A three-dimensional point cloud plant segmentation method common throughout the whole growth period of soybean

Abstract

The present application relates to the field of wisdom agriculture and digital twin technology, and discloses a three-dimensional point cloud plant segmentation method which is universal in the whole growth period of soybeans. By adopting dynamic multi-level sampling, the sampling rate can be adaptively adjusted according to the spatial sparsity and local complexity of the plant. While greatly reducing the amount of point cloud data and improving the processing efficiency, the key geometric structure features are completely retained. The multi-scale feature extraction and geometric perception attention mechanism can strengthen the feature expression of the regions with significant geometric changes of leaf overlap and stem-leaf junction. The segmentation accuracy and robustness are significantly improved. The encoder-decoder structure combined with the feature pyramid realizes efficient fusion of multi-scale features. Not only excellent performance is achieved in the whole period segmentation of soybeans, but also the method can be directly migrated to the seedling stage scenes of crops such as corn and tomato, and has strong generalization ability. It provides stable and reliable, and universal three-dimensional vision core technology support for soybean phenotype analysis, intelligent breeding, growth monitoring and precision agriculture.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and digital twin technology, and in particular to a three-dimensional point cloud plant segmentation method applicable across the entire growth period of soybeans. Background Technology

[0002] Crop 3D reconstruction and organ-level segmentation technology are the core support for intelligent breeding and precision agriculture. They can accurately characterize the spatial structure and developmental dynamics of crops, providing key phenotypic information for the screening of breeding materials and the optimization of agronomic measures. As an important food and oil crop, soybean has 12 key growth stages throughout its entire growth period (V1-R7). The plant structure gradually develops from simple seedlings to high-density complex canopies, which places extremely high demands on the adaptability and robustness of segmentation methods.

[0003] Existing technologies related to soybean 3D reconstruction and organ-level segmentation have significant core defects, specifically: lack of a general segmentation framework, insufficient adaptability of sampling strategies, weak segmentation ability in areas with significant geometric changes, and limited cross-crop generalization ability. Existing models are mostly designed for specific growth stages or single crops, making it difficult to cover the entire growth period of soybean. In particular, the segmentation accuracy drops significantly in high-density canopy scenarios in the later stages, and they cannot flexibly adapt to plant morphological variations at different growth stages. Areas with significant geometric changes, such as stem-leaf junctions and overlapping leaves, are prone to problems such as blurred segmentation boundaries and classification errors. At the same time, cross-crop transfer applications are limited, making it difficult to meet the high-precision and widely adaptable technical requirements of intelligent breeding and precision agriculture for soybean 3D reconstruction and organ-level segmentation. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period.

[0005] This invention provides a three-dimensional point cloud plant segmentation method applicable across the entire growth period of soybeans, comprising the following steps: Input the original three-dimensional point cloud data of soybeans throughout the entire growth period from V1 to R7, and run the preprocessing script to calculate the bounding box volume, local point density, and coefficient of variation of the point cloud for each plant. A dynamic multi-level sampling strategy is adopted to downsample the preprocessed original point cloud. The first step is to retain the original point cloud at a preset ratio through random sampling. Then, the sampling ratio of the farthest point is adaptively adjusted based on the joint criterion of bounding box volume and local point density to generate a lightweight point set that retains the geometric structure. Multi-scale feature extraction is performed on the sampling point set. Three neighborhoods with different radii are constructed through spherical query. The relative geometric position and feature differences of points in each neighborhood are encoded. The results are then processed by a multilayer perceptron and spliced ​​to form a multi-scale feature representation. A geometric perception attention mechanism is introduced, which defines an attention score function with decreasing distance based on the Euclidean distance between points, and performs weighted aggregation of neighborhood features to enhance the feature representation of regions with significant geometric changes such as stem-leaf junctions and leaf overlaps. Multi-scale features are fused with geometrically aware features and input into the encoder-decoder segmentation network. High-level features are extracted through downsampling, and multi-scale feature fusion is achieved by combining the encoder-side feature pyramid. Then, the resolution is restored through an upsampling layer, and finally, the stem / leaf semantic segmentation results are output through a fully connected classification head.

