Research on Individual Tree Canopy Segmentation of Camellia oleifera Based on a UAV-LiDAR System

Abstract

In consideration of the limited accuracy of individual tree canopy segmentation algorithms due to the diverse canopy structure and complex environments in mountainous and hilly areas, this study optimized the segmentation parameters of three algorithms for individual tree canopy segmentation of Camellia oleifera in such environments by analyzing their respective parameters. Utilizing an Unmanned Aerial Vehicle-Light Detecting and Ranging (UAV-LiDAR) system, we obtained Canopy Height Models (CHM) of Camellia oleifera canopies based on Digital Elevation Models (DEM) and Digital Surface Models (DSM). Subsequently, we investigated the effects of CHM segmentation, point cloud clustering segmentation, and layer stacking fitting segmentation on Camellia oleifera canopies across different research areas. Additionally, combining ground survey data from forest lands with visual interpretation of UAV orthophoto images, we evaluated the performance of these three segmentation algorithms in terms of the F-score as an evaluation indicator for individual tree canopy segmentation accuracy. Combined with the Cloth Simulation Filter (CSF) filtering algorithm after removing the ground point cloud, our findings indicate that among different camellia densities and terrain environments, the point cloud clustering segmentation algorithm achieved the highest segmentation accuracy at 93%, followed by CHM segmentation at 88% and the layer stacking fitting segmentation method at 84%. By analyzing the data from UAV-LiDAR technology involving various land and Camellia oleifera planting types, we verified the applicability of these three segmentation algorithms for extracting camellia canopies. In conclusion, this study holds significant importance for accurately delineating camellia canopies within mountainous hilly environments while providing valuable insights for further research in related fields.

1. Introduction

Camellia oleifera is a significant indigenous oil plant with diverse applications such as edible and medicinal uses, light chemical raw materials, soil and water conservation, and ecological climate regulation, and it is typically cultivated in subtropical alpine and hilly regions [1,2]. As the world’s population grows with a large demand for food, obtaining edible oil from Camellia oleifera can reduce the acreage of oil crops such as soybeans, so that more land can be used to grow food crops. At present, the planting area of Camellia oleifera in China has reached 70 million hectares, and it is increasing year by year. Canopy extraction plays a pivotal role in monitoring the health of Camellia oleifera trees, estimating yields, and facilitating mechanical harvesting. Using UAV and remote sensing technology can enable the precise and efficient extraction of individual tree canopy information for Camellia oleifera and aid in implementing precision operations based on plant variations [3,4,5], thereby significantly enhancing the level of intelligent management in Camellia oleifera plantations.

Owing to its advantages in data resolution, flight flexibility, high efficiency, and convenience, UAV-based proximal sensing technology has emerged as a prominent trend for acquiring high-resolution remote sensing image data using LiDAR [6,7,8,9]. By capturing images, the UAV can improve detection efficiency and accuracy [10,11] by analyzing and extracting information, such as topography and the growth status of the canopy, in order to achieve precise management and monitoring of an orchard [12,13,14]. LiDAR enables the extraction of canopy information by transmitting a pulsed laser and receiving echo signals, resulting in the generation of laser point clouds with wide coverage, high precision, and efficient processing capabilities [15,16,17]. Obtaining individual tree canopy parameters such as plant density, diameter at breast height, tree height, and crown size and position is a crucial step toward achieving precise individual tree canopy segmentation. However, when utilizing airborne LiDAR data for segmentation purposes, challenges related to over/under segmentation may arise owing to adjacent or overlapping canopies as well as variations in their shapes. To address these issues comprehensively, numerous methods have been developed for canopy-based individual tree canopy segmentation using airborne LiDAR data; these primarily include approaches based on canopy height models and normalized point cloud analysis.

