Rubber tree segmentation method based on RubberTreeNet

Abstract

The invention discloses a rubber tree segmentation method based on RubberTreeNet, through a point cloud adaptive transform attention (PTCA) mechanism, local detail capture and global background information fusion capability are enhanced, and high-precision segmentation of a single rubber tree is further realized, the method comprises the following steps: 1, collecting rubber tree point cloud data, storing the data in a ply format, and performing rendering and primary processing by using Cloud Compare software; comprising the steps of removing noise and non-target objects. Then, through manual selection and marking, an independent rubber tree in the point cloud data is separated from a background, and the independent rubber tree is exported as a txt format for subsequent processing; step 2, feature extraction and fusion are carried out on the point cloud data through a point cloud adaptive transformation attention (PTCA) mechanism so as to enhance local detail capture and global background information fusion capability; and step 3, selecting representative points by using farthest point sampling (FPS), obtaining neighborhood points through K-nearest neighbor (KNN) search, finally recovering the original resolution through feature aggregation, pooling operation and a decoding mapping layer, predicting each point by using a classifier, and outputting a segmentation result of a single rubber tree. The method provided by the invention has the characteristic of high recognition accuracy, and the recognition precision reaches 92.45%.

Description

Technical Field

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

[0002] As an important economic crop, rubber tree (Hevea brasiliensis) can bring economic benefits and improve management efficiency when planted on a large scale, but it also faces challenges brought by inaccurate pest management and growth monitoring. The traditional method of obtaining rubber tree structural parameters relies on manual field measurement, which is not only time-consuming and labor-intensive, but also limited by geographical accessibility. To improve this, researchers began to use laser scanning technology and point cloud data to improve the efficiency and accuracy of parameter acquisition of individual rubber trees. Through the individual rubber tree segmentation technology, the structural parameters of each rubber tree can be accurately monitored and managed, which is crucial for forest inventory and health monitoring. However, in actual operation, there are problems such as serious intersection between rubber trees, ground interference, insufficient capture of local details, and difficulty in fusing global background information, which bring challenges to the accurate segmentation of rubber trees. To solve these problems, researchers have developed a variety of 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, watershed transform, etc., 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 effects.

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

[0004] The technical problems to be solved by the present invention mainly have two points. The first point is that the amount of data is insufficient: the existing public data sets often contain a limited number of samples, which is not enough to train a segmentation model with excellent performance. At the same time, the quality of annotations is uneven, and there may be errors or omissions, which affects the effect of model training. The second point is that in the rubber tree segmentation task, the problem of feature extraction, especially the problem of insufficient fusion of local details of rubber trees and global background information. For the first problem, the present invention constructs a new single rubber tree data set. High-resolution 3D point cloud data of Hainan rubber forests are collected by laser radar to ensure that the crown, vertical structure and lower canopy information of rubber trees are captured. The data is saved in ply format, and CloudCompare software is used to render and manually segment an independent single rubber tree. We selected 1000 rubber trees for analysis, and finally selected 600 for experiments to ensure data diversity and representativeness. Through self-built data sets, we effectively solved the problems of insufficient data volume and annotation quality, and improved the model training effect. In response to the second problem, the present invention proposes a PTCA module, which effectively improves the local detail capture and global background information fusion capabilities of rubber trees in 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, and uses the self-attention mechanism of Point Transformer to finely extract local features of 3D point clouds, ensuring that each point can fully capture the geometric information of its neighborhood. At the same time, the Criss-Cross Attention mechanism captures long-distance dependencies in the spatial dimension, enhancing global relevance and selectivity. The PTCA module significantly reduces information loss and enhances detail capture capabilities through dual feature extraction and fusion, thereby improving the feature expression capability and robustness of the model.

