Rapeseed whole growth cycle monitoring method and system based on three-dimensional point cloud reconstruction

Abstract

The application discloses a rape whole growth cycle monitoring method and system based on three-dimensional point cloud reconstruction, relates to the field of intelligent agriculture and computer vision technology, collects three-dimensional point cloud data at multiple key growth periods of rape, constructs a semantic growth graph after pretreatment; the graph is input into a pre-trained time sequence model, the energy flow among source nodes, library nodes and channel nodes is simulated based on the energy dynamic balance principle, the future growth state of the plant is deduced, and growth monitoring data is finally output. The application realizes nondestructive monitoring of rape from the bud stage to the silique stage through three-dimensional point cloud technology, and overcomes the problems of low efficiency and strong destructiveness of traditional methods. Through the construction of the semantic growth graph, the organ recognition and segmentation problem is solved. Further, a calculable model is constructed by combining the source-library theory, photosynthetic product flow and organ competition are simulated, the biomass is accurately predicted before the silique forms, and the phenomenon of abortion is recognized, so that the leap from morphological monitoring to growth mechanism deduction is realized.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and computer vision technology, specifically to a method and system for monitoring the entire growth cycle of rapeseed based on 3D point cloud reconstruction. Background Technology

[0002] With the rapid development of modern agriculture towards precision, intelligence, and high throughput, crop phenotypic monitoring technology is becoming a core support for smart agriculture. This trend shows great potential in improving breeding efficiency, achieving precision cultivation, and predicting crop yields. However, the accompanying need for technology to conduct full-cycle, high-throughput, and non-destructive precision monitoring of crop growth status projections is becoming increasingly urgent and complex.

[0003] Rapeseed, as an important oilseed crop, ultimately determines its yield during the silique development stage. The number, spatial distribution, and developmental status (such as abortion) of siliques directly determine the final yield. In traditional breeding and cultivation research, the investigation of key traits such as siliques in rapeseed mainly relies on manual field sampling and destructive laboratory measurements. This method is not only inefficient and highly subjective, but also cannot achieve continuous tracking of the entire growth cycle of the same plant, especially failing to capture the dynamic growth process from the budding stage, flowering stage to the silique stage, severely restricting the breeding process of high-yielding and superior varieties.

[0004] In the crop phenotypic monitoring technology system, 3D point cloud reconstruction technology occupies an important position due to its ability to non-destructively acquire the three-dimensional spatial structural information of plants. Currently, most existing crop monitoring methods based on 3D point clouds focus on relatively simple structural stages such as the seedling stage and vegetative growth stage, or only extract overall morphological parameters such as plant height and canopy width. Undeniably, these methods can achieve a certain degree of macroscopic assessment of crop growth status projections, but they have significant limitations:

[0005] First, it cannot solve the problem of point cloud segmentation caused by the complex plant structure and severe shading and adhesion between organs during the silique stage. It is difficult to accurately extract semantic information of organs such as leaves, stems, and siliques from three-dimensional point clouds, and it is impossible to construct a structured map that can characterize the functional attributes of organs.

[0006] Secondly, it lacks high-throughput population monitoring and parallel processing capabilities, and cannot make early and accurate predictions of the developmental fate (such as normal development or abortion) and final yield based on the energy competition relationship between organs before the silique is fully formed. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction, so as to overcome the shortcomings of the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction, comprising:

[0009] At least three key growth stages of rapeseed, obtain multi-view three-dimensional point cloud data of the same plant or plant population.

[0010] The three-dimensional point cloud data is preprocessed to obtain clean point cloud data after denoising and registration;

[0011] A semantic growth graph is constructed based on the pure point cloud data, wherein the nodes in the semantic growth graph are assigned type attributes, and the type attributes include at least source nodes representing energy supply sites, library nodes representing energy consumption sites, and channel nodes representing energy transmission channels.

[0012] The constructed semantic growth graph is input into a pre-trained temporal prediction model;

[0013] Using the aforementioned time-series prediction model, based on the principles of energy allocation and dynamic balance, the energy flow between the source node, the sink node, and the channel node is dynamically simulated to predict the growth state of the plant from the current moment to the future target moment.

[0014] Based on the growth status projection results, growth monitoring data is output, including silique development prediction results, organ biomass distribution, and overall plant yield prediction. The growth monitoring data also includes rapeseed line yield potential assessment indicators.

[0015] In a preferred embodiment, the step of constructing the semantic growth graph includes:

[0016] Extract the topological skeleton of the plant from the pure point cloud data;

[0017] Semantic classification of skeleton nodes: nodes corresponding to leaf regions are classified as source nodes, nodes corresponding to growth points, flowers, and siliques are classified as library nodes, and nodes corresponding to stems are classified as channel nodes.

[0018] Establish the mapping relationship between the skeleton nodes and the original 3D point cloud.

[0019] In a preferred embodiment, the time series prediction model is a differential equation-driven model based on physical mechanisms, with its core being a source-sink dynamic system;

[0020] The operating mechanism of the source-sink dynamics system is as follows: the plant structure represented by the semantic growth map is regarded as a dynamic flow network. The source node generates energy, and the sink node consumes energy to grow. The energy is dynamically distributed throughout the network according to the sink attraction of each sink node and the path conduction efficiency between it and the source node.

[0021] In a preferred embodiment, the time-series prediction model employs a hybrid architecture that combines a spatiotemporal graph neural network with the source-reservoir dynamics system;

[0022] The spatiotemporal graph neural network takes a sequential semantic growth graph as input and its output is used to estimate the key physiological parameters required for the operation of the source-sink dynamic system.

[0023] In a preferred embodiment, the step of inferring the plant growth status is specifically as follows:

[0024] By using key physiological parameters estimated by a spatiotemporal graph neural network, the source-sink dynamic system is driven to simulate the dynamic biomass accumulation trajectory of each sink organ from the current moment to the future target moment through numerical integration.

[0025] In a preferred embodiment, the silique development prediction results include a list of siliques identified as aborted, and the criteria for determining aborted siliques are: during the projection process, their biomass growth rate is continuously lower than a preset abortion threshold for a continuous period of time.

[0026] In a preferred embodiment, the energy flow allocation is achieved in the following manner:

[0027] On the semantic growth graph, calculate the path propagation efficiency from each source node to each library node;

[0028] The proportion of energy flow allocated to a reservoir node is directly proportional to its own reservoir attraction, directly proportional to the path conduction efficiency from the source node to the reservoir node, and inversely proportional to the total competitive attraction of all reservoir nodes. The competitive attraction is the sum of the reservoir attraction parameters of all reservoir nodes.

[0029] In a preferred embodiment, during the step of constructing the semantic growth graph, nodes are also assigned a resource competition index feature; the resource competition index is obtained by analyzing the canopy point cloud of adjacent plants, simulating the solar trajectory, and calculating the daily cumulative value of photosynthetically active radiation.

[0030] This invention also provides a rapeseed full growth cycle monitoring system based on three-dimensional point cloud reconstruction, including:

[0031] The data acquisition module is used to acquire multi-view three-dimensional point cloud data of rapeseed plants;

[0032] The graph construction module is used to process point cloud data and construct the semantic growth graph.

[0033] The dynamic simulation engine has the aforementioned time-series prediction model built-in, which is used to perform growth state simulation results based on the principles of energy allocation and dynamic balance.

[0034] A monitoring output module is used to output the growth monitoring data;

[0035] The breeding decision support module is configured to rank different rapeseed lines by yield potential based on the growth monitoring data.

