \text{(Linearity)}
$$
$$
\text{DA}_3 = \frac{\lambda_3}{\lambda_1} \quad \text{(Sphericity)}
$$
DA1 measures the compactness of the local point distribution, with higher values indicating more concentrated distributions. DA2 captures the degree of linear alignment, with higher values indicating stronger directional structure. DA3 quantifies the spherical uniformity of the distribution, with higher values indicating more isotropic configurations. Together, these descriptors provide a concise characterization of local geometry that is sensitive to differences in branch structure, leaf arrangement, and stem morphology.
The LiDAR intensity feature records the backscattered signal strength, which is related to the reflectance properties of the target surface at the laser wavelength. Different tree species exhibit characteristic intensity responses due to differences in bark texture, leaf water content, and surface roughness. While intensity values are influenced by sensor configuration and acquisition conditions, they provide valuable complementary information when acquisition conditions are consistent across the dataset.
Our final input representation concatenates the spatial coordinates, geometric descriptors, and intensity into a 7-dimensional feature vector:
$$
\mathbf{F} = [x, y, z, \text{DA}_1, \text{DA}_2, \text{DA}_3, I]
$$
where $I$ is the normalized LiDAR intensity. This enriched representation enables the network to jointly reason about geometric structure and material properties, leading to more robust species discrimination. The ablation studies presented in Section 4.3 quantify the individual and combined contributions of these features.
Network Architecture Summary
Building upon the PointConT baseline, TreeConT introduces two key modifications: (1) replacement of standard FPS with Tree-FPS in all progressive down-sampling stages, and (2) expansion of the input feature dimension from 3 (xyz only) to 7 (xyz + DA1-DA3 + intensity). The overall architecture retains the five-stage Inception Feature Aggregator design with content-based attention, maintaining the same number of parameters and computational complexity as the baseline. This allows for fair comparison and isolates the contributions of our modifications.
During training, each individual tree point cloud is uniformly sampled to 4,096 points using standard FPS for input alignment. The Tree-FPS is applied internally within the network stages, operating on the feature-transformed point sets at each resolution level. This design ensures that the benefits of vertical structure preservation are realized at multiple scales throughout the feature hierarchy.
Experimental Setup
Dataset Description
Our experiments are based on the SYSSIFOSS dataset, which comprises UAV drone laser scanning data collected over 12 forest plots in Baden-Württemberg, Germany. Each plot measures 1 hectare (100 m × 100 m), and the data were acquired across multiple phenological phases: August-September 2019 (leaf-on), December 2020 (leaf-off), and March-April 2021 (early leaf-on). The UAV drone platform was equipped with a RIEGL miniVUX-1UAV laser scanner operating at a pulse repetition frequency of 100 kHz, with a flight altitude of 60-75 m and a flight speed of 3-5 m/s. The resulting point density ranges from 797 to 1,554 points per square meter, providing high-resolution structural information.
From the raw UAV drone point clouds, we applied a series of preprocessing steps including noise filtering, ground point removal using cloth simulation filtering, and individual tree segmentation using a region-growing algorithm. Each segmented tree instance was manually verified and corrected to ensure label reliability. After quality control, we retained 18,749 individual trees across 11 dominant species. The dataset, which we call TreeNetXplorer, exhibits natural class imbalance reflecting the actual species distribution in the study area. To mitigate bias during training, we employed a stratified random split that preserves the class proportions across training, validation, and test sets. Table 1 summarizes the sample distribution for each species.
The 11 species included in our study represent the dominant and ecologically significant taxa in Central European mixed forests. They encompass a range of growth forms, from conifers with distinct apical dominance to broadleaves with diverse crown architectures. This diversity poses a challenging test for classification methods, requiring the network to capture both coarse morphological differences and fine structural variations.