[0006] Furthermore, in the dynamic multi-level sampling strategy, the joint criterion includes: Spatial sparsity weight is determined by dividing the point cloud space into voxels and calculating the coefficient of variation of the number of points within each voxel. The larger the coefficient of variation, the more uneven the spatial distribution, and the sampling rate needs to be reduced to preserve the features of sparse regions. The local complexity weight is determined by calculating the coefficient of variation of the density of k nearest neighbors. The higher the coefficient of variation, the more complex the local structure, and the more the sampling rate needs to be increased. The dynamic sampling ratio is calculated based on the weighted fusion of the spatial sparsity weight and the local complexity weight.

[0007] Furthermore, in the multi-scale feature extraction, the three spherical queries with different radii are normalized according to the plant height and set as the first radius, the second radius, and the third radius, respectively, corresponding to the leaf level, branch level, and plant level scales. The relative geometric position of the neighborhood points is encoded as a point difference vector, which is mapped to a position embedding through a multilayer perceptron.

[0008] Furthermore, in the geometric perception attention mechanism, the distance-decreasing attention score function is defined as: , in, The Euclidean distance between points. To adjust the parameters.

[0009] Furthermore, the encoder-decoder segmentation network includes: The encoder contains multiple downsampling layers. Each layer downsamples by sampling the farthest point and combines it with a multilayer perceptron to extract features, thus constructing a multi-scale feature pyramid. The decoder contains multiple upsampling layers, recovers resolution through distance-weighted upsampling, and fuses features from each layer of the encoder through skip connections.

[0010] Furthermore, the method employs the AdamW optimizer for training, combines a learning rate decay strategy with an early stopping mechanism, and supports an automatic mixed-precision training mode.

[0011] Furthermore, the method achieves an average crossover ratio of a first predetermined threshold in the stem-leaf segmentation during the entire growth period of soybean (V1 to R7 stages), an average crossover ratio of a second predetermined threshold in the high-density canopy stage (R6-R7), and an average crossover ratio of a third predetermined threshold in cross-crop segmentation tests during the seedling stages of corn and tomato.

[0012] Furthermore, the 3D point cloud data is acquired through multi-view stereo vision, LiDAR, or structured light scanner, with a spatial resolution of ≤0.2 mm, supporting high-density point clouds containing millions of points.

[0013] Furthermore, the data augmentation strategies employed during the training process include one or more of the following: center offset along the three coordinate axes, random scaling within a preset range, and random flipping along a specified plane.

[0014] A three-dimensional point cloud plant segmentation system applicable across the entire soybean growth period includes: The data preprocessing module is used to receive the original three-dimensional point cloud data of soybeans throughout the entire growth period from V1 to R7, and run the preprocessing script to calculate the bounding box volume, local point density and coefficient of variation of the point cloud for each plant. The dynamic multi-level sampling module is used to perform two-step progressive downsampling: first, random sampling retains the original point cloud at a preset ratio, and then adaptively adjusts the sampling ratio of the farthest point based on the joint criterion of bounding box volume and local point density to output a lightweight point set. The multi-scale feature extraction module is used to construct a multi-scale neighborhood through spherical queries with three different radii, encode the relative geometric position of points within the neighborhood, and generate a multi-scale feature representation through a multilayer perceptron; The geometry-aware attention module is used to calculate attention scores based on the Euclidean distance between points, perform weighted aggregation of neighborhood features, and output geometry-aware enhanced features. The encoder-decoder segmentation network module is used to fuse multi-scale features and geometrically enhanced features. It extracts high-level features through downsampling, combines the encoder-side feature pyramid to achieve multi-scale feature fusion, and then restores the resolution through an upsampling layer to output point-by-point stem / leaf semantic segmentation results.

[0015] Compared with existing technologies, the technical solution provided in this invention has the following advantages: This invention employs a dynamic multi-level sampling strategy, multi-scale feature extraction, and a geometric perception attention mechanism to create a three-dimensional point cloud plant segmentation method spanning the entire soybean growth cycle. This method achieves stable and high-precision semantic segmentation of stems and leaves throughout the entire growth cycle from V1 to R7, effectively solving technical problems such as the inability of traditional models to adapt to plant morphological differences throughout the entire growth cycle, easy missegmentation of high-density canopies, and blurred boundaries in areas with significant geometric changes. Specifically, the dynamic multi-level sampling adaptively adjusts the sampling rate based on the spatial sparsity and local complexity of the plant, significantly reducing the amount of point cloud data and improving processing efficiency while ensuring complete... The system retains key geometric structural features, and multi-scale feature extraction and geometric perception attention mechanisms can enhance the feature representation of areas with significant geometric changes in leaf overlap and stem-leaf junctions, significantly improving segmentation accuracy and robustness. The encoder-decoder structure combined with feature pyramids enables efficient fusion of multi-scale features. It not only achieves excellent performance in soybean full-cycle segmentation but can also be directly transferred to seedling stages of crops such as corn and tomatoes, demonstrating strong generalization ability. This enables efficient processing of high-density point cloud data, providing stable, reliable, and versatile 3D vision core technology support for soybean phenotyping, intelligent breeding, growth monitoring, and precision agriculture. It has significant practical value and broad prospects for promotion. Attached Figure Description