[0036] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0037] This invention utilizes 3D point cloud reconstruction technology to replace traditional, manual, and destructive field sampling and measurement methods, enabling continuous and non-destructive tracking of the same plant or population from the budding stage, flowering stage to the pod-forming stage. This addresses the fundamental shortcomings of traditional methods, such as low efficiency, strong subjectivity, and inability to obtain dynamic growth data, laying a solid foundation for high-throughput phenotypic identification in the breeding process.

[0038] This invention transforms disordered 3D point clouds into topological structures rich in biological semantics by constructing a semantic growth map. This map not only clearly characterizes the skeletal structure of the plant but also performs functional classification of nodes, thereby enabling accurate identification, segmentation, and functional labeling of organs such as leaves, stems, and siliques, providing a structured data foundation for subsequent mechanism simulation.

[0039] The core innovation of this invention lies in modeling and computing the source-sink theory of plant physiology. Through a hybrid architecture combining spatiotemporal graph neural networks and source-sink dynamic differential equations, the model can simulate the dynamic allocation and competition of photosynthetic products within the source-channel-sink network. This makes the system not only a data fitter but also a digital twin capable of revealing the intrinsic growth patterns of plants. Based on this, this invention can accurately predict the final biomass of siliques before they are fully formed and identify siliques that will abort due to energy competition failure at an early stage, providing unprecedented accuracy and foresight for yield prediction. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0041] Figure 1 This is a flowchart of the method of the present invention.

[0042] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Semantic growth graph: A structured graph consisting of plant topological skeleton nodes, connecting edges between nodes, and node attribute information. Node attribute information includes node type attributes and resource competition index features, while connecting edge attribute information includes path transmission efficiency.

[0045] Source node / library node / channel node: The source node is the topological skeleton node corresponding to the leaf region of the plant, the library node is the topological skeleton node corresponding to the growth point, flower, and silique region of the plant, and the channel node is the topological skeleton node corresponding to the stem region of the plant.

[0046] Source-sink dynamics: The plant is regarded as a dynamic flow network, and a dynamic model is used to describe the energy generation of source nodes, energy consumption of sink nodes, and energy transfer of channel nodes based on differential equations.

[0047] Path transmission efficiency: In the semantic growth graph, it is the product of the transmission efficiencies of all channel nodes on the path from the source node to the library node. The channel node transmission efficiency is calculated by the stem diameter, length and preset species specificity coefficient.

[0048] Resource competition index: A quantitative indicator representing the daily cumulative value of photosynthetically active radiation at a node, calculated by analyzing the canopy point clouds of adjacent plants and simulating the solar trajectory.

[0049] Example 1, please refer to Figure 1 As shown in this embodiment, the method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction includes:

[0050] S1. Obtain multi-view three-dimensional point cloud data of the same plant or plant population during at least three key growth stages of rapeseed.

[0051] S2. Preprocess the three-dimensional point cloud data to obtain clean point cloud data after denoising and registration;

[0052] S3. Construct a semantic growth graph based on the pure point cloud data, wherein the nodes in the semantic growth graph are assigned type attributes, and the type attributes include at least source nodes representing energy supply sites, library nodes representing energy consumption sites, and channel nodes representing energy transmission channels.

[0053] S4. Input the constructed semantic growth graph into the pre-trained temporal prediction model;

[0054] S5. Using the time-series prediction model, based on the principle of energy distribution and dynamic balance, the energy flow between the source node, the sink node and the channel node is dynamically simulated to deduce the growth state of the plant from the current time to the future target time.

[0055] S6. Based on the growth state projection results, output growth monitoring data including silique development prediction results, organ biomass distribution, and overall plant yield prediction.

[0056] As described in S1-S6 above, rapeseed, as an important oilseed crop, ultimately determines its yield during the silique development stage. The number, spatial distribution, and developmental status (such as abortion) of siliques directly determine the final yield. In traditional breeding and cultivation research, the investigation of key traits such as rapeseed siliques mainly relies on manual field sampling and destructive laboratory measurements. This method is not only inefficient and highly subjective, but also cannot achieve continuous tracking of the entire growth cycle of the same plant, especially making it difficult to capture the dynamic growth process from the budding stage, flowering stage to the silique stage, which seriously restricts the breeding process of high-yielding and superior varieties.

[0057] In the crop phenotypic monitoring technology system, 3D point cloud reconstruction technology occupies an important position due to its ability to non-destructively acquire the three-dimensional spatial structural information of plants. Currently, most existing crop monitoring methods based on 3D point clouds focus on relatively simple stages such as the seedling stage and vegetative growth stage, or only extract overall morphological parameters such as plant height and crown width. Admittedly, these methods can achieve a certain degree of macroscopic assessment of the results of crop growth status projection, but they have obvious limitations: First, they cannot solve the problem of point cloud segmentation caused by the complex plant structure and severe shading and adhesion between organs during the silique stage, making it difficult to accurately extract semantic information of organs such as leaves, stems, and siliques from 3D point clouds, and unable to construct structured maps that can characterize organ functional attributes; Second, they lack high-throughput population monitoring and parallel processing capabilities, and cannot make early and accurate predictions of the developmental fate (such as normal development or abortion) and final yield of organs based on the energy competition relationship between organs before the siliques are fully formed.

[0058] By employing 3D point cloud reconstruction technology, traditional manual and destructive field sampling methods are replaced, enabling continuous and non-destructive tracking of the same plant or population from the budding stage, flowering stage to the silique stage. This addresses the fundamental pain points of traditional methods—low efficiency, strong subjectivity, and inability to obtain dynamic growth data—laying a solid foundation for high-throughput phenotypic identification in the breeding process. By constructing a semantic growth map, the disordered 3D point cloud is transformed into a topological structure rich in biological semantics. This map not only clearly characterizes the plant's skeletal structure but also functionally classifies nodes, thus achieving accurate identification, segmentation, and functional labeling of organs such as leaves, stems, and siliques, providing a structured data foundation for subsequent mechanism simulation. Through a hybrid architecture of spatiotemporal graph neural networks and source-sink dynamic differential equation systems, the model can simulate the dynamic allocation and competition of photosynthetic products in the source-channel-sink network. This makes the system not only a data fitter but also a digital twin capable of revealing the plant's intrinsic growth patterns. Based on this, the present invention can accurately predict the final biomass of siliques before they are fully formed, and identify siliques that will fail to develop due to energy competition failure at an early stage, providing unprecedented accuracy and foresight for yield prediction.

[0059] In one embodiment, step S3 specifically includes:

[0060] S31: Extract the topological skeleton of the plant from pure point cloud data;

[0061] The purified point cloud data is a collection of point clouds containing only the three-dimensional spatial coordinates of the target plant after denoising and background removal processing. When extracting the topological skeleton, a combined method based on distance transformation and thinning algorithms is used.

[0062] First, the pure point cloud data is spatially gridded. The shortest distance from each point in the point cloud to the plant surface is calculated by distance transformation. Points whose distance values ​​meet the preset threshold are retained as skeleton candidate points.

[0063] Secondly, an iterative refinement algorithm is used to optimize the candidate points, remove redundant points and retain key connection structures, and finally obtain a topological skeleton composed of nodes (including endpoints, branch points and intermediate points) and connecting edges. The spatial coordinate error of the nodes is controlled within ±0.2mm, and the deviation of the direction of the connecting edges from the actual growth direction of the plant branches does not exceed 5°, so as to accurately reflect the connection relationship of the plant branches, spatial growth posture and overall morphological structure.

