Rice early growth activity prediction method, device, equipment and medium

Abstract

The invention relates to a rice early-stage growth activity prediction method, device, equipment and medium, and relates to the field of data processing, and the method comprises the steps: generating a canopy height model of single-plant rice according to laser radar point cloud data to extract the plant height of the single-plant rice, based on a structural feature extraction model, according to the laser radar point cloud data corresponding to the preset number of days after rice transplanting, determining the leaf age and the tiller number of the single-plant rice, and according to the plant height, the leaf age and the tiller number, constructing structural feature data of the single-plant rice; obtaining a first sample data set; and inputting the current spectral feature data and the current structural feature data of the single-plant rice of the target variety into an early growth activity prediction model trained by adopting the first sample data set to predict an early growth activity score of the single-plant rice of the target variety in a preset number of days after rice transplanting. According to the method, quantitative evaluation of the early-stage growth activity of the rice in a single plant scale in a field environment is realized, and a traditional mode depending on manual judgment is broken through.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method for predicting the early growth vigor of rice, a corresponding device, electronic equipment, and a computer-readable storage medium. Background Technology

[0002] As the staple food source for half the world's population, rice's early growth vigor directly impacts photosynthetic efficiency and nutrient accumulation capacity, making it a key trait determining yield. This is especially true in double-cropping rice systems, where varieties with strong early growth vigor can significantly shorten the growth cycle and achieve breakthroughs in annual yield. Early growth vigor is a complex trait, a comprehensive reflection of changes in phenotypic parameters such as greening, tillering, and biomass over time. Currently, due to methodological limitations, the evaluation of rice early growth vigor relies primarily on the manual judgment of experienced breeders and cultivation experts, lacking standardized evaluation systems and corresponding quantitative indicators. This process is time-consuming and labor-intensive, making large-scale evaluation in field settings difficult. This severely restricts the breeding of rice varieties with strong early growth vigor and the integration and matching of high-yield and efficient cultivation techniques.

[0003] In recent years, UAV remote sensing has become an effective tool for monitoring crop phenotypic traits. UAV platforms can carry multiple sensors simultaneously to invert or directly measure crop phenotypic traits in the field. Among these, the red-edge (RE) and near-infrared (NIR) bands provided by multispectral sensors have been proven to generate various spectral indices related to phenotypic parameters such as crop vigor, morphology, density, and chlorophyll content. However, optical orthophoto images only provide two-dimensional planar features of the crop canopy, limiting their ability to extract phenotypic information in three-dimensional space. Therefore, many researchers have used the Structure of Motion (SfM) algorithm combined with the Multi-View Stereo (MVS) algorithm to achieve three-dimensional reconstruction of the study area. However, optical remote sensing data is prone to saturation in areas with high biomass or leaf area index (LAI), and digital cameras cannot penetrate the vegetation canopy to obtain data from within the canopy. The reconstructed point cloud model has low density and the acquisition method is cumbersome and complex, making it unsuitable for large-scale extraction of fine-grained three-dimensional crop features.

[0004] Tillering is one of the core indicators for evaluating early growth vigor in rice, and its accurate monitoring has long faced technical bottlenecks. Currently, sensor-based tillering monitoring research can be mainly divided into three categories: direct measurement based on image analysis, indirect measurement based on crop spatial morphology, and inversion based on vegetation indices. Image-based methods have strict requirements on image quality, resulting in low data collection efficiency and hindering large-scale application in field environments. Measurement methods based on crop spatial morphology mainly use ground-based lidar or indoor 3D scanners to acquire high-precision raw point cloud data, but the feasibility of applying this method in field environments still needs experimental verification. While multispectral indices can indirectly reflect the tillering trend of the population, they cannot achieve precise analysis at the individual plant scale. This contradiction of "insufficient population inversion accuracy and low efficiency of individual plant measurement" has prevented tiller numbers from being included in large-scale phenotypic analysis systems.

