Rubber tree segmentation method based on rubber tree net

Abstract

The application discloses a rubber tree segmentation method based on RubberTreeNet, which enhances the local detail capture and global background information fusion capability through a point cloud adaptive transformation attention (PTCA) mechanism, thereby realizing high-precision segmentation of single rubber trees, step one, collecting rubber tree point cloud data and saving in a ply format, and using CloudCompare software for rendering and preliminary processing; the independent rubber trees in the point cloud data are separated from the background; step two, feature extraction and fusion are performed on the point cloud data through the point cloud adaptive transformation attention (PTCA) mechanism; step three, the original resolution is restored through feature aggregation, pooling operation and decoding mapping layer, and a classifier is applied to predict each point, and the segmentation result of the single rubber tree is output. The method has the characteristics of high recognition accuracy, and the recognition accuracy reaches 92.45%.

Description

Technical Field

[0001] This invention relates to a rubber tree segmentation method based on RubberTreeNet. Background Technology

[0002] Rubber tree (Hevea brasiliensis), as an important economic crop, brings economic benefits and improved management efficiency when planted on a large scale. However, it also faces challenges due to inaccurate pest and disease management and growth monitoring. Traditional methods for obtaining rubber tree structural parameters rely on manual on-site measurements, which are not only time-consuming and labor-intensive but also limited by geographical accessibility. To improve this, researchers have begun to use laser scanning technology and point cloud data to enhance the efficiency and accuracy of obtaining parameters for individual rubber trees. Individual rubber tree segmentation technology enables precise monitoring and management of the structural parameters of each tree, which is crucial for forest surveys and health monitoring. However, in practice, problems such as severe overlap between rubber trees, ground interference, insufficient capture of local details, and difficulties in fusing global background information all pose challenges to accurate rubber tree segmentation. To address these issues, researchers have developed various segmentation methods, including different algorithms based on airborne laser scanning (ALS) and terrestrial laser scanning (TLS / MLS) data, such as K-means clustering, region growing, and watershed transformation, aiming to improve the accuracy and efficiency of segmentation. With the development of deep learning technology, deep learning-based methods have also been introduced into rubber tree segmentation in order to achieve higher precision segmentation results.

[0003] In conclusion, it is necessary to design a method for rubber tree segmentation using deep learning. Summary of the Invention

[0004] The technical problems this invention aims to solve are mainly twofold. First, insufficient data volume: existing public datasets often contain a limited number of samples, insufficient to train high-performance segmentation models, and inconsistent annotation quality may contain errors or omissions, affecting model training effectiveness. Second, in rubber tree segmentation tasks, there is insufficient feature extraction, particularly in capturing local details and fusing global background information. To address the first problem, this invention constructs a new single-tree rubber tree dataset. High-resolution 3D point cloud data of Hainan rubber plantations is collected using LiDAR to ensure the capture of the rubber tree's canopy, vertical structure, and lower canopy information. The data is saved in PLY format, rendered using CloudCompare software, and manually segmented into individual rubber trees. We selected 1000 trees for analysis, ultimately choosing 600 for experiments to ensure data diversity and representativeness. By constructing this self-built dataset, we effectively solve the problems of insufficient data volume and annotation quality, improving model training performance. To address the second issue, this invention proposes the PTCA module. This method effectively improves the ability to capture local details and fuse global background information in rubber tree segmentation tasks by combining Point Transformer and Criss-Cross Attention mechanisms. The PTCA module combines the advantages of Point Transformer and Criss-Cross Attention modules. It utilizes the self-attention mechanism of Point Transformer to finely extract local features from 3D point clouds, ensuring that each point can fully capture the geometric information of its neighborhood. Simultaneously, the Criss-Cross Attention mechanism captures long-distance dependencies in the spatial dimension, enhancing global relevance and selectivity. Through dual feature extraction and fusion, the PTCA module significantly reduces information loss and enhances detail capture capabilities, thereby improving the model's feature representation ability and robustness.

[0005] The technical solution of the invention is as follows:

[0006] A rubber tree segmentation method based on RubberTreeNet is characterized by employing a self-built training set and an adaptive transformation attention mechanism to improve the accuracy and robustness of single rubber tree segmentation in complex backgrounds. The overall process includes the following steps:

[0007] Step 1: Collect and preprocess point cloud data of rubber trees;

[0008] Point cloud data of rubber trees was collected and saved in .ply format. CloudCompare software was used for rendering and preliminary processing, including noise removal and removal of non-target objects. Subsequently, individual rubber trees in the point cloud data were separated from the background by manual selection and labeling, and exported in .txt format for further processing.