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

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

[0007] Step 1: Collect and pre-process rubber tree point cloud data;

[0008] The collected rubber tree point cloud data was saved in ply format and rendered and preliminarily processed using CloudCompare software, including the removal of noise and non-target objects. Subsequently, the independent rubber trees in the point cloud data were separated from the background by manual selection and marking, and exported to txt format for subsequent processing;

[0009] Step 2: Use point cloud adaptive transformation attention mechanism (PTCA);

[0010] The Point Cloud Adaptive Transformation Attention (PTCA) mechanism is adopted, which combines the advantages of Point Transformer and Criss-Cross Attention. Point Transformer finely extracts local features of 3D point clouds through the self-attention mechanism to ensure that each point fully captures the geometric information of the neighborhood; Criss-Cross Attention captures long-distance dependencies in the spatial dimension and enhances global relevance and selectivity. The PTCA module solves the problem of insufficient fusion of local detail capture and global background information through dual feature extraction and fusion, significantly reduces information loss, and enhances detail capture capabilities;

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

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

[0013] The specific steps of step one are:

[0014] First, a new dataset of individual rubber trees was constructed to better train the model. Data collection was conducted in Hainan rubber forests, and high-resolution 3D point cloud data was obtained by using LiDAR technology. This method ensures that the data of each rubber tree can clearly capture the vegetation canopy, the vertical structural characteristics of the tree, and the information of the lower part of the canopy (such as the trunk and leaves). The data collection covered a large area of ​​rubber trees and was saved in PLY file format. During the data collection process, the background included fallen leaves, ground, and other large areas of heavily crossed rubber trees. In order to facilitate subsequent processing, we used CloudCompare software to quickly and efficiently render a large amount of point cloud data, allowing users to intuitively view and interact with the data on the screen. Through a manual selection method, the severely crossed rubber tree point cloud data was segmented into independent individual rubber trees and marked separately. Independent individual rubber trees were marked as 1, while non-independent rubber trees (including ground, fallen leaves, and crossed leaves) were marked as 0. This segmentation can be based on spatial location, color, or other attributes. Next, we exported the segmented point cloud data into TXT file format to facilitate subsequent network processing. We used CloudCompare software to segment thousands of rubber trees from the large rubber forests we collected, and selected 1,000 rubber trees for detailed analysis to ensure the diversity and representativeness of the data, covering different growth stages and complex background conditions. Finally, 600 rubber trees were selected for the experiment to ensure the feasibility and statistical reliability of the experiment, while facilitating resource management and analysis. These datasets include not only ground data, but also independent rubber trees and some rubber tree point cloud data that intersect with independent rubber trees. Such datasets help the model learn complete rubber tree features. Through the above steps, we ensure the quality and diversity of the dataset, providing a solid foundation for subsequent model training.

[0015] In step 2, the module uses the sparsity and irregularity of the point cloud to weight the neighborhood of each point through the self-attention mechanism, thereby effectively extracting local and global features, which is expressed as follows:

[0016]

[0017] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. k is the dimension of the key vector, K Tis the transpose of the key matrix. First, the input feature X is linearly transformed to generate Q, K, and V. Then, the attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid the problem of vanishing or exploding gradients. The scaling and normalization steps ensure the stability of the calculation and the rationality of the numerical range. Then, the Softmax function is applied to normalize the scaled attention scores to obtain the attention weights. Next, the relative position vector p between each pair of points is calculated i -p j , and processes it through a multi-layer perceptron (MLP) to generate a positional encoding.

[0018] Combining position encoding with self-attention output enhances the model’s understanding of spatial relationships. Then, entering the cross-channel attention (CCTA) module, the input features are first convolved and pooled, and normalized. 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, the attention weights in the horizontal and vertical directions are calculated, respectively:

[0019]

[0020] X h : Horizontal attention weight, W h The horizontal weight matrix, g i,j represents the eigenvalue of the input feature map at position (i, j), b h Represents the horizontal bias. v : vertical attention weight, W v The weight matrix in the vertical direction, g j,i represents the eigenvalue of the input feature map at position (j,i), b v Represents the horizontal bias.