[0005] Unmanned aerial vehicle (UAV) lidar (LiDAR) offers a new possibility for overcoming this bottleneck. LiDAR, as an active sensing technology, measures the distance from the sensor to the target by emitting short-wavelength lasers and recording the laser velocity and flight time using a timer. Because short-wavelength lasers have a certain penetrating ability into vegetation canopy, compensating for the shortcomings of traditional optical imaging, LiDAR is considered the most promising technology for quantifying vegetation structure parameters. LiDAR technology has already achieved successful applications in high-throughput crop phenotypic parameter extraction, such as plant height, leaf area index, and biomass. However, existing research mainly focuses on simple morphological parameters, and there is still a lack of effective point cloud deep learning algorithms for fine structural features such as tillers. Traditional rasterization methods result in the loss of three-dimensional features, while segmentation networks based on regular voxels are difficult to adapt to the topological heterogeneity of rice tillers. This has resulted in current LiDAR technology not yet achieving high-throughput analysis of tiller numbers.

[0006] The fusion of LiDAR point clouds and spectral images provides information combining the 3D structure of vegetation and its spectral properties. Currently, most fusion methods involve rasterizing the point cloud and aligning it spatially with the spectral data. While this approach initially solves the problem of fusing point cloud and spectral data, it results in a significant loss of feature information from the point cloud data.

[0007] In summary, while multispectral indices in existing technologies can indirectly reflect the tillering trend of a population, they cannot achieve precise analysis at the single-plant scale. Furthermore, traditional rasterization methods result in the loss of three-dimensional features, and segmentation networks based on regular voxels are difficult to adapt to the topological heterogeneity of rice tillers. This has led to problems such as the current LiDAR technology not yet achieving high-throughput analysis of tiller numbers. The applicant has made corresponding explorations to address these issues. Summary of the Invention

[0008] The purpose of this application is to solve the above-mentioned problems by providing a method, device, electronic equipment and computer-readable storage medium for predicting early growth vigor of rice.

[0009] To achieve the various objectives of this application, the following technical solution is adopted: A method for predicting early growth vigor of rice, proposed to meet one of the purposes of this application, includes: Acquire multispectral image data and lidar point cloud data of multiple individual rice plants of the target variety at a preset number of days after rice transplanting, and determine the spectral feature data of the individual rice plants based on the multispectral image data. The canopy height model of a single rice plant is generated based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to a convergent state, the leaf age and number of tillers of the single rice plant are determined according to the lidar point cloud data corresponding to the preset number of days after rice transplanting. The structural feature data of the single rice plant is constructed based on the plant height, the leaf age, and the number of tillers. Obtain a first sample dataset, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent spectral feature time series data constructed from spectral feature data collected from a single rice plant at a preset number of days after rice transplanting, and structural feature time series data constructed from structural feature data. The first sample labels represent the early growth vigor score of the rice. The current spectral and structural feature data of a single rice plant of the target variety are input into the early growth vigor prediction model trained using the first sample dataset to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, thus completing the prediction of the early growth vigor of rice.

[0010] Optionally, the spectral feature data includes normalized vegetation index, chlorophyll index, green normalized vegetation index, soil-adjusted vegetation index, normalized red edge difference index, and canopy coverage. The basic network architecture of the structural feature extraction model is a point cloud deep learning network, which includes an input layer, an encoder, and a decoder. The encoder includes two T-Net mini-networks and two multilayer perceptrons. The encoder includes a first T-Net mini-network, a first multilayer perceptron, a second T-Net mini-network, a second multilayer perceptron, and a max pooling layer connected in sequence. The decoder includes a third multilayer perceptron and a fourth multilayer perceptron connected in sequence. The basic network architecture of the early growth vigor prediction model is a Long Short-Term Memory (LSTM) network based on a time attention mechanism. The LSTM network based on the time attention mechanism includes an input layer, a dual-branch LSTM module, a feature fusion module, and a prediction layer connected in sequence. The dual-branch LSTM module includes a first LSTM branch module and a second LSTM branch module. The first LSTM branch module is used to process spectral feature time series data, and the second LSTM branch module is used to process structural feature time series data. The first LSTM branch module and the second LSTM branch module each include multiple LSTM processing units corresponding to multiple time steps. Each LSTM processing unit corresponding to a time step is connected to a time attention mechanism module.

