A Novel Adaptive Cuboid Regional Growth Algorithm for Trunk–Branch Segmentation of Point Clouds from Two Fruit Tree Species

Abstract

Accurate acquisition of the phenotypic information of trunk-shaped fruit trees plays a crucial role in intelligent orchard management, pruning during dormancy, and improving fruit yield and quality. However, the precise segmentation of trunks and branches remains a significant challenge, limiting the accurate measurement of phenotypic parameters and high-precision pruning of branches. To address this issue, a novel adaptive cuboid regional growth segmentation algorithm is proposed in this study. This method integrates a growth vector that is adaptively adjusted based on the growth trend of branches and a growth cuboid that is dynamically regulated according to branch diameters. Additionally, an innovative reverse growth strategy is introduced to enhance the efficiency of the growth process. Furthermore, the algorithm can automatically and effectively identify the starting and ending points of growth based on the structural characteristics of fruit tree branches, solving the problem of where to start and when to stop. Compared with PointNet++, PointNeXt, and Point Transformer, ACRGS achieved superior performance, with F₁-scores of 95.75% and 96.21% and mIoU values of 0.927 and 0.933 for apple and cherry trees. The results show that the method enables high-precision and efficiency trunk–branch segmentation, providing data support for fruit tree phenotypic parameter extraction and pruning.

1. Introduction

Deciduous fruit trees represent the most widely cultivated and productive fruit-bearing species worldwide, characterized by a comprehensive industrial supply chain and considerable market demand. These species play a crucial role in ensuring international fruit supply security [1,2]. Among the common deciduous fruit trees, apple and cherry trees are extensively cultivated, with trunk-shape cultivation and management practices being predominantly employed [3,4,5,6]. The architecture of the trunk shape is vital for enhancing growth, fruit production, and wind resistance, significantly affecting the overall development, yield, and quality of deciduous fruit trees. The manual measurement of traditional fruit tree phenotypic parameters presents numerous challenges, including time consumption, labor intensity, subjectivity, inefficiency, and high labor costs [7,8]. Furthermore, achieving real-time, rapid, non-destructive, and accurate management of phenotypic information remains difficult. These limitations hinder the refinement and efficiency of fruit tree management [9,10,11].

To address these challenges and foster the development of smart orchards, remote sensing measurement technologies such as thermal infrared cameras, light detection and ranging (LiDAR), depth cameras, hyperspectral imagers, and RGB cameras can be utilized to measure high-throughput and accurate phenotypic parameters of fruit trees, providing data support for automatic high-precision pruning of fruit trees during the dormant period [12,13,14,15,16]. LiDAR, as an active remote sensing technology, not only realizes high-precision characterization of forest spatial structure and sub-canopy topography information by rapidly acquiring high-density and high-precision 3D point cloud data but also provides data support for non-destructive and precise extraction of morphological parameters of trees and fruit trees, which promotes the development of scientific and intelligent fruit tree management [17,18,19]. The current research primarily focuses on identifying key challenges in processing methods for 3D point clouds of trees, particularly addressing the segmentation of individual trees within complex forest and orchard environments and the segmentation of tree organs from the point cloud data of individual trees [20,21,22].

1.1. Related Work

In complex forest and orchard environments, the primary 3D point cloud scanning technologies utilized for individual tree segmentation include airborne laser scanning (ALS) and terrestrial laser scanning (TLS). While ALS offers advantages such as extensive coverage and rapid data acquisition, its accuracy can be affected by factors including flight altitude and vegetation density. Yu [23] employed the DJI Zenmuse L1 drone, which is equipped with an integrated Livox ALS module, to collect samples from four distinct forest types and terrains. An adaptive tree crown seed point detection algorithm was proposed, focusing on tree crown surface points to enhance detection precision. Subsequently, a top-down regional growth algorithm was applied to segment individual trees from the normalized forest point cloud, thereby improving the accuracy of single-tree segmentation. Similarly, Li et al. [24] used a UAV-mounted RIEGL VUX-1UAV laser scanner to collect urban road point cloud data. They proposed a segmentation method based on non-photosynthetic component clustering and stepwise refinement, enabling effective separation of individual trees with varying shapes and sizes while removing understory vegetation and trunks.

In contrast, TLS, despite its limited coverage area, provides higher accuracy for individual tree scanning, rendering it suitable for detailed measurements in complex environments. Liu [25] collected 3D point cloud data from natural forests using TLS and proposed a single-tree point cloud segmentation method based on tree trunk growth, which enabled efficient segmentation of individual trees exhibiting varying growth patterns within complex natural forests. Yang [26] utilized TLS to scan point cloud data from six forest types of differing complexities, introducing a topology-based single-tree segmentation method that enhances the versatility and performance of superpoint segmentation algorithms based on point clouds.

The aforementioned studies primarily concentrate on single-tree segmentation within forested areas, with insufficient attention given to organ-level segmentation of individual trees. Organ-level segmentation entails the meticulous separation of tree organs, such as leaves, branches, and trunks, thereby enabling the comprehensive and precise extraction and analysis of tree characteristics. This methodology provides critical insights for various domains, including forest ecology, environmental monitoring, and forest management.

Currently, individual tree organ segmentation based on LiDAR point cloud data predominantly emphasizes branch–leaf and trunk–branch segmentation. Li [17] employed TLS to collect point cloud data from pear trees, proposing an automated segmentation pipeline that integrates the PointNet++ model and the mean shift clustering algorithm. This approach yielded accurate measurements of leaf inclination, length, width, and area. Wu [27] utilized TLS to acquire 3D point cloud data from broadleaf forests (Cyclocarya paliurus and Populus adenopoda) and coniferous forests (Pinus massoniana), developing a convolutional neural network (CNN)-based model to differentiate leaves and branches utilizing local point geometry and laser return intensity (LRI) information. Liu [28] employed a Kinect camera to capture color and depth image data of poplar seedlings subjected to heavy metal stress, proposing a segmentation method for poplar leaves that combines Mask R-CNN with the DBSCAN algorithm. Zhang [29] captured multi-view images of maize tassels during the loose pollen stage using the MVS-Pheno V2 platform, presenting an automatic segmentation method, MaizeTasselSeg, based on the PointNet++ network and the shortest path algorithm for organ segmentation and phenotype extraction. However, this method necessitates extensive datasets and enhancements in robustness to effectively handle highly adherent organ components.