[0009] Step 2: Employ the Point Cloud Adaptive Transform Attention (PTCA) mechanism;

[0010] A Point Cloud Adaptive Transform Attention (PTCA) mechanism is employed, combining the advantages of Point Transformer and Criss-Cross Attention. Point Transformer refines local features of 3D point clouds through self-attention, ensuring each point fully captures the geometric information of its neighborhood; Criss-Cross Attention captures long-range dependencies in the spatial dimension, enhancing global relevance and selectivity. The PTCA module, through dual feature extraction and fusion, addresses the problem of insufficient local detail capture and global background information fusion, significantly reducing information loss and enhancing detail capture capabilities.

[0011] Step 3: Multi-level feature extraction and segmentation process;

[0012] Next, farthest point sampling (FPS) is used to select representative points, and K nearest neighbor (KNN) search is used to obtain the neighborhood point information of these representative points. Finally, the features are restored to the original resolution through multi-scale feature aggregation (MSFA), pooling operation and decoding mapping layer, and a classifier is applied to predict each point to achieve accurate segmentation of a single rubber tree.

[0013] The specific steps of step one are as follows:

[0014] First, a new dataset of individual rubber trees was constructed to better facilitate model training. Data collection was conducted in rubber plantations in Hainan Province, using LiDAR technology to acquire high-resolution 3D point cloud data. This method ensures that the data for each rubber tree clearly captures its canopy, vertical structural features, and information about the lower parts of the canopy (such as the trunk and leaves). Data collection covered a large area of ​​rubber tree groves, and the collected data was saved in PLY file format. During data collection, the background included fallen leaves, the ground, and other extensively intersecting rubber trees. To facilitate subsequent processing, CloudCompare software was used to quickly and efficiently render the large amount of point cloud data, allowing users to intuitively view and interact with the data on the screen. Severely intersecting rubber tree point cloud data was segmented into individual rubber trees using a manual selection method, and each tree was labeled. Individual rubber trees were labeled as 1, while non-individual rubber trees (including the ground, fallen leaves, and intersecting leaves) were labeled as 0. This segmentation can be based on spatial location, color, or other attributes. Next, we exported the segmented point cloud data as TXT files to facilitate subsequent network processing. We used CloudCompare software to segment thousands of rubber trees from a large rubber plantation, and selected 1000 trees for detailed analysis to ensure data diversity and representativeness, covering different growth stages and complex background conditions. Finally, we selected 600 rubber trees for experiments to ensure feasibility and statistical reliability, while also facilitating resource management and analysis. These datasets include not only ground data but also point cloud data from individual rubber trees and portions of the tree data intersecting with individual trees. This type of dataset helps the model learn the complete features of the rubber tree. Through these steps, we ensured the quality and diversity of the dataset, providing a solid foundation for subsequent model training.

[0015] In step two, this module leverages the sparsity and irregularity of point clouds, using a self-attention mechanism to weight the neighborhood of each point, thereby effectively extracting local and global features, as shown in the following formula.

[0016]

[0017] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, d k K is the dimension of the key vector. TThis is the transpose of the key matrix. First, the input feature X undergoes a linear transformation to generate Q, K, and V. Then, attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid gradient vanishing or exploding problems. The scaling and normalization steps ensure computational stability and reasonable numerical range. Next, the Softmax function is applied to normalize the scaled attention scores, yielding the attention weights. Finally, the relative position vector p between each pair of points is calculated. i -p j The data is then processed by a multilayer perceptron (MLP) to generate a location code.

[0018] Combining positional encoding with self-attention output enhances the model's understanding of spatial relationships. Then, the Cross-Channel Attention (CCTA) module first performs convolution and pooling operations on the input features for normalization. It then calculates the cross-similarity between input features, combining the maximum value with the original input to generate an enhanced feature representation. Within the CCTA module, attention weights in the horizontal and vertical directions are calculated as follows:

[0019]

[0020] X h Horizontal attention weights, W h The horizontal weight matrix, g i,j b represents the feature value of the input feature map at position (i,j). h This represents the horizontal offset term. X v Vertical attention weights, W v The weight matrix in the vertical direction, g j,i b represents the feature value of the input feature map at position (j,i). v The term represents the horizontal offset.