[0064] S32: Perform semantic classification on skeleton nodes;

[0065] Based on the physiological structure and functional characteristics of the plant, semantic annotation was performed on the skeleton nodes extracted from S31, specifically including:

[0066] Source node classification: By analyzing the geometric features of the point cloud regions associated with the skeleton nodes (the point cloud in the leaf region is usually flat, with concentrated normal vector distribution and an angle greater than 60° with the branch, and the point cloud density is 2-5 times that of the stem region), and combining the preset leaf morphology template (such as leaf area threshold, aspect ratio range), the nodes in the corresponding leaf region (including the connection point between the leaf and petiole, the end point of the leaf midrib, etc.) are marked as source nodes (the source nodes are the main production sites of photosynthetic products).

[0067] Library node classification: For parts such as growth points, flowers, and siliques that store or consume nutrients, their associated point cloud regions usually exhibit a compact cluster structure (point cloud volume less than 5cm³, spatial aggregation degree more than 30% higher than that of the stem region), and growth points have a pointed shape (end point curvature radius less than 1mm), and flowers / siliques have specific color characteristics (which can be filtered through the RGB information attached to the point cloud, such as the RGB value range of flower color). Nodes that meet the above characteristics are marked as library nodes.

[0068] Channel node classification: The remaining nodes all correspond to stem parts, and their associated point cloud regions exhibit a slender columnar structure (length to diameter ratio greater than 10:1). They connect source nodes and sink nodes or different sink nodes, and are responsible for nutrient transport; therefore, they are labeled as channel nodes. During the classification process, machine learning models (such as a random forest classifier based on point cloud features, with a training sample size of no less than 50 labeled plants of the same type) can be combined to improve classification accuracy, ensuring that the recall rate of semantic classification is no less than 90%.

[0069] S33: Establish the mapping relationship between skeleton nodes and the original 3D point cloud;

[0070] The original 3D point cloud is initial point cloud data containing complete plant surface information without skeleton extraction (it originates from the same source as the pure point cloud data in S11, retaining all original coordinates and attribute information). The mapping relationship is achieved through spatial region association:

[0071] First, construct a spherical neighborhood with radius r centered on each skeleton node (r is half the average diameter of the part to which the node belongs, such as r = stem diameter / 2 for stem nodes and r = twice the leaf thickness for leaf nodes).

[0072] Secondly, calculate the Euclidean distance between each point in the original 3D point cloud and the skeleton node, and mark the point cloud region that falls within the spherical neighborhood as the associated region of that node.

[0073] Finally, an index table is used to record the ID of each skeleton node and the coordinate range, number of points, and attributes of the corresponding associated point cloud region, establishing a one-to-one mapping relationship to ensure that any skeleton node can be traced back to the specific spatial region in the original point cloud.

[0074] As described in S31-S33 above, topological skeleton extraction simplifies the three-dimensional structure of the plant while accurately preserving key morphological features such as branch connections and growth postures, providing an efficient structural carrier for subsequent analysis. Node classification based on plant physiological functions associates abstract skeleton nodes with specific physiological parts, providing a semantic basis for functional studies such as plant nutrient transport simulation and growth dynamic analysis. The established mapping relationship enables bidirectional association between skeleton nodes and the original point cloud, ensuring that skeleton analysis results can be verified through the original point cloud and that key areas in the original point cloud can be quickly located through the skeleton, improving the accuracy and efficiency of data utilization. Specific ranges or bases are given for parameter settings and algorithm selection in each sub-step, and these are combined with the geometric and physiological characteristics of the plant. Those skilled in the art can adjust the parameters according to the actual plant species to ensure the scheme is repeatable.

[0075] In one embodiment, step S5 specifically includes:

[0076] S51: Define the overall framework and core objectives of the model:

[0077] The time-series prediction model is a differential equation-driven model based on physical mechanisms. Its core is the source-sink dynamics system. It aims to dynamically predict the biomass change trend of each sink organ (such as the growing point, flower, silique, etc.) at different growth stages by quantifying the material and energy flow patterns between the plant's source organs (the parts that produce photosynthetic products), sink organs (the parts that store / consume products), and channels (the parts that transport products).

[0078] The core logic of the model is based on the source-sink relationship in plant physiology: source organs produce assimilates through photosynthesis, which are transported to sink organs via pathways. The biomass growth of sink organs depends on the difference between the amount of products allocated and their own maintenance consumption. The model inputs are initial plant morphological parameters (such as the number of source / sink organs and initial biomass) and environmental parameters (such as light and temperature), and the output is the predicted biomass values ​​of each sink organ over future time series (such as hourly and daily).

[0079] S52: Decomposing the core differential equations of the source-sink dynamics system:

[0080] The operating mechanism of the source-sink dynamics system is as follows: the plant structure represented by the semantic growth map is regarded as a dynamic flow network. Source nodes generate energy, and sink nodes consume energy for growth. Energy is dynamically distributed throughout the network according to the sink attraction of each sink node and the path conduction efficiency between it and the source node.

[0081] For any reservoir organ i, the instantaneous growth rate of its biomass Mi is:

[0082] =η· · (1);

[0083] in:

[0084] : The biomass growth rate of the i-th reservoir organ, positive values ​​represent growth, and negative values ​​represent consumption;

[0085] η: Global assimilation and conversion efficiency coefficient, which characterizes the efficiency of source organ assimilation products to sink organ biomass. The value is 0.3-0.7 (depending on the crop type, such as about 0.5 during the wheat heading stage and about 0.6 during the rapeseed pod stage). It can be determined by experiments on the ratio of source organ assimilation to sink organ weight gain.

[0086] Energy flow allocation coefficient: reflects the priority of channel supply to organ i. ∈ And the sum is 1, which is regulated by the growth stage and spatial location of the deposit organs (such as young deposit organs). (Higher levels), which can be determined by isotope labeling;

[0087] Total source intensity of the plant represents the total amount of assimilated products per unit time of all source organs, and is positively correlated with the photosynthetic capacity and area of ​​the source organs;

[0088] The sink attraction of sink organ i reflects its ability to compete for assimilates and is positively correlated with metabolic activity, size, and growth potential.

[0089] The sum of the attractiveness of all library organs, used to normalize the attractiveness percentage of a single library;

[0090] The maintenance respiration consumption function of organ i is related to the current biomass Mi (e.g., the consumption of the growing point is higher than that of the mature silique), and can be determined by the rate of biomass reduction under dark conditions.

[0091] S53: Clarify the operating mechanism of the source-liquid dynamics system:

[0092] Based on equation (1), the system operation mechanism includes three core processes:

[0093] Source organ product supply: Source organs produce assimilates through photosynthesis, with a total source intensity of [missing information]. It is affected by the environment (such as increased light). (increase) and the state of the source organ (such as leaf senescence) (Reduce) Dynamic regulation;

[0094] Inter-liquidity distribution of products: total source intensity First, by the proportion of attractiveness of the warehouse Allocation, and then through Secondary regulation is used to ultimately determine the actual amount of organ i obtained from the depot.

[0095] Net increase in organ biomass: Actual gain minus respiration consumption after η conversion. To obtain the net growth rate Positive values ​​represent accumulation, while negative values ​​represent consumption.

