A single-plant organ point cloud segmentation method based on heterogeneous topological graph attention

Abstract

The application discloses a single-plant organ point cloud segmentation method based on a heterogeneous topological graph attention, relates to the technical field of single-plant three-dimensional point cloud semantic segmentation and instance segmentation, and has the advantages that by introducing a superpoint grouping and superpoint level aggregation mechanism in the network, the model can adapt to point cloud inputs of different point number scales during reasoning, the dependence on fixed point number sampling or forced blocking is reduced, and therefore the loss of original point cloud structure information is reduced. Meanwhile, by introducing the skeleton topological structure information of the plant in the form of a heterogeneous graph structure into the segmentation network and utilizing weighted graph attention to perform feature propagation under topological constraints, the segmentation process has structure consistency constraint basis, and the structure consistency and interpretability of the segmentation result are improved. Further, the topological prior can enhance the discrimination ability for organ junctions, mutual occlusion and overlapping regions, reduces the misclassification, missed classification and instance mixed classification phenomenon, and therefore the precision and stability of the plant organ point cloud segmentation are improved.

Description

Technical Field

[0001] This invention relates to the field of semantic segmentation and instance segmentation of 3D point clouds of single plants, specifically to a method for segmenting plant organ point clouds based on skeleton topology priors and graph attention mechanisms, used to achieve semantic segmentation and instance segmentation of point clouds of single plants. Background Technology

[0002] Digital plant 3D phenotypic analysis is a technique that uses 3D data to quantitatively measure and evaluate plant morphology and structure. In recent years, thanks to the development of 3D data acquisition technologies such as 3D reconstruction, structured light scanning, and LiDAR, it has become possible to acquire large amounts of 3D point cloud data of single plants containing detailed organ structures such as stems, leaves, fruits, and flowers. As a crucial step in the phenotypic analysis process, organ-level point cloud segmentation is the foundation for subsequent organ trait extraction and structural parameter statistics; its segmentation quality directly determines the accuracy and reliability of phenotypic index analysis.

[0003] Existing methods often employ deep learning techniques to identify organ categories or organ instances. However, to adapt to the input scale and computational resources of the network architecture, most solutions require fixed-point sampling or block processing of the point cloud, along with preprocessing and post-processing to complete the segmentation. However, plant point clouds often exhibit significant differences in point count, uneven point density, and the coexistence of noise and missing data under different acquisition conditions and plant morphologies. While fixed-point downsampling or block processing facilitates network input, it often leads to the loss or fragmentation of information in slender organs and organ boundary regions, compromising the integrity of the original structural information. Furthermore, most existing methods rely solely on local geometric neighborhood features for discrimination, neglecting the inherent growth topology of plants (such as branching and hierarchy). This results in insufficient structural consistency and interpretability of the segmentation results, particularly prone to misclassification, omissions, or mixing of different organ instances at boundaries, limiting the application of existing technologies in accurate organ-level phenotypic analysis. Summary of the Invention

[0004] To address the shortcomings of existing plant organ point cloud segmentation methods, such as insufficient adaptability to input scale when the point cloud size varies significantly, and the tendency for segmentation instability and instance mixing in organ boundaries and overlapping occlusion regions, this invention proposes a method for simultaneous semantic and instance segmentation of single plant organ point clouds that can utilize plant skeleton topological structure information to improve segmentation consistency and interpretability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for point cloud segmentation of single plant organs based on heterogeneous topological graph attention, the method being as follows:

[0007] Step 1: Data Acquisition and Preprocessing: The original 3D point cloud data of a single plant is obtained by reconstructing the surrounding photographed image using the structure-of-motion motion reconstruction-multi-view stereo vision 3D reconstruction method; the point cloud is filtered and denoised, and then single-plant segmentation and coordinate normalization are performed to form point cloud samples to be segmented.

[0008] Step 2: Supervised data annotation (for training): Perform organ-level manual annotation on the point cloud samples to obtain point-level semantic labels and organ instance labels, which are used to construct training sets, validation sets and test sets;

[0009] Step 3: Unsupervised skeleton extraction: For each point cloud sample, a point cloud skeleton extraction method based on Laplace shrinkage transform is used to obtain a skeleton point cloud, so that the skeleton point cloud represents the centerline and branch structure of the plant organ, and a topological structure (edges between skeleton points) is constructed based on the skeleton point cloud.