[0011] Optionally, the steps for training the structural feature extraction model include: Obtain a second sample dataset, wherein the second sample dataset includes multiple second training samples and their corresponding second sample labels. The second training samples represent the lidar point cloud data of a single rice plant corresponding to a preset number of days after rice transplanting. The second sample labels represent the leaf age and tiller number of a single rice plant. The multiple second training samples and their corresponding second sample labels are input into a preset structural feature extraction model. A 3×3 affine transformation matrix is ​​generated in the first T-Net micro-network to perform three-dimensional spatial alignment on the lidar point cloud data. In the first multilayer perceptron, the three-dimensionally aligned lidar point cloud data is upsized from 3D to 64D to extract local features. In the second T-Net micro-network, a 64×64 affine transformation matrix is ​​generated to align the 64-dimensional local features in the feature space. In the second multilayer perceptron, the features after feature space alignment are increased to 1024 dimensions. In the max pooling layer, the 1024-dimensional features are aggregated to generate a global feature vector. After receiving the global feature vector, the decoder reduces the 1024-dimensional global feature vector to 2-dimensional in the third multilayer perceptron, and then fuses the 2-dimensional global feature vector with the preset number of days after rice transplanting in the fourth multilayer perceptron to perform a nonlinear transformation, so as to output the predicted leaf age and tiller number of the single rice plant. Repeat the above steps until the structural feature extraction model is trained to a convergent state, in order to determine the structural feature extraction model that has been trained to a convergent state.

[0012] Optionally, the step of inputting the current spectral feature data and current structural feature data of a single rice plant of the target variety into an early growth vigor prediction model trained using the first sample dataset to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting includes: The current spectral feature time series data and current structural feature time series data of a single rice plant of the target variety are input into the input layer of the early growth vigor prediction model; In the first LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the spectral feature time series data, and in the second LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the structural feature time series data. In the temporal attention mechanism module corresponding to each time step, the weights of the spectral feature data or structural feature data output by the corresponding LSTM processing unit are calculated to obtain the spectral feature weighted vector or structural feature weighted vector corresponding to the LSTM processing unit at each time step. The spectral feature data of all time steps are weighted and aggregated into a global weighted vector of spectral features after the weights are calculated by their respective time attention mechanism modules. The structural feature data of all time steps are weighted and aggregated into a global weighted vector of structural features after the weights are calculated by their respective time attention mechanism modules. In the feature fusion module, the global weighted vector of the spectral features and the global weighted vector of the structural features are concatenated and fused to obtain a global feature vector. In the prediction layer, the global feature vector is nonlinearly mapped by a fully connected layer and a Sigmoid activation function to output the early growth vitality score of a single rice plant of the target variety at a preset number of days after rice transplanting.

[0013] Optionally, the steps for determining an early growth viability score include: The goal is to obtain the number of days required for the target rice variety to reach the effective number of tillers, the maximum number of days required for all rice varieties to reach the effective number of tillers, the minimum number of days required for all rice varieties to reach the effective number of tillers, the number of days required for the target rice variety to reach the time node with the fastest leaf age increase, the maximum number of days required for all rice varieties to reach the time node with the fastest leaf age increase, and the minimum number of days required for all rice varieties to reach the time node with the fastest leaf age increase. Calculate a first difference between the number of days required for the target rice variety to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; calculate a second difference between the maximum number of days required for all rice varieties to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; and calculate a first ratio between the first difference and the second difference. Calculate the third difference between the number of days required for the target rice variety to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; calculate the fourth difference between the maximum number of days required for all rice varieties to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; and calculate the second ratio between the third difference and the fourth difference. A first sum between the first ratio and the second ratio is calculated, a third ratio between the first sum and the second value is calculated, and the early growth viability score is determined based on the difference between the first value and the third ratio.