Based on the CHM individual tree canopy segmentation method [18,19,20], the crown height model was initially constructed using ground elevation data and vegetation elevation information. By applying a threshold, areas exceeding this value in the CHM were identified as non-canopy regions. Subsequently, morphology, regional growth, and other algorithms were employed to further partition the vegetation area into individual tree canopies [18]. Ding [19] and colleagues have achieved promising outcomes in segmenting tomato canopy multispectral image leaves by integrating wavelet transform and a watershed algorithm with an average error rate below 8%. However, when dealing with complex backgrounds and varying light intensities, the average error rate for tomato canopy leaves increases to 21%. In order to address the challenge of a complex background on crown overlap, Liu et al. [20] proposed the ITCD algorithm integrated with CHM and introduced the multi-scale local maximum (LM) algorithm to enhance the segmentation accuracy of individual tree canopies. However, this method tends to misidentify crown edges as separable areas, leading to excessive segmentation. Ayry et al. [21] developed the layer stacking algorithm to improve the segmentation accuracy of individual tree canopies in deciduous or leafless conditions. Considering the difficulties in detection and information loss in complex forested areas, Paris et al. [22] proposed a method that combines CHM with point cloud spatial clustering for delineating dominant tree crowns’ boundaries, achieving significant results with 97% accuracy using high-density point cloud data and 92% accuracy using low-density data, thereby greatly improving understory vegetation detection and segmentation accuracy.

The Normalized Point Cloud Individual-Tree Segmentation [23,24,25] technique normalizes the point cloud data by mapping the coordinates of all points to a uniform range. This approach employs clustering algorithms such as density-based DBSCAN and connectivity-based methods to group the normalized point cloud data into clusters, thereby enabling the identification and segmentation of individual tree canopies based on their shape, density, and other characteristics. Unlike the CHM segmentation algorithm, this method directly analyzes the original point cloud data without relying on elevation information. Hence, it was well-suited for analyzing point cloud data generated by vegetation types with complex canopies and minimal height differences. Yang et al. [26,27] conducted high-resolution UAV image experiments in larch forests using the Structure-from-Motion (SfM) method, yielding outstanding results in the visual interpretation of orthophoto images and various automatic segmentation methods based on images and point clouds. The overall detection rate exceeded 91%, while the accuracy of crown width extraction surpassed 81%. Fu et al. [28] implemented tree segmentation from Terrestrial Laser Scanning (TLS) data using DBSCAN and an improved Distance Distribution Matrix (DDM). Through a combination of detection, correction, and layer-by-layer clustering, the evaluation results on plantation and mixed forest datasets demonstrate the superiority of the proposed method over traditional DBSCAN in terms of recall, accuracy, and F1 score. The method automatically extracts optimal parameters and accurately segments small trees under tall canopies. Yan et al. [29,30] employed an adaptive mean shift segmentation approach that divides the three-dimensional space into sectors from global maximum points to simulate canopy surfaces, iteratively identifying potential boundaries within specified areas. Results indicate accurate segmentation rates of 95% for simple samples and 80% for complex environments.

From the referenced literature, various approaches were studied for individual tree canopy segmentation; however, a solution for segmentation of the Camellia oleifera tree canopy is still lacking because of the challenges encountered in previous studies, such as uneven distribution of the canopy, complex terrain, and a significant overlap phenomenon. This study employs an airborne LiDAR data acquisition system to acquire point cloud data for Camellia oleifera. Furthermore, typical research areas considering terrain and planting type were selected, and algorithm parameters were optimized separately to improve the segmentation accuracy of the Camellia oleifera tree canopy. It integrates the CSF filtering algorithm for CHM segmentation and performs point cloud clustering segmentation. Additionally, the impacts of different segmentation parameters and growth environments on the accuracy of the Camellia oleifera canopy were analyzed to determine optimal parameters that were suitable for individual tree canopy segmentation in mountainous and hilly regions.

2. Materials and Methods

2.1. Experimental Materials

The airborne LiDAR data were acquired using the LiAir VH2 LiDAR scanning system mounted on a DJI M300 RTK UAV (DJI Technology Co., Ltd., Shenzhen, China), which was equipped with an RTK-GNSS system and BMI088 IMU sensor. According to

Table 1. Parameters of airborne equipment for data acquisition.