[0010] Step 4: Unsupervised superpoint extraction: Unsupervised geometric oversegmentation is performed on the original point cloud of the single plant from Step 1 to obtain several non-overlapping subsets of the point cloud. Each subset consists of spatially adjacent points with similar geometric properties, forming a set of superpoint regions. Each point is assigned a superpoint region identifier, and a representative point (or center point) is calculated for each superpoint region for subsequent superpoint-level aggregation calculation.

[0011] Step 5: Offline topological heterogeneous graph structure generation: Based on the skeleton point cloud from Step 3 and the set of superpoint regions from Step 4, the skeleton topological heterogeneous graph structure is generated offline.

[0012] Step Six: Network Architecture Construction (Topology Graph Attention Learning): Construct a segmentation network that includes a point cloud feature encoding module, a super-point pooling module, a topology graph attention learning module, and a decoding prediction module. The point cloud feature encoding module is used to extract the initial features of the original point cloud, the super-point pooling module is used to aggregate point-level features into super-point-level features, and the topology graph attention learning module uses the heterogeneous topology graph obtained in Step Five as a structural constraint to perform weighted attention aggregation and message passing on the super-point node features, and outputs super-point-level representation features that include topological consistency.

[0013] Step 7: Joint prediction of semantics and instances: The super-point level representation features enhanced by the topological graph attention learning module are input into the decoding prediction module; the decoding prediction module adopts a query-based Transformer decoder structure, which includes multiple stacked decoding layers;

[0014] Step 8: Training Optimization: The network is trained under the supervised label constraints of Step 2. During training, constraints related to topological consistency are introduced to ensure that the network learning results are consistent with the skeleton topology and to improve the segmentation stability of organ junctions and occluded regions. The joint prediction in Step 7 is performed in both the training and inference phases; the training optimization in Step 8 is performed only in the training phase, and the loss is calculated and the network parameters are updated based on the prediction results from Step 7 and the supervised labels from Step 2.

[0015] Further, in step four, the geometric attributes include at least one of spatial coordinates, normal vector direction, curvature, color, or combinations thereof; the attribute similarity includes: the geometric feature difference between adjacent points does not exceed a preset threshold or the feature similarity is greater than a preset threshold, wherein the threshold is a fixed value or adaptively determined according to the point cloud density, the average neighborhood distance between points, or the statistical distribution of geometric feature differences.

[0016] Further, in step five, the heterogeneous graph includes: a first type of node: skeleton nodes (composed of the skeleton point cloud or its key points); a second type of node: supernodes (each supernode region corresponds to one node); a first type of edge: skeleton node-skeleton node connections, used to characterize the skeleton topological connectivity; a second type of edge: supernode-skeleton node connections, used to establish the association between supernodes and the topological structure; and self-loop edges: self-loop edges are added to each supernode to preserve the node's own characteristics and improve the stability of weighted message passing.

[0017] To adapt to network learning, adjacency lists and edge weights for various edge types are generated and stored offline. The edge weights can be calculated based on Euclidean distance and edge type. By setting edge weights offline, the skeleton topological constraints can be strengthened and noise propagation caused by irrelevant adjacencies can be suppressed during graph attention learning. (No supernode-supernode connections are established in the heterogeneous graph structure of this invention.)

[0018] Furthermore, in step seven, the decoder is set to a 3-layer Transformer decoding layer. Each decoding layer includes at least: a self-attention sub-layer for the query token, a cross-attention sub-layer between the query token and the super-point level features, and a feedforward network sub-layer. Residual connections and layer normalization can be used to stabilize training. The decoder uses the super-point level features as the keys and values ​​of the cross-attention to update the query token and outputs the prediction results for semantic segmentation and instance segmentation. The super-point level prediction is then backfilled into the point-level output according to the point-to-super-point mapping relationship.

[0019] The advantages of this invention compared to existing technologies are as follows: By introducing superpoint grouping and superpoint-level aggregation mechanisms into the network, this invention enables the model to adapt to point cloud inputs with different point sizes during inference, reducing reliance on fixed-point sampling or forced block division, thereby reducing the loss of original point cloud structural information. Simultaneously, by introducing the plant skeleton topological structure information into the segmentation network in the form of a heterogeneous graph structure and utilizing weighted graph attention for feature propagation under topological constraints, the segmentation process is equipped with structural consistency constraints, improving the structural consistency and interpretability of the segmentation results. Furthermore, topological priors can enhance the discrimination ability at organ boundaries, mutually occluding, and overlapping regions, reducing misclassification, missed classification, and instance mixing, thereby improving the accuracy and stability of plant organ point cloud segmentation. Attached Figure Description

[0020] Figure 1 This is the overall flowchart of the method;

[0021] Figure 2 The image shows the data preprocessing (denoising, single-plant segmentation);

[0022] Figure 3 Manually labeled organ diagrams;

[0023] Figure 4 The images show skeleton extraction (left) and superpoint extraction (right) based on Laplacian contraction.