[0016] Figure 1 A schematic diagram outlining the SOY3DSEG workflow provided in this embodiment of the invention; Figure 2 Detailed diagram of the SOY3DSEG network architecture provided in this embodiment of the invention; Figure 3 This is a line graph comparing mAcc values ​​for various models throughout the entire soybean growth period, provided in an embodiment of the present invention. Figure 4 This is a trend chart of mIoU for various models throughout the entire growth period of soybean provided in this embodiment of the invention; Figure 5 This is a comparison chart of the segmentation results of SOY3DSEG and the baseline model provided in an embodiment of the present invention; Figure 6 Enlarged view of the V4 and R6 stage segmentation details provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the time sequence segmentation results of corn and tomato seedlings provided in an embodiment of the present invention; Figure 8 This is a comparison chart of the sampling effects of DMSS, RS, and FPS provided in an embodiment of the present invention. Detailed Implementation

[0017] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0018] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0019] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0020] Example 1 like Figures 1-8 As shown, a three-dimensional point cloud plant segmentation method applicable across the entire soybean growth cycle is presented. The model is trained and validated using the publicly available soybean-MVS dataset. This dataset contains 102 three-dimensional point cloud samples of soybean plants, fully covering 12 growth stages from V1 to R7. The point cloud data is acquired using a multi-view stereo vision system with a spatial resolution ≤0.2 mm, ensuring accurate capture of the fine geometric structures of organs such as stems and leaves. The method includes the following steps: Input the original three-dimensional point cloud data of soybeans throughout the entire growth period from V1 to R7, and run the preprocessing script to calculate the bounding box volume, local point density, and coefficient of variation of the point cloud for each plant. Specifically, the original soybean point cloud data is input, and each original point cloud sample is standardized and converted into a .txt format, which includes X, Y, Z coordinates and semantic labels (stem=1, leaf=0). At the same time, the bounding box volume, local point density and coefficient of variation of each plant are calculated to provide prior parameters for the subsequent dynamic sampling module. The volume of the point cloud bounding box (Vbbox) for each plant is calculated using axis-aligned bounding boxes (AABB): , Where xmax, xmin, ymax, ymin, zmax, and zmin are the extreme values ​​of the point cloud along the three coordinate axes.

[0021] Local point density Defined as: , Where Nk(i) is the set of k nearest neighbors of point i; The coefficient of variation includes: The coefficient of variation for spatial sparsity is: , The coefficient of variation of local complexity is: , Where σ represents the standard deviation and μ represents the mean.

[0022] The dataset is divided into spatiotemporally decoupled sets based on year and variety: the training set contains point cloud data of the entire growth period of 5 different varieties, the validation set is the DN252 variety (covering the V3-R6 stages), and the test set is the full growth period data of the DN251 variety. This division method ensures that the model can learn the general features across varieties and avoids overfitting to specific varieties.

