The precision monitoring and sustainable management of forest resources, the main body of terrestrial ecosystems, are of paramount importance for maintaining ecological balance, assessing carbon stocks, and guiding sustainable forestry practices. As the fundamental structural unit of a forest, the individual tree provides key parameters—such as its spatial location, dimensions, and biophysical attributes—that are essential for forest inventory, ecological modeling, and management decision-making. Traditionally, acquiring this information has relied on labor-intensive, time-consuming, and weather-dependent manual field measurements, which are impractical for large-scale applications. The advancement of high-resolution remote sensing technologies offers a promising alternative. Unlike conventional optical remote sensing, which primarily provides two-dimensional spectral information, Light Detection and Ranging (LiDAR) technology can directly and accurately capture three-dimensional point cloud data, effectively characterizing the vertical structure of forests. In recent years, the rapid development of Unmanned Aerial Vehicle (UAV) LiDAR systems, known for their operational flexibility, strong penetration capability through canopies, and relatively low cost, has demonstrated immense potential for the rapid and precise acquisition of detailed structural information at both the stand and individual tree levels, with China UAV drone platforms becoming increasingly prevalent in forestry research.
Accurate individual tree crown (ITC) delineation is a critical step in deriving tree-level metrics from LiDAR data. Numerous algorithms have been developed and compared, which can be broadly categorized into two groups: two-dimensional methods based on the Canopy Height Model (CHM) and three-dimensional methods operating directly on the normalized point cloud. Common CHM-based methods include the watershed algorithm, the marker-controlled watershed segmentation (MWS) method, and region-growing algorithms. For instance, the MWS method, which identifies local maxima in the CHM as tree apexes and then uses a watershed transform constrained by these markers to delineate crowns, has shown effectiveness in various forest types. Conversely, three-dimensional point-cloud-based methods, such as distance-dependent clustering (e.g., Point Cloud-based Cluster Segmentation, PCS) and K-means clustering, utilize the full 3D spatial distribution of points. The PCS method typically starts by identifying local maxima in the point cloud and then employs a top-down region-growing process with a distance threshold to cluster points belonging to the same tree, often showing robust performance in complex canopies.
While both categories of methods have their respective strengths and limitations, many existing studies have focused on evaluating them within single-species stands. There is a need to assess their adaptability across different tree species and, crucially, under varying stand densities, which significantly influence canopy complexity and overlap. This study, therefore, aims to evaluate and compare the performance of two representative algorithms—the MWS (2D) and PCS (3D) methods—for individual tree segmentation in two major plantation types in southern China: Slash Pine (Pinus elliottii) and Eucalyptus (Eucalyptus spp.). Specifically, we established sample plots with low, medium, and high densities for each species. We collected high-density UAV LiDAR data using a mainstream China UAV drone platform, generated corresponding CHMs and normalized point clouds, and applied both segmentation methods. The segmentation accuracy was rigorously evaluated against high-precision ground survey data. This work provides insights into selecting appropriate ITC delineation methods for intensive forest management and resource monitoring in the region’s fast-growing plantations.
1. Materials and Methodology
1.1 Study Area and Field Survey
The study was conducted in the Yangjiang State-Owned Forest Farm, located in the southwestern coastal region of Guangdong Province, China (approximately 21°42’10″–22°15’33” N, 111°45’45″–112°21’16” E). The area experiences a south subtropical maritime monsoon climate characterized by distinct seasons, long sunshine hours, abundant heat, and plentiful rainfall. The topography is generally higher in the east, north, and west, sloping towards the sea. The primary soil types are lateritic red earth and coastal sandy soil, mostly acidic. The native vegetation is typical south subtropical evergreen broad-leaved forest, with major plantation species including Eucalyptus, Masson Pine (Pinus massoniana), Slash Pine, and Acacia.
Field surveys were carried out in October 2023. A total of six square sample plots, each 30 m × 30 m (0.09 ha), were established: three in Slash Pine plantations and three in Eucalyptus plantations, representing a gradient of stand densities. The corner and center points of each plot were accurately georeferenced using Real-Time Kinematic (RTK) GPS. Within each plot, a full tree census was performed, recording species, diameter at breast height (DBH, at 1.3 m), total tree height (H), and positional coordinates for every tree using diameter tapes, hypsometers, and RTK. A summary of the field data is presented in Table 1.