| Species Code | Scientific Name | Common Name | Training | Validation | Test | Total |
|---|---|---|---|---|---|---|
| AcePse | Acer pseudoplatanus | European Sycamore | 500 | 150 | 200 | 850 |
| CarBet | Carpinus betulus | European Hornbeam | 775 | 250 | 275 | 1,300 |
| FagSyl | Fagus sylvatica | European Beech | 1,550 | 512 | 537 | 2,599 |
| JugReg | Juglans regia | English Walnut | 1,175 | 375 | 425 | 1,975 |
| LarDec | Larix decidua | European Larch | 400 | 100 | 100 | 600 |
| PicAbi | Picea abies | Norway Spruce | 1,575 | 525 | 525 | 2,625 |
| PinSyl | Pinus sylvestris | Scots Pine | 400 | 125 | 150 | 675 |
| PruAvi | Prunus avium | European Cherry | 550 | 150 | 250 | 950 |
| PseMen | Pseudotsuga menziesii | Douglas Fir | 1,725 | 600 | 625 | 2,950 |
| QuePet | Quercus petraea | European Oak | 1,200 | 400 | 425 | 2,025 |
| QueRub | Quercus rubra | Red Oak | 1,325 | 450 | 425 | 2,200 |
Implementation Details
All experiments were conducted on a workstation equipped with an Intel Core i7-12700K processor, an Nvidia RTX 3090 GPU (24 GB VRAM), and 64 GB of system memory, running Ubuntu 18.04. The code was implemented in PyTorch. For the TreeConT network, we trained with a batch size of 64 for 300 epochs, using the AdamW optimizer with an initial learning rate of 1 × 10-4. We employed a cosine annealing learning rate schedule with warm restarts to promote convergence. During training, we applied data augmentation including random rotation around the Z-axis and random scaling, which improved generalization by simulating variations in tree orientation and size.
For all baseline comparisons, we adopted the same training protocol and hyperparameters to ensure fair evaluation. Each model’s architecture-specific parameters (e.g., number of layers, channel widths, neighborhood sizes) were set according to the original papers or official implementations, with necessary adaptations for the input feature dimension and number of classes. Table 2 summarizes the key hyperparameters used in our experiments.
| Parameter | Value |
|---|---|
| Input points per tree | 4,096 |
| Batch size | 64 |
| Number of epochs | 300 |
| Initial learning rate | 1 × 10-4 |
| Optimizer | AdamW |
| Weight decay | 1 × 10-4 |
| Learning rate schedule | Cosine annealing with warm restarts |
| Number of layers (k) in Tree-FPS | 50 |
| Decay parameter (α) in Tree-FPS | 1.0 |
| Mixing factor (λ) in Tree-FPS | 0.7 |
| Data augmentation | Z-axis rotation [0, 360°], scaling [0.8, 1.2] |
| Input feature dimension | 7 (xyz + DA1-DA3 + intensity) |
Evaluation Metrics
We evaluate classification performance using five complementary metrics: Overall Accuracy (OA), Macro-Precision (MP), Macro-Recall (MR), Macro-F1 score, and Cohen’s Kappa coefficient. These metrics provide a comprehensive assessment of model performance, accounting for class imbalance and providing insights into per-class discrimination ability. OA measures the proportion of correctly classified samples across all classes. MP and MR compute the average precision and recall across classes, giving equal weight to each class regardless of sample size. Macro-F1 is the harmonic mean of MP and MR, providing a balanced measure of precision and recall. The Kappa coefficient quantifies the agreement between predicted and true labels beyond chance, with values above 0.8 indicating strong agreement and values above 0.9 indicating near-perfect agreement.
Results and Analysis
Overall Classification Performance
We compared TreeConT against seven state-of-the-art methods: PointNet++, DGCNN, PointTransformer, PointNeXt, PointMLP, Mamba3D, and PointConT. All models were trained and evaluated under identical conditions, with the same data split, input point count, and augmentation strategy. Table 3 presents the comprehensive comparison results.
| Method | OA (%) | MP (%) | MR (%) | Macro-F1 (%) | Kappa |
|---|---|---|---|---|---|
| PointNet++ | 66.75 | 63.06 | 52.77 | 52.11 | 0.6080 |
| DGCNN | 78.98 | 77.73 | 66.66 | 66.52 | 0.7555 |
| PointTransformer | 86.78 | 85.66 | 82.32 | 83.46 | 0.8475 |
| PointNeXt | 84.29 | 82.91 | 76.60 | 78.55 | 0.8183 |
| PointMLP | 89.56 | 86.27 | 83.88 | 84.22 | 0.8791 |
| Mamba3D | 89.67 | 89.12 | 88.60 | 87.91 | 0.8809 |
| PointConT | 91.31 | 91.02 | 90.95 | 90.56 | 0.9000 |
| TreeConT (ours) | 93.12 | 92.45 | 91.03 | 91.66 | 0.9207 |
TreeConT achieves the best performance across all evaluation metrics, with an OA of 93.12% and a Kappa of 0.9207. Compared to the strong baseline PointConT, which already demonstrates competitive performance (OA = 91.31%, Kappa = 0.9000), our method improves OA by 1.81% and Kappa by 0.0207. The improvement in Macro-F1 (from 90.56% to 91.66%) is particularly noteworthy, as it indicates consistent gains across all classes rather than improvement driven by majority classes alone.