[0023] A dynamic multi-level sampling strategy is adopted to downsample the preprocessed original point cloud. The first step is to retain the original point cloud at a preset ratio through random sampling. Then, the sampling ratio of the farthest point is adaptively adjusted based on the joint criterion of bounding box volume and local point density to generate a lightweight point set that retains the geometric structure. Specifically, for the input point cloud data, this embodiment first performs random sampling. Through ablation experiments comparing random sampling ratios of 5%, 10%, and 15%, it was found that when 10% of the original point cloud is retained, the computational load is reduced by approximately 65%-75% while the mIoU decreases by less than 1%, achieving an optimal balance between efficiency and accuracy. Retaining 10% of the original point cloud data achieves initial lightweight data processing. Subsequently, the sampling ratio of the farthest point is adaptively adjusted based on the volume-density joint criterion. Spatial sparsity weight calculation: Divide the point cloud bounding box into uniform voxels with a side length of 5cm, count the number of points in each voxel, and calculate the coefficient of variation of the number of points in each voxel: , in, The standard deviation of the number of points within the voxel. The mean number of points within a voxel; A larger value indicates a more uneven spatial distribution, requiring a reduction in the sampling rate to preserve the characteristics of sparse regions. Local complexity weight calculation: setting =50 nearest neighbors (k is preferably 50, but can be 30-50 when the point cloud density is low, and 50-64 when the point cloud density is high, to balance the ability to depict local structure and computational complexity), calculate the density of the neighborhood points of each point, and calculate the coefficient of variation of the density: , in, The standard deviation of the nearest neighbor density. The mean density of nearest neighbors; A higher value indicates a more complex local structure, requiring a higher sampling rate to preserve detailed information; Determining the dynamic sampling ratio: Calculating the spatial sparsity weights and local complexity weights Set the baseline sampling ratio =0.5, final dynamic sampling ratio ; During the seedling stage (V1-V3), the plant structure is simple. When the value is low, the dynamic sampling ratio is automatically reduced to approximately 0.4 to avoid oversampling; in the R6-R7 high-density canopy stage, leaf overlap is severe. As the value increases, the dynamic sampling ratio automatically increases to approximately 0.7, ensuring that key geometric features are fully preserved.

[0024] Multi-scale feature extraction is performed on the sampling point set. Three neighborhoods with different radii are constructed through spherical query. The relative geometric position and feature differences of points in each neighborhood are encoded. The results are then processed by a multilayer perceptron and spliced ​​to form a multi-scale feature representation. Specifically, for the sampled point set, spherical neighborhoods at three scales are constructed for feature extraction: Neighborhood construction: After normalization by plant height, the spherical query radius is normalized based on the maximum plant height Hmax. rs = αs × Hmax, Where α1=0.05, α2=0.10, α3=0.15, the spherical query radius is set to 0.05 (leaf level), 0.10 (branch level), and 0.15 (plant level), respectively. For each center point i, the neighborhood point set Ni is searched within each radius range. Geometric encoding: For each neighboring point j, calculate the relative geometric position. = - And calculate the Euclidean norm. The relative position vector is input into a three-layer multilayer perceptron (with hidden layer dimensions of 64, 128, and 256, respectively) to generate a learnable position embedding. Feature aggregation: Attention-weighted summation is performed on the position embeddings within the neighborhood of each scale to obtain three scale feature vectors h1, h2, and h3. These vectors are then concatenated through channels to form a multi-scale feature representation. .

[0025] A geometric perception attention mechanism is introduced, which defines an attention score function with decreasing distance based on the Euclidean distance between points, and performs weighted aggregation of neighborhood features to enhance the feature representation of regions with significant geometric changes such as stem-leaf junctions and leaf overlaps. Specifically, to enhance the feature representation of regions with significant geometric changes at the stem-leaf junction and leaf overlap, this embodiment introduces an attention mechanism based on geometric similarity: Distance calculation: For each center point i and its neighboring points j, calculate the Euclidean distance. ; Attention Score: Define an attention score function with decreasing distance. ,in, To adjust the parameters, this function ensures that nearby points receive higher attention, while the weight of distant points automatically decreases. Weight normalization: The scores are normalized using the Softmax function to obtain the attention weights. ; Feature Refinement: Computational Geometry-Aware Enhanced Features , where fj is the original feature of the neighboring point j, and this feature has a stronger discriminative ability in areas where the stem-leaf junction and the overlapping leaves have significant geometric changes.

[0026] Multi-scale features are fused with geometric perception enhancement features and input into the encoder-decoder segmentation network. High-level features are extracted through downsampling, and multi-scale feature fusion is achieved by combining the encoder-side feature pyramid. Then, the resolution is restored through an upsampling layer, and finally, the stem / leaf semantic segmentation results are output through a fully connected classification head. Specifically, this embodiment adopts an encoder-decoder architecture that includes 4 layers of downsampling and 4 layers of upsampling: The encoder contains four transition down layers. Each layer downsamples by sampling the farthest point (sampling rate 0.5) and combines it with a multilayer perceptron to extract high-level features. The outputs of each layer form a feature pyramid to achieve multi-scale feature alignment. Feature fusion: Multi-scale features and geometrically perceived features are concatenated through channels and then linearly projected to obtain fused features. ,in, For multi-scale local geometric features, This is a global geometric perception feature. It is a linear projection matrix. This represents the final feature representation after fusion. Decoder: Contains 4 Transition Up layers, recovers point cloud resolution through distance-weighted upsampling, combines skip connections to fuse features from each layer of the encoder, and finally outputs the segmentation result (stem = 1, leaf = 0) through a 2-class fully connected layer.