[0014] Optionally, the step of generating a canopy height model of the single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant includes: Acquire lidar point cloud data of a single rice plant of the target variety at a preset number of days after rice transplanting; A digital surface model and a digital elevation model of a single rice plant are generated based on the lidar point cloud data. The canopy height model of the single rice plant is determined based on the difference between the digital surface model and the digital elevation model. The plant height of a single rice plant is extracted based on the canopy height model of the single rice plant.

[0015] Optionally, the step of determining the spectral characteristic data of the single rice plant based on the multispectral image data includes: Acquire multispectral image data of a single rice plant of the target variety at a preset number of days after rice transplanting. Preprocess the multispectral image data to determine the multispectral reflectance of the single rice plant at the preset number of days after rice transplanting. The multispectral reflectance includes green band reflectance, red band reflectance, red edge band reflectance, and near-infrared band reflectance. The spectral characteristic data of a single rice plant corresponding to a preset number of days after rice transplanting are determined based on the multispectral reflectance.

[0016] A rice early growth vigor prediction device provided for another purpose of this application includes: The spectral feature determination module is configured to acquire multispectral image data and lidar point cloud data of multiple individual rice plants of the target variety corresponding to a preset number of days after rice transplanting, and determine the spectral feature data of the individual rice plants based on the multispectral image data. The structural feature determination module is configured to generate a canopy height model of a single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to a convergent state, the module determines the leaf age and tiller number of the single rice plant based on the lidar point cloud data corresponding to a preset number of days after rice transplanting. The module then constructs the structural feature data of the single rice plant based on the plant height, leaf age, and tiller number. The dataset acquisition module is configured to acquire a first sample dataset, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent spectral feature time series data constructed from spectral feature data collected from a single rice plant after a preset number of days following rice transplanting, and structural feature time series data constructed from structural feature data. The first sample labels represent the early growth vigor score of the rice. The growth vigor prediction module is configured to input the current spectral feature data and current structural feature data of a single rice plant of the target variety into the early growth vigor prediction model trained using the first sample dataset, so as to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, thereby completing the prediction of the early growth vigor of rice.

[0017] An electronic device provided for another purpose of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the rice early growth vigor prediction method of this application.

[0018] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the method for predicting early growth vigor of rice, which, when invoked by a computer, performs the steps included in the corresponding method.

[0019] Compared to existing technologies, this application addresses the following issues: while multispectral indices can indirectly reflect the tillering trend of a population, they cannot achieve precise analysis at the single-plant scale; traditional rasterization methods result in the loss of three-dimensional features; and regular voxel-based segmentation networks are ill-suited to the topological heterogeneity of rice tillers. These issues prevent current LiDAR technology from achieving high-throughput analysis of tiller numbers. This application offers the following benefits, including but not limited to: Firstly, the rice early growth vigor prediction method proposed in this application addresses the difficulty in extracting fine structural features such as tiller number and leaf age. It designs an end-to-end point cloud deep learning network, using a T-Net micro-network to achieve spatial alignment and normalization of the point cloud, combined with a multilayer perceptron for feature dimensionality upscaling and downscaling, and incorporating time parameters. This overcomes the bottleneck in extracting tiller number and leaf age at the single-plant scale. The network achieves a determination coefficient (R²) of 0.91 for tiller number prediction and 0.97 for leaf age prediction, resolving the contradiction of insufficient population inversion accuracy and low efficiency in single-plant measurement in traditional methods, and achieving high-throughput and accurate analysis of key growth parameters. Simultaneously, it innovatively employs a two-level feature fusion structure LSTM network to integrate multimodal time series data, combining an attention mechanism to quantify the importance of data at different growth stages, overcoming the shortcomings of insufficient stability and interpretability in traditional static phenotypic analysis.