[0021] Then, the cosine similarity of 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. Represents the value of the i-th feature in the vertical vector. n represents the dimension of the vector.

[0024] Filter out important features based on the set threshold and discard irrelevant information:

[0025] When cos h ≤thr, reserved 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 are the linear transformation weight matrices of 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, the weighted sum is used to obtain the attention output result, and finally the output transformation W is used O The result of Point Transformer is obtained. Through the above steps, the PTCA module realizes efficient feature extraction and aggregation of point cloud data, which is particularly suitable for target recognition and segmentation tasks in complex backgrounds.

[0030] In step 3, the local center point is first selected from the original point set by farthest point sampling (FPS), reducing the number of points to half. Next, the k-nearest neighbor (kNN) algorithm is used to group k spatial neighbors for each center point to form a local 3D region. Then, feature aggregation is performed to calculate the feature vector for each center point and its neighbor points, extract geometric features using learnable operators, and aggregate features through maximum pooling and average pooling. Next, the dimension of the feature is further reduced by the pooling operation while retaining important information. Subsequently, decoding mapping is performed to gradually restore the spatial distribution of features through multi-layer decoders, map high-dimensional features back to the original point cloud space, and restore the spatial structure of the point cloud. Finally, the decoded features are input into the classifier, each point is classified, and the category label is output to complete 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 point cloud segmentation network based on RubberTreeNet of the present invention provides comprehensive technical support for rubber tree segmentation in complex scenes by performing careful preprocessing of a self-built data set and introducing a PTCA (Point Transformer Criss-Cross Attention) module, which significantly improves the overall performance of the rubber tree segmentation task.

[0033] In the first aspect, the present invention makes the model adaptability of the present invention design wide through meticulous data preprocessing steps in order to make data closer to the real situation. We have built a high-quality single rubber tree data set. Utilizing CloudCompare software, the collected point cloud data are efficiently rendered and segmented, and the severely crossed rubber tree point cloud data are segmented into independent single rubber trees, and marked respectively. This process not only ensures the diversity and representativeness of the data set, but also provides clear and accurate data support for model training, further improving the generalization ability and segmentation effect of the model.

[0034] Secondly, in order to enhance the model's ability to extract features from rubber tree point cloud data in complex scenes and improve segmentation accuracy, the PTCA (Point Transformer Criss-Cross Attention) module was introduced to significantly enhance the model's ability to extract features 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 point cloud data, and improving the model's ability to capture geometric information and spatial relationships of rubber trees, thereby improving segmentation accuracy.

[0035] The example of the present invention aims to solve the problem of insufficient local detail capture and global background information fusion in the rubber tree segmentation task by constructing a high-quality single rubber tree dataset and a point cloud adaptive transform attention (PTCA) mechanism, thereby achieving high-precision single rubber tree segmentation and significantly improving the segmentation effect and robustness of the model under complex background conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the overall framework diagram of the rubber tree segmentation of the present invention;

[0037] Figure 2 It is a network framework diagram for the segmentation of rubber trees of the present invention;

[0038] Figure 3 It is a PTCA module framework diagram for rubber tree segmentation of the present invention;

[0039] Figure 4 It is a labeled diagram of the rubber tree segmentation dataset of the present invention. DETAILED DESCRIPTION

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

[0041] Figure 1The overall flow chart of the present invention, the overall process of RubberTreeNet is, firstly, the collection and pre-processing of high-resolution 3D point cloud data are carried out, the data of rubber forests are obtained by laser radar, and CloudCompare software is used to segment and mark; then, the point cloud adaptive transformation attention mechanism (PTCA) is adopted to enhance the local detail capture and global relevance; finally, the representative point is selected using the farthest point sampling (FPS), and the neighborhood point is obtained by the K nearest neighbor (KNN) search, the original resolution is restored by feature aggregation, pooling operation and decoding mapping layer, and the classifier is applied to predict each point, and the segmentation result of the single rubber tree is output. Improve the convergence speed and stability of the model, so as to achieve the high-precision segmentation of the single rubber tree under complex background.