The comparison with PointTransformer (OA = 86.78%, Macro-F1 = 83.46%) and Mamba3D (OA = 89.67%, Macro-F1 = 87.91%) further highlights the effectiveness of our approach. These methods represent the latest advances in point cloud processing with attention mechanisms and state-space models, respectively. Yet, they do not account for the specific structural characteristics of UAV drone point clouds, particularly the vertical density heterogeneity that our Tree-FPS strategy addresses. The superior performance of TreeConT demonstrates that targeted architectural modifications, informed by domain-specific knowledge, can yield significant improvements over general-purpose architectures.
We also observe that PointNet++ and DGCNN, while pioneering in their respective domains, achieve substantially lower performance (OA of 66.75% and 78.98%, respectively). This gap can be attributed to their limited capacity for capturing long-range dependencies and global structure, which are critical for tree species classification where crown morphology and stem architecture provide essential cues. The Transformer-based methods (PointTransformer, PointConT, and TreeConT) consistently outperform these earlier approaches, underscoring the importance of global context modeling for this task.
Confusion Matrix Analysis
To gain deeper insights into the classification behavior, we analyzed the confusion matrices of the four best-performing methods: PointMLP, Mamba3D, PointConT, and TreeConT. The analysis reveals several important patterns regarding inter-class confusion and the impact of our design choices.
First, we observe that the most common confusions occur between species with similar crown morphology or growth form. For instance, CarBet (European Hornbeam) is frequently misclassified as FagSyl (European Beech) or PruAvi (European Cherry), reflecting the similarity in their broadleaf crown structures. Similarly, LarDec (European Larch) and PinSyl (Scots Pine) exhibit mutual confusion, as both are conifers with needle-like leaves, though the larch is deciduous while the pine is evergreen. These confusions are expected given the structural similarities and highlight the need for fine-grained feature extraction.
Second, we note a marked reduction in confusion for several hard classes when moving from PointConT to TreeConT. Specifically, the recall for CarBet improves from 53.82% to 66.18%, and for FagSyl from 86.51% to 90.51%, when replacing FPS with Tree-FPS while keeping other factors constant. This improvement is directly attributable to the enhanced representation of stem and lower crown structures enabled by Tree-FPS. These regions provide critical distinguishing features for species with similar crown appearances, such as differences in bark texture, branch angle, and stem taper.
Third, we find that TreeConT achieves perfect classification (100% recall) for certain species, including QueRub (Red Oak) and AcePse (European Sycamore), under the controlled comparison setting. This suggests that these species have particularly distinctive vertical structure characteristics that are well captured by the Tree-FPS sampling strategy. The consistent presence of stem points in the sampled set allows the network to learn robust features associated with these species.
The confusion matrix analysis also reveals remaining challenges. Species with high intra-class variability, such as JugReg (English Walnut) and PruAvi (European Cherry), still exhibit moderate confusion rates even with TreeConT. This may be due to the diverse growth conditions and management histories that lead to variable tree forms within the same species. Addressing this residual confusion may require additional information sources, such as spectral data from hyperspectral sensors, or more sophisticated modeling of within-species variation.
Ablation Studies
Sampling Strategy Comparison
To isolate the contribution of the Tree-FPS sampling strategy from other design elements, we conducted a controlled comparison between PointConT (which uses standard FPS) and a variant with Tree-FPS, keeping all other factors including input features identical. Table 4 presents the results of this comparison.
| Metric | FPS (PointConT) | Tree-FPS (TreeConT) | Improvement |
|---|---|---|---|
| Overall Accuracy (%) | 78.85 | 80.23 | +1.38 |
| CarBet Recall (%) | 53.82 | 66.18 | +12.36 |
| FagSyl Recall (%) | 86.51 | 90.51 | +4.00 |
| QueRub Recall (%) | 94.12 | 100.00 | +5.88 |
| Macro-F1 (%) | 69.56 | 70.84 | +1.28 |
| Kappa | 0.7513 | 0.7648 | +0.0135 |
The results demonstrate that Tree-FPS alone provides a meaningful improvement over standard FPS, with OA increasing from 78.85% to 80.23% and Macro-F1 from 69.56% to 70.84%. The improvement is particularly pronounced for species with strong vertical structural features, such as CarBet and QueRub, where recall increases by 12.36% and 5.88%, respectively. This confirms that the vertical structure-aware sampling strategy helps the network maintain access to discriminative stem and lower crown features throughout the multi-stage feature abstraction process.