Parameters	Value
FOV	70.4° (horizontal) × 4.5° (vertical)
accuracy (elevation)	5 cm@70 m
distance measurement accuracy	2 cm (1σ@20 m)
maximum operating height	120 m
inertial guidance system parameters	GPS, GLONASS, BeiDou
RTK position accuracy	1 cm + 1 ppm (horizontal), 1.5 cm + 1 ppm (vertical)
IMU data frequency	200 HZ (Max 2000 HZ)
camera pixels, image size	2430 W, 6000 × 4000
flight altitude	82 m
flight speed	3 m/s
overlap rate in side and heading	90%

, the LiDAR field of view (FOV) is 70.4° horizontally and 4.5° vertically, with an accuracy of 5 cm@70 m. The RTK-GNSS achieved 1 cm + 1 ppm horizontally and 1.5 cm + 1 ppm vertically, and the IMU had a 200 HZ data sampling frequency up to 2000 HZ. The LiDAR worked with a multi-thread and repeated scanning mode, recording three echoes for each pulse. In complex Camellia oleifera environments, the use of multi-echo LiDAR proved suitable for acquiring point cloud information, effectively enhancing data quality, increasing point cloud density, and enabling improved information fusion capabilities. As the terrain of the Camellia oleifera forest is undulating and the highest elevation is close to 30 m, in order to ensure flight safety and data uniformity in each research area, the flight altitude was set to 83 m. During data collection, the RTK position information and IMU attitude information of the UAV were recorded at the same time. In order to improve the fusion and matching accuracy of the information, the IMU frequency was set to 200 HZ. At the same time, a relatively low speed of 3 m/s was adopted to improve the attitude stability of the UAV during the flight process and increase the amount of data in the point cloud, so as to further improve the point cloud matching accuracy. An overlap rate of 90% can improve the amount of point cloud data, obtain more comprehensive ground details, overcome the influence of occlusion, and improve the matching accuracy. In this research, the flight altitude was set at 82 m, with an IMU data frequency of 200 Hz and a flight speed of 3 m/s. Figure 1 illustrates the airborne data acquisition system, while Table 1 presents the specific parameters.

2.2. Data Acquisition

The research area is situated in Xinyu City, Jiangxi Province, the main growing area of Camellia oleifera in China. Considering the growth characteristics of Camellia oleifera, the branches and leaves of Camellia oleifera were the most lush in June, and the data collection period was from 27 May to 2 June 2023. Within this forest land, a diverse range of growth differentiation can be observed among Camellia oleifera trees. This study focused on investigating the sparsity between Camellia oleifera trees, canopy size variations, terrain slope characteristics, and different spatial grid sizes as the primary research objectives. The aim was to analyze the applicability of three segmentation algorithms for effectively delineating Camellia oleifera tree boundaries.

In order to investigate the impact of different terrains and Camellia oleifera planting type on segmentation effectiveness, three representative areas were selected within the forest land for study. Among them, research area 1 was characterized by a flat terrain, uniform canopy width, and consistent spacing (approximately 1.0 m) between Camellia oleifera plants and consisted of 88 samples covering an area of 1082.20 m². Research area 2 was characterized by an irregular and uneven distribution of canopy widths with sparse growth of Camellia oleifera plants on a gentle slope ranging from 15° to 25°, and the spacing between Camellia oleifera plants was approximately 1.2 m, consisting of 94 samples covering an area of 1194.13 m². Research area 3 featured overlapping occurrences in the Camellia oleifera canopy width on steeper slopes ranging from 25° to 60°, along with significant variations in canopy height as well as inconsistent and uneven planting spacing. The spacing between Camellia oleifera plants was approximately 1.2 m and consisted of 95 samples spanning an area measuring approximately 1610.66 m².

In this study, tree numbers and heights were obtained by field investigations and measurements. As the canopy width is difficult to directly measure in the field, we used the “Measure” tool of the CloudCompare (version 2.13.0) software to measure the results on a stitched point cloud map. Notably, our investigation on canopy segmentation of Camellia oleifera trees in mountainous and hilly regions served as a representative case. The specific research areas are illustrated in Figure 2, while

Table 2. Growth characteristics and canopy parameters of Camellia oleifera in each research area.

Research Area	Type of Area	Number of Samples (Tree)	Camellia Height (m)			Canopy Width Diameter (m)
			Min	Max	Ave	Min	Max	Ave
area 1	flat terrain, even distribution, consistent canopy width	88	1.50	3.30	2.40	1.60	5.00	3.30
area 2	gentle inclination, various levels of sparsity, haphazard distribution	94	1.02	2.52	1.77	0.97	5.29	3.13
area 3	terraced field environment, large difference in tree height, canopy overlap	95	1.80	3.30	2.55	0.50	7.00	3.75

presents the corresponding statistical data for Camellia oleifera trees within each area.