| Plot ID | Species | Tree Count | Stand Density (trees/ha) | Mean DBH (cm) | Std. DBH (cm) | Mean Height (m) | Std. Height (m) |
|---|---|---|---|---|---|---|---|
| 1 | Slash Pine | 67 | 744 | 19.5 | 6.1 | 13.4 | 2.2 |
| 2 | Slash Pine | 103 | 1144 | 14.3 | 3.4 | 12.2 | 1.6 |
| 3 | Slash Pine | 136 | 1511 | 15.0 | 2.9 | 14.6 | 1.8 |
| 4 | Eucalyptus | 99 | 1100 | 12.2 | 3.2 | 16.6 | 2.5 |
| 5 | Eucalyptus | 141 | 1567 | 13.4 | 2.1 | 17.4 | 1.6 |
| 6 | Eucalyptus | 184 | 2044 | 12.9 | 3.3 | 18.3 | 2.7 |
1.2 UAV LiDAR Data Acquisition and Preprocessing
UAV LiDAR data were acquired concurrently with the field survey. A DJI Matrice 300 RTK UAV, a widely used China UAV drone, was equipped with a DJI Zenmuse L1 sensor. The sensor specifications include a scan rate of 160 Hz, a laser emission frequency of 240 kHz, a range accuracy of ±5 cm, and a field of view of 70.4° (horizontal) × 4.5° (vertical). Flights were conducted at altitudes between 50 m and 120 m above ground, with a speed of 7–10 m/s and both forward and side overlap exceeding 50%, ensuring high-density point cloud coverage (>200 points/m² on average).
The raw LiDAR data were processed to generate georeferenced point clouds in LAS format. The preprocessing workflow involved several key steps:
- Clipping: Point clouds were clipped to the precise boundaries of each 30 m × 30 m sample plot.
- Denoising: A spatial distribution-based denoising algorithm, supplemented by manual removal, was applied to eliminate low (e.g., ground clutter) and high (e.g., bird strikes) outliers.
- Filtering: An Improved Progressive TIN Densification (IPTD) algorithm was used to classify ground points from non-ground (vegetation) points.
- DEM/DSM Generation: The ground points were interpolated using Inverse Distance Weighting (IDW) to create a Digital Elevation Model (DEM) with a 0.2 m resolution. Similarly, a Digital Surface Model (DSM) was generated from the first-return points.
- Normalization: The raw point cloud height values (ellipsoidal) were normalized to heights above ground by subtracting the DEM value for each point’s XY location, producing a normalized LiDAR point cloud.
- CHM Generation: The Canopy Height Model (CHM) was computed as the difference between the DSM and the DEM (CHM = DSM – DEM), also at a 0.2 m resolution.
1.3 Individual Tree Segmentation Methods
Two established and widely used segmentation algorithms were implemented and compared: the Marker-Controlled Watershed Segmentation (MWS) and the Point Cloud-based Cluster Segmentation (PCS).
1.3.1 Marker-Controlled Watershed Segmentation (MWS)
This is a 2D method operating on the CHM. The procedure is as follows:
- Local Maxima Detection: A variable window size, determined from the lower prediction interval of a regression between crown width and tree height, is used to search for local maxima in the CHM. These identified points serve as initial tree apex markers, creating a Canopy Maxima Model (CMM).
- Smoothing and Marker Creation: The CMM is smoothed using a Gaussian filter to reduce noise, and the local maxima are refined and marked.
- Watershed Transform: The marked CMM is inverted, and the watershed transform is applied. Conceptually, water is flooded from the marked minima (the inverted maxima). As the “water level” rises, “catchment basins” form around each marker. Boundaries (“watershed lines”) emerge where basins meet.
- Distance Transform and Crown Delineation: A distance transform is applied to the resulting segmentation to refine and extract the final polygon boundaries for each individual tree crown.
The process relies heavily on the accurate identification of tree tops from the CHM, which can be challenging in dense, overlapping canopies.
1.3.2 Point Cloud-based Cluster Segmentation (PCS)
This is a 3D method that operates directly on the normalized point cloud. The algorithm is based on a top-down region-growing principle combined with a distance threshold. The main steps are:
- Identification of Seed Points: The highest point within the normalized point cloud (or a local neighborhood) is identified as a seed point, representing a tree apex.
- Distance-Based Clustering: Starting from the seed point, the algorithm iteratively evaluates all other points. A horizontal distance threshold ($d_{th}$) is crucial. If a point’s horizontal distance to the current seed is less than $d_{th}$, it is considered a candidate for belonging to the same tree. Further rules (e.g., vertical proximity) may be applied to finalize the assignment.
- Cluster Finalization and Iteration: The process continues, growing the cluster until no more points within the distance threshold can be added. This cluster is labeled as one tree. The algorithm then selects the next highest unassigned point as a new seed and repeats the process until all points are assigned to a cluster or rejected as noise.