[0027] Model training and optimization This embodiment uses the following training configuration: Optimizer: AdamW optimizer, initial learning rate 0.003, weight decay 0.01, using cosine learning rate decay strategy, minimum learning rate set to 1e-5; Training parameters: Batch size is set to 2, automatic mixed precision training is enabled to accelerate computation, and an early stopping mechanism is set to terminate training when the validation set loss shows no improvement for 5 consecutive epochs. Data augmentation: During training, online data augmentation strategies are implemented, including random center offset of ±0.5 meters along the three coordinate axes, random scaling in the range of [0.8, 1.2], and random flipping along the XZ plane, which effectively improves the model's generalization ability. Hardware environment: Training was completed on a single NVIDIA RTX 4090 GPU (24GB VRAM), with an AMD 9950X CPU and PyTorch 2.1.0 + CUDA 11.8 software environment.

[0028] Performance verification On the test set, the model in this embodiment (SOY3DSEG) performs excellently: Segmentation accuracy throughout the entire growth period: The average crossover and union ratio of stem and leaf segmentation during the entire growth period of soybean V1 to R7 reached 93.34%, and the average pixel accuracy exceeded 95%. In the high-density canopy stage of R6-R7, the average crossover and union ratio remained above 89.26%, which is significantly better than mainstream models such as Point Transformer (improved by 14.18%) and RandLA-Net. Sampling efficiency: Compared with fixed-ratio sampling at the furthest point, the dynamic multi-level sampling module improves sampling efficiency by 44.8% (soybeans) and 89.6% (corn), and model training takes only 622 minutes per 100 batches; Cross-crop generalization ability: Zero-sample transfer tests were conducted on point cloud data of corn and tomato seedlings, with average crossover and union ratios of 98.7% and 98.1% respectively, achieving effective transfer without additional fine-tuning; Segmentation effect: In the enlarged segmentation details of V4 and R6 stages, the boundary at the junction of stem and leaf is clear, the phenomenon of missegmentation due to adhesion in the overlapping area of ​​leaves is greatly reduced, and the detailed structure of petioles, branches and other details can be preserved in detail.

[0029] Cross-crop application examples For point cloud data of crops such as corn and tomatoes during the seedling stage, segmentation can be achieved by directly loading the pre-trained weights of the model in this embodiment. Taking corn seedling point cloud as an example, after inputting the untuned model, the average intersection-union ratio of the stem region in the output segmentation result reaches 98.7%. If it is necessary to further improve the segmentation accuracy of the mature stage, 50 corn mature stage point clouds can be used for fine-tuning (fine-tuning learning rate 0.0003, training for 15 epochs), and the average intersection-union ratio is improved to 99.2%.

[0030] This embodiment successfully achieves high-precision organ-level segmentation and effective cross-crop migration of soybeans throughout their entire growth period through a collaborative design of dynamic sampling, multi-scale feature fusion, and geometric perception attention enhancement, providing key technical support for precision agriculture and intelligent breeding.

[0031] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period, characterized in that, Includes the following steps: Input the original three-dimensional point cloud data of soybeans throughout the entire growth period from V1 to R7, and run the preprocessing script to calculate the bounding box volume, local point density, and coefficient of variation of the point cloud for each plant. A dynamic multi-level sampling strategy is adopted to downsample the preprocessed original point cloud. The first step is to retain the original point cloud at a preset ratio through random sampling. Then, the sampling ratio of the farthest point is adaptively adjusted based on the joint criterion of bounding box volume and local point density to generate a lightweight point set that retains the geometric structure. Multi-scale feature extraction is performed on the sampling point set. Three neighborhoods with different radii are constructed through spherical query. The relative geometric position and feature differences of points in each neighborhood are encoded. The results are then processed by a multilayer perceptron and spliced ​​to form a multi-scale feature representation. A geometric perception attention mechanism is introduced, which defines an attention score function with decreasing distance based on the Euclidean distance between points, and performs weighted aggregation of neighborhood features to enhance the feature representation of regions with significant geometric changes such as stem-leaf junctions and leaf overlaps. Multi-scale features are fused with geometrically aware features and input into the encoder-decoder segmentation network. High-level features are extracted through downsampling, and multi-scale feature fusion is achieved by combining the encoder-side feature pyramid. Then, the resolution is restored through an upsampling layer, and finally, the stem / leaf semantic segmentation results are output through a fully connected classification head.

2. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, In the dynamic multi-level sampling strategy, the joint criteria include: Spatial sparsity weight is determined by dividing the point cloud space into voxels and calculating the coefficient of variation of the number of points within each voxel. The larger the coefficient of variation, the more uneven the spatial distribution, and the sampling rate needs to be reduced to preserve the features of sparse regions. The local complexity weight is determined by calculating the coefficient of variation of the density of k nearest neighbors. The higher the coefficient of variation, the more complex the local structure, and the more the sampling rate needs to be increased. The dynamic sampling ratio is calculated based on the weighted fusion of the spatial sparsity weight and the local complexity weight.

3. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, In the multi-scale feature extraction, the three spherical queries with different radii are normalized according to the plant height and set as the first radius, the second radius, and the third radius, respectively, corresponding to the leaf level, branch level, and plant level scales. The relative geometric position of the neighborhood points is encoded as a point difference vector, which is mapped to a position embedding through a multilayer perceptron.

4. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, In the geometric perception attention mechanism, the distance-decreasing attention score function is defined as: ， in, The Euclidean distance between points. To adjust the parameters.

5. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, The encoder-decoder segmentation network includes: The encoder contains multiple downsampling layers. Each layer downsamples by sampling the farthest point and combines it with a multilayer perceptron to extract features, thus constructing a multi-scale feature pyramid. The decoder contains multiple upsampling layers, recovers resolution through distance-weighted upsampling, and fuses features from each layer of the encoder through skip connections.

6. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, The method employs the AdamW optimizer for training, combines a learning rate decay strategy with an early stopping mechanism, and supports an automatic mixed-precision training mode.

7. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, The method achieves an average crossover ratio of a first predetermined threshold in the stem-leaf segmentation during the entire growth period of soybean (V1 to R7 stages), an average crossover ratio of a second predetermined threshold in the high-density canopy stage (R6-R7), and an average crossover ratio of a third predetermined threshold in cross-crop segmentation tests during the seedling stages of corn and tomato.

8. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, The 3D point cloud data is acquired through multi-view stereo vision, LiDAR, or structured light scanner, with a spatial resolution of ≤0.2 mm, supporting high-density point clouds containing millions of points.

9. The three-dimensional point cloud plant segmentation method applicable across the entire soybean growth period as described in claim 1, characterized in that, The data augmentation strategies employed during training include one or more of the following: center offset along the three coordinate axes, random scaling within a preset range, and random flipping along a specified plane.

10. A three-dimensional point cloud plant segmentation system applicable across the entire soybean growth period based on the method of any one of claims 1-9, characterized in that, include: The data preprocessing module is used to receive the original three-dimensional point cloud data of soybeans throughout the entire growth period from V1 to R7, and run the preprocessing script to calculate the bounding box volume, local point density and coefficient of variation of the point cloud for each plant. The dynamic multi-level sampling module is used to perform two-step progressive downsampling: first, random sampling retains the original point cloud at a preset ratio, and then adaptively adjusts the sampling ratio of the farthest point based on the joint criterion of bounding box volume and local point density to output a lightweight point set. The multi-scale feature extraction module is used to construct a multi-scale neighborhood through spherical queries with three different radii, encode the relative geometric position of points within the neighborhood, and generate a multi-scale feature representation through a multilayer perceptron; The geometry-aware attention module is used to calculate attention scores based on the Euclidean distance between points, perform weighted aggregation of neighborhood features, and output geometry-aware enhanced features. The encoder-decoder segmentation network module is used to fuse multi-scale features and geometrically enhanced features. It extracts high-level features through downsampling, combines the encoder-side feature pyramid to achieve multi-scale feature fusion, and then restores the resolution through an upsampling layer to output point-by-point stem / leaf semantic segmentation results.

Images