[0024] Figure 5 The graph is composed of heterogeneous topological graphs;

[0025] Figure 6 This is a diagram of the overall network architecture.

[0026] Figure 7 This is a partial screenshot showing the segmentation effect;

[0027] Figure 8 This is a diagram showing the partial segmentation results of the method on other plants. Detailed Implementation

[0028] The specific embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art regarding the sequence of steps, implementation details, and parameters without departing from the spirit of the present invention should all fall within the scope of protection of the present invention.

[0029] Example 1: Simultaneous organ-level semantic and instance segmentation for point clouds of single wheat plants

[0030] Step 1: Point Cloud Acquisition and Preprocessing

[0031] An image-based structure-of-motion reconstruction method with multi-view stereo vision was used to obtain the original 3D point cloud data of potted plants. For example... Figure 2 As shown in the flowchart, statistical filtering and radius filtering are applied to the point cloud for noise reduction, color filtering is used to remove points related to flower pots and soil, and the point cloud is then segmented into individual plant samples. Coordinate normalization is then performed to obtain the point cloud samples to be segmented, denoted as [sample name missing]. , where each point It must contain at least three-dimensional coordinates. It can also optionally include attributes such as color, reflectivity, and normal.

[0032] Step 2: Organ Instance-Level Supervised Annotation

[0033] All individual point cloud samples were manually annotated at the organ level to obtain point-level semantic labels. (Corresponding to stem, leaf, and ear of wheat organ categories) and organ instance identifier (Instances of the same organ share the same instance number). Training, validation, and test sets are constructed based on labeled data; in this embodiment, they are divided according to a preset ratio (7:2:1).

[0034] Step 3: Unsupervised skeleton extraction based on Laplace transform

[0035] For each point cloud sample X, a point cloud skeleton extraction method based on Laplace shrinkage transform is used to obtain a skeleton point cloud S, which represents the central line and branching structure of the plant organ. A topological structure (edges connecting skeleton points) is then constructed based on the skeleton point cloud. The basic process is as follows: a point cloud adjacency graph is constructed based on K-nearest neighbors, forming a Laplace operator L. An optimization problem is solved by iterative shrinkage to converge the point set towards the central line structure, for example, minimizing the following energy function:

[0036]

[0037] in, This is the shrunk point position matrix. This represents the graph Laplacian operator constructed from the nearest neighbor graph of the point cloud. Describing the Frobenius norm, For constraint coefficients, This is the original point location matrix. Iterative shrinking followed by downsampling yields a sparse skeleton point cloud. .

[0038] Step 4: Unsupervised superpoint extraction (point cloud over-segmentation)

[0039] The single-plant point cloud sample Unsupervised geometric oversegmentation is performed, dividing the data into several spatially adjacent point sets with similar geometric properties. Each sub-block is defined as a "hyperpoint region". Let the resulting set of hyperpoint regions be... ,satisfy And each They don't interact with each other.

[0040] Hyperpoint regions are obtained using a partitioning strategy based on geometric homogeneity constraints (region growing or graph segmentation based on features such as local normal, curvature, color, or flatness). For each hyperpoint region... Calculate its geometric center and in the region Selecting from and The closest point is used as the representative point of the superpoint region (representative superpoint). All representative points together constitute the superpoint representative point set. , Simultaneously, record the mapping relationship of which superpoint region each point belongs to. .

[0041] Step 5: Offline skeleton topological heterogeneous graph construction and edge weight setting

[0042] like Figure 5 Based on skeleton point cloud Set of super-point regions Offline construction of skeleton topology heterogeneous graph The node set includes two categories: skeleton nodes and supernodes. Nodes are composed of skeleton points or skeleton key points; their feature is their 3D coordinates. Supernodes. Each superpoint region corresponds to one superpoint node.

[0043] Edge sets include at least the following three types:

[0044] 1. Skeleton - Skeleton Edge : Used to represent the skeleton connectivity. Preferably, candidate adjacencies of skeleton points are first constructed based on the K-nearest neighbor algorithm, and then a minimum spanning forest is used to filter from the candidate edges to obtain a connected and sparse set of skeleton backbone edges.