[0021] Then, the cosine similarity between the horizontal and vertical features is calculated.

[0022]

[0023] Where cosθ represents the cosine similarity between two vectors. The value of the i-th feature in the horizontal vector. This represents the value of the i-th feature in the vertical vector. n represents the dimension of the vector.

[0024] Select important features based on a set threshold and discard irrelevant information:

[0025] When cos h ≤thr, keep as 0;

[0026] When cos h>thr, keep it as 1.

[0027] Finally, the features processed by CCTA are combined with the original input features to generate the final output features:

[0028]

[0029] Among them, W Q W K and W V These are the linear transformation weight matrices for the query, key, and value, respectively. The self-attention mechanism uses the dot product similarity formula... To measure the relationship between input features, after softmax normalization, a weighted sum is obtained to get the attention output, and finally, the output is transformed W. O The Point Transformer result is obtained. Through the above steps, the PTCA module achieves efficient feature extraction and aggregation of point cloud data, making it particularly suitable for target recognition and segmentation tasks in complex backgrounds.

[0030] In step three, firstly, local center points are selected from the original point set using farthest point sampling (FPS), reducing the number of points to half. Next, the k-nearest neighbor (kNN) algorithm is used to group each center point with k spatial neighbors, forming a local 3D region. Then, feature aggregation is performed, calculating feature vectors for each center point and its neighbors, extracting geometric features using learnable operators, and aggregating features through max pooling and average pooling. Next, pooling operations further reduce the dimensionality of the features while retaining important information. Subsequently, decoding mapping is performed, gradually restoring the spatial distribution of features through a multi-layer decoder, mapping high-dimensional features back to the original point cloud space, and recovering the spatial structure of the point cloud. Finally, the decoded features are input into a classifier to classify each point, outputting class labels, and completing the point cloud segmentation task. Through this series of steps, the entire process can effectively process point cloud data, extract and aggregate features, and ultimately achieve accurate point cloud segmentation.

[0031] Beneficial effects:

[0032] The RubberTreeNet-based point cloud segmentation network of this invention provides comprehensive technical support for rubber tree segmentation in complex scenarios by performing meticulous preprocessing on a self-built dataset and introducing the PTCA (Point Transformer Criss-Cross Attention) module, which significantly improves the overall performance of the rubber tree segmentation task.

[0033] Firstly, to make the data more closely resemble reality and to broaden the adaptability of the model designed in this invention, we constructed a high-quality dataset of individual rubber trees through meticulous data preprocessing steps. Using CloudCompare software, we efficiently rendered and segmented the collected point cloud data, separating severely overlapping rubber tree point cloud data into independent individual rubber trees, which were then labeled separately. This process not only ensured the diversity and representativeness of the dataset but also provided clear and accurate data support for model training, further improving the model's generalization ability and segmentation performance.

[0034] Secondly, to enhance the model's feature extraction capability from rubber tree point cloud data in complex scenes and improve segmentation accuracy, the PTCA (Point Transformer Criss-Cross Attention) module was introduced. This significantly enhanced the model's feature extraction capability from rubber tree point cloud data in complex scenes. The PTCA module combines the advantages of the Point Transformer's self-attention mechanism and the Criss-Cross Attention mechanism, effectively capturing local details and global background information in the point cloud data. This improves the model's ability to capture the geometric information and spatial relationships of rubber trees, thereby enhancing segmentation accuracy.

[0035] This invention aims to address the problem of insufficient local detail capture and global background information fusion in rubber tree segmentation tasks by constructing a high-quality single rubber tree dataset and employing a point cloud adaptive transform attention (PTCA) mechanism. This results in high-precision single rubber tree segmentation and significantly improves the model's segmentation performance and robustness under complex background conditions. Attached Figure Description

[0036] Figure 1 This is a general framework diagram of rubber tree segmentation according to the present invention;

[0037] Figure 2 This is a network framework diagram of rubber tree segmentation according to the present invention;

[0038] Figure 3 This is a framework diagram of the PTCA module for rubber tree segmentation in this invention;

[0039] Figure 4 This is a labeled illustration of the rubber tree segmentation dataset of the present invention. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0041] Figure 1The overall flowchart of RubberTreeNet is as follows: First, high-resolution 3D point cloud data is acquired and preprocessed. Data on rubber forests is obtained using LiDAR, and segmentation and labeling are performed using CloudCompare software. Next, Point Cloud Adaptive Transform Attention (PTCA) is used to enhance local detail capture and global correlation. Finally, representative points are selected using Farthest Point Sampling (FPS), and neighboring points are obtained through K-Nearest Neighbor (KNN) search. The original resolution is restored through feature aggregation, pooling operations, and a decoding mapping layer. A classifier is applied to predict each point, outputting the segmentation result of a single rubber tree. This improves the convergence speed and stability of the model, thereby achieving high-precision segmentation of single rubber trees in complex backgrounds.