2.3. Data Preprocessing

The quality of data preprocessing significantly influences subsequent processing and outcomes. Canopy segmentation preprocessing primarily involves denoising, resampling, and ground point separation steps. When the laser scanning system acquires point cloud data, various types of noise are inevitably introduced, including instrument noise errors, environmental noise, and irregular reflective surfaces. Eliminating these noises can establish a more accurate foundation for data analysis and provide support for subsequent analyses. To ensure the density distribution of point cloud data adheres to specifications and is better suited for accurate segmentation and feature extraction of camellia canopies, we can enhance the calculation efficiency by resampling while retaining sufficient information to accurately represent the target object. For camellia canopy segmentation, filtering algorithms were employed to eliminate interference from ground point cloud data, including DEM extraction and ground point cloud classification methods. By removing the ground points, the remaining point cloud data primarily captures the structural characteristics of camellia trees’ canopies, thereby facilitating high-precision canopy segmentation.

Furthermore, the integration of DEM, DSM, and CHM plays a pivotal role in the preprocessing stage of Camellia oleifera canopy segmentation. DEM provides crucial ground elevation information for eliminating ground point cloud data, while DSM offers comprehensive vegetation elevation details that aid in the identification and segmentation of vegetation. Additionally, CHM supplies essential vertical structure information to facilitate accurate identification and segmentation of the vegetation canopy. The combined utilization of DEM, DSM, and CHM enables a more thorough analysis of vertical distribution patterns and growth status within the Camellia oleifera canopy segmentations by providing fundamental data. Insufficient removal of ground point clouds may occur owing to lower DEM resolution, thereby impacting segmentation accuracy; conversely, higher DSM smoothing factor may result in the loss of vegetation elevation information, consequently affecting the accuracy of vegetation segmentation. Additionally, CHM resolution directly influences canopy segmentation accuracy; excessively low resolution leads to detail loss, while excessively high resolution introduces noise. Field detection and analysis revealed that setting the DEM resolution at 1 m yielded optimal results in this study. When the parameter was set below 1 m, complete removal of non-canopy point clouds could not be achieved; on the other hand, when the parameter exceeded 1 m, small areas of Camellia oleifera canopy were mistakenly identified as ground point clouds and removed, thus further increasing the error.

The parameters for DSM smoothing ranged from 0.3 to 1.0. Excessively high parameter values resulted in a rough TIN model, leading to over-segmentation, while excessively low values caused the TIN model to become overly smooth, resulting in potential under-segmentation during the segmentation process. The CHM grid varied between 0.3 m and 0.7 m. Comparative analysis of camellia canopy segmentation accuracy using different grid sizes (0.3 m, 0.5 m, and 0.7 m) revealed that the optimal segmentation accuracy was achieved with a grid size of 0.5 m. If the value is smaller than the threshold, it may lead to a reduced segmentation area, resulting in potential under-segmentation; conversely, if the value exceeds this threshold, it may cause an expanded segmentation area and potential over-segmentation.

2.4. Individual Tree Canopy Segmentation Method

Based on the diversity of topography and canopy structure, three segmentation methods were employed in this study (CHM segmentation, point cloud clustering segmentation, and layer stacking fitting segmentation) by optimizing parameters to achieve high-precision segmentation.

The CHM segmentation algorithm [27] achieves precise canopy segmentation of Camellia oleifera by calculating the height difference between the crown center vertex and the ground. This algorithm effectively partitions the CHM into distinct vegetation objects through threshold settings or watershed algorithms, making it suitable for low-complexity segmentation tasks specific to Camellia oleifera. Moreover, crucial parameters such as minimum tree height, Gaussian smoothing coefficient, and Gaussian smoothing radius significantly impact the performance of this method. In this study, field measurements provided tree height data, and a minimum tree height of 1.2 m was set to exclude Camellia oleifera trees below this threshold during segmentation. Additionally, a Gaussian radius coefficient ranging from 0.5 to 1.0 was selected. Adjusting these parameters helps alleviate issues related to excessive or inadequate segmentation in different sample environments; smaller values may result in overly refined segmentations, while larger values may lead to insufficient differentiation among various vegetation objects.