[0020] Secondly, the rice early growth vigor prediction method of this application, which integrates LiDAR point cloud deep learning features and multispectral data, achieves an R² of 0.79 and an RMSE of only 0.07 for early growth vigor prediction, significantly outperforming models using LiDAR alone (R²=0.72, RMSE=0.08) or multispectral data alone (R²=0.54, RMSE=0.11). The three-dimensional structural parameters provided by LiDAR compensate for the shortcomings of spectral indices in the three-dimensional representation of the canopy, while the physiological parameters of multispectral data enhance the model's understanding of biochemical processes. Furthermore, the two are dynamically complementary at different growth stages, with structural parameters contributing more significantly within 12 days after transplanting, and spectral indices contributing more prominently after canopy closure. Further optimization of the input parameter combination revealed that the highest prediction accuracy (R²=0.84) was achieved when "tiller number + leaf age + plant height" was fused with multispectral indices, fully validating the complementary effect of multimodal data.

[0021] Third, the method for predicting early growth vigor in rice, as described in this application, visualizes the process through a time attention mechanism, accurately pinpointing 12 to 20 days after transplanting as the critical period influencing early growth vigor prediction. During this stage, canopy phenotypic data contributes the most weight, shortening the window of opportunity for traditional empirical judgment. It was also found that leaf age has the highest predictive value among single parameters, while the dynamic synergy between structural parameters and spectral indices is the core influencing factor on early growth vigor. Furthermore, the relationship between early growth vigor and yield was clarified. High-yielding varieties generally possess high early growth vigor, but high-vigor varieties do not necessarily have high yields, confirming that early growth and rapid development are a necessary but not sufficient condition for high yield. Moreover, varieties exhibiting "high yield + high phenotypic" or "low yield + low phenotypic" characteristics account for over 75%, providing a new perspective for yield formation mechanism research.

[0022] Fourth, the method for predicting early growth vigor of rice in this application achieves quantitative evaluation of early growth vigor of rice at the single-plant scale in field conditions, breaking the traditional model that relies on manual judgment and providing a standardized technical means for breeding rice varieties with strong early growth vigor. By comparing the early growth vigor scores of different rice varieties, data support is provided for variety adaptability assessment. Simultaneously, the study clarifies the management priorities for varieties with different growth types, such as strengthening late-stage regulation for "high-scoring + low-yielding" varieties and optimizing early-stage management for "low-scoring + high-yielding" varieties, providing personalized solutions for precision agriculture management. Furthermore, the constructed technical system can be extended to the field of crop phenotyping, laying a solid foundation for the application of multimodal remote sensing technology in intelligent agriculture. Attached Figure Description