[0045] 2. Superpoint - Skeleton Edge Used to establish the association between superpoints and the topology. Each superpoint represents the coordinates of a point. Connections are generated by matching the nearest neighbor of a skeleton point (or a skeleton key point); in this embodiment, each superpoint is connected to its nearest skeleton node.

[0046] 3. Exceeding the boundary of a loop Add self-loop edges to each supernode to preserve the node's own features and stabilize training during graph propagation.

[0047] Furthermore, assign edge weights to various offline edges. In this embodiment, the method of "by edge type and Euclidean distance" is adopted, for example:

[0048] skeleton-skeleton edge :

[0049]

[0050] For superpoint-skeleton edge :

[0051]

[0052] Self-loop edge of super point :

[0053]

[0054] in, Represents skeleton nodes coordinates Represents the three-dimensional coordinates of the skeleton node v. This represents the weight of the edge connecting skeleton point u to skeleton point v. This represents the weight of the edge connecting the superpoint m to the skeleton point u. This represents the weight of the self-loop edge from the superpoint m to itself. The superpoint represents the coordinates of the point. , , and , These are preset parameters used to highlight the role of skeleton topological constraints in information propagation. The adjacency list, edge types, and edge weights are saved offline for direct loading during training and inference phases.

[0055] (Note: In this embodiment, the heterogeneous graph does not establish super-super-point connections.) , , and , In this embodiment, the values ​​are 2, 1, 0.1, 10, and 10, respectively.

[0056] Step Six: Network Architecture

[0057] like Figure 6 As shown, a segmentation network is constructed for semantic segmentation and instance segmentation of point clouds of single plant organs. The network input includes: single plant point cloud samples. Step 4 yields the superpoint partitioning. (Including point-to-superpoint mapping) (and step five, offline generation of a heterogeneous topology graph carrying edge type and edge weight) The network output includes point-level semantic labels and point-level organ instance segmentation results. The network modules are as follows.

[0058] 1. Sparse 3D Convolutional U-Net Point-Level Feature Encoding

[0059] (1) Voxelization and sparse tensor construction: for input points Perform voxelization, and convert the coordinates of each point... Mapping to 3D voxel index (Voxel resolution is) For points falling into the same voxel, their input attribute features are pooled to obtain voxel features, and a sparse voxel feature tensor is constructed as the input of the sparse convolutional network.

[0060] (2) Sparse 3D convolutional U-Net encoding: The above sparse voxel tensors are input into the sparse three-dimensional convolutional U-Net network. The network includes downsampling encoding path and upsampling decoding path. Multi-scale features are fused through skip connections to output sparse voxel-level features.

[0061] (3) Point-level feature backfilling: Based on the mapping relationship from points to voxels, voxel features are backfilled onto points to obtain the point-level feature matrix. ,in It represents the dimension of point-level features. The point-level feature vector representing the i-th point is obtained by backfilling the point with 3D-Unet encoding.

[0062] 2. Super-pooling generates super-point level features.

[0063] Based on the set of super-point regions and point-to-superpoint mapping Point-level features Aggregates into super-point level features This embodiment uses average pooling: ,in As a super node initial characteristics, , .

[0064] 3. Topology Graph Attention Learning Module

[0065] Using the heterogeneous topology graph from step five To address structural constraints, weighted attention aggregation and updates are performed on the node features in the graph to enhance topological consistency. The node set in the graph includes skeleton nodes and supernodes, and the edge set includes skeleton-skeleton edges, supernode-skeleton edges, and supernode self-loop edges. Edge weights are stored offline for each edge. .

[0066] Let the first Layer input node features are For any node Let its neighborhood be For adjacent edges Calculate attention coefficient and edge weights Introducing a aggregation coefficient allows different types of edges to contribute differently to propagation, for example:

[0067]

[0068] in, For learnable parameters, This is a non-linear activation function. After several layers, the output corresponding to the supernode is taken as the topology-enhanced supernode feature: , This is used for subsequent query interactions with the decoder. This represents the set of supernodes in a heterogeneous topological graph G. It is the set of skeleton nodes in graph G.

[0069] 4. Transformer Decoding and Prediction (Semantic + Instance Joint, 3-Layer Transformer)

[0070] Topology-enhanced superpoint features As a key across attention in the decoder AND value The decoder employs a query-based Transformer decoding structure, constructing instance query tokens. With semantic query token In this embodiment, the topology token is used. and , Concatenate to form the decoder input token sequence:

[0071]

[0072] The decoder contains three Transformer layers, each including: token self-attention: for information interaction between instance / semantic / topological tokens; and cross-attention from token to superpoint feature: for information interaction from... Read the context information related to the segmentation.