[0096] S54: Model initialization and solution methods:

[0097] Initialization conditions: Bank organ biomass at initial time t=0 (0) Estimated using 3D point cloud data (e.g., combining volume and density); initial parameters were determined experimentally or calibrated using literature (default values ​​were used when no measured data were available, such as η=0.5);

[0098] Numerical solution: The fourth-order Runge-Kutta method is used to solve equation (1), with a time step of 1 hour and a solution interval of [0, T] (e.g., 7 days); the solution is updated in real time during the solution process. , , To ensure that the parameters match the results of the plant growth state projection;

[0099] Model validation: The parameters were calibrated by measuring organ biomass in the experimental bank, so that the root mean square error (RMSE) of the prediction was ≤10%.

[0100] As described in S51-S54 above, based on the physiological mechanism of plant source-sink, energy flow is quantified through core differential equations to avoid the black box problem, and the results can be interpreted in conjunction with physiological laws; key parameters have clear physical meanings and measurement methods, and those skilled in the art can adjust them according to crop type to ensure repeatability; parameters are dynamically updated with the results of plant growth state extrapolation, which can simulate the differences in biomass allocation at different growth stages and improve prediction accuracy.

[0101] In one embodiment, step S5 involves constructing a hybrid architecture time-series prediction model based on a spatiotemporal graph neural network and a source-liquid dynamics system, specifically including:

[0102] S55: Define the overall framework and core objectives of the hybrid architecture:

[0103] The time-series prediction model adopts a hybrid architecture of data-driven and mechanism-constrained dual modules, consisting of a spatiotemporal graph neural network module and a source-sink dynamics system module. Its core objective is to integrate the spatiotemporal dynamic characteristics and physiological mechanism laws of plant growth to achieve high-precision time-series prediction of changes in the biomass of sink organs (prediction granularity of 12 hours / step, covering key growth periods of crops such as rapeseed from flowering to silique maturity).

[0104] The two modules work together: the spatiotemporal graph neural network learns latent spatiotemporal features from plant dynamic observation data and outputs key physiological parameters required by the source-sink dynamics system; based on these parameters, the source-sink dynamics system calculates the biomass growth rate of sink organs through physical mechanism equations, and finally outputs time-series prediction results. The overall input is a serialized semantic growth map, and the output is the predicted biomass values ​​of each sink organ at the next T time steps (e.g., 7 days).

[0105] S56: Constructing the network structure of the spatiotemporal graph neural network module:

[0106] Spatiotemporal graphical neural networks are used to model the spatial topological relationships and temporal dynamic characteristics of plant growth. The structure is as follows:

[0107] Graph structure foundation: Based on semantic growth graph, nodes correspond to source (leaf), pool (growth point, flower, silique), and channel (stem) nodes, and edges correspond to physical connection relationships between nodes;

[0108] Core network layer:

[0109] Graph convolutional layer (2 layers, 64 hidden dimensions): Aggregates neighbor node information through adjacency matrix and extracts spatial association features (e.g., edge weights are greater if stalk transportation efficiency is high).

[0110] Temporal convolutional layers (3 layers, 128 hidden dimensions): 1D convolutional kernels (kernel size 3) are used to expand the receptive field through dilated convolution and capture the historical dynamics of 8 time steps;

[0111] Spatiotemporal attention layer: Spatial attention focuses on key growth nodes (such as young organoids), temporal attention highlights important time periods (such as moments of strong light), and outputs fused features;

[0112] Output layer: The fully connected layer maps to estimates of key physiological parameters. The output dimension is m+1, where m is the number of reservoir organs, corresponding to reservoir attractiveness (A1, ..., Am) and total source intensity. .

[0113] S57: Define the input for the serialized semantic growth graph:

[0114] A serialized semantic growth map is a temporal organization of the semantic features of a plant at different time points. The construction method is as follows:

[0115] Single-moment map: includes node features (source nodes: leaf area, photosynthetic rate, etc.; sink nodes: biomass, volume, etc.; channel nodes: diameter, length, etc.), edge features (physical distance, transport resistance), and global features (light, temperature, etc.).

[0116] Serialization processing: Single-time maps of 4 consecutive days (8 time points) are collected at 12-hour intervals to form a sequence {G1, ..., G8}; missing data are completed by linear interpolation, and outliers are filtered and corrected by the 3σ criterion.

[0117] S58: Define the parameter mapping relationship:

[0118] The output of the spatiotemporal graph neural network is directly used as the parameter input of the source-liquid dynamics system, with the following mapping rule:

[0119] Library attractiveness The network outputs the i-th component as The estimated value must satisfy ( >0) (non-negativity is ensured by the activation function), and it is positively correlated with the library node volume and activity level;

[0120] Total source intensity The network outputs the (m+1)th component as The estimated value must meet the following requirements. ≧0 (photosynthetic products are non-negative) and are positively correlated with the total leaf area of ​​the source organ and light intensity.

[0121] S59: Establish a collaborative operation mechanism:

[0122] The two modules collaborate through real-time parameter supply and prediction result feedback. The process is as follows:

[0123] Preprocess the serialized graph and input it into the spatiotemporal graph neural network;

[0124] Network output (A1, ..., Am) and ;

[0125] Substitute the parameters into the differential equation of the source-sink dynamic system and solve for the biomass of the sink organs after 12 hours using the fourth-order Runge-Kutta method;

[0126] Rolling prediction: Update library node features to construct a new graph, repeat steps 2-3 until prediction is completed for T time steps;

[0127] Feedback adjustment: Compare the measured and predicted biomass every 3 time steps, and fine-tune the network weights to correct the deviation when the error exceeds 5%.

[0128] S510: Model training and optimization methods:

[0129] Data-driven training and model optimization, method:

[0130] Training data: Time-series data of at least 3 varieties (50 plants each), including semantic growth maps at 12-hour intervals and measured parameters. , and reservoir organ biomass;

[0131] Loss function: Multi-objective loss (parameter estimation loss + biomass prediction loss), with an emphasis on biomass prediction accuracy;

[0132] Training process: Adam optimizer (learning rate 0.001), batch size 16, 200 training epochs, early stopping method to avoid overfitting;

[0133] Evaluation criteria: mean absolute error (MAE) of biomass prediction for the test set ≤ 0.2 g, and relative error of parameter estimation ≤ 8%.

[0134] As described in S55-S510 above, by combining the high-dimensional feature learning capability of spatiotemporal graph neural networks with the physiological mechanism constraints of source-liquid dynamics systems, the black-box problem of purely data-driven models is avoided, improving prediction reliability and interpretability; the network dynamically estimates changes with growth stages through spatiotemporal feature learning. and This solves the adaptability problem of fixed parameters in traditional mechanistic models; the rolling prediction + feedback adjustment mechanism can adapt to environmental fluctuations and changes in plant status, enabling long-term prediction of key growth periods and improving model accuracy.

[0135] In one embodiment, the deduced plant growth status result in step S5 specifically includes:

[0136] S511: Preprocessing and validity verification of key physiological parameters:

[0137] Key physiological parameters (pool attraction of each reservoir organ) were estimated using a spatiotemporal graph neural network. ) and total plant source intensity Afterwards, pretreatment and verification are performed to ensure that the data conforms to plant physiological laws, providing a reliable input for the source-sink dynamics system.

[0138] Range constraint verification: Check the reasonableness of parameters: ( >0) (Vigorous growth reservoir organs such as young growth points) It is 1.5-3 times that of aging organs); ≥0 (positive during the day, approximately 0 at night). Corrections are made by truncation when values ​​exceed the range (e.g., ...). ≦0 is corrected to 0.1).

[0139] Correlation verification: Verify the correlation between parameters and plant status: It is positively correlated with the volume of the deposit organ and the relative growth rate (correlation coefficient r≧0.7). It is positively correlated with the total leaf area of ​​the source organ and light intensity (r≧0.6).