[0042] Figure 2 The PTCA (Point Transformer Criss-Cross Attention) module combines the advantages of the PointTransformer and Criss-Cross Attention mechanisms, aiming to enhance the capture of local details and the fusion of global background information. Specifically: Point Transformer: Use the self-attention mechanism to weight the neighborhood of each point cloud point to ensure that each point can fully capture the geometric information within its neighborhood. By calculating the dot product between the query, key, and value, the Point Transformer can capture the complex geometric relationships in the point cloud data, thereby better understanding the local features. Criss-Cross Attention: Capture long-distance dependencies in the spatial dimension and enhance global relevance and selectivity. Criss-Cross Attention calculates the attention weights in the horizontal and vertical directions, allowing the model to better understand and process complex point cloud structures, thereby enhancing the fusion of global features. The PTCA module performs best at the front end of the network for two main reasons: first, the PTCA module effectively models the spatial relationship between points in the point cloud through the point conversion layer and the cross-channel attention mechanism, which is very important in the front-end feature extraction process and helps to better understand the geometric structure of the input data; second, the PTCA module can start to fuse global context information at the front end, so that the network can retain more global information in the subsequent processing process, thereby improving the performance of the overall network. The present invention improves PointMLP and uses the PTCA module to play an important role in RubberTreeNet, significantly improving the model's understanding and segmentation capabilities of rubber tree point cloud data, and ensuring the model's efficiency and accuracy in complex backgrounds.

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

[0044] Step 1: In order to enable the model to better learn the characteristics of individual rubber trees, the severely crossed rubber tree point cloud data needs to be segmented into independent individual rubber trees. This process is completed by manual selection, marking each independent rubber tree as 1, and non-independent rubber trees (including ground, fallen leaves, and crossed leaves) as 0. Such 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 annotation. CloudCompare is a powerful open source point cloud processing software that can quickly and efficiently render a large amount of point cloud data to the screen, making it convenient for users to interactively view and select point clouds. Through manual selection methods, the point cloud data is segmented into independent individual rubber trees and marked separately. In order to ensure the diversity and representativeness of the data, 1,000 rubber trees were selected from the large amount of data collected for detailed analysis. Finally, 600 rubber trees were selected for experiments. These data sets contain ground data, independent rubber trees, and some rubber tree point cloud data that intersect with independent rubber trees, which helps the model learn complete rubber tree characteristics. The ratio of the training set to the test set is 8:2, that is, 480 trees are used for training and 120 trees are used for testing. Through the above detailed preprocessing steps, the quality of the data is ensured, making the model perform better in the task of segmenting a single rubber tree in a complex background. Each step of the preprocessing is to maximize the value of the data and ensure that the model can learn useful features from it, thereby providing more accurate rubber tree segmentation results in practical applications.

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

[0046]

[0047] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. k is the dimension of the key vector. First, the input feature X is linearly transformed to generate Q, K, and V. Then, the attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid the problem of gradient vanishing or exploding. The scaling and normalization steps ensure the stability of the calculation and the rationality of the numerical range. In addition, Point Transformer also combines the position encoding extracted directly from the point coordinates to effectively modulate the attention mechanism given by (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 i -p j , and then input this relative position vector into the MLP, and after a series of fully connected layers and nonlinear activation functions, a new feature vector is obtained as the output of the position encoding. The core purpose of position encoding is to use the relative position information between points to enhance the model's understanding of spatial relationships, which is particularly important when processing point cloud data without a clear order and topological structure.