[0023] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the method for predicting early growth vigor of rice in the embodiments of this application; Figure 2 This is an exemplary network architecture for a point cloud deep learning network in the embodiments of this application; Figure 3 This is a scatter plot of the tiller number prediction of the point cloud deep learning network in the embodiments of this application; Figure 4 This is a scatter plot of leaf age prediction using a point cloud deep learning network in the embodiments of this application; Figure 5 This is a time-series variation diagram of leaf age prediction using a point cloud deep learning network in an embodiment of this application; Figure 6 This is a time-series variation diagram of the tiller number prediction of the point cloud deep learning network in the embodiments of this application; Figure 7 This is a schematic diagram illustrating the number of days required for Hemei Simiao to reach the effective number of tillers in the embodiments of this application; Figure 8 This is a flowchart illustrating the number of days required for the Hemei Silk seedling to reach the point where the leaf age increases most rapidly, as shown in the embodiments of this application. Figure 9 This is an exemplary network architecture for a long short-term memory network based on a temporal attention mechanism in the embodiments of this application; Figure 10 This is a scatter plot of the predicted and actual values ​​of the multispectral and lidar fusion model in the embodiments of this application; Figure 11 This is the error histogram predicted by the multispectral and lidar fusion model in the embodiments of this application; Figure 12 This is a schematic diagram of the rice early growth vigor prediction device in the embodiments of this application; Figure 13 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0025] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0026] Please see Figure 1 In one embodiment of the method for predicting early growth vigor of rice according to this application, the method includes: Step S10: Obtain multispectral image data and lidar point cloud data of multiple single rice plants of the target variety corresponding to a preset number of days after rice transplanting, and determine the spectral feature data of the single rice plant based on the multispectral image data. The rice early growth vigor prediction system in the terminal device can acquire multispectral image data and lidar point cloud data of multiple individual rice plants of a target variety corresponding to a preset number of days after rice transplanting. Based on the multispectral image data, the system determines the spectral characteristic data of each individual rice plant. The preset number of days can be 4 days, 8 days, 12 days, or 16 days, etc., and those skilled in the art can determine the preset number of days as needed according to business scenario requirements. The spectral characteristic data includes Normalized Difference Vegetation Index (NDVI), Chlorophyll Index (LCI), Green Normalized Difference Vegetation Index (GNDVI), Optimized Soil Adjusted Vegetation Index (OSAVI), Normalized Difference in Red Edge Index (NDRE), and Canopy Coverage. In some embodiments, the formula for calculating the Normalized Difference Vegetation Index (NDVI) is expressed as follows: , Wherein, NDVI represents the Normalized Difference Vegetation Index; NIR represents the Near Infrared Reflectance; and R represents the Red Reflectance.

[0027] The formula for calculating the chlorophyll index (LCI) is as follows: , Wherein, LCI represents chlorophyll index; NIR represents near-infrared reflectance; RE represents red-edge reflectance; and R represents red-band reflectance.

[0028] The formula for calculating the Normalized Green Vegetation Index (GNDVI) is as follows: , Wherein, GNDVI represents the normalized green vegetation index; NIR represents the near-infrared reflectance; and G represents the green reflectance.

[0029] The formula for calculating the Optimized Soil Adjusted Vegetation Index (OSAVI) is as follows: , Among them, OSAVI represents the optimized soil-adjusted vegetation index; NIR represents near-infrared reflectance; and R represents red reflectance.

[0030] The formula for calculating the Normalized Red Edge Difference Index (NDRE) is as follows: , Wherein, NDRE represents the normalized red-edge difference index; NIR represents the near-infrared band reflectance; and RE represents the red-edge band reflectance.

[0031] The formula for calculating canopy coverage is as follows: , When using UAV multispectral data to calculate spectral feature data to analyze vegetation growth, soil pixels (i.e., bare soil areas) and shadow pixels (such as shadow areas formed by objects) are present. These pixels are not vegetation themselves, and including them in vegetation index calculations will interfere with the results, producing confounding effects, meaning the vegetation index cannot accurately reflect the actual vegetation growth. Therefore, a Normalized Difference Vegetation Index (NDVI) threshold is set, where the NDVI threshold can be 0.45. Pixels with an NDVI less than 0.45 are marked as background pixels. This means that in the subsequent vegetation index calculation process, these background pixels (soil, weeds, shadows, and other non-canopy pixels) will be excluded from the calculation, allowing the vegetation index to more accurately reflect the actual growth status of the vegetation (rice), improving the accuracy of the analysis results. After removing soil, weeds, shadows, and other non-canopy pixels, the effective number of canopy pixels per rice plant is determined.

[0032] In some embodiments, strong early growth vigor refers to the characteristic of rice to grow vigorously during its vegetative growth period after transplanting. This characteristic increases the duration and efficiency of rice photosynthesis and is an important basis for rice variety breeding and high-yield, high-efficiency cultivation, especially in double-cropping rice areas. However, the current lack of precise quantitative descriptions of strong early growth vigor and efficient evaluation techniques in field settings restricts the efficiency of rice variety breeding and the development of high-yield cultivation techniques.