[0042] Figure 2 The PTCA (Point Transformer Criss-Cross Attention) module combines the advantages of Point Transformer and Criss-Cross Attention mechanisms, aiming to enhance the capture of local details and the fusion of global contextual information. Specifically: Point Transformer: Utilizes a self-attention mechanism to weight the neighborhood of each point in the point cloud, ensuring that each point can fully capture the geometric information within its neighborhood. By calculating the dot product between the query, key, and value, Point Transformer can capture complex geometric relationships in point cloud data, thereby better understanding local features. Criss-Cross Attention: Captures long-distance dependencies in the spatial dimension, enhancing global relevance and selectivity. By calculating attention weights in the horizontal and vertical directions, Criss-Cross Attention enables the model to better understand and process complex point cloud structures, thereby enhancing the fusion of global features. The PTCA module performs best in the network front-end for two main reasons: First, the PTCA module effectively models the spatial relationships between points in the point cloud through point transformation layers and cross-channel attention mechanisms. This is crucial in the front-end feature extraction process, helping to better understand the geometric structure of the input data. Second, the PTCA module can integrate global contextual information at the front end, allowing the network to retain more global information in subsequent processing, thereby improving the overall network performance. This invention improves PointMLP by using the PTCA module, which plays a vital role in RubberTreeNet, significantly enhancing the model's understanding and segmentation capabilities of rubber tree point cloud data, ensuring the model's efficiency and accuracy in complex contexts.

[0043] like Figures 1-3The rubber tree segmentation method based on RubberTreeNet of the present invention includes the following steps:

[0044] Step 1: To enable the model to better learn the characteristics of individual rubber trees, the point cloud data of severely overlapping rubber trees needs to be segmented into independent individual rubber trees. This process is done manually, labeling each independent rubber tree as 1 and non-independent rubber trees (including the ground, fallen leaves, and overlapping leaves) as 0. This data segmentation helps the model better understand the structural characteristics of individual rubber trees during training. CloudCompare software is used to assist in data segmentation and labeling. CloudCompare is a powerful open-source point cloud processing software that can quickly and efficiently render large amounts of point cloud data to the screen, allowing users to interactively view and select point clouds. The point cloud data is segmented into independent individual rubber trees through manual selection and labeled accordingly. To ensure data diversity and representativeness, 1000 rubber trees were selected from the large amount of collected data for detailed analysis. Ultimately, 600 rubber trees were selected for the experiment. These datasets contain ground data, independent rubber trees, and point cloud data of rubber trees overlapping with independent rubber trees, which helps the model learn the complete characteristics of rubber trees. The training set to test set ratio was 8:2, meaning 480 trees were used for training and 120 trees for testing. These detailed preprocessing steps ensured data quality, enabling the model to perform better in single rubber tree segmentation tasks with complex backgrounds. Each step of the preprocessing was designed to maximize the value of the data and ensure the model could learn useful features, thereby providing more accurate rubber tree segmentation results in practical applications.

[0045] Step 2: The self-attention mechanism is the core of the Point Transformer, which aggregates features by calculating the importance of each point in its neighborhood. Specifically, the self-attention mechanism is implemented through the following formula (1):

[0046]

[0047] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, d k It is the dimension of the key vector. First, the input feature X is transformed linearly to generate Q, K, and V. Then, attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid the problems of gradient vanishing or exploding. The scaling and normalization steps ensure the stability of the computation and the reasonableness of the numerical range. In addition, the Point Transformer also incorporates positional encoding extracted directly from the point coordinates to effectively modulate the attention mechanism, as given in (2):

[0048] Positional Encoding(p i ,p j ) = MLP(p i -p j )#(2)

[0049] Calculate point p i and p j The relative position vector p between them i -p j This relative position vector is then input into the MLP, where it undergoes a series of fully connected layers and nonlinear activation functions to obtain a new feature vector, which serves as the output of the positional encoding. The core purpose of positional encoding is to enhance the model's understanding of spatial relationships by utilizing the relative positional information between points, which is particularly important when processing point cloud data without a clear order or topological structure.