[0140] Normalization processing: for Proportional normalization (i.e., dividing by the sum of the attractiveness of all reservoir organs) facilitates subsequent calculation of inter-reservoir allocation ratios and reduces numerical errors.

[0141] S512: Set the integration interval and initial conditions:

[0142] Defining the time range for the derivation and the initial state for solving the system of differential equations provides boundary conditions for numerical integration:

[0143] Integration interval: current time step Up to the future target time T{n+1}, the time interval ∆T is set according to the crop growth rhythm (24h for seedling stage, 12h for flowering / horn fruit stage, and higher resolution is required for rapid growth stage).

[0144] Initial biomass: Initial biomass of each reservoir organ at time ( Volume estimation based on semantic segmentation using 3D point cloud: ( ), combined with organ density (e.g., rapeseed growing point p=0.85g / c) ,calculate ( )= · ( The error is controlled within ±5% (by weighing calibration).

[0145] Environmental parameters: Light intensity, temperature, etc. within the integration interval are based on short-term forecast values ​​or constant values ​​(e.g., 800 μmol / (m²・s) indoors) to reduce environmental fluctuations.

[0146] S513: Numerical integration of differential equations based on the fourth-order Runge-Kutta method:

[0147] The fourth-order Runge-Kutta method (RK4) was used to solve the differential equations of the source-sink dynamic system to calculate the changes in sink organ biomass.

[0148] Core equation: The biomass growth rate of sink organ i is:

[0149] =η· · ;

[0150] Wherein, η (conversion efficiency), The allocation coefficient is a species characteristic constant (e.g., η=0.6 for rapeseed). To maintain respiratory expenditure.

[0151] Step size and iteration: Divide ∆T into N small step sizes h= (e.g., taking a 12-hour interval (h=1h)), calculate the biomass at each step using RK4 iteration: based on the current biomass. ( The rate equation is used to calculate the increment, and the biomass at time t{k+1} is finally updated (synchronously updated). (Ensure the allocation ratio is accurate).

[0152] Coupling process: Each reservoir organ is processed through... They are mutually coupled and need to be calculated synchronously within the same time step to avoid timing deviations.

[0153] S514: Generation and analysis of dynamic biomass accumulation trajectory:

[0154] The integral results are organized into a continuous trajectory, and key features are extracted to reflect the growth process of the organ deposits.

[0155] Trajectory construction: Biomass at each time step is recorded to form a discrete sequence, and a continuous function is obtained through cubic spline interpolation. (t).

[0156] Feature extraction: including cumulative increment ∆ = (Total growth), maximum growth rate and its occurrence time (active period), growth stagnation point (resource balance point).

[0157] Collaborative analysis: Comparing the trajectories of different reservoir organs to identify growth centers (e.g., the ∆ of flower organs during flowering). (significantly higher), parsing source-library allocation priority.

[0158] S515: Validation and Correction of Trajectory Validity

[0159] Multiple verifications ensure the trajectory conforms to actual growth patterns, and deviations are corrected accordingly.

[0160] Physiological rationality test: The growth point Mi(t) should continuously increase (∆) >0), aging organs may decline; growth rate change rate ≤20%, avoid mutations.

[0161] Historical trend comparison: Compared with previous trajectories, the ∆ of three consecutive cycles Change rate ≤15% (excluding sudden environmental changes), backtracking correction when deviation is too large. or .

[0162] Actual measurement calibration: Calculate the relative error for the actual measured biomass of 10%-20% of the bank organs.

[0163] e= ;

[0164] If (e>10%), fine-tune η or Until the error is controllable.

[0165] As described in S511-S515 above, constraint and correlation verification ensures that parameters conform to physiological laws, laying a high-quality foundation for the extrapolation. The integration interval and initial conditions are clearly defined, and biomass is estimated using point cloud data to reduce errors caused by initial state ambiguity. The RK4 method is employed, offering higher accuracy than lower-order methods. Step size settings and coupling processing ensure solution stability and adaptability to dynamic source-sink coupling systems. The generated trajectories and features provide quantitative basis for analyzing growth rhythms and prioritizing source-sink allocation, with information value far exceeding that of single endpoint prediction. Multi-dimensional verification corrects biases, improving the consistency between the extrapolation results and the actual growth state extrapolation results.

[0166] In one embodiment, step S6, generating silique development prediction results and determining aborted siliques, specifically includes:

[0167] S61: Determine the sterility threshold :

[0168] It is the critical value of biomass growth rate for distinguishing whether a silique has failed. Physically, it is the minimum growth rate at which a silique can maintain normal development. When the growth rate is consistently below this value, the silique enters the process of failure due to insufficient source supply or metabolic abnormalities (such as the cessation of seed development).

[0169] Basis for value selection: Based on crop physiological characteristics and measured data, the baseline value for cruciferous crops such as rapeseed is set at 0.0005 g / h (the growth rate of aborted siliques is usually less than or equal to this value, or even negative); different varieties or environments (such as low light) can be adjusted through calibration experiments (e.g., adjusted to 0.0003 g / h under stress conditions). Specifically, the average growth rate of 30 artificially marked aborted siliques during the critical period is measured, and the upper limit of the 95% confidence interval is taken.

[0170] Threshold storage: Associated crop varieties, growth stages and environmental tags are stored in the parameter library (e.g., rapeseed - 10 days after flowering - normal light corresponds to 0.0005g / h) for easy retrieval and matching.

[0171] S62: Define the extrapolation time period [Tm, T{m+1}] for determining abortion. This time period must cover the abortion-sensitive period of silique development to ensure the timeliness and accuracy of the determination.

[0172] Starting point Tm: 3-5 days after silique formation (rapid accumulation period; if growth stagnates, the probability of abortion is high), or automatically triggered based on the young fruit stage tag in the semantic growth map;

[0173] Endpoint T{m+1}: Tm+3-5 days (e.g., before rapeseed pods are fully formed, sufficient to capture the characteristic of insufficient continuous growth);

[0174] Resolution: Consistent with the granularity of biomass growth rate calculation (e.g., once every 12 hours), ensuring at least 6 data points (to meet the continuous judgment sample size).

[0175] S63: Extracted biomass growth rate sequence of the target silique:

[0176] Growth rate data of siliques were selected from the projection results to construct a rate sequence over a time period:

[0177] Silique screening: Organ IDs of the silique library were extracted based on semantic growth maps, and mature / withered individuals were excluded;

[0178] Rate calculation: for time points The central difference method is used for calculation:

[0179] = ;

[0180] (in, (to ensure a smooth and noise-free experience).

[0181] Anomaly handling: Extreme values ​​(more than 10 times the normal range) are replaced with the average of the two points before and after to avoid interference with the judgment.

[0182] S64: Determining aborted siliques (continuously meeting the growth rate) The minimum growth rate for siliques to maintain normal development ):

[0183] Identification of aborted siliques using continuous threshold testing:

[0184] Continuous satisfaction of the definition: The percentage of points whose growth rate meets the threshold within a time period is ≥90% (one short-term fluctuation point is allowed, excluding instantaneous environmental interference);

[0185] Angle-by-angle test: Count the number of points that satisfy the threshold. Calculate the percentage:

[0186] r= If r ≥ 90%, it is marked as suspected sterility;

[0187] Secondary verification: For suspected aborted pods, if If the total biomass is ≤0 (no increase or decrease), it is ultimately determined to be sterile, thus avoiding misjudgment of weak growth.