Visual inspection of the down-sampled point clouds at various sampling rates provides further insights. At high sampling rates (e.g., 4,096 points), both FPS and Tree-FPS produce point sets that adequately represent the tree structure. However, at lower sampling rates (e.g., 256 or 128 points), the differences become striking: FPS tends to concentrate points in the crown region, with the stem becoming increasingly fragmented or completely disappearing. In contrast, Tree-FPS maintains stem continuity even at very low sampling rates, preserving the critical vertical structure that anchors the overall tree morphology. This structural preservation is essential for downstream feature extraction, as the stem provides a reference axis for interpreting branch arrangements and crown architecture.
We also note that the improvement from Tree-FPS is not uniform across all species. Species with relatively uniform vertical density distributions, or those where the crown provides the dominant discriminative features, show smaller gains. This is expected, as the benefits of structure-preserving sampling are most pronounced when there is significant density heterogeneity between stem and crown regions. The adaptive nature of Tree-FPS, controlled by the α and λ parameters, allows it to adjust the compensation strength based on the actual density distribution, ensuring robustness across diverse tree forms.
Feature Effectiveness Analysis
We conducted a series of ablation experiments to quantify the contribution of each input feature component. Four configurations were compared: (1) xyz only, (2) xyz + intensity, (3) xyz + geometric descriptors (DA1-DA3), and (4) xyz + DA1-DA3 + intensity (full configuration). All experiments were conducted with the TreeConT architecture and the same training protocol. Table 5 summarizes the results.
| Input Features | OA (%) | MP (%) | MR (%) | Macro-F1 (%) | Kappa |
|---|---|---|---|---|---|
| xyz only | 78.85 | 81.45 | 67.44 | 69.56 | 0.7513 |
| xyz + Intensity | 86.87 | 88.30 | 78.49 | 78.49 | 0.8470 |
| xyz + DA1-DA3 | 84.29 | 86.31 | 76.40 | 78.18 | 0.8166 |
| xyz + DA1-DA3 + Intensity (Full) | 93.12 | 92.45 | 91.03 | 91.66 | 0.9207 |
The results reveal several important findings. First, the xyz-only baseline achieves an OA of 78.85%, which, while modest, demonstrates that spatial coordinates alone contain significant discriminative information for tree species classification. This is consistent with the fact that different species exhibit characteristic three-dimensional forms that can be captured from point cloud geometry alone.
Second, adding either intensity or geometric descriptors leads to substantial improvements over the xyz-only baseline. The xyz + Intensity configuration achieves an OA of 86.87% (an improvement of 8.02%), while xyz + DA1-DA3 achieves 84.29% (an improvement of 5.44%). The fact that intensity provides a larger boost than geometric descriptors in this dataset suggests that material properties (captured by backscatter intensity) are particularly discriminative for the species present in our study area. This could be due to differences in bark texture, leaf water content, or surface roughness that produce distinct intensity signatures.
Third, the full configuration (xyz + DA1-DA3 + Intensity) achieves the best performance with an OA of 93.12%, representing a cumulative improvement of 14.27% over the xyz-only baseline. Importantly, the improvement from adding both feature types (8.02% + 5.44% = 13.46%) is slightly less than the sum of individual improvements (14.27%), suggesting some overlap in the information captured by the two feature types. Nonetheless, the complementary nature of geometric descriptors and intensity is evident: while both capture aspects of tree structure and material, they provide non-redundant information that together enables more robust species discrimination.
The increases in Macro-F1 and Kappa follow similar patterns, confirming that the benefits of multi-feature fusion are consistent across evaluation metrics. The substantial improvement in Macro-F1 (from 69.56% to 91.66%) is particularly noteworthy, as it indicates that all classes benefit from the enriched input representation, not just the majority classes.