[0073] A prediction head is set at the decoder output to obtain super-point level semantic category prediction and super-point level instance mask prediction.

[0074] 5. Point-level output generation

[0075] The super-point level semantic prediction results and super-point level instance mask prediction results are backfilled to the point level based on the super-point partition π to obtain the point-level semantic segmentation results and the point-level organ instance segmentation results.

[0076] 6. Training and topology consistency constraints (explicit loss function)

[0077] The network is trained end-to-end under supervised label constraints. This embodiment uses the following total loss function:

[0078]

[0079] in:

[0080] (1) Semantic segmentation loss : Point-level semantic prediction results and point-level semantic labels Calculate the cross-entropy loss:

[0081]

[0082] This represents the cross-entropy loss function commonly used in semantic segmentation tasks, used to measure the difference between the predicted result and the true label. This represents the semantic prediction result output by the network for the i-th point. This represents the true semantic label of the i-th point.

[0083] (2) Instance segmentation loss It consists of instance class loss and instance mask loss.

[0084]

[0085] Among them, category loss Cross-entropy loss and mask loss are used. We use a weighted sum of binary cross-entropy loss and Dice loss:

[0086]

[0087] (Note: The correspondence between the predicted instances and the real instances can be determined using the conventional training method of the query-based instance segmentation framework. This embodiment does not impose any restrictions on the specific matching implementation.)

[0088] (3) Topological consistency constraints are based on the set of superpoints and skeleton edges constructed offline in step five. and its offline set and stored edge weights Regularization is applied to the superpoint semantic prediction results to enhance the consistency of organ segmentation in terms of skeleton topology. Let the superpoint-level semantic prediction output of the decoder be... The following is a super node. In category On the predicted probability ,in This embodiment only applies to... Superpoints-skeleton edges existing in Calculate the topological loss and edge weights. Take the superpoint-skeleton edge weights generated and saved offline in step five (and load them directly during the training and inference phases).

[0089] To characterize the consistency between the superpoint prediction and the category of its associated skeleton node, the superpoint prediction probability is first arranged along... Project onto the skeleton node. For any skeleton node In category The projection probability on is defined as:

[0090]

[0091] Let the total edge weight Define edge consistency loss. for:

[0092]

[0093] This represents the class c probability obtained by "converging / projecting" the predictions of the skeleton node u from its connected superpoints, used in conjunction with the superpoint predictions. Alignment with topological consistency constraints.

[0094] Define edge cutting loss for:

[0095]

[0096] The topology consistency constraint in this embodiment is defined as follows:

[0097]

[0098] When a certain skeleton node exist If there are no associated edges or the sum of the weights of the associated edges is 0, the skeleton node can be excluded from the above process. The calculations are performed to ensure numerical stability during the training process.

[0099] Step 7: Joint prediction of semantics and instances

[0100] The topology-enhanced superpoint-level representation features obtained in step six are input into the decoding and prediction module. A query-based Transformer decoder outputs superpoint-level semantic predictions and superpoint-level instance predictions. Based on the point-to-superpoint mapping relationship, the superpoint-level predictions are backfilled to obtain point-level semantic segmentation results and point-level instance segmentation results. Step seven is performed in both the training and inference phases.

[0101] Step 8: Training Optimization

[0102] Under the supervised annotation constraints of step two, the supervised loss is calculated based on the semantic prediction and instance prediction output from step seven, and a constraint term related to the consistency of the skeleton topology is introduced to form the total loss function. The network parameters are then updated through backpropagation. Step eight is performed only during the training phase.

[0103] The segmentation results of this invention on single wheat plant data:

[0104] Table 1. Organ-level semantic segmentation results

[0105]

[0106] Table 2. Results of Organ-Level Instance Analysis

[0107]

[0108] Example 2: Qualitative generalization verification of point clouds of other plants

[0109] Following the process and network structure of Example 1, a new training set was constructed and the model trained on other plant point cloud data (another wheat seedling dataset, corn seedling data, and sorghum seedling data). Inference was then performed on the corresponding test samples, outputting point-level semantic segmentation results and point-level instance segmentation results to demonstrate the qualitative segmentation effect of the method of the present invention on point clouds of different crops / different growth stages. Figure 8 .