The performance of PCS is sensitive to the choice of the distance threshold $d_{th}$. In this study, for each plot, $d_{th}$ was set to the average crown width calculated from the field data, an approach shown to yield good results.
1.4 Accuracy Assessment
The segmentation results from both methods were compared to the ground reference data. Each detected tree crown was considered correctly segmented (True Positive, $N_t$) if its position corresponded to a reference tree and its crown polygon intersected significantly with the estimated location of that tree’s crown. Detections not matching any reference tree were counted as commissions ($N_c$), and reference trees not detected by the algorithm were counted as omissions ($N_o$). Three standard metrics were calculated for each plot and method:
- Crown Detection Rate (Recall, $r$): Measures the algorithm’s ability to find reference trees.
$$ r = \frac{N_t}{N_t + N_o} $$ - Crown Accuracy (Precision, $p$): Measures the correctness of the detected trees.
$$ p = \frac{N_t}{N_t + N_c} $$ - Overall F-score ($F$): The harmonic mean of recall and precision, providing a single measure of overall accuracy.
$$ F = \frac{2 \cdot (r \cdot p)}{r + p} $$
2. Results
The individual tree segmentation results for both Slash Pine and Eucalyptus plantations across the density gradient are summarized in Table 2. The visual segmentation effects for one sample plot from each species are illustrated, showing the 3D point cloud clusters from PCS and the 2D crown boundaries from MWS.
| Plot ID | Species | Density (trees/ha) | Method | $N_t$ | $N_o$ | $N_c$ | $r$ | $p$ | $F$ |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Slash Pine | 744 (Low) | PCS | 54 | 13 | 16 | 0.81 | 0.77 | 0.79 |
| MWS | 55 | 12 | 24 | 0.82 | 0.70 | 0.75 | |||
| 2 | 1144 (Medium) | PCS | 82 | 21 | 27 | 0.80 | 0.75 | 0.77 | |
| MWS | 82 | 21 | 24 | 0.80 | 0.77 | 0.78 | |||
| 3 | 1511 (High) | PCS | 102 | 34 | 28 | 0.75 | 0.78 | 0.77 | |
| MWS | 93 | 43 | 28 | 0.68 | 0.77 | 0.72 | |||
| 4 | Eucalyptus | 1100 (Low) | PCS | 82 | 17 | 16 | 0.83 | 0.84 | 0.83 |
| MWS | 75 | 24 | 17 | 0.76 | 0.82 | 0.79 | |||
| 5 | 1567 (Medium) | PCS | 121 | 20 | 29 | 0.86 | 0.81 | 0.83 | |
| MWS | 103 | 38 | 29 | 0.73 | 0.78 | 0.75 | |||
| 6 | 2044 (High) | PCS | 144 | 40 | 43 | 0.78 | 0.77 | 0.78 | |
| MWS | 97 | 87 | 24 | 0.53 | 0.80 | 0.64 |
The key findings from the results are as follows:
1. Superiority of PCS over MWS: The PCS method demonstrated equal or higher overall accuracy ($F$-score) compared to the MWS method across all six sample plots. The $F$-scores for PCS ranged from 0.77 to 0.83, showing consistent and high performance. In contrast, the MWS method showed a wider and lower range of $F$-scores (0.64 to 0.79), with notably poorer performance in the highest density plots.
2. Effect of Stand Density: A clear trend of decreasing segmentation accuracy with increasing stand density was observed for both methods and both species. In low and medium-density plots (Plots 1, 2, 4, 5), both methods achieved relatively good results ($F$ > 0.75). However, in high-density plots (Plots 3 and 6), accuracy declined, and the performance gap between the two methods widened significantly. This was most extreme in the highest-density Eucalyptus plot (Plot 6, 2044 trees/ha), where the MWS method’s $F$-score plummeted to 0.64, primarily due to a very high omission error ($N_o$ = 87, omission rate ~47%).
3. Effect of Tree Species: At low and medium densities, the segmentation accuracy for Eucalyptus plots was generally higher than for Slash Pine plots of comparable density (e.g., Plot 4 vs. Plot 2, Plot 5 vs. Plot 3). However, in high-density conditions, this interspecies difference diminished. For the PCS method, accuracy became similar between the high-density plots of both species. Interestingly, for the MWS method, the high-density Slash Pine plot (Plot 3, $F$=0.72) actually outperformed the high-density Eucalyptus plot (Plot 6, $F$=0.64).