Combined Effect of Sampling and Feature Enhancements
To understand the interaction between the two main contributions of our work—Tree-FPS sampling and multi-feature fusion—we conducted a factorial ablation study. We compared four configurations representing all combinations of sampling strategy (FPS vs. Tree-FPS) and input features (xyz-only vs. full features). The results are summarized in Table 6.
| Sampling Strategy | Input Features | OA (%) | Macro-F1 (%) | Kappa |
|---|---|---|---|---|
| FPS | xyz only | 78.85 | 69.56 | 0.7513 |
| Tree-FPS | xyz only | 80.23 | 70.84 | 0.7648 |
| FPS | Full features | 91.31 | 90.56 | 0.9000 |
| Tree-FPS | Full features | 93.12 | 91.66 | 0.9207 |
The results demonstrate that the two contributions are complementary and additive. When using xyz-only features, switching from FPS to Tree-FPS improves OA by 1.38% (from 78.85% to 80.23%). When using full features, the same switch improves OA by 1.81% (from 91.31% to 93.12%). The slightly larger improvement in the full-feature setting suggests that the enriched input representation and the structure-preserving sampling strategy reinforce each other: better features provide more discriminative information for the structure-preserved point sets to exploit.
Similarly, the benefit of full features is larger when combined with Tree-FPS (improvement of 12.89% over xyz-only with Tree-FPS) compared to FPS (improvement of 12.46% over xyz-only with FPS). This indicates that the full features are more effectively utilized when the sampling strategy ensures that stem and lower crown regions are well represented. The interaction effect, while modest, is consistent across metrics and suggests that the two design choices target complementary aspects of the classification challenge: Tree-FPS addresses the sampling bias that limits structural representation, while multi-feature fusion enhances the discriminative capacity of the input representation.
Discussion
Interpretation of Key Findings
Our experimental results establish that TreeConT significantly outperforms existing methods for tree species classification from UAV drone point clouds. The improvements stem from two complementary mechanisms: the Tree-FPS sampling strategy that preserves vertical structure information, and the multi-feature input representation that enriches the network’s perceptual capabilities. The combination of these two design elements yields a system that is particularly well-suited to the unique characteristics of UAV drone data, where vertical density heterogeneity and structural complexity pose significant challenges.
The success of Tree-FPS can be attributed to its principled handling of the density disparity between canopy and stem regions. By allocating more samples to low-density but structurally critical regions, Tree-FPS ensures that stem, branch, and lower crown features are consistently represented throughout the feature abstraction hierarchy. This is particularly important for species discrimination, as many tree species have similar crown morphologies but distinct stem characteristics, such as bark texture, stem taper, and branching angle. Without explicit structure preservation, these features may be diluted or lost during progressive down-sampling.
The multi-feature fusion strategy addresses a different aspect of the classification challenge. While spatial coordinates capture overall geometry, geometric descriptors provide localized shape information that is sensitive to fine-scale structural variations, and intensity encodes material properties that complement shape cues. The geometric descriptors—pointness, linearity, and sphericity—each capture different aspects of local structure: pointness reflects the compactness of local point distributions, linearity captures directional alignment (important for stem and branch detection), and sphericity indicates isotropic scattering (useful for distinguishing dense foliage from open branches). The intensity channel, meanwhile, provides information about surface reflectance that can help differentiate species with similar geometry but different bark or leaf optical properties.
The complementary nature of these features is evident from the ablation studies. While both geometric descriptors and intensity individually improve performance, their combination yields results that exceed the sum of their individual contributions. This synergy suggests that the network learns to exploit the non-redundant information in each feature type, using geometric cues for structural discrimination and intensity cues for material-based discrimination.
Comparison with Previous Work
Our work advances beyond existing tree species classification methods in several ways. Compared to feature engineering approaches, which rely on hand-crafted parameters that may not capture the full complexity of tree structure, our deep learning approach learns hierarchical features directly from the data. This allows the network to discover discriminative patterns that may not be apparent to human experts or captured by predefined features.
Compared to standard deep learning models for point cloud classification, our work introduces domain-specific inductive biases that are tailored to the characteristics of UAV drone data of trees. While models like PointNet++, DGCNN, and PointTransformer have achieved impressive results on general point cloud benchmarks, they do not account for the unique challenges of tree point clouds, particularly the vertical density heterogeneity. By incorporating vertical structure awareness through Tree-FPS, we provide the network with a built-in understanding of the importance of preserving stem and lower crown information.
The comparison with PointConT is particularly informative, as it shares the same backbone architecture and differs only in the sampling strategy and input representation. The clear improvement of TreeConT over PointConT highlights the importance of these targeted modifications. It also demonstrates that significant gains can be achieved through thoughtful architectural design rather than simply increasing model capacity or computational complexity.