[0050] Cross-Attention is an enhanced attention mechanism used to combine and process multiple different types of input features to capture the relationships between different features, thereby improving the model's understanding and performance. Cross-Attention generates weights in both vertical and horizontal directions, allowing the structured graph convolution output feature map to... Where H×W represents the size of the feature map. The input is fed into the horizontal attention unit, and the weight direction at the horizontal level and its calculation formula can be defined as shown in (3):

[0051]

[0052] Among them W h and b h It is the weight parameter, X h These are the coefficients for horizontal attention. The weights in the vertical direction are calculated by inverting the feature matrix, and the process is represented by the following equation (4):

[0053]

[0054] in This represents the transposed feature map. Therefore, we can obtain the vertical attention weights in the same way, as shown in formula (5):

[0055]

[0056] Among them W v and b v It is the weight parameter, X vIt is the coefficient along the vertical attention. In order to further process the enhanced feature matrix and discard any irrelevant information, we added a threshold filtering network after downsampling the output results to filter out noise or redundant features of non-rubber trees. The main function can be mapped to two levels of importance: important and secondary. The main principle is to map the composite enhanced matrix to the spatial domain and use attention to distinguish the cosine values ​​of the horizontal and vertical features. The calculation method is shown in equation (6). When the two functions are close to zero, it indicates that the angle is close to 90 degrees. At this time, it is determined to be a vector that is not related to a single independent rubber tree. By setting a threshold, useful information is enhanced and useless information is discarded. The specific method is as follows: when the threshold is lower than thr, the cosine distance is large, indicating that the rubber tree point cloud information obtained from feature points in different directions is not highly similar; when the threshold is higher than thr, that is, the higher the cosine similarity, the more obvious the rubber tree point cloud information is retained, as shown in equation (7). Based on the two scenarios above, we retain 0 for values ​​below the threshold and 1 for values ​​above the threshold to achieve better feature selection. This allows the attention network to focus on features with high activation values, reducing the misleading effect of complex backgrounds on the model, improving the sparsity of the feature matrix, and reducing unnecessary computational overhead.

[0057]

[0058]

[0059] The PTCA (Point Transformer Criss-Cross Attention) module combines the advantages of PointTransformer and CCTA modules, achieving efficient feature extraction and aggregation of point cloud data. PointTransformer captures the global context of input features through a self-attention mechanism, while CCTA further processes these features, enhancing the interaction of local features. This module is designed to effectively identify and segment target point clouds in complex backgrounds, such as segmenting a single rubber tree in a heavily intersecting background. The specific operation of the PCTA module can be represented by the following formula (8).

[0060]

[0061] In this formula, the input feature XX is first self-attention calculated using the Point Transformer module. Here, W Q W K and W V These are the linear transformation weight matrices for the query, key, and value, respectively. The self-attention mechanism uses the dot product similarity formula... To measure the relationship between input features, after softmax normalization, a weighted sum is obtained to get the attention output, and finally, the output is transformed W. O The Point Transformer result is obtained. Subsequently, the CCTA module further processes the Point Transformer output. CCTA uses a cross-channel attention mechanism, normalizing the input through convolution and pooling operations, calculating the cross-similarity X between two inputs, and then combining the maximum value with the original input to obtain the final feature CCTA(X+Maximum+inputs). The PTCA module effectively combines local and global features through the above process. It meticulously extracts local features from the 3D point cloud through a self-attention mechanism, ensuring that each point fully captures the geometric information of its neighborhood. Simultaneously, the module captures long-range dependencies in the spatial dimension, enhancing global relevance and selectivity.

[0062] To demonstrate the validity of this invention patent, the following experiment was conducted.

[0063] The dataset used in this invention consists of thousands of rubber trees segmented from a large rubber plantation collected using CloudCompare software. 1000 trees were selected for detailed analysis to ensure data diversity and representativeness, covering different growth stages and complex background conditions. Finally, 600 trees were chosen for experiments to ensure feasibility and statistical reliability, while also facilitating resource management and analysis. Each dataset contains ground data, individual rubber trees, and point cloud data of partial rubber trees intersecting with the individual trees. This data helps the model learn the complete characteristics of the rubber trees. Furthermore, this paper selected two common point cloud datasets containing complex scenes and conditions.