The point cloud clustering segmentation method [31] primarily extracts features such as spatial relationships and colors from point cloud data and utilizes clustering or segmentation algorithms to partition the data into distinct regions or individual tree canopies. The results of the clustering segmentation process were controlled by adjusting the segmentation parameters. Setting a value that is too small may result in noise points or small non-individual tree canopy areas being clustered as individual tree canopies, while setting a value that is too large may lead to incorrect clustering of small individual tree canopy areas. In this study, we set the threshold for the average sample distance of Camellia oleifera trees between 0.3 and 0.7, with a field search radius of 0.5 and an average density range of Camellia oleifera point clouds at 15–30%.

The layer stacking fitting segmentation method [32] mainly combines the point cloud clustering algorithm and the area growing method. In the point cloud data, some seed points representing individual tree canopies were selected by manual selection or feature extraction. Starting from the seed point, the area growing algorithm was used to gradually add adjacent points to the same area until the growth conditions were met. In the segmentation of the Camellia oleifera canopy, the area growth can be judged according to the spatial relationship and color similarity between points. The size of the CHM spatial grid used for the layer stacking algorithm was 0.3−0.5 m, and the Gaussian smoothing coefficient and Gaussian radius were consistent with the coefficients of the CHM segmentation algorithm, which were 0.3–1.0 and 0.5–1.0.

2.5. Evaluation of Segmentation Accuracy

The integrity of Camellia oleifera canopy segmentation can be assessed through the visual interpretation of high-resolution UAV images, three-dimensional morphological analysis of laser point clouds, and ground survey data from forestry lands. Each canopy point cloud segmentation algorithm exhibits unique performance characteristics and adaptability. In the case of orthophoto images obtained by UAV, areas exceeding 80% of the canopy diameter at breast height were typically classified as correct detection and segmentation (true positive, TP), while areas surpassing the threshold but not accurately detected and segmented were considered false positives (FP), and those below the threshold that remain undetected and unsegmented were regarded as false negatives (FN). To evaluate the accuracy of results produced by three different segmentation algorithms, we calculate the weighted harmonic mean value F based on the detection recall (r) and precision (p) and evaluated the accuracy of the segmentation results. The calculation of r and p are shown in Equations (1) and (2) [33]. Figure 3 illustrates the corresponding segmentation criteria. The weighted harmonic mean F was calculated according to Equation (3):

$$\mathrm{r}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\times 100\%,$$

(1)

$$\mathrm{p}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}\times 100\%,$$

(2)

$$\mathrm{F}=\frac{2\left(\mathrm{rp}\right)}{\mathrm{r}+\mathrm{p}}\times 100\%,$$

(3)

where TP represents the number of correctly detected and segmented Camellia oleifera canopies, FP represents the number of incorrectly detected Camellia oleifera canopies, and FN represents the number of undetected Camellia oleifera canopies.

3. Results

In this research, the accuracy of three canopy segmentation algorithms for Camellia oleifera was compared in three distinct research areas, and the impact of the CSF filtering algorithm on canopy segmentation accuracy was analyzed. The findings revealed that the performance of canopy segmentation algorithms for Camellia oleifera was primarily influenced by sample plant density, canopy width, spatial grid size, threshold setting, and edge overlap. Among these factors, the point cloud clustering segmentation algorithm exhibited the highest overall accuracy, followed by CHM segmentation and layer stacking segmentation.

In the three research areas, when the grid size was smaller than 0.3 m or bigger than 0.7 m, the accuracies of the three segmentation algorithms in the canopy segmentation of Camellia oleifera were significantly low. When the grid was 0.3 m, the F-score for the approaches of CHM segmentation, point cloud cluster segmentation, and layer stacking fitting segmentation in area 1 were 75%, 82%, and 69%, respectively; those in area 2 were 70%, 82%, and 69%, respectively; and those in area 3 were 58%, 66%, and 60%, respectively. When the grid was 0.5 m, the segmentation evaluation indicators in area 1 were 88%, 93%, and 79%, respectively; in area 2 they were 88%, 88%, and 83%, respectively; and in area 3 they were 85%, 90%, and 84%, respectively. When the grid was 0.7 m, the evaluation indexes in area 1 were 78%, 91%, and 72%, respectively; in area 2 they were 70%, 82%, and 71%, respectively; and the evaluation indexes in area 3 were 70%, 83%, and 80%, respectively. When the grid was 0.5 m, the segmentation accuracy achieved the highest level of 90%.