[0188] S65: Generate a list of aborted siliques and additional information:

[0189] The results of the judgment are organized into a structured list, containing:

[0190] Core fields: Silique ID, spatial coordinates, determination time period, average growth rate, first abortion time point;

[0191] Visualization: In the 3D model, red highlights indicate the difference between the average rate and the threshold.

[0192] Data format: Stored in JSON format, supporting integration with agricultural management systems (such as precision fruit thinning decisions).

[0193] S66: Validation and threshold optimization of sterility determination results:

[0194] The threshold was verified and optimized through actual testing.

[0195] Field verification: Collect predicted aborted and normal pods, and statistically analyze the recall rate (≥85%) and precision rate (≥80%).

[0196] Threshold adjustment: If the recall rate is too low (<80%), lower the threshold by 5%-10%; if the precision rate is too low (<75%), raise the threshold by 5%-10% until the target is met.

[0197] As described in S61-S66 above, dynamic adjustment is made by combining physiological characteristics and measured data. It avoids the limitations of fixed thresholds and provides clear judgment criteria; it focuses on setting windows during the abortion sensitive period to balance the timeliness and accuracy of judgment and avoid judging too early or too late; outlier correction, 90% percentage standard and secondary verification effectively distinguish abortion from weak growth, reducing the misjudgment rate by more than 20%; the list contains spatial and temporal details and can be directly used for three-dimensional labeling and agricultural decision-making, enhancing its practical value.

[0198] In one embodiment, in step S5, the energy flow distribution coefficient is calculated. numerical value

[0199] S516: Clear The physical meaning and core function of:

[0200] Energy flow distribution coefficient It is a key parameter for quantifying the proportion of source organ assimilation products allocated to the i-th bank organ (0 < 1). <1), and all bank organs The sum is 1). Its core function is to coordinate the allocation of total source intensity according to sink attraction and transport efficiency in the source-sink dynamics system—reflecting both the sink organ's own competitive ability (sink attraction). This also reflects the convenience of transportation between the source and the warehouse (path transmission efficiency). This allows the allocation pattern to align with the actual material transport in plants (e.g., sink organs that are close to the source and have active metabolism receive more products).

[0201] S517: Acquiring the library attraction of library organs :

[0202] The parameters for organ competition capacity (estimated through prior spatiotemporal graph neural networks or experimentally determined) must satisfy the following:

[0203] Physical properties: >0 (non-negative), positively correlated with sink organ volume and relative growth rate (e.g., young siliques). (2-3 times the size of mature siliques)

[0204] Exception handling: If =0 (e.g., organs from a completely aged organ bank), directly label them. =0 (not involved in allocation).

[0205] S518: Calculate path propagation efficiency :

[0206] The smoothness of transport from source to depot organ i is represented by the product of the efficiencies of all channel nodes on the connecting path in the semantic growth graph:

[0207] Channel node efficiency Reflects the transport capacity of the stem segment, based on the stem diameter. ,length And the species-specific coefficient k is quantified (the thicker and shorter the stem, the better). The higher the value, the higher the value (0, 1).

[0208] Path extraction: Select the shortest path from source to channel to library (fewest channel nodes); when there are multiple paths, select the path with the highest overall efficiency.

[0209] calculate: = (All on the path) The product), if the path contains ( =0) (if the stem breaks), then =0.

[0210] S519: Calculation The allocation coefficient; this coefficient is based on the library attractiveness of the library organs. and the path propagation efficiency between it and the source node To be determined jointly;

[0211] Calculated based on the core formula: = ;

[0212] Numerator: (Comprehensive allocation weight of organ i in the library);

[0213] Denominator: All library organs The sum (total comprehensive allocation weight);

[0214] Normalization: If Deviation 1 (±0.01) = Perform normalization corrections to ensure that the sum of the energy flow allocation coefficients of all reservoir organs is 1, i.e. (Note: Here) (This represents the final energy flow distribution coefficient after normalization).

[0215] S520: Verification Physiological rationality:

[0216] If the rules are not met, the process is backtracked and corrected.

[0217] Range verification: ∈ and ;

[0218] Correlation verification: and Positive correlation (correlation coefficient ≥ 0.6), negative correlation with source-sink distance (correlation coefficient ≤ -0.5);

[0219] Growth center validation: growth center sink organs (such as young siliques) It is at least twice that of other libraries.

[0220] S521: Storage And it updates dynamically:

[0221] Data storage: Information such as organ ID, location, and growth stage is associated with the organ and stored in a table format (including organ ID, location, growth stage, etc.). , , (and other fields);

[0222] Dynamic update: Recalculated every 24 hours (to match the growth rhythm) (because , (As the plant grows, its condition changes), ensuring that the plant's status is reflected in real time.

[0223] As described in S516-S521 above, the attractiveness of the fusion library With path transmission efficiency This overcomes the limitations of traditional reliance solely on warehouse attraction and better aligns with the actual laws governing material transportation. This can be obtained through models or experiments. The calculation is based on the physical parameters of the stem, the method for obtaining the parameters is clear, and it is easy to repeat the implementation; periodic updates are possible. The response to organ growth and stem development ensures that the allocation coefficient is synchronized with the plant status; multi-dimensional verification ensures... It conforms to physiological laws and provides high-quality input for the source-liquidity system.

[0224] In one embodiment, in step S3, the node is assigned a resource competition index characteristic. Specifically, it includes:

[0225] S34: Clear The physical meaning and core function of:

[0226] Resource Competition Index It is a characteristic parameter that quantifies the intensity of light resource competition among plant nodes (especially source nodes such as leaves). Its physical meaning is the cumulative value of photosynthetically active radiation (PAR, 400-700nm) received by the node each day.

[0227] Its core function is to indirectly characterize the photosynthetic potential of a node by reflecting the shading effect of the canopy of adjacent plants: The higher the value, the more abundant the light resources, and the higher the source intensity (such as net photosynthetic rate) may be; conversely, Too low a shade level (strong shading) will limit the production of photosynthetic products, thereby affecting the allocation of source and sink materials.

[0228] S35: Obtain canopy point cloud data of the target plant and adjacent plants:

[0229] To accurately analyze the light competition environment, a complete three-dimensional point cloud of the canopy, including the target plant and surrounding adjacent plants, was collected:

[0230] Collection area: All plants within a 5m radius of the target plant (covering any adjacent plants that may be obscured).

[0231] Point cloud attributes: include three-dimensional spatial coordinates and reflectivity information (to distinguish between leaves and non-photosynthetic organs);

[0232] Data acquisition equipment: Ground-based lidar or multi-view stereo vision system, point cloud density ≥100 points / cm² (to capture fine leaf structures).

[0233] Data collection time: Clear, cloudless morning (9:00-11:00) to reduce noise caused by leaf movement.

[0234] S36: Preprocessing and semantic segmentation of canopy point clouds:

[0235] The collected point cloud data is processed to extract the effective canopy structure and distinguish organ types:

[0236] Noise reduction: Outliers were removed using the 3σ criterion, and points with a distance of ≤2mm from the plant surface were retained;

[0237] Point cloud registration: When acquiring data from multiple perspectives, the coordinate system is unified using the ICP algorithm (with the base of the target plant as the origin), and the registration error is ≤ ±1mm;

[0238] Canopy segmentation: Canopy point cloud is extracted based on height threshold, and then leaves (photosynthetic organs) and stems / flowers and fruits (non-photosynthetic organs) are segmented by reflectance and morphological features, with a segmentation accuracy of ≥95%.