[0050] Cross-Attention is an enhanced attention mechanism that combines and processes multiple different types of input features to capture the relationship between different features, thereby improving the model's understanding and performance. Cross-Attention generates vertical and horizontal weights, allowing the structured graph convolution to output a feature map of Where H×W represents the size of the feature map, Input to the horizontal attention unit, the weight direction on the calculation level and its calculation formula can be defined as shown in (3):

[0051]

[0052] Where W h and b h is the weight parameter, X h is the coefficient of attention along the horizontal direction. The weight in the vertical direction is calculated by inverting the feature matrix, and the process is expressed by the following equation (4):

[0053]

[0054] in represents the transposed feature map. Therefore, we can obtain the vertical attention weight in the same way, and its formula (5) is as follows:

[0055]

[0056] Where W v and b v is the weight parameter, X vis the coefficient of vertical attention. In order to further process the enhanced feature matrix and discard the irrelevant information that may exist, we added a threshold screening network to filter out the noise or redundant features of non-rubber trees after downsampling the output results. The main function can be mapped into two levels of importance: important and secondary. The main principle is to map the composite enhanced matrix to the spatial domain, and take attention to distinguish the horizontal feature cosine value and the vertical feature cosine value. The calculation method is shown in formula (6). When the two functions are close to zero, it means that the angle is close to 90 degrees. At this time, it is judged as a vector that is irrelevant to a single independent rubber tree. By setting the threshold, useful information can be enhanced and useless information can be 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 formula (7). According to the above two situations, we retain values ​​below the threshold as 0 and values ​​above the threshold as 1, so as to obtain better feature screening effects. This allows the attention network to focus on features with higher activation values, reduces the misleading effects of complex backgrounds on the model, improves the sparsity of the feature matrix, and reduces unnecessary computational overhead.

[0057]

[0058] The PTCA (Point Transformer Criss-Cross Attention) module combines the advantages of the PointTransformer and CCTA modules to achieve efficient feature extraction and aggregation of point cloud data. The PointTransformer captures the global context of the input features through the self-attention mechanism, while CCTA further processes these features and enhances the interaction of local features. The design of this module can effectively identify and segment the target point cloud under complex backgrounds, such as segmenting a single rubber tree under severe cross backgrounds. The specific operation of the PCTA module can be expressed by the following formula (8).

[0059]

[0060] In this formula, the input feature XX is first self-attentioned through the Point Transformer module. Here, W Q , W K and W V are the linear transformation weight matrices of 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, the weighted sum is used to obtain the attention output result, and finally the output transformation W is used OThe result of Point Transformer is obtained. Subsequently, the CCTA module further processes the output of Point Transformer. CCTA uses a cross-channel attention mechanism to normalize the input through convolution and pooling operations, and calculates the cross similarity X between the two inputs, and then combines the maximum value and 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. The local features of the 3D point cloud are carefully extracted through the self-attention mechanism to ensure that each point can fully capture the geometric information of its neighborhood. At the same time, the module captures long-distance dependencies in the spatial dimension to enhance global relevance and selectivity.

[0061] In order to prove the validity of the present invention, the following experiments were conducted.

[0062] The data set used in the present invention is thousands of rubber trees that we segmented in the large rubber forests collected using CloudCompare software. 1000 rubber trees were selected for detailed analysis to ensure the diversity and representativeness of the data, covering different growth stages and complex background conditions. Finally, 600 rubber trees were selected for experiment to ensure the feasibility and statistical reliability of the experiment, while facilitating resource management and analysis. Each data set includes ground data, independent rubber trees and some rubber tree point cloud data that intersect with independent rubber trees. These data contribute to the complete rubber tree features of model learning. In addition, this paper selects the public point cloud data set containing complex scenes and complex conditions in two.

[0063] The invention patent was experimented in the hardware environment of Intel(R) Xeon(R) Gold 6330CPU and NVIDIA RTX 3090GPU. The deep learning framework based on Pytorch is Pytorch1.8.1, the Python environment is 3.8, and the system environment is Ubuntu 18.04.6LTS. The number of training times is 540. The batch-size is 10. The learning rate is 0.001, and the learning rate is decayed to one-half every 40 epochs. The ratio of training set to test set is 8:2. The training set data contains 480 trees and the test set contains 120 trees.