Branch–leaf segmentation has indeed made strides in the precise differentiation of tree leaves and branches through the application of computer vision and machine learning techniques. However, trunk–branch segmentation is essential for a comprehensive analysis of tree structures, especially regarding the relationship between the trunk and its branches. This task is more complex than branch–leaf segmentation, as it requires higher precision and often employs advanced methods such as instance segmentation and 3D point cloud processing to capture detailed structural information.

Recent studies have highlighted various approaches to trunk–branch segmentation using different technologies and methodologies. For instance, Niknejad [30] leveraged 3D imaging devices to analyze torch pine trunks and branches. Their research introduced deep learning-based instance segmentation techniques alongside image and point cloud processing methods, enabling them to quantitatively assess trunk diameter, branch angles, and branch diameters. Their findings demonstrated that these automated measurements surpassed those obtained through traditional manual methods, showcasing the potential for improved accuracy in tree structure analysis. Similarly, Majeed [31] utilized Kinect V2 sensors to capture RGB and point cloud data from apple trees. Their approach combined color and depth information with a deep learning-based semantic segmentation technique, facilitating the automatic segmentation of trunks, branches, and gridlines. This integration of multi-modal data allowed for a more nuanced understanding of the tree structure. Dong [32] employed TLS technology to gather information about wood and leaves. They proposed an unsupervised automatic semantic segmentation method that further categorizes wood data into trunks, larger branches, and smaller branches, highlighting the capability of TLS in providing detailed tree structure insights. Ma [33] took a different approach by utilizing a robotic platform equipped with dual RGB-D cameras to collect three-dimensional point cloud data from dormant jujube trees. They developed a three-dimensional reconstruction algorithm based on deep learning and the SPGNet model, achieving effective automatic segmentation of trunks and branches.

Deep learning methods have achieved high-precision trunk–branch segmentation by leveraging complex structural features and multimodal data fusion. However, these methods depend on large-scale datasets and substantial computational resources, and they still encounter challenges in managing occlusions and overlapping regions.

In contrast, 3D point cloud-based segmentation techniques, which utilize laser scanning and other methods to capture high-precision spatial data, effectively address these challenges. They efficiently capture tree geometry, mitigate perspective and occlusion issues, reduce data requirements, and lower computational demands, thereby offering a more efficient solution for trunk and branch segmentation. For example, Mirande [34] proposed a graph-based two-stage method that employs laser-scanned 3D point cloud data to automatically and accurately segment plant organs. This approach ensures overall consistency by utilizing local geometric and spectral features. He [35] optimized tree skeleton extraction by harnessing the geometric features of branches and local properties, achieving a high-precision characterization of branch structures. Liu [36] developed a 3D point cloud segmentation method based on Euclidean clustering and multi-plane extraction to address the challenges posed by hanging branches in complex terrains, significantly enhancing both accuracy and efficiency in outdoor mobile robot navigation.

Despite some progress in recent years, existing methods still have significant limitations in adapting to the structural complexity and variability of fruit tree point clouds. Many methods rely on hand-crafted rules or fixed geometric assumptions, which make them poorly effective in the face of occlusion, natural curvature of trunks, or densely entangled branches. These problems often lead to confusion between trunks and branches and inaccurate instance segmentation. Moreover, most existing algorithms are designed for specific tree species and lack generalizability to different tree shapes. In addition, current evaluation schemes rely heavily on coarse global metrics that cannot reflect the segmentation quality of key areas at trunk junctions or overlapping branch structures.

1.2. Goals and Objectives

This study aims to develop a region-growing-based segmentation system capable of accurately segmenting trunks and branches in apple and cherry trees under commercial orchard conditions. The research addresses key challenges including the structural complexity of fruit trees, limited cross-species adaptability of existing methods, and the inadequacy of current evaluation frameworks. The main objectives are as follows:

(1) To validate the cross-species adaptability and segmentation accuracy of the proposed algorithm: Apple and cherry trees, which exhibit distinct differences in branch density and growth morphology, were selected as representative species. Experimental results demonstrate that the proposed region-growing segmentation algorithm achieves high segmentation accuracy and structural adaptability across both species, indicating its robustness in handling different tree architectures. It is acknowledged, however, that the current study is limited to two species, and further validation on a broader range of fruit trees is planned for future work;

(2) To establish a multi-level, fine-grained performance evaluation framework: To address the limitations of existing evaluation systems, a comprehensive assessment framework is proposed that includes both global and structural-level metrics. In addition to traditional P, R, and F₁ metrics, the framework introduces independent quantification of trunk and branch segmentation performance, along with the inclusion of instance-level metrics such as mean intersection over union (mIoU). This allows a more accurate and practical reflection of segmentation performance, particularly for individual branch instances in complex tree structures.

In this study, we not only provide a high-precision data foundation for the phenomics research of fruit trees but also offer strong support for the intelligent management of orchards. To enhance the transparency and reproducibility of the research, we have made publicly available both the original datasets and the associated source code. Three deep learning models—PointNet++, PointNeXt, and Point Transformer—were improved, parameterized, and code-modified for the task of segmenting fruit tree point clouds and used in comparative experiments. These models were implemented on a Windows 10 system equipped with an Intel i9-9900K CPU and an NVIDIA RTX 3070 GPU using CUDA 11.3. Specifically, PointNeXt was developed using PyTorch 1.10.1 and Python 3.7, while PointNet++ and Point Transformer were implemented using PyTorch 1.12.1 and Python 3.8. All materials have been uploaded to a public GitHub repository (https://github.com/henry3539/ACRGS, accessed on 28 June 2025). By open-sourcing relevant datasets and algorithms, we hope to provide references and bases for subsequent research and promote technological progress in this field.

2. Materials and Methods

2.1. Study Area

TLS data collection was conducted in a commercial orchard located in Shunyi District, Beijing, China (40°21′ N, 116°54′ E), on 9 March 2024. The research subjects included dormant Asentech Red Fuji apple trees (Malus domestica Borkh., 6 years old) and dormant Kordia cherry trees (Prunus avium L., 5 years old). Apple trees (3.0–3.5 m tall) had thin, dense branches with smooth surfaces, growing in a uniform, slightly spiraled upward pattern, forming rounded or ovate crowns (Figure 1a). In contrast, cherry trees (3.2–3.7 m tall) had thick, short, and slightly twisted trunks with grey-brown to dark grey bark, often featuring small cracks or holes. Their branches were thick, dense, and slightly drooping, forming conical or flattened crowns (Figure 1b).