[0033] In some embodiments, the experimental area can be a double-cropping rice planting area in a certain region. To ensure the diversity of experimental data, 200 international rice varieties from around the world and 100 rice varieties from different years in a certain region were selected, totaling 300 rice varieties, and a replicate area was set up. Therefore, there are a total of 600 plots. Each plot is planted with 7 rows and 8 columns of 56 plants each, with a row spacing of 6 inches and a plant spacing of 5 inches. One seedling is planted per hole to facilitate tillering statistics later. Cultivation measures follow local traditional methods, applying urea, potassium chloride, and superphosphate fertilizers. The transplanting date is August 6, 2024. Multispectral and lidar data from UAVs are collected every 4 days for 40 days after transplanting, and the tillering and leaf age of 3 rice plants at fixed locations in the plot are manually counted, for a total of 10 data collection periods.

[0034] In some embodiments, the step of determining the spectral feature data of the single rice plant based on the multispectral image data includes: Step S1001: Obtain multispectral image data of a single rice plant of the target variety corresponding to a preset number of days after rice transplanting; preprocess the multispectral image data to determine the multispectral reflectance of the single rice plant corresponding to a preset number of days after rice transplanting; wherein the multispectral reflectance includes green band reflectance, red band reflectance, red edge band reflectance and near-infrared band reflectance. Step S1002: Calculate and determine the spectral characteristic data of a single rice plant corresponding to a preset number of days after rice transplanting based on the multispectral reflectance.

[0035] Specifically, the DJI Mavic 3M Multispectral Edition was used for multispectral image data acquisition. The drone integrates a camera with both a multispectral sensor and a visible light sensor, capable of simultaneously acquiring 5-megapixel multispectral image data including green (560±16nm), red (650±16nm), red-edge (730±16nm), and near-infrared (860±26nm) wavelengths, as well as 20-megapixel visible light images. Furthermore, the drone integrates a Real-Time Kinematic (RTK) differential positioning module, enabling spatial positioning within 1 cm horizontally and 1.5 cm vertically. To balance reconstruction success rate, ground resolution, and acquisition efficiency, the flight altitude was 12 meters during data acquisition, with a visible light ground resolution of 0.36 cm and a multispectral resolution of 0.62 cm. The forward overlap of the images was 80%, and the lateral overlap was 70%. All data was acquired between 11:00 AM and 2:00 PM in clear weather. After data acquisition, the images captured by the drone were stitched together into an orthophoto using the preset 3D modeling software (DJITerra). Radiometric correction was performed using three standard reflectance gray panels with reflectances of 25%, 50%, and 75% and an area of ​​20cm*20cm, and the grayscale values ​​(DN) were converted into reflectance.

[0036] In some embodiments, the spectral feature data includes normalized vegetation index, chlorophyll index, green normalized vegetation index, optimized soil-adjusted vegetation index, normalized red edge difference index, and canopy coverage; the basic network architecture of the structural feature extraction model is a point cloud deep learning network, wherein the point cloud deep learning network includes an input layer, an encoder, and a decoder, the encoder includes two T-Net micronetworks and two multilayer perceptrons, wherein the encoder includes a first T-Net micronetwork, a first multilayer perceptron, a second T-Net micronetwork, a second multilayer perceptron, and a max pooling layer connected in sequence; the decoder includes a third multilayer perceptron and a fourth multilayer perceptron connected in sequence; The basic network architecture of the early growth vigor prediction model is a Long Short-Term Memory (LSTM) network based on a time attention mechanism. The LSTM network based on the time attention mechanism includes an input layer, a dual-branch LSTM module, a feature fusion module, and a prediction layer connected in sequence. The dual-branch LSTM module includes a first LSTM branch module and a second LSTM branch module. The first LSTM branch module is used to process spectral feature time series data, and the second LSTM branch module is used to process structural feature time series data. The first LSTM branch module and the second LSTM branch module each include multiple LSTM processing units corresponding to multiple time steps. Each LSTM processing unit corresponding to a time step is connected to a time attention mechanism module.