[0064] Table 1

[0065] Network Model Acc mIou PointMLP 74.65％ 73.48％ SPOTR 88.65％ 79.75％ RubberTreeNet 92.45％ 86.35％

[0066] The experimental results are shown in Table 1, which compares 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 mIoU of 79.75%. PointMLP performs relatively poorly with an accuracy of 74.65% and mIoU of 73.48%. This shows that RubberTreeNet has higher classification accuracy and better segmentation effect in this task.

[0067] The present invention discloses a rubber tree segmentation method based on RubberTreeNet. The method, the present invention introduces a PTCA (Point Transformer Criss-Cross Attention) module in the backbone network, and utilizes the self-attention mechanism and Criss-Cross Attention mechanism of PointTransformer to enhance local detail capture and global background information fusion, reduce information loss and enhance detail capture capability. High-precision single rubber tree segmentation is achieved, and in terms of accuracy, RubberTreeNet reaches 92.45%, showing excellent performance under limited computing resources. The present invention provides reliable technical support for rubber tree segmentation in complex scenarios, and significantly improves 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 can effectively identify and segment individual rubber trees from complex point cloud data through the point cloud adaptive transformation attention (PTCA) mechanism and multi-level feature extraction and segmentation process. The overall process includes the following steps: Step 1: Collect and pre-process rubber tree point cloud data; The collected rubber tree point cloud data was saved in ply format and rendered and preliminarily processed using CloudCompare software, including the removal of noise and non-target objects. Subsequently, the independent rubber trees in the point cloud data were separated from the background by manual selection and marking, and exported to txt format for subsequent processing; Step 2: Use point cloud adaptive transformation attention mechanism (PTCA); The Point Cloud Adaptive Transformation Attention (PTCA) mechanism is adopted, which combines the advantages of Point Transformer and Criss-CrossAttention. Point Transformer finely extracts local features of 3D point clouds through the self-attention mechanism to ensure that each point fully captures the geometric information of the neighborhood; Criss-Cross Attention captures long-distance dependencies in the spatial dimension and enhances global relevance and selectivity. The PTCA module solves the problem of insufficient fusion of local detail capture and global background information through dual feature extraction and fusion, significantly reduces information loss, and enhances detail capture capabilities; Step 3: Multi-level feature extraction and segmentation process; Then, the farthest point sampling (FPS) is used to select representative points, and the neighborhood point information of these representative points is obtained through K nearest neighbor (KNN) search; finally, the features are restored to the original resolution through multi-scale feature aggregation (MSFA), pooling operation and decoding mapping layer, and the classifier is applied to predict each point to achieve accurate segmentation of single rubber trees.

2. The rubber tree segmentation method based on RubbertreeNet according to claim 1, characterized in that, The specific steps of step one are: First, a new dataset of individual rubber trees was constructed to better train the model. Data collection was conducted in Hainan rubber forests, and high-resolution 3D point cloud data was obtained by using LiDAR technology. This method ensures that the data of each rubber tree can clearly capture the vegetation canopy, the vertical structural characteristics of the tree, and the information of the lower part of the canopy (such as the trunk and leaves). The data collection covered a large area of ​​rubber trees and was saved in PLY file format. During the data collection process, the background included fallen leaves, ground, and other large areas of heavily crossed rubber trees. In order to facilitate subsequent processing, we used CloudCompare software to quickly and efficiently render a large amount of point cloud data, allowing users to intuitively view and interact with the data on the screen. Through a manual selection method, the severely crossed rubber tree point cloud data was segmented into independent individual rubber trees and marked separately. Independent individual rubber trees were marked as 1, while non-independent rubber trees (including ground, fallen leaves, and crossed leaves) were marked as 0. This segmentation can be based on spatial location, color, or other attributes. Next, we exported the segmented point cloud data into TXT file format to facilitate subsequent network processing. We used CloudCompare software to segment thousands of rubber trees from the large rubber forests we collected, and selected 1,000 rubber trees for detailed analysis to ensure the diversity and representativeness of the data, covering different growth stages and complex background conditions. Finally, 600 rubber trees were selected for the experiment to ensure the feasibility and statistical reliability of the experiment, while facilitating resource management and analysis. These datasets include not only ground data, but also independent rubber trees and some rubber tree point cloud data that intersect with independent rubber trees. Such datasets help the model learn complete rubber tree features. Through the above steps, we ensure the quality and diversity of the dataset, providing a solid foundation for subsequent model training.