Both species exhibit a central trunk but differ significantly in branch morphology, affecting TLS data collection. Apple trees’ dense branching leads to cross-obscuration and potential data loss, while cherry trees’ sparser branches enable more comprehensive scanning. These morphological differences make them ideal for validating the generalization and robustness of the proposed trunk–branch segmentation method.

2.2. Point Cloud Collection Device

The TX8 features a broad viewing angle of 360° × 317°, enabling it to complete typical survey tasks in approximately 3 min. Its accuracy is maintained over a distance of 120 m, with extended capabilities up to 340 m. Engineered for durability, the TX8 boasts an IP54 rating, ensuring resistance to dust and water, which is essential for outdoor scanning environments. Additionally, the scanner utilizes a Class 1 invisible laser, guaranteeing safe operation at all times. This combination of features makes the Trimble TX8 an optimal choice for accurate and reliable data acquisition in the study of trunk–branch segmentation in apple and cherry trees, as summarized in

Table 1. Performance parameters of Trimble TX8 LiDAR.

Parameters	Value	Parameters	Value
Scanning speed	1 million points/s	Power wastage	72 W
Laser class	1 (human eye safety)	Scanning density	1/2/3 level, expansion mode
Laser wavelength	1.5 μm	Measuring range	0.6–120 m
Goniometric accuracy	16″	Viewing angle	360° × 317°
Working temperature	0~40 °C	Measurement accuracy	2 mm

.

2.3. TLS Data Collection

Before commencing the scanning operation, multiple target balls were positioned within the data acquisition area. To enhance the accuracy of point cloud registration, at least three identical, non-collinear target balls were incorporated into the point cloud data for each collection experiment. The Trimble TX8 LiDAR was mounted on a stabilizing tripod, which ensured that the device remained perpendicular to the ground by adjusting the three leveling screws at its base. After setting the scanning parameters, the scan was initiated.

In the apple tree section of the orchard, where slender and dense branches posed a risk of occlusion, a multi-station approach was adopted to collect point cloud data from a long row of apple trees. The 3D point cloud data underwent an orientation operation to ensure proper alignment, with the x-axis, y-axis, and z-axis of the point cloud corresponding to the rows, columns, and vertical plumb line of the fruit trees, respectively.

To achieve uniform scanning and comprehensive coverage of both the apple and cherry tree areas, eight symmetrically arranged scanning stations were established in each data collection zone. The stations were spaced 8 m apart for the apple trees and 10 m apart for the cherry trees. The scanning stations, labeled S1 through S8, are illustrated in Figure 1a,b.

2.4. Overall Point Cloud Processing Method

Trimble RealWorks^® software (Version 12.0, Trimble Inc., Sunnyvale, CA, USA) [37] was utilized for the registration and segmentation of the fruit tree point cloud data. The development environment was set up using Visual Studio 2019 and also required the installation of Point Cloud Library (PCL) version 1.12.1 and CMake version 3.26.6. C++ programming language is used to achieve tree trunk and branch segmentation. The overall point cloud data processing method is illustrated in Figure 2. The segmentation process of the fruit tree point cloud data into trunks and branches involved three primary stages: data preprocessing, preliminary segmentation, and adaptive cuboid regional growth segmentation.

2.5. Data Preprocessing

The first step in Figure 2 involves data preprocessing, which includes data import, target ball registration, point cloud segmentation, and data read–write operations. This preprocessing phase is essential for accurately isolating individual fruit trees within the intricate orchard environment, serving as a foundation for subsequent segmentation processes, including preliminary segmentation and regional growth segmentation.

The initial task involved performing data import and target ball registration. Point cloud data from both apple and cherry trees was imported into Trimble RealWorks software, where site clouds were generated in registration mode. To facilitate the registration of point cloud data, an automatic target registration method was employed to extract and register the target ball with a diameter of 0.145 m. This process is illustrated in Figure 3a.

Following the target ball registration, the next step involved obtaining point cloud data for individual fruit trees without any extraneous elements, such as brackets. This was achieved through a manual segmentation process, as depicted in Figure 3b. The registered point cloud data was segmented in analysis and modeling mode, allowing for the extraction of multiple rows of fruit tree point clouds. Using the clipping box mode, the point cloud data for each single fruit tree was carefully isolated from these rows. To facilitate subsequent algorithmic processing, a bracket removal procedure was conducted on each individual fruit tree’s point cloud using the manual shear function. Following this, spatial sampling was conducted with a resolution of 10 mm, and the resulting sampled point cloud data was extracted and saved in the *.las format.

To further enhance processing efficiency, compatibility, and performance for applications utilizing the PCL, the point cloud data format was converted from *.las to *.pcd. This conversion not only streamlined data storage but also significantly improved processing speed, facilitating more effective analysis and modeling of the fruit tree point cloud data in subsequent steps.

2.6. Preliminary Segmentation

To achieve accurate point cloud segmentation of the complete trunk and all branches of a single tree, this study proposes a method that combines preliminary segmentation with adaptive cuboid regional growth segmentation. The preliminary segmentation serves as a foundational step that facilitates the subsequent adaptive cuboid regional growth segmentation. Its primary goal is to coarsely separate the trunk from the branches, thereby establishing an initial division of the tree structure. As shown in Figure 2, preliminary segmentation includes plane segmentation, statistical filtering, multi-layer cylindrical segmentation, and Euclidean clustering segmentation. Using apple tree point clouds as an example Figure 4, the raw data (Figure 4a) underwent plane segmentation via RANSAC, removing ground points (Figure 4b). Next, statistical filtering (KNN = 10, standard deviation = 0.6) eliminated noise while preserving tree structure (Figure 4c).

In the subsequent steps, a preliminary segmentation of the trunk and branch point clouds is performed to enhance the accuracy of later segmentation. A key challenge is ensuring the segmented trunk point cloud fully encompasses the entire trunk. Traditional cylindrical fitting algorithms often struggle with this, as tree trunks are rarely perfectly vertical. If the cylinder radius is too large, excess branch points are included, compromising segmentation accuracy. Conversely, a smaller radius risks truncating the trunk, leading to an incomplete representation that disrupts subsequent algorithms. The curvature variability among fruit tree trunks further complicates segmentation, making a fixed parameter setting impractical for effective trunk extraction.