Claims

1. A method for segmenting point clouds of single plant organs based on heterogeneous topological graph attention, characterized in that: The method is as follows: Step 1: Data Acquisition and Preprocessing: The original 3D point cloud data of a single plant is reconstructed using the structure-of-motion reconstruction method with multi-view stereo vision; the point cloud is filtered and denoised, and then segmented and normalized to form a point cloud sample to be segmented. Step 2: Supervised data annotation (for training): Perform organ-level manual annotation on the point cloud samples to obtain point-level semantic labels and organ instance labels, which are used to construct training sets, validation sets and test sets; Step 3: Unsupervised skeleton extraction: For each point cloud sample, a point cloud skeleton extraction method based on Laplace shrinkage transform is used to obtain a skeleton point cloud, so that the skeleton point cloud represents the centerline and branch structure of the plant organ, and a topological structure (edges between skeleton points) is constructed based on the skeleton point cloud. Step 4: Unsupervised superpoint extraction: Unsupervised geometric oversegmentation is performed on the original point cloud of the single plant from Step 1 to obtain several non-overlapping subsets of the point cloud. Each subset consists of spatially adjacent points with similar geometric properties, forming a set of superpoint regions. Each point is assigned a superpoint region identifier, and a representative point (or center point) is calculated for each superpoint region for subsequent superpoint-level aggregation calculation. Step 5: Offline topological heterogeneous graph structure generation: Based on the skeleton point cloud from Step 3 and the set of superpoint regions from Step 4, the skeleton topological heterogeneous graph structure is generated offline. Step Six: Network Architecture Construction (Topology Graph Attention Learning): Construct a segmentation network that includes a point cloud feature encoding module, a super-point pooling module, a topology graph attention learning module, and a decoding prediction module. The point cloud feature encoding module is used to extract the initial features of the original point cloud, the super-point pooling module is used to aggregate point-level features into super-point-level features, and the topology graph attention learning module uses the heterogeneous topology graph obtained in Step Five as a structural constraint to perform weighted attention aggregation and message passing on the super-point node features, and outputs super-point-level representation features that include topological consistency. Step 7: Joint prediction of semantics and instances: The super-point level representation features enhanced by the topological graph attention learning module are input into the decoding prediction module; the decoding prediction module adopts a query-based Transformer decoder structure, which includes multiple stacked decoding layers; Step 8: Training Optimization: Train the network under the supervised label constraints of Step 2. During the training process, introduce constraints related to topological consistency to ensure that the network learning results are consistent with the skeleton topology and improve the segmentation stability of organ junctions and occluded regions.

2. The method for segmenting point clouds of single plant organs based on heterogeneous topological graph attention according to claim 1, characterized in that: In step four, the geometric attributes include at least one of spatial coordinates, normal vector direction, curvature, color, or combinations thereof; the attribute similarity includes: the geometric feature difference between adjacent points does not exceed a preset threshold or the feature similarity is greater than a preset threshold, wherein the threshold is a fixed value or adaptively determined according to the point cloud density, the average neighborhood distance between points, or the statistical distribution of geometric feature differences.

3. The method for segmenting point clouds of single plant organs based on heterogeneous topological graph attention according to claim 1, characterized in that: In step five, the heterogeneous graph includes: a first type of node: skeleton nodes (composed of the skeleton point cloud or its key points); a second type of node: supernodes (each supernode region corresponds to one node); a first type of edge: skeleton node-skeleton node connections, used to characterize the topological connectivity of the skeleton; a second type of edge: supernode-skeleton node connections, used to establish the association between supernodes and the topological structure; and self-loop edges: self-loop edges are added to each supernode to preserve the node's own characteristics and improve the stability of weighted message passing. To adapt to network learning, adjacency lists and edge weights for various types of edges are generated and stored offline. The edge weights can be calculated according to Euclidean distance and edge type. By setting edge weights offline, the skeleton topological constraints can be strengthened and noise propagation caused by irrelevant adjacencies can be suppressed during graph attention learning.

4. The method for segmenting point clouds of single plant organs based on heterogeneous topological graph attention according to claim 1, characterized in that: In step seven, the decoder is set to a 3-layer Transformer decoding layer. Each decoding layer includes at least: a self-attention sub-layer for the query token, a cross-attention sub-layer between the query token and the super-point level features, and a feedforward network sub-layer. Residual connections and layer normalization can be used to stabilize training. The decoder uses the super-point level features as the keys and values ​​of the cross-attention to update the query token and outputs the prediction results for semantic segmentation and instance segmentation. The super-point level prediction is then backfilled into the point-level output according to the point-to-super-point mapping relationship.

Images