[0239] S37: Time grid for simulating solar trajectory and radiation calculations:

[0240] Based on geographical location and time parameters, the time nodes for radiation calculation are determined as follows:

[0241] Input parameters: latitude and longitude of the target plot, altitude (used to calculate solar declination angle, hour angle, etc.);

[0242] Trajectory simulation: The solar altitude angle α and azimuth angle γ from 0:00 to 24:00 daily are calculated using the SPA model, with a time resolution of 10 minutes (144 nodes per day).

[0243] Time period filtering: Only retain the effective radiation period when the solar altitude angle α is greater than 5° (ignoring atmospheric scattering error).

[0244] S38: Calculate direct radiation flux :

[0245] Direct radiation is the radiant flux of parallel sunlight reaching the node directly; the effects of shading must be considered.

[0246] Unobstructed intensity: calculated based on solar constant (1361 W / m²), atmospheric transmittance (0.7 in clear weather), and solar altitude angle;

[0247] Shading detection: If a ray emitted along the direction of solar incidence intersects with the dot cloud of adjacent plant leaves (distance ≤ 0.5 mm), it is considered shading. =0; when there is no obstruction, it is converted into photosynthetically active radiation (PAR accounts for 45%).

[0248] S39: Calculate the scattered radiation flux :

[0249] Scattered radiation is diffuse radiation that has been scattered by the atmosphere or canopy and is not affected by direct sunlight.

[0250] Basic atmospheric scattering values: Atmospheric scattering radiation without a canopy (PAR accounts for 30%) calculated using an empirical model.

[0251] Canopy scattering correction: Corrected by combining leaf average reflectance (0.25) and canopy porosity (obtained by estimating leaf area index from point cloud);

[0252] Final flux: The corrected value is (Converted to PAR).

[0253] S310: Cumulative Computing Resource Competition Index :

[0254] Summing the direct and scattered radiation within the effective time period yields the node's value. :

[0255] Radiation accumulation: Traverse all valid time points and calculate the resource competition index. ,set up For direct radiation flux, Let be the diffuse radiation flux, and PAR be the photosynthetically active radiation flux, then:

[0256] = · ;

[0257] ( =10) minutes, units converted to mol / m²·d);

[0258] Node association: Based on the spatial coordinates of the nodes, Assign values ​​to the corresponding nodes to ensure that each node (especially the source node) is associated with a unique index.

[0259] S311: Validation and anomaly correction:

[0260] Accuracy was verified through actual testing and logical checks, and outliers were corrected.

[0261] Actual measurement comparison: Continuous recording was performed using a photosynthetically active radiation sensor, and the relative error between the simulated and measured values ​​was ≤15%.

[0262] Spatial consistency: adjacent nodes Difference ≤10%, bottom node of the canopy ≤1 / 3 of the top of the canopy;

[0263] Anomaly correction: When the error exceeds the standard, recalculate the solar trajectory or adjust the atmospheric transmittance (e.g., 0.5 on cloudy days) until the error is within the standard.

[0264] As described in S34-S311 above, the abstract resource competition is transformed into a calculable cumulative radiation value, reflecting the canopy shading effect in the population environment and providing a physical basis for source intensity estimation; based on the SPA model and radiative transfer theory, the measured error is ≤15%, avoiding the subjectivity of empirical indices; point cloud acquisition, preprocessing, and parameter selection all have clear standards, and those skilled in the art can adjust the parameters according to crop type and environment, which is convenient for reproduction; environmental resource competition information is integrated into the semantic growth map, providing key input for the spatiotemporal graph neural network to learn the environment-growth relationship, and improving the model's adaptability to complex field environments.

[0265] Example 2, please refer to Figure 2 As shown in this embodiment, the rapeseed full growth cycle monitoring system based on three-dimensional point cloud reconstruction includes:

[0266] The data acquisition module is used to acquire multi-view three-dimensional point cloud data of rapeseed plants;

[0267] The graph construction module is used to process point cloud data and construct the semantic growth graph.

[0268] The dynamic simulation engine has the aforementioned time-series prediction model built-in, which is used to perform growth state simulation results based on the principles of energy allocation and dynamic balance.

[0269] A monitoring output module is used to output the growth monitoring data;

[0270] The breeding decision support module is configured to rank different rapeseed lines by yield potential based on the growth monitoring data.

[0271] (1) Data acquisition module: continuously acquire multi-view three-dimensional point cloud data and environmental parameters of rapeseed plants throughout the entire growth cycle, providing raw data support for subsequent analysis.

[0272] Specific implementation:

[0273] It employs ground-based lidar, multispectral cameras, and environmental sensors.

[0274] Collection scope: Cover all target lines in the experimental field (e.g., 100 rapeseed lines, with 3 plants selected for each line for replication), with a single plant collection radius of ≥1.5m to ensure complete capture of the plant canopy and the part above the root system.

[0275] Collection cycle: dynamically adjusted according to the growth period: seedling stage (1-30 days after sowing) once every 7 days; budding stage (31-50 days) once every 5 days; flowering stage (51-70 days) once every 3 days; pod stage (71-120 days) once every 4 days until maturity, ensuring coverage of key growth stages.

[0276] Data preprocessing: After acquisition, point cloud denoising (statistical filtering to remove outliers), multi-view registration (ICP algorithm to unify coordinate system, registration error ≤ ±1mm) and format conversion (outputting in .ply format, retaining 3D coordinates and reflectivity information) are automatically performed.

[0277] (2) Map construction module: Semantic parsing is performed on the collected point cloud data to construct a semantic growth map, which transforms the three-dimensional morphological data into a structured plant growth feature map.

[0278] Specific implementation:

[0279] Point cloud semantic segmentation: An improved PointNet++ deep learning model is used to perform organ-level segmentation on single plant point clouds, and to identify and extract point cloud subsets of source organs (leaves), sink organs (growing points, flowers, siliques), and channel organs (stems).

[0280] Node and edge definitions:

[0281] Nodes: Each organ is an independent node (e.g., leaf L1, silique K3), and associated features include morphological parameters (leaf: area, perimeter, angle of inclination; silique: volume, length, number; stem: diameter, height) and physiological parameters (leaf chlorophyll content, inverted from multispectral reflectance; silique growth stage labels, such as young fruit, mature fruit).

[0282] Edge: Characterizes the physical connection between organs (such as leaf L1-stem segment S2, stem segment S2-silique K3), and the edge features are connection distance (mm) and transport resistance coefficient (calculated based on stem diameter).

[0283] Temporal map generation: Construct a temporal semantic growth map sequence {G1, ..., Gn} according to the collection time sequence (e.g., day 1, day 3, ...). Each map is associated with the environmental parameters (light, temperature) at the corresponding time, forming a dynamic data structure that integrates morphology, physiology and environment.

[0284] (3) Dynamic inference engine: Built-in hybrid time-series prediction model based on spatiotemporal graph neural network and source-sink dynamic system, infers the growth status of rapeseed throughout the entire growth cycle (such as organ biomass and silique development dynamics) based on semantic growth graph.

[0285] Specific implementation:

[0286] Model call: Load the pre-trained hybrid architecture model (including spatiotemporal graph neural network and source-library dynamics system module), and input the current temporal semantic growth graph sequence.

[0287] Parameter estimation and extrapolation:

[0288] The spatiotemporal graph neural network module learns spatiotemporal features from the graph and outputs key physiological parameters (library attraction). Total source intensity Energy flow distribution coefficient ).

[0289] Based on the above parameters, the source-Cootta dynamics system module solves the differential equations using the fourth-order Runge-Kutta method (e.g., =η· · To predict changes in biomass of various organs, silique development trajectory, and risk of abortion over the next 15 days.