Limitations and Future Directions
Despite the promising results, our work has several limitations that point to directions for future research. First, the Tree-FPS strategy involves hyperparameters (k, α, λ) that require tuning for optimal performance. While our default settings work well for the current dataset, different forest types or acquisition conditions may require adjustments. Developing adaptive methods that automatically set these parameters based on data characteristics would enhance the practical applicability of our approach.
Second, the intensity features, while beneficial in our experiments, are known to be sensitive to acquisition conditions. Variations in flight altitude, sensor calibration, and atmospheric conditions can cause intensity shifts that compromise model generalization. Future work could explore intensity normalization techniques or develop invariant representations that are robust to such variations. Alternatively, integrating spectral information from hyperspectral sensors could provide more stable material-related features.
Third, our method assumes that individual trees have been successfully segmented from the UAV drone point cloud, which itself is a challenging problem in dense forests. Errors in tree segmentation—such as under-segmentation (merging adjacent trees) or over-segmentation (splitting a single tree into multiple parts)—will propagate to the classification stage and degrade performance. Developing integrated segmentation and classification approaches that jointly address both tasks is an important direction for future work.
Fourth, our evaluation is limited to 11 species in Central European mixed forests. Extending the evaluation to other geographic regions, forest types, and species assemblages would test the generalizability of our approach. Different forest ecosystems may present unique structural characteristics that require adaptation of our methods.
Finally, while our work focuses on classification, the same principles could be applied to other tasks such as tree species segmentation (assigning species labels to individual points within a tree), biomass estimation (where species-specific allometric equations are used), or phenological monitoring (where species-specific responses to seasonal changes are tracked). Extending our approach to these related tasks would broaden its impact and utility.
Conclusion
In this study, we proposed TreeConT, a novel deep learning network for tree species classification from UAV drone laser scanning point clouds. The network introduces two key innovations designed to address the unique challenges of tree point cloud analysis: the significant vertical density heterogeneity between canopy and stem regions, and the need for multi-modal features that capture both geometric and material properties.
The first innovation, Tree-FPS, is a vertical structure-aware progressive down-sampling strategy that replaces standard farthest point sampling in the network’s feature abstraction stages. By performing height-wise stratification and allocating sampling quotas using an inverse-density principle, Tree-FPS ensures that structurally critical but sparse regions—particularly stems and lower branches—remain well represented throughout the multi-stage feature learning process. This preserves the vertical continuity of tree structure and provides the network with stable access to discriminative features that would otherwise be diluted or lost.
The second innovation is a multi-feature input representation that fuses spatial coordinates (xyz) with three geometric descriptors (pointness, linearity, sphericity) and LiDAR intensity. The geometric descriptors capture local shape information that is sensitive to fine-scale structural variations, while the intensity channel provides complementary material-related information. Together, these features enable the network to discriminate between species based on both structural and optical properties, leading to more robust classification performance.
Experimental results on the TreeNetXplorer dataset, comprising 18,749 individual trees across 11 species, demonstrate that TreeConT achieves state-of-the-art performance with an overall accuracy of 93.12%, a Macro-F1 of 91.66%, and a Kappa coefficient of 0.9207. These results significantly outperform strong baselines including PointNet++, DGCNN, PointTransformer, PointMLP, Mamba3D, and the baseline PointConT. Controlled ablation studies confirm that both Tree-FPS and multi-feature fusion contribute substantially to the observed improvements, with the combined effect exceeding the sum of individual contributions.
The findings of this work have practical implications for forest monitoring and management. Accurate species classification from UAV drone data can support a range of applications, from biodiversity assessment and carbon stock estimation to sustainable forest management and rare species monitoring. By improving the accuracy and reliability of species classification, our method can help reduce the need for costly and labor-intensive field surveys, enabling more frequent and extensive forest assessments.
From a methodological perspective, our work demonstrates the value of incorporating domain-specific structural priors into deep learning architectures. While general-purpose point cloud networks have achieved remarkable success across diverse tasks, tailoring them to the specific characteristics of a given application domain can yield substantial improvements. The Tree-FPS strategy, in particular, embodies a general principle—preserving structure in regions with high information density but low point density—that may be applicable to other point cloud tasks involving heterogeneous density distributions, such as building facade analysis, archaeological feature detection, or medical image processing.
In conclusion, TreeConT represents a significant step forward in tree species classification from UAV drone laser scanning data. By combining vertical structure-preserving sampling with multi-modal feature fusion, our approach achieves robust performance in challenging mixed-forest conditions and provides a solid foundation for future research in this important area.