To address this, a multi-layer cylindrical segmentation method is proposed. The apple tree point cloud is first segmented along the z-axis into multiple layers via passthrough filtering, followed by cylindrical fitting in each layer. However, the number of layers significantly impacts extraction accuracy. Fewer layers reduce completeness for curved trunks, while more layers increase failure rates in top-layer fitting, hindering trunk extraction. To balance these factors, the point cloud was divided into four layers (a = 1, 2, 3, and 4). The random sample consensus (RANSAC) algorithm estimated parameters for each layer, followed by cylindrical fitting. To ensure complete trunk segmentation, each cylinder’s radius was expanded by 0.07 m, forming the reference radius (R_r), which better accommodates trunk shape variability, particularly at curved or branching regions.

The inlier points from all cylindrical segments were then merged, forming an initial segmented trunk point cloud that included small branches (Figure 4d). These branches were later separated through regional growth segmentation and assigned accordingly. This approach improved trunk completeness and accuracy, facilitating subsequent branch segmentation. For branch extraction, the overall branch point cloud was divided into independent segments (Figure 4e). Some branches, split during layering due to passthrough filtering, were reconstructed using a point cloud merging algorithm, ensuring completeness. Finally, the Euclidean clustering segmentation algorithm, with a 0.01 m distance threshold, was applied to distinguish adjacent branches while maintaining internal coherence. Short branches near the trunk were extracted via passthrough filtering, sequentially indexed based on their minimum z-axis values, and labeled j = 1 to j_max, where j_max is the total number of segmented branches.

2.7. Adaptive Cuboid Regional Growth Segmentation

Due to variations in trunk curvature and branch growth patterns among different fruit trees, the initial segmentation algorithm demonstrates insufficient accuracy when handling branch–trunk segmentation. To address this issue, this study proposes the adaptive cuboid regional growth segmentation (ACRGS) algorithm to achieve precise segmentation of trunk–branch point clouds in trunk-shaped fruit trees, thereby improving segmentation accuracy, as shown in the third part of Figure 2.

The core contribution of the ACRGS algorithm lies in proposing an adaptive regional growth segmentation method driven by both growth vectors and growth cuboids, which can adaptively adjust according to the growth direction of branches. This enables the dynamic identification of small branch point clouds within a complete trunk point cloud, extracting and integrating them into their corresponding branch point clouds.

The algorithm innovatively integrates three steps (Figure 5): regional growth initialization to determine the starting point, adaptive cuboid regional growth segmentation to dynamically adjust regional growth and stop, and segmentation optimization to refine the boundary and improve accuracy. It aims to solve three core problems: where to start, how to grow, and when to stop.

2.7.1. Regional Growth Initialization

To obtain the initial growth cuboid closest to the trunk within the branch point cloud and determine its dimensions along the x-, y-, and z-axes as well as the initial growth vector, the branch point cloud must first be initialized. Based on the ordering of the branch point cloud described earlier, the jth branch point cloud is imported sequentially (shown in Figure 5a).

The minimum value along the z-axis of the branch point cloud in the nearest neighborhood to the trunk is extracted to apply the pass-through filter corresponding to the branch’s layer. To avoid confusion and incorrect model parameter retrieval, the cylindrical fitting model corresponding to the branch’s layer is identified, and the associated model parameters are obtained.

The growth cuboid and its length, width, and height are defined as GU_i (x_i′, y_i′, and z_i′), and the coordinates of the growth center point are defined as C_i (x_i, y_i, and z_i), where i = 0, 1, 2, 3… Given the significant variation in trunk radii among different trunk-shaped fruit trees, a fixed reference radius (R_r) cannot adequately address all cases. To address this limitation, a radial growth coefficient (R_gc) is introduced to dynamically adjust the radius range, enabling the cylindrical model to accommodate various tree types.

Specifically, the initial filtering process employs R_r + R_gc as the passthrough filter parameter to preliminarily filter branch point clouds. This step yields the first growth cuboid (GU₁ (x₁′, y₁′, and z₁′)) and calculates the corresponding growth center coordinates (C₁ (x₁, y₁, and z₁)), as illustrated in Figure 5b. Subsequently, the radius range is expanded further by using R_r + 2R_gc as the filtering parameter for a second passthrough process. This results in a growth cuboid (GU₀ (x₀′, y₀′, and z₀′)) aligned along the branch growth direction, with the growth center coordinates (C₀ (x₀, y₀, and z₀)) being calculated. The region-growing algorithm is then initiated from the growth center (C₀ (x₀, y₀, and z₀)) and iteratively extends to neighboring point clouds, as shown in Figure 5c.

2.7.2. Adaptive Cuboid Regional Growth

For refined segmentation of tree trunks and branches, the trunk point cloud must be segmented using regional growth to extract and merge the branch point cloud into the original branch point cloud. The question thus arises: how does regional growth iterate and expand. A key parameter in this process is the adaptive growth vector V_i (x_vi, y_vi, and z_vi), calculated from the growth center point. The calculation of this vector is represented in Equation (1).

$$\left\{\begin{array}{l}{x}_{vi}={\displaystyle \frac{x_i-x_{i-1}}{\left|V_i\right|}}\times b\\ y_{vi}={\displaystyle \frac{y_i-y_{i-1}}{\left|V_i\right|}}\times b\\ z_{vi}={\displaystyle \frac{z_i-z_{i-1}}{\left|V_i\right|}}\times b\\ \hspace{0.17em}\hspace{0.17em}i\in N^{+}\end{array}\right.$$

(1)

To ensure consistent growth distances, the growth step size is set to b, with its formula shown in Equation (2).

$$b={\displaystyle \frac{1}{2}}\sqrt{x_{v1}^2+y_{v1}^2+z_{v1}^2}$$

(2)

Based on Equation (1), the adaptive growth vector V_i (x_vi, y_vi, and z_vi) from C_i−1 to C_i can be calculated. This vector is computed between adjacent growth center points with a fixed magnitude b, but its direction must be dynamically adjusted to follow the branch growth direction. Once V_i+1 is computed and the recursive relationship of the growth vector is established, adaptive growth can be realized.