[0290] Dynamic updates: Whenever new point cloud data is acquired (such as after the 3rd day of collection), the semantic growth map is automatically updated, historical predictions are re-analyzed and corrected to ensure that the simulation results are synchronized with the actual growth.

[0291] (4) Monitoring output module: integrates growth projection results with measured data, and outputs full-cycle monitoring information in multiple forms to meet visualization and data analysis needs.

[0292] Specific implementation:

[0293] Real-time visualization: Displays plant morphological dynamics (such as plant height growth and canopy expansion), organ biomass distribution (silique biomass is marked with color gradients), and the location of aborted siliques (highlighted in red) in a 3D interactive interface, and supports sliding the timeline to view historical status.

[0294] Growth Indicator Reports: Automatically generates time-series curves for key indicators, including: total biomass per plant, leaf area index (LAI), number of siliques / total biomass, and predicted average number of siliques. Data can be exported to Excel format.

[0295] Anomaly warning: When the simulation results show that a certain strain has an anomaly (such as silique abortion rate >30% or biomass growth rate 20% lower than that of similar strains), an alarm is triggered and possible causes are marked (such as insufficient light resource competition or low source intensity).

[0296] (5) Breeding decision support module: Based on the monitoring data of the whole growth cycle, the yield potential of different rapeseed lines is quantitatively evaluated and ranked to assist breeders in screening superior lines.

[0297] Specific implementation:

[0298] Key indicator extraction: Features strongly correlated with yield were extracted from the monitoring data, including: number of effective siliques per plant at maturity (excluding aborted siliques), average silique biomass, thousand-seed weight (estimated based on silique volume and density), and harvest index (economic yield / biomass).

[0299] Yield potential scoring model: The weighted summation method is used to calculate the overall score of the lines.

[0300] ;

[0301] Where the weight ( - Determined based on the correlation between historical breeding data and actual yield (e.g.) =0.35, =0.25, =0.2, =0.2).

[0302] Sorting and Recommendation: All lines are sorted in descending order according to their comprehensive scores, and a list of the top 20% of superior lines is output, along with radar charts of various indicators (such as a line with an outstanding number of effective siliques but a low thousand-seed weight), providing a basis for targeted selection in hybridization breeding.

[0303] As mentioned above, by using adaptive acquisition cycles and high-precision point cloud technology, continuous monitoring from the seedling stage to maturity is achieved, overcoming the limitations of traditional destructive sampling (such as periodic harvesting) in capturing growth dynamics and comprehensively recording the entire rapeseed growth process. Three-dimensional morphology, physiological parameters, and environmental factors are integrated into a structured map, overcoming the problem that simple point cloud data is sufficient for morphological description but lacks sufficient physiological significance, providing rich feature support for growth projection. The dynamic projection engine combines the data analysis capabilities of spatiotemporal graph neural networks with the physiological mechanism constraints of the source-sink dynamics system, achieving a silique biomass prediction error of ≤10% and an aborted silique identification accuracy of ≥85%, providing a scientific basis for yield assessment. Through key indicator extraction and scoring models, subjective breeding experience is transformed into objective data ranking, increasing the efficiency of superior line selection by more than 40% and reducing breeding cycle and labor costs.

[0304] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring the entire growth cycle of rapeseed based on 3D point cloud reconstruction, characterized in that, include: At least three key growth stages of rapeseed, obtain multi-view three-dimensional point cloud data of the same plant or plant population. The three-dimensional point cloud data is preprocessed to obtain clean point cloud data after denoising and registration; A semantic growth graph is constructed based on the pure point cloud data, wherein the nodes in the semantic growth graph are assigned type attributes, and the type attributes include at least source nodes representing energy supply sites, library nodes representing energy consumption sites, and channel nodes representing energy transmission channels. The steps for constructing the semantic growth graph include: Extract the topological skeleton of the plant from the pure point cloud data; Semantic classification of skeleton nodes: nodes corresponding to leaf regions are classified as source nodes, nodes corresponding to growth points, flowers, and siliques are classified as library nodes, and nodes corresponding to stems are classified as channel nodes. Establish the mapping relationship between the skeleton nodes and the original 3D point cloud; The constructed semantic growth graph is input into a pre-trained temporal prediction model; Using the aforementioned time-series prediction model, based on the principles of energy allocation and dynamic balance, the energy flow between the source node, the sink node, and the channel node is dynamically simulated to predict the growth state of the plant from the current moment to the future target moment. Based on the growth status projection results, growth monitoring data is output, including silique development prediction results, organ biomass distribution, and overall plant yield prediction. The growth monitoring data also includes rapeseed line yield potential assessment indicators.

2. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 1, characterized in that: The time series prediction model is a differential equation-driven model based on physical mechanisms, and its core is a source-sink dynamic system. The operating mechanism of the source-sink dynamics system is as follows: the plant structure represented by the semantic growth map is regarded as a dynamic flow network. The source node generates energy, and the sink node consumes energy to grow. The energy is dynamically distributed throughout the network according to the sink attraction of each sink node and the path conduction efficiency between it and the source node.

3. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 2, characterized in that: The time-series prediction model adopts a hybrid architecture that combines a spatiotemporal graph neural network with the source-liquid dynamics system; The spatiotemporal graph neural network takes a sequential semantic growth graph as input and its output is used to estimate the key physiological parameters required for the operation of the source-sink dynamic system.

4. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 3, characterized in that: The specific steps for predicting the plant growth status are as follows: By using key physiological parameters estimated by a spatiotemporal graph neural network, the source-sink dynamic system is driven to simulate the dynamic biomass accumulation trajectory of each sink organ from the current moment to the future target moment through numerical integration.

5. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 4, characterized in that: The silique development prediction results include a list of siliques identified as aborted. The criteria for determining aborted siliques are: during the simulation process, their biomass growth rate is consistently lower than the preset abortion threshold for a continuous period of time.

6. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 2, characterized in that: The energy is dynamically allocated across the network based on the pool attraction of each pool node and the path conduction efficiency between it and the source node. Energy flow allocation is achieved through the following methods: On the semantic growth graph, calculate the path propagation efficiency from each source node to each library node; The proportion of energy flow allocated to a reservoir node is directly proportional to its own reservoir attraction, directly proportional to the path conduction efficiency from the source node to the reservoir node, and inversely proportional to the total competitive attraction of all reservoir nodes. The competitive attraction is the sum of the reservoir attraction parameters of all reservoir nodes.

7. The method for monitoring the entire growth cycle of rapeseed based on three-dimensional point cloud reconstruction according to claim 1, characterized in that: In the step of constructing the semantic growth graph, nodes are also assigned a resource competition index feature; the resource competition index is obtained by analyzing the canopy point cloud of adjacent plants, simulating the solar trajectory and calculating the daily cumulative value of photosynthetically active radiation.

8. A rapeseed full-life cycle monitoring system based on three-dimensional point cloud reconstruction, used to implement the rapeseed full-life cycle monitoring method based on three-dimensional point cloud reconstruction as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire multi-view three-dimensional point cloud data of rapeseed plants; The graph construction module is used to process point cloud data and construct the semantic growth graph. The dynamic simulation engine has the aforementioned time-series prediction model built-in, which is used to perform growth state simulation results based on the principles of energy allocation and dynamic balance. A monitoring output module is used to output the growth monitoring data; The breeding decision support module is configured to rank different rapeseed lines by yield potential based on the growth monitoring data.

Images