In addition, after CSF filtering processing of the ground point cloud, the accuracy of the canopy segmentation of Camellia oleifera was significantly improved. According to the actual measurement results, when the threshold coefficient of the CSF algorithm was set to 0.55 and the number of iterations was set to 1000, the accuracy of the canopy segmentation of Camellia oleifera was most improved. Considering the use of the CSF filtering algorithm, the use of this algorithm not only improves the segmentation efficiency but also increases the segmentation accuracy by 21%, which verifies the positive role of the CSF filtering algorithm in the canopy segmentation of Camellia oleifera.

4. Discussion

4.1. Comparison of the Canopy Segmentation Accuracy of Camellia oleifera in Different Research Areas Using the Same Segmentation Algorithm

In order to investigate the impact of different sample environments on the performance of the same segmentation algorithm in the research area, as well as the influence of crown shape, density, and height variations on segmentation accuracy, we conducted separate analyses for the three segmentation algorithms. Based on the results presented in Figure 4 and statistical analysis of the sample segmentation outcomes shown in Figure 5, it was evident that the point cloud clustering algorithm and layer stacking segmentation method both exhibit higher detection rates and improved accuracy compared with the CHM segmentation algorithm. There were significant variations in canopy shape across the different research areas and sample environments, some of which exhibited complex features such as branching, concavity, and convexity, resulting in substantial changes in the canopy surface. These differences exacerbate the issue of canopy surface overlap and make segmentation more challenging. However, clustering segmentation based on point clouds can leverage three-dimensional information to cluster points according to their spatial position and relationships with neighboring points, enabling more accurate classification of points belonging to the same canopy class and improving detection rates and accuracy. Layer stacking segmentation can process point clouds through varying scale windows that better capture undulations on the canopy surface while accommodating diverse sizes and shapes of canopies, further enhancing detection rates and accuracy.

4.2. Comparison of Segmentation Accuracy Differences among the Three Algorithms in the Same Research Area

In the same research area, a comparison and analysis of canopy segmentation accuracy was conducted among the three segmentation algorithms. It was observed that over-segmentation tended to occur in cases of large canopies with sparse structures and multiple tree vertices, while threshold settings could not be applied to different canopy shapes, resulting in over/under segmentation for objects with uneven canopy height and multiple connected canopies. The point cloud clustering algorithm demonstrated superior performance across all three segmentation research areas owing to its flexible adjustment and optimization capabilities that enabled it to adapt to various Camellia oleifera canopy characteristics and better handle irregular shapes, crossing points, and fuzzy boundaries, thereby providing more accurate results. The proposed method utilizes 3D information to effectively capture the spatial structure of Camellia oleifera canopies, resulting in more accurate segmentation by considering the position and adjacent relationships of each point in 3D space. By employing a point cloud clustering segmentation approach, coordinate information for each point in the point cloud data was utilized to accurately represent the shape and structure of the Camellia oleifera canopies. Through consideration of both the position and surrounding relationships of points in 3D space, canopy points were classified with higher accuracy, leading to improved canopy segmentation results (Figure 6). To evaluate the accuracy of canopy segmentation within our research area using three different algorithms, we conducted a linear regression fitting analysis between field measurements and extracted data. Our findings demonstrated that as the grid size increased, there was an initial improvement followed by a subsequent decrease in the fitting effect. Figure 7 shows regression analysis of the canopy width of areas 1, 2, and 3 at 0.3 m, 0.5 m, and 0.7 m. For three research areas, a grid size of 0.5 m achieved maximum RMSE values of 0.92, 0.91, and 0.86 for research area 1, research area 2, and research area 3 respectively.

4.3. Influence of the CSF Filtering Algorithm on Segmentation Accuracy

For research areas 1, 2, and 3, we employed a consistent grid size and resolution to investigate the impact of the CSF filtering algorithm on camellia canopy segmentation accuracy. The CSF filtering algorithm leveraged the inherent continuity of ground points and fitted them based on physical constraints. The experimental results demonstrated that after applying the CSF filtering algorithm to remove the ground point cloud, there was an initial increase followed by a subsequent decrease in the overall segmentation accuracy of the camellia canopy. Notably, in research area 1, significant improvement in segmentation accuracy was observed owing to uniform distribution and minimal overlap among canopy widths. The application of the CSF filtering algorithm for mitigating ground disturbance can effectively minimize its impact on camellia canopy segmentation results. However, it is important to note that the parameters of the filtering algorithm should be tailored to suit specific research areas and camellia canopy distributions in order to achieve optimal outcomes. This study further confirms that the CSF filtering algorithm has a positive effect on enhancing both accuracy and efficiency in camellia canopy segmentation.