Initially, the dynamic center point coordinates C_i′ (x_i″, y_i″, and z_i″) are calculated based on the original growth vector V_i and the growth center point C_i, with the calculation formula provided in Equation (3).

$$\left\{\begin{array}{l}{x_i}^{″}=x_i+x_{vk}\\ {y_i}^{″}=y_i+y_{vk}\\ {z_i}^{″}=z_i+z_{vk}\end{array}\right.$$

(3)

Subsequently, the center point of the next branch point cloud is determined, and the growth vector is adjusted to align its direction with the branch growth trajectory. By utilizing the dynamic center point C_i′ (x_i″, y_i″, and z_i″) as the geometric center and the growth cuboid GU_i (x_i′, y_i′, and z_i′) as a parameter, the maximum external cuboid of the point cloud within the dynamic cuboid is computed, yielding the growth cuboid GU_i+1 (x_i+1′, y_i+1′, and z_i+1′). The center of this growth cuboid becomes the new growth center point C_i+1. It is critical to note that while the dynamic center point C_i′ and the growth center point C_i+1 are in close proximity, they are distinct. Thereafter, substituting C_i and C_i+1 into Equation (1) yields the growth vector V_i+1, as illustrated in Figure 5d. Following this, the trunk point cloud within growth cuboid GU_i is extracted and integrated into the jth branch point cloud, and the remaining trunk point cloud is updated. This method effectively addresses the iteration problem in the growth process, but the regional growth of a single branch cannot proceed indefinitely. Therefore, appropriate termination conditions must be established to ensure effective segmentation of the point cloud.

To effectively halt the regional growth of an individual branch and ensure it is not interfered with by other branches, a judging cuboid JU_i (x_ui′, y_ui′, and z_ui′) is placed directly beneath each growth cuboid. Since the branches are introduced into the regional growth process from the bottom up, the base of the branch serves as the criterion for determining the termination of growth, effectively avoiding interference from upper branches. Moreover, the uppermost branches will not reduce the number of point clouds within the judgment cuboid due to the absence of trunks above, thus preventing excessive regional growth. This design ensures both the accuracy and stability of the regional growth process.

The judging cuboid JU_i (x_ui′, y_ui′, and z_ui′) is calculated using Equation (4), based on the growing cuboid GU_i (x_i′, y_i′, and z_i′).

$$\left\{\begin{array}{l}{x_{ui}}^{\prime }=x_i\\ {y_{ui}}^{\prime }=y_i\\ {z_{ui}}^{\prime }=z_i\end{array}\right.$$

(4)

Subsequently, the judging center point J_i (x_ui, y_ui, and z_ui) is computed using on Equation (5), considering the growing center point C_i (x_i, y_i, and z_i) and the growing cuboid GU_i (x_i′, y_i′, and z_i′).

$$\left\{\begin{array}{l}{x}_{ui}=x_i\\ y_{ui}=y_i\\ z_{ui}=z_i+{z_i}^{\prime }\end{array}\right.$$

(5)

In this study, the growth convergence ratio (G_cr) and its threshold (G_cr′) are introduced as key parameters for evaluating the termination conditions of the growth process. This ratio measures the degree of convergence by comparing the number of point clouds within the growth cuboid (GU_i) and the judgment cuboid (JU_i). Within the algorithm, all point coordinates are iteratively examined to determine whether they fall inside GU_i or JU_i. The number of point clouds within these cuboids is recorded as N_GUi, and N_JUi, respectively. The growth convergence ratio (G_cr) is then calculated using Equation (6).

$$G_{cr}={\displaystyle \frac{N_{J{U}_i}}{N_{G{U}_i}}}$$

(6)

The appropriate selection of the threshold (G_cr′) is critical for achieving accurate and detailed segmentation of trunks and branches. If G_cr′ is set too high, the growth process may terminate prematurely, leading to incomplete segmentation and the inability to capture all trunk and branch details. Conversely, an excessively low G_cr′ may result in overgrowth, causing trunk point clouds to be misclassified as branches and thereby reducing segmentation accuracy. Thus, the careful determination of G_cr′ is essential for ensuring optimal segmentation performance. As shown in Figure 5e, the definition and computation of G_cr provide an effective method for evaluating and controlling the convergence of the growth process. This approach enhances algorithm performance and ensures precise segmentation of trunks and branches.

During the growth process, if the stopping condition is not satisfied, growth continues until the established threshold for growth number is attained. Once the regional growth fulfills the stopping condition—the G_cr or the growth number exceeds the threshold—the jth branch regional growth concludes, with j incremented by 1. If j = j_max, perform the next step: segmentation optimization; otherwise, perform the previous step: regional growth initialization, followed by the inclusion of the (j + 1)th branch. The trajectory of the branch adaptive cuboid regional growth is illustrated in Figure 5f–h, where each point cloud cuboid block represents a distinct regional growth process.

2.7.3. Segmentation Optimization

After the initialization of regional growth and adaptive rectangular regional growth segmentation, scattered branch point clouds remain around the trunk point cloud. To improve segmentation accuracy and recall, these residual branch point clouds must be merged with the corresponding branch point clouds.

To extract residual branches from the trunk point cloud (Figure 5i), the Euclidean clustering algorithm is used with a distance threshold of 0.015 m and a minimum of 10 points, ensuring effective extraction despite the small number of residual points. Subsequently, all branch point clouds are traversed to merge residual branches with the nearest corresponding branches, as shown in Figure 5j. If no corresponding branch is found, and the residual branches are too numerous, their distance to the trunk point cloud is calculated. If this distance is small, the residual branches are merged with the trunk point cloud and saved as a new trunk point cloud; otherwise, the original trunk point cloud is retained. The optimized segmentation result is shown in Figure 5k.

In summary, the adaptive cuboid regional growth segmentation method effectively segments trunks and branches. Its primary advantages and contributions are as follows:

(1) The algorithm significantly enhances the accuracy and efficiency of trunk–branch segmentation through reverse regional growth and the strategic configuration of growth termination conditions;

(2) To accommodate various fruit tree species, the algorithm proposes an automatic method for setting the initial point of the regional growth. It also integrates adaptively adjusted growth vector directions and cuboid dimensions. This approach demonstrates strong adaptability to curved trunks and branches of different shapes.

2.7.4. Manual Point Cloud Segmentation for Ground Truth Generation

To ensure the accurate evaluation of the proposed segmentation algorithm, a high-quality ground truth dataset was manually constructed using CloudCompare software (Version 2.13) [38]. The manual segmentation process was carefully designed to ensure objectivity, precision, and cross-species consistency. The procedure is described in three key steps:

(1)

Manual segmentation protocol and annotation criteria.