3. The rubber tree segmentation method based on RubberTreeNet according to claim 1, is characterized in that, The specific steps of step 2 are: This module uses the sparsity and irregularity of point clouds to weight the neighborhood of each point through a self-attention mechanism, thereby effectively extracting local and global features, which is expressed as follows: Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. k is the dimension of the key vector. First, the input feature X is linearly transformed to generate Q, K, and V. Then, the attention scores are obtained by calculating the dot product of the query and the key, and these scores are scaled to avoid the problem of vanishing or exploding gradients. The scaling and normalization steps ensure the stability of the calculation and the rationality of the numerical range. Then, the Softmax function is applied to normalize the scaled attention scores to obtain the attention weights. Next, the relative position vector p between each pair of points is calculated i -p j , and processes it through a multi-layer perceptron (MLP) to generate a positional encoding. Combining position encoding with self-attention output enhances the model’s understanding of spatial relationships. Then, entering the cross-channel attention (CCTA) module, the input features are first convolved and pooled, and normalized. 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, the attention weights in the horizontal and vertical directions are calculated, respectively: X h : Horizontal attention weight, W h The horizontal weight matrix, g i,j represents the eigenvalue of the input feature map at position (i, j), b h Represents the horizontal bias. v : vertical attention weight, W v The weight matrix in the vertical direction, g j,i represents the eigenvalue of the input feature map at position (j,i), b v Represents the horizontal bias. Then, the cosine similarity of the horizontal and vertical features is calculated. where cosθ represents the cosine similarity between two vectors. The value of the i-th feature in the horizontal vector. Represents the value of the i-th feature in the vertical vector. n represents the dimension of the vector. Filter out important features based on the set threshold and discard irrelevant information: When cos h ≤thr, reserved as 0; When cos h >thr, keep it as 1. Finally, the features processed by CCTA are combined with the original input features to generate the final output features: Where W Q , W K and W V are the linear transformation weight matrices of 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, the weighted sum is used to obtain the attention output result, and finally the output transformation W is used O The result of Point Transformer is obtained. Through the above steps, the PTCA module realizes efficient feature extraction and aggregation of point cloud data, which is particularly suitable for target recognition and segmentation tasks in complex backgrounds.

4. The rubber tree segmentation method based on RubberTreeNet according to claim 1, is characterized in that, The specific steps of step three are as follows: First, select the local center point from the original point set through the farthest point sampling (FPS) to reduce the number of points to half. Next, use the k nearest neighbor (kNN) algorithm to group k spatial neighbors for each center point to form a local 3D region. Then, feature aggregation is performed to calculate the feature vector for each center point and its neighbor points, extract geometric features using learnable operators, and aggregate features through maximum pooling and average pooling. Next, the dimension of the feature is further reduced by the pooling operation while retaining important information. Subsequently, decoding mapping is performed to gradually restore the spatial distribution of features through multi-layer decoders, map high-dimensional features back to the original point cloud space, and restore the spatial structure of the point cloud. Finally, the decoded features are input into the classifier to classify each point and output the category label to complete 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