After applying the CSF filtering algorithm to remove the ground point cloud, the segmentation accuracy of the Camellia oleifera canopy was significantly improved. Directly segmenting the Camellia oleifera canopy without separating the ground point cloud yielded overall accuracy F values of 76%, 66%, and 58% for different samples with a grid size of 0.5 m. However, when segmenting the canopy after removing the ground point cloud using CSF, the overall accuracy F values improved to 88%, 88%, and 85% under the same parameters. The main reasons were as follows: (1) Directly applying the segmentation algorithm to the canopy of Camellia oleifera trees often results in a ground point cloud with lower elevation values and flatter features, which can interfere with the segmentation results. Failure to remove the ground point cloud may introduce interference to the segmentation algorithm when processing the canopy point cloud, thereby reducing the segmentation accuracy. (2) By removing the ground point cloud using a filtering algorithm, the canopy segmentation algorithm can focus more effectively on capturing characteristics specific to Camellia oleifera tree canopies, thus enhancing segmentation accuracy. (3) The CSF filtering algorithm facilitates the provision of clearer and more accurate data by reducing noise and interference. Therefore, when utilizing CSF for the removal of ground point clouds in post-processing data, canopy segmentation algorithms can achieve enhanced accuracy. By employing CSF-filtered preprocessed data instead of directly adopting segmentation algorithms, higher levels of segmentation accuracy can be attained, as demonstrated in Figure 8, which showcases the impact on accuracy and contrasting segmentation results.

4.4. Influence of Grid Size on Canopy Segmentation Accuracy

Adjusting the spatial grid size of the canopy model can enhance the recognition and accurate segmentation of the Camellia oleifera canopy during canopy segmentation across different scenes and phenotypic characteristics. This phenomenon was validated in previous studies by Yin et al. [34,35,36]. In this study, we investigated the segmentation accuracy at spatial grid sizes of 0.3 m, 0.5 m, and 0.7 m by analyzing DEM, DSM, and CHM information along with extraction data from actual Camellia oleifera detection. The results demonstrated that the optimal grid parameter for achieving the highest image accuracy, recognition accuracy, and segmentation accuracy of Camellia oleifera was 0.5 m. Lower resolutions resulted in a smoother canopy height model, leading to the omission of important canopy information and under-segmentation phenomena. Conversely, higher resolutions caused artificial bumps [37] in the Camellia oleifera height model that did not correspond to the actual surface, resulting in an over-segmentation phenomenon. Figure 9 and Figure 10 show the effects of high or low spatial resolution on the accuracy of individual tree canopy segmentation. Figure 11 illustrates the DEM, DSM, and CHM of Camellia oleifera generated using a curvature radius of 1 m, smoothing factor of 0.5, and grid size of 0.5 m.

5. Conclusions

In this study, we used an UAV-LiDAR to collect laser point cloud information for Camellia oleifera. Subsequently, according to different terrains and planting types, three research areas were selected, and three different canopy segmentation algorithms (CHM segmentation, point cloud clustering segmentation, and layer stacking fitting segmentation) were evaluated under different parameters. Evaluation indexes were assessed using F-scores to determine the accuracy and effectiveness of the experimental protocol. The experimental results show that the point cloud clustering segmentation algorithm has the highest performance, followed by CHM segmentation and layer stacking segmentation, and the segmentation accuracy was highest when the grid size was 0.5 m, and the CSF filtering algorithm also has a positive effect on the canopy segmentation of Camellia oleifera.

By comparing and analyzing the segmentation accuracy of different segmentation algorithms, we determined the best segmentation parameters for segmentation. In order to further enhance the accuracy of canopy segmentation algorithms for Camellia oleifera, future research should focus on optimizing the sampling method for plant density, improving canopy width measurement technology, and fine-tuning threshold settings and edge overlap parameters. Additionally, exploring the application of alternative filtering algorithms may also contribute to enhancing the effectiveness of Camellia oleifera canopy segmentation. These endeavors will facilitate a deeper understanding of the structure and characteristics of Camellia oleifera canopies while providing a scientific foundation for managing and conserving Camellia oleifera.