The pre-processed 3D point clouds of apple and cherry trees were imported into CloudCompare. Under the guidance of experienced orchard researchers, manual segmentation was conducted using the “Polygonal Selection Tool” in a slice-by-slice manner from the tree base upwards. The trunk was defined as the thickest vertically or near-vertically oriented central structure extending from the base, while branches were defined as lateral structures that deviate from the trunk axis. Classification of each point into trunk or branch was based on geometric features (position, orientation, diameter) and topological context (connectivity and branching angles).

(2)

Instance-level branch labeling and 3D visual inspection.

Beyond binary classification, each branch was assigned a unique instance ID, enabling the use of instance-aware evaluation metrics such as mIoU. Segmentation was performed entirely within the 3D environment, allowing the annotators to freely rotate, zoom, and slice the point cloud to identify complex areas, particularly around trunk–branch junctions and overlapping branches.

(3)

Quality assurance via cross-validation.

To reduce subjectivity, each manually segmented tree was annotated by one operator and independently reviewed by a second observer. Discrepancies in ambiguous regions were resolved through consensus. This cross-validation process ensured the reliability and consistency of the ground truth annotations.

The final labeled point clouds were saved in *.pcd and *.txt formats, with separate files for trunk and branch points, and these serve as the reference dataset for algorithm performance evaluation.

2.8. Evaluation Indicators

The automatic and manual segmentation of trunk–branch structures in the 3D point clouds of apple and cherry trees were compared and analyzed to evaluate the accuracy of the novel method proposed in this study for stem–branch segmentation. This detailed analysis focused on the branches, trunks, and overall structure of the fruit tree point clouds as subjects for quantitative assessment. Four evaluation indicators—precision (P), recall (R), F₁-score (F₁), and mean intersection over union (mIoU)—were selected to assess the effectiveness of trunk–branch segmentation in the fruit tree point clouds, as illustrated in the third section of Figure 2. The calculation formulas for these metrics are provided in Equations (7)–(10), respectively.

$$P={\displaystyle \frac{TP}{TP+FP}}$$

(7)

$$R={\displaystyle \frac{TP}{TP+FN}}$$

(8)

$$F_1={\displaystyle \frac{2P\times R}{P+R}}$$

(9)

$$mIoU={\displaystyle \frac{1}{m}}{\displaystyle {\sum }_{i=1}^m\frac{TP}{TP+FP+FN}}$$

(10)

where m is the number of classes. The TP, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively.

3. Results

3.1. Overall Results of Trunk–Branch Segmentation of Trunk-Shaped Fruit Trees

This study focused on two trunk-shaped fruit tree species: apple and cherry trees. For each species, 20 trees were randomly selected, and their point cloud data underwent trunk–branch segmentation using an adaptive cuboid regional growth method. The segmentation result is shown in Figure 6, where the average P, R, F₁, and mIoU for apple trees were 96.05%, 95.45%, 95.75%, and 0.927, respectively, while those for cherry trees were slightly higher at 96.52%, 95.91%, 96.21%, and 0.938. These indicators remained at a high level throughout—the F₁ of both tree species exceeded 95.75%, and the mIoU exceeded 0.927—highlighting the robustness and versatility of our method.

It is worth noting that the mIoU indicator is introduced on the basis of the existing P, R, and F₁ global evaluation indicators. By calculating the intersection over union ratio between the predicted area and the actual area of a single branch, mIoU can effectively reflect the over-segmentation and under-segmentation of local areas, thereby providing a more comprehensive and rigorous performance evaluation. Experimental results show that this method exhibits good spatial adaptability and robustness in complex branch structures, not only achieving high-precision and high-stability trunk–branch segmentation but also showing excellent segmentation performance at the local branch level.

For apple trees, the lowest P, R, F₁, and mIoU were 94.15%, 92.71%, 93.80%, and 0.896, respectively, with segmentation results depicted in Figure 7a–c. The highest scores, 97.39%, 96.60%, 96.86%, and 0.954, are shown in Figure 7a,d. It should be noted that the same fruit tree may have multiple indicators reaching the maximum or minimum values. The lowest P and F₁ occurred in Figure 7b, where segmentation failed to distinguish upper branches due to their close attachment to the trunk and small branching angle, causing misclassification as trunk points and preventing regional growth activation. The lowest R in Figure 7c resulted from incomplete branch segmentation due to high branch density, leading to premature termination of the adaptive cuboid regional growth algorithm. Despite these issues, overall segmentation remained effective, confirming the method’s suitability for dormant apple trees.

Cherry trees exhibited superior segmentation, with maximum P, R, F₁, and mIoU reaching 98.26%, 97.88%, 98.07%, and 0.969 (Figure 8a,b). The lowest values, recorded in Figure 8c were 93.45%, 92.68%, 93.06%, and 0.901. The localized enlargement in Figure 8c reveals the presence of twigs, scabs, and overgrowth, which prevented small branches and nodules within the cylindrical fitting radius from activating regional growth, limiting segmentation accuracy improvements. Despite these challenges, the overall metrics confirm the method’s effectiveness in segmenting dormant cherry trees.

Both species exhibited overgrowth issues, where trunks were misclassified as branches, as well as incomplete growth, leaving residual short branches on the trunk. However, overall evaluation metrics alone do not provide detailed insights into issues like overgrowth, incomplete growth, or misclassification. To refine the assessment, P, R, F₁, and mIoU were separately calculated for branches and trunks, enabling a more detailed evaluation of segmentation performance and areas for improvement.

3.2. Analysis of Evaluation Indicators for Trunk and Branches of Apple Trees

As shown in

Table 2. Evaluation indicators for trunk and branch in apple tree segmentation.

Division	Branch			Trunk
	Min	Max	Mean	Min	Max	Mean
P	95.54%	99.34%	97.90%	82.18%	94.17%	89.01%
R	94.22%	98.12%	96.57%	83.06%	97.83%	91.01%
F1	96.23%	98.07%	97.22%	83.60%	94.12%	89.91%

, P, R, and F₁ were separately calculated for apple tree branches and trunks. The average values for branches exceeded 96.57%, while those for trunks were only above 89.01%, indicating significantly better segmentation performance for branches. To investigate this disparity, cases with the lowest P, R, and F₁ were analyzed, identifying factors such as overgrowth, misclassification, and incomplete segmentation that affect trunk accuracy.