[0064] This invention was tested on a hardware environment consisting of an Intel(R) Xeon(R) Gold 6330 CPU and an NVIDIA RTX 3090 GPU. The PyTorch deep learning framework was used (PyTorch 1.8.1), Python 3.8 was used, and the system environment was Ubuntu 18.04.6LTS. The training iterations were 540. The batch size was 10. The learning rate was 0.001, and it was halved every 40 epochs. The ratio of training set to test set was 8:2. The training set contained 480 trees, and the test set contained 120 trees.

[0065] Table 1

[0066] Network model Acc mIou PointMLP 74.65％ 73.48％ SPOTR 88.65％ 79.75％ RubberTreeNet 92.45％ 86.35％

[0067] The experimental results are shown in Table 1, comparing the performance of three network models in the classification task: PointMLP, SPOTR, and RubberTreeNet. The results show that RubberTreeNet performs best in both accuracy (Acc) and mean intersection-over-union (mIoU), reaching 92.45% and 86.35%, respectively. SPOTR is second, with an accuracy of 88.65% and an mIoU of 79.75%. PointMLP performs relatively poorly, with an accuracy of 74.65% and an mIoU of 73.48%. This indicates that RubberTreeNet has higher classification accuracy and better segmentation performance in this task.

[0068] This invention discloses a rubber tree segmentation method based on RubberTreeNet. The method introduces a PTCA (Point Transformer Criss-Cross Attention) module into the backbone network, leveraging the self-attention mechanism of PointTransformer and the Criss-Cross Attention mechanism to enhance local detail capture and global background information fusion, reducing information loss and improving detail capture capabilities. High-precision single rubber tree segmentation is achieved, with RubberTreeNet achieving an accuracy of 92.45%, demonstrating excellent performance under limited computational resources. This invention provides reliable technical support for rubber tree segmentation in complex scenarios, significantly improving the segmentation effect and stability of the model in practical applications.

Claims

1. A rubber tree segmentation method based on RubberTreeNet, characterized in that, The network effectively identifies and segments individual rubber trees from complex point cloud data through a point cloud adaptive transformation attention mechanism and a multi-level feature extraction and segmentation process. The overall process includes the following steps: Step 1: Collect and preprocess point cloud data of rubber trees; Rubber tree point cloud data was collected and saved in .ply format. CloudCompare software was used for rendering and preliminary processing, including noise removal and removal of non-target objects. Subsequently, individual rubber trees in the point cloud data were separated from the background by manual selection and labeling, and exported in .txt format for further processing. Step 2: Employ the Point Cloud Adaptive Transformation Attention (PTCA) mechanism; The Point Cloud Adaptive Transformation Attention (PTCA) mechanism combines the advantages of Point Transformer and Criss-Cross Attention. Point Transformer extracts local features of 3D point clouds through self-attention, ensuring that each point fully captures the geometric information of its neighborhood. Criss-Cross Attention captures long-distance dependencies in the spatial dimension, enhancing global relevance and selectivity. The PTCA module solves the problem of insufficient local detail capture and global background information fusion through dual feature extraction and fusion, significantly reducing information loss and enhancing detail capture capabilities. The specific calculations of the PCTA module are represented by the following formula: ; in and These are the linear transformation weight matrices for the query, key, and value; the self-attention mechanism uses the dot product similarity formula. To measure the relationship between input features, after Softmax normalization, a weighted sum is obtained to get the attention output, and finally, the output is transformed. The Point Transformer results are obtained. Through the above steps, the PTCA module achieves efficient feature extraction and aggregation of point cloud data, which is suitable for target recognition and segmentation tasks in complex backgrounds. CCTA is a cross-channel attention module that uses a cross-channel attention mechanism to normalize the input through convolution and pooling operations, calculate the cross similarity X between two inputs, and then combine the maximum value and the original input to obtain the final feature CCTA. Step 3: Multi-level feature extraction and segmentation process; Next, representative points are selected using farthest point sampling, and neighborhood information of these representative points is obtained through K-nearest neighbor search. Finally, features are restored to the original resolution through multi-scale feature aggregation, pooling operations, and decoding mapping layer, and a classifier is applied to predict each point to achieve accurate segmentation of a single rubber tree.