4. Error Pattern Analysis: The MWS method consistently suffered from higher omission rates ($N_o$) as density increased, particularly for Eucalyptus, indicating its difficulty in detecting trees within dense, overlapping canopies. The PCS method maintained a better balance between omission and commission errors across all densities, contributing to its higher and more stable $F$-scores. The use of advanced China UAV drone LiDAR systems was instrumental in providing the high-density point clouds necessary for the effective application of the 3D PCS method.
3. Discussion
The findings of this study underscore the critical importance of selecting an appropriate individual tree segmentation algorithm based on stand structure, particularly density. The consistent outperformance of the 3D PCS method over the 2D MWS method aligns with several previous studies conducted in complex forest environments. The robustness of PCS can be attributed to its direct utilization of the three-dimensional spatial information inherent in the LiDAR point cloud. By employing a distance threshold in the horizontal plane during the clustering process, it can better dissociate interlocking crowns that appear merged in a 2D projection like the CHM. This capability is increasingly valuable as data from modern China UAV drone platforms offer ever-higher point densities, making detailed 3D analysis more feasible.
The sharp decline in the performance of the MWS method in high-density stands, characterized by severe under-segmentation (high omissions), highlights a fundamental limitation of CHM-based approaches. The CHM is inherently a smoothed 2.5D representation where complex, multi-layered canopy structures are compressed. The algorithm’s first step—detecting local maxima as tree tops—becomes highly unreliable when crowns extensively overlap, as peaks may be suppressed or shifted. Furthermore, the watershed transform often fails to correctly locate the boundary between adjacent trees when the “ridge” between them in the CHM is not well-defined. This effect was more pronounced in the dense Eucalyptus canopy, likely due to its denser foliage and more continuous crown layer compared to the more open structure of Slash Pine at similar densities.
The species-dependent results at lower densities are insightful. The higher segmentation accuracy for Eucalyptus in low/medium density plots may be linked to its more uniform, dome-shaped crown architecture, which creates distinct, regularly spaced maxima in the CHM and compact point clusters. Slash Pine, with its potentially more irregular and layered crown form, might present a noisier signal for both algorithms. However, as density increases, the confounding factor of intense crown interaction and occlusion becomes the dominant driver of error, overshadowing the intrinsic species differences. This explains the convergence of PCS accuracy and the reversal in MWS performance at the highest densities.
The practical implication for forest management using China UAV drone technology is clear. For routine inventory in young or low-to-medium density plantations, both MWS and PCS can provide satisfactory results, with PCS holding a slight edge. However, for mature, high-density, or structurally complex stands—common in intensive forestry—the PCS method is unequivocally recommended to ensure reliable tree count and crown delineation. The parameterization of the distance threshold ($d_{th}$) in PCS remains crucial; using a plot-specific value based on prior knowledge or field-derived mean crown width, as done here, is a sound strategy.
Future work could explore hybrid methods that combine the strengths of both approaches, perhaps using a CHM-based pre-segment to guide a subsequent 3D point cloud refinement. Additionally, investigating deep learning-based segmentation methods trained on high-quality China UAV drone LiDAR data may offer a path to even more robust and automated individual tree detection across diverse forest conditions.
4. Conclusion
This study evaluated the performance of two individual tree segmentation methods—Marker-Controlled Watershed Segmentation (MWS) and Point Cloud-based Cluster Segmentation (PCS)—applied to UAV LiDAR data from Slash Pine and Eucalyptus plantations with varying stand densities in southern China. The comprehensive analysis leads to the following conclusions:
- The three-dimensional PCS method consistently achieved higher overall segmentation accuracy ($F$-score) than the two-dimensional MWS method across all stand densities and for both tree species. The PCS method demonstrated superior robustness, particularly in handling dense canopies with extensive crown overlap.
- Stand density is a primary factor influencing segmentation accuracy. Both methods exhibited a decline in performance as density increased. The MWS method was especially sensitive to high density, suffering from significant under-segmentation (high omission errors).
- Tree species characteristics influence segmentation outcomes at low to medium densities, with Eucalyptus’s more regular crown form facilitating higher accuracy. However, in high-density conditions, the effect of crown interaction becomes dominant, reducing interspecies differences in accuracy. In the most challenging high-density scenario, the MWS method performed better in Slash Pine than in Eucalyptus, while PCS maintained similar accuracy for both.
The results provide practical guidance for forest resource monitoring applications leveraging efficient China UAV drone LiDAR systems. For accurate individual tree detection and crown delineation in southern China’s intensively managed plantations, especially at high stand densities, the adoption of point cloud-based segmentation methods like PCS is strongly advocated. This approach ensures more reliable data extraction to support precision forestry, carbon accounting, and sustainable management practices.