For branch segmentation, the minimum P, R, and F₁ were 95.54%, 94.22%, and 96.23%, respectively, with results in Figure 7b,c,f. The maximum values reached 99.34%, 98.12%, and 98.07%, shown in Figure 7a,d,f. The method consistently achieved high segmentation accuracy, with F₁ exceeding 96%, demonstrating its robustness. However, challenges remain: overgrowth (Figure 7f) led to erroneous trunk segmentation as multiple branches, while incomplete branch growth (Figure 7g) reduced accuracy. These issues highlight the need for further refinement in complex growth scenarios.

For trunk segmentation, all P, R, and F₁ values exceeded 82.00% (Table 2). The lowest P, R, and F₁ were 82.18%, 83.06%, and 83.60% (Figure 7b,h), while the highest reached 94.17%, 97.83%, and 94.12% (Figure 7d,e,g). Variability in trunk features and occlusion led to false identification and missed detections, impacting accuracy. Further optimization is planned to improve trunk segmentation under diverse conditions.

In conclusion, while the method performs well, particularly for branch segmentation, challenges remain in handling structural complexity and occlusions in trunk segmentation. Continuous refinement is necessary to enhance accuracy and robustness in practical applications.

3.3. Analysis of Evaluation Indicators for Trunk and Branches of Cherry Trees

As shown in

Table 3. Evaluation indicators for trunk and branch segmentation in cherry trees.

Division	Branch			Trunk
	Min	Max	Mean	Min	Max	Mean
P	95.97%	99.77%	98.67%	87.34%	95.97%	92.18%
R	90.13%	98.01%	95.27%	90.27%	99.06%	96.98%
F1	94.36%	98.56%	96.93%	90.50%	96.98%	94.50%

, the P, R, and F₁ for branches and trunks of cherry trees were calculated separately. The average values for branches exceeded 95.27%, while those for trunks were above 92.18%, indicating that branch segmentation performed significantly better than trunk segmentation.

The minimum P, R, and F₁ for branch segmentation were 95.97%, 90.13%, and 94.36%, respectively, as illustrated in Figure 8b,e. The maximum values reached 99.77%, 98.01%, and 98.56%, as shown in Figure 8a,f. Despite occlusions causing fragmentation in the 3D point cloud during LiDAR scanning, the proposed method effectively segmented trunks and branches, achieving satisfactory results. Notably, the cherry tree in Figure 8b attained P and F₁ of 99.77% and 98.56%, respectively, demonstrating the algorithm’s strong performance and practical potential. Overall, P, R, and F₁ for branches exceeded 90.00% for all samples, confirming the method’s accuracy and robustness.

For trunks, the P, R, and F₁ (Table 3) were all above 87.00%. The minimum values were 87.34%, 90.27%, and 90.50%, as shown in Figure 8c,f, while the maximum values reached 95.97%, 99.06%, and 96.98%, presented in Figure 8a,d,g. However, challenges remain, as seen in Figure 8f, where the upper trunk was misclassified as a branch due to insufficient point cloud density, causing segmentation errors. Enhancing accuracy in trunk segmentation requires adjustments and a comprehensive analysis of point cloud data, integrating branch segmentation results for more reliable outcomes.

4. Discussion

4.1. Comparison with Geometry-Based 3D Point Cloud Branch Segmentation Methods

To address the structural segmentation of two representative trunk-shaped fruit trees, this study proposes an ACRGS method for trunk–branch segmentation. We conducted a systematic comparative analysis with several representative geometry-based 3D point cloud branch segmentation methods that are most relevant to our task.

Compared to recent advancements in the field, Zhang [39] employed BLS to randomly select eight apple trees from three orchards and introduced a hierarchical growth approach suitable for edge deployment. Their method achieved semantic segmentation, instance segmentation, and phenotypic feature extraction of apple tree branch point clouds, with semantic segmentation precision and recall reaching 85.80% and 99.50%, respectively. Li [40] proposed a branch segmentation and phenotypic parameter extraction method for apple trees based on an improved Laplacian algorithm, which combines CH-GMM registration and multi-scale skeleton extraction. The method achieved precision, recall, and F₁-score of 93.7%, 96.2%, and 92.6%, respectively. Yrttimaa [41] proposed a segmentation method based on Cartesian coordinate transformation and morphological filtering for Scots pine, reporting an average precision, recall, and F₁-score of 92%, 75%, and 82%, respectively. However, most methods based on geometric attempts do not release their source code, which makes them difficult to reproduce, yielding a generalization performance. These limitations significantly hinder their widespread application in real-world orchard environments. Therefore, this study focused primarily on comparing segmentation accuracy with these methods.

Compared with the above methods, the proposed ACRGS method demonstrates superior performance in trunk–branch structural segmentation tasks for both apple and cherry trees. Specifically, the average precision for branch segmentation reached 97.90% and 98.67%, respectively, significantly outperforming existing approaches. The average recall rates for trunk segmentation were 89.01% and 92.18%, respectively. This method not only markedly improves the accuracy of branch segmentation but also exhibits strong generalization capability and robustness. Nonetheless, there remains room for improvement in trunk segmentation accuracy, which could be further enhanced in the future by integrating structural priors or morphological modeling.

4.2. Comparison with Deep Learning-Based Instance Segmentation Models

In recent years, deep learning methods have achieved significant progress in point cloud segmentation tasks. In order to comprehensively evaluate the effectiveness of our proposed ACRGS model in complex tree structure instance segmentation, this study selected three of the most representative mainstream open-source deep learning models in the field of point cloud segmentation—PointNet++ [42], PointNeXt [43], and Point Transformer [44]—to conduct segmentation comparison experiments on point cloud data of apple trees and cherry trees.

During model selection and optimization, this study referred to several research works that have proposed improvements for fruit tree scenarios. Targeted model reproduction and training were conducted, including parameter configuration, structural adjustments, and hyperparameter tuning. Sun [45] proposed a PointNet++-based method for segmenting main branches of apple trees during the leaf-off season. Their method performs semantic segmentation to classify the point cloud into trunk, main branches, and endpoints and then applies skeleton extraction and path analysis for instance-level branch segmentation. The overall segmentation accuracy (OA) reached 0.84. Jiang [46] employed PointNeXt for organ-level instance segmentation of apple trees. By combining DBSCAN clustering with graph structure analysis, they achieved high-precision segmentation of complex multi-branch structures. Based on the semantic segmentation results from PointNeXt and SoftGroup++, their method achieved mAP_50 scores of 0.842 and 0.815, respectively, demonstrating strong robustness and adaptability. In addition, Guo [47] proposed the CTpoint architecture, which integrates local convolution operations with a global Transformer mechanism. By introducing a feature transmission module, it effectively fuses local details with global structural information, significantly improving segmentation performance in tasks based on the Point Transformer framework.