2. The rubber tree segmentation method based on RubberTreeNet according to claim 1, characterized in that, The specific steps for step one are as follows: First, a new dataset of individual rubber trees was constructed to better facilitate model training. Data collection was conducted in rubber plantations in Hainan Province, using LiDAR technology to acquire high-resolution 3D point cloud data. This method ensured that the data for each rubber tree clearly captured its canopy, vertical structural features, and information about the lower part of the canopy. Data collection covered a large area of ​​rubber tree groves, and the collected data was saved in PLY file format. During the data collection process, the background included fallen leaves, the ground, and other severely overlapping rubber trees. CloudCompare software was used to quickly and efficiently render the large amount of point cloud data, allowing users to intuitively view and interact with the data on the screen. Severely overlapping rubber tree point cloud data were segmented into independent individual rubber trees through manual selection and labeled accordingly. Points were labeled as 1, while non-independent rubber trees were labeled as 0; this segmentation was based on spatial location, color, or other attributes; the segmented point cloud data was exported as a TXT file; using CloudCompare software, thousands of rubber trees were segmented from a large area of ​​rubber forest, and 1000 rubber trees were selected for detailed analysis to ensure data diversity and representativeness, covering different growth stages and complex background conditions; finally, 600 rubber trees were selected for experiments to ensure the feasibility and statistical reliability of the experiments, while also facilitating resource management and analysis; these datasets include not only ground data, but also point cloud data of independent rubber trees and rubber trees intersecting with independent rubber trees, such a dataset helps the model learn complete rubber tree features; ensuring the quality and diversity of the dataset provides a solid foundation for subsequent model training.

3. The rubber tree segmentation method based on RubberTreeNet according to claim 1, characterized in that, The specific steps for step two are as follows: This module leverages the sparsity and irregularity of point clouds, employing a self-attention mechanism to weight the neighborhood of each point, thereby effectively extracting local and global features, as shown in the following equation: ； Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. It is the dimension of the key vector; first, the input feature X is transformed linearly to generate Q, K and V; then, attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid the problems of gradient vanishing or exploding. Scaling and normalization steps ensure the stability of the calculation and the reasonableness of the numerical range; Then, the Softmax function is applied to normalize the scaled attention scores to obtain the attention weights; Next, calculate the relative position vector between each pair of points. The data is then processed by a multilayer perceptron to generate a location code. By combining positional encoding with self-attention output, the model's understanding of spatial relationships is enhanced. Then, the Cross-Channel Attention (CCTA) module is entered. First, convolution and pooling operations are performed on the input features for normalization. The cross-similarity between input features is calculated, and the maximum value and the original input are combined to generate an enhanced feature representation. In the CCTA module, attention weights in the horizontal and vertical directions are calculated as follows: ； ； Horizontal attention weights Horizontal weight matrix, This represents the feature value of the input feature map at position (i,j). The term representing the horizontal offset; Vertical attention weights The weight matrix in the vertical direction, This represents the feature value of the input feature map at position (j,i). The term representing the horizontal offset; Then, the cosine similarity between the horizontal and vertical features is calculated; ； in Cosine similarity is used to represent the cosine similarity between two vectors. The value of the i-th feature in the horizontal vector; This represents the value of the i-th feature in the vertical vector; Indicates the dimension of the vector; Select important features based on a set threshold and discard irrelevant information: when Reserved as 0; when Reserved as 1; Finally, the features processed by CCTA are combined with the original input features to generate the final output features.

4. The rubber tree segmentation method based on RubberTreeNet according to claim 1, characterized in that, Step 3 consists of the following steps: First, local center points are selected from the original point set using farthest point sampling, reducing the number of points to half. Next, the k-nearest neighbor algorithm is used to group each center point with k spatial neighbors, forming a local 3D region. Then, feature aggregation is performed, calculating feature vectors for each center point and its neighbors, extracting geometric features using learnable operators, and aggregating features through max pooling and average pooling. Next, pooling operations are used to reduce the dimensionality of the features while retaining important information. Subsequently, decoding and mapping are performed, gradually restoring the spatial distribution of features through a multi-layer decoder, mapping high-dimensional features back to the original point cloud space, and restoring the spatial structure of the point cloud. Finally, the decoded features are input into a classifier to classify each point, outputting category labels, and completing the point cloud segmentation task. Through this series of steps, the entire process can effectively process point cloud data, extract and aggregate features, and ultimately achieve accurate point cloud segmentation.

Images