Building on these foundations, the present study introduces the ACRGS method and compares its performance against the aforementioned learning-based models. The results are presented in

Table 4. Post-processed instance segmentation performance of different models on apple and cherry tree datasets.

Division	PointNet++	PointNext	Point Transformer	ACRGS (Ours)
P_apple	78.09%	85.41%	89.14%	96.05%
R_apple	76.58%	84.73%	87.57%	95.45%
F1_apple	77.33%	85.07%	88.35%	95.75%
mIoU_apple	0.658	0.814	0.832	0.927
P_cherry	79.81%	86.22%	91.38%	96.52%
R_cherry	78.33%	85.81%	90.04%	95.91%
F1_cherry	79.07%	86.02%	90.71%	96.21%
mIoU_cherry	0.674	0.821	0.857	0.933
Model size (MB)	20.4	11.6	222.2	0.274

Note: Bold indicates the best performance.

.

According to Table 4, PointNet++ has the lowest segmentation performance, with F₁ of 77.33% on the apple tree dataset and 79.07% on the cherry tree dataset and corresponding mIoU values of 0.658 and 0.674, respectively. This poor performance is clearly visible in Figure 9I(a,c). The poor segmentation performance is due to the limited ability of the model to capture local structural details and the dense branch structure. Multiple branches are either merged or completely missed, and there are a large number of errors in the segmentation of local junctions between trunks and branches. PointNeXt shows moderate improvement, reaching F₁ of 85.07% and 86.02% and mIoU of 0.814 and 0.821, respectively. However, it still struggles in more complex scenes like Figure 9II(b,c), often resulting in single branches being segmented as trunks. Point Transformer further improves global reasoning, achieving F₁ of 88.35% and 90.71% with mIoU of 0.832 and 0.857. Yet, as shown in Figure 9III, it occasionally merges adjacent but structurally distinct branches, especially in complex branch structure regions such as Figure 9III(a,b).

In contrast, ACRGS significantly outperforms all baselines, with F₁ of 95.75% on the apple dataset and 96.21% on the cherry dataset and the highest mIoU of 0.927 and 0.933. In Figure 9IV, its advantages are evident across all samples, particularly in Figure 9IV(a,c), where fine, overlapping branches are accurately segmented without over-segmentation or under-segmentation. The model consistently preserves instance boundaries, even in regions with occlusion or high structural complexity, demonstrating strong robustness and precise structural understanding.

Beyond accuracy, ACRGS offers a dramatic advantage in efficiency. With a model size of just 0.274 MB, it is significantly more lightweight than PointNeXt (11.6 MB) and Point Transformer (222.2 MB). More importantly, ACRGS is a non-learning, rule-based algorithm that requires no annotated datasets, no pretraining, and no high-performance hardware. This makes it ideally suited for deployment on low-power edge computing devices such as mobile scanning units or field robots. Its minimal resource footprint and ease of deployment enhance its practical applicability in orchard automation systems, where both precision and efficiency are critical.

4.3. Adaptability and Limitations in Trunk–Branch Structure Segmentation

Beyond overall instance segmentation performance, this study further investigated the model’s structural-level capability in distinguishing between trunks and branches. Results demonstrate that the proposed ACRGS method exhibits high adaptability and structural recognition accuracy in trunk–branch separation tasks for trunk-structured fruit trees, showing particular effectiveness on structurally complex and densely branched apple trees.

At the species level, the segmentation performance on cherry trees was generally superior to that on apple trees. This discrepancy primarily stems from intrinsic biological and morphological differences. Apple tree trunks often exhibit curvature, displacement, and torsion, coupled with a higher density and number of closely spaced branches, which increases spatial ambiguity between trunks and branches. In contrast, cherry trees tend to have more upright trunks and thicker, fewer branches. Additionally, modern orchard management practices, such as branch training to promote horizontal branch orientation, create clear structural layering, thereby simplifying the segmentation task significantly.

From the perspective of segmentation targets, branch segmentation generally outperforms trunk segmentation. This can be attributed to the outside-in region-growing strategy employed by ACRGS. In cases where termination criteria are insufficiently precise, core trunk areas may be partially uncovered or mistakenly classified as branch extensions. Moreover, factors such as data quality, point cloud density, and orchard management conditions also influence performance. Especially in cases involving irregular tree forms or significant structural occlusions, trunk identification becomes more challenging, occasionally resulting in overfitting or underfitting.

In summary, the ACRGS method demonstrates strong adaptability and generalization capability when dealing with diverse trunk–branch structures and complex tree architectures. However, in extreme cases (e.g., heavy overlap or adhesion between trunks and branches), further refinement is needed. Future research may consider integrating maximum inscribed sphere fitting for trunk skeleton extraction to enhance central point localization and improve hierarchical structural recognition, particularly in regions of severe structural overlap. Additionally, expanding experimental validation across more species and varied morphologies will be essential to developing a more generalizable segmentation framework for fruit tree point clouds. This, in turn, will support structural modeling, phenotypic analysis, and precision orchard management across diverse fruit tree types, thereby providing a robust foundation for intelligent agriculture applications.

5. Conclusions

This study presents an adaptive cuboid regional growth segmentation (ACRGS) algorithm for high-precision segmentation of trunks and branches in trunk-shaped fruit trees. The method was specifically validated on dormant apple and cherry trees using TLS data. The specific conclusions of this study include the following:

(1) This study innovatively proposes a regional reverse growth strategy, incorporating the growth convergence rate (G_cr) as a dynamic optimization parameter to regulate the termination conditions of growth. This approach significantly enhances regional growth efficiency and segmentation accuracy. Compared with PointNet++, PointNeXt, and Point Transformer deep learning models, ACRGS demonstrated superior trunk–branch segmentation performance in both F₁ and mIoU metrics, confirming its effectiveness for apple and cherry tree data;

(2) The algorithm integrates adaptively adjusted growth vectors and dynamically scaled growth cuboids, along with a radial growth coefficient (R_gc)-based mechanism for automatic growth initiation. These innovations enable the method to effectively simulate branch growth for instance segmentation, significantly improving segmentation accuracy. Future work will expand its applicability to other tree species and morphological types through further experiments and algorithm improvements.