[0037] Step S20: Generate a canopy height model of the single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to convergence state, determine the leaf age and number of tillers of the single rice plant based on the lidar point cloud data corresponding to the preset number of days after rice transplanting. Construct the structural feature data of the single rice plant based on the plant height, the leaf age, and the number of tillers. The process involves acquiring multispectral image data and lidar point cloud data of multiple individual rice plants of a target variety at preset numbers of days after transplanting. After determining the spectral feature data of the individual rice plants based on the multispectral image data, a canopy height model of the individual rice plants is generated based on the lidar point cloud data to extract the plant height. A structure feature extraction model, trained to convergence, is used to determine the leaf age and tiller number of the individual rice plants based on the lidar point cloud data at preset numbers of days after transplanting. The structure feature data of the individual rice plants is then constructed based on the plant height, leaf age, and tiller number. The basic network architecture of the structure feature extraction model is a point cloud deep learning network, which includes an input layer, an encoder, and a decoder. The encoder includes two T-Net micronetworks and two multilayer perceptrons, where the encoder comprises a first T-Net micronetwork, a first multilayer perceptron, a second T-Net micronetwork, a second multilayer perceptron, and a max pooling layer connected sequentially. The decoder includes a third multilayer perceptron and a fourth multilayer perceptron connected sequentially.

[0038] In some embodiments, the step of generating a canopy height model of a single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant includes: Step S201: Obtain the lidar point cloud data of a single rice plant of the target variety corresponding to a preset number of days after rice transplanting; Step S202: Generate a digital surface model and a digital elevation model of the single rice plant based on the lidar point cloud data, and determine the canopy height model of the single rice plant based on the difference between the digital surface model and the digital elevation model. Step S203: Extract the plant height of the single rice plant based on the canopy height model of the single rice plant.

[0039] Specifically, a DJI M350 RTK drone equipped with a ZENMUSE L2 LiDAR can be used to collect point cloud data. This system integrates a Livox LiDAR module, a high-precision inertial navigation system (IMU), a differential positioning module (RTK), a mapping camera, and a three-axis gimbal. During data acquisition, the flight speed is 0.7 m / s, the flight altitude is 20 m, and the point cloud density is 70452 points / m². 3D modeling software (DJI Terra) is used to read the LiDAR point cloud data, IMU records, track files, and other related data, performing POS calculation, point cloud coloring, and stitching operations to generate a LAS point cloud file. The LAS file data includes the 3D coordinates of each point, scan angle, echo number, return pulse intensity, and GPS timestamp. LiDAR360 software is used for preprocessing such as cropping, denoising, point cloud registration, and ground filtering to generate a Digital Surface Model (DSM), a Digital Elevation Model (DEM), and a Canopy Height Model (CHM).

[0040] After generating the digital surface model (DSM), digital elevation model (DEM), and canopy height model (CHM), the canopy height model of the single rice plant is determined based on the difference between the digital surface model and the digital elevation model, and the plant height of the single rice plant is extracted based on the canopy height model of the single rice plant.

[0041] In some embodiments, the step of training a structural feature extraction model includes: Step S2001: Obtain the second sample dataset, wherein the second sample dataset includes multiple second training samples and their corresponding second sample labels. The second training samples represent the lidar point cloud data of a single rice plant corresponding to a preset number of days after rice transplanting. The second sample labels represent the leaf age and tiller number of a single rice plant. Step S2002: Input the multiple second training samples and their corresponding second sample labels into a preset structural feature extraction model. Generate a 3×3 affine transformation matrix in the first T-Net micro-network to perform three-dimensional spatial alignment on the lidar point cloud data. In the first multilayer perceptron, upgrade the three-dimensionally aligned lidar point cloud data from 3D to 64D to extract local features. Step S2003: Generate a 64×64 affine transformation matrix in the second T-Net micro-network to perform feature space alignment on the 64-dimensional local features. In the second multilayer perceptron, increase the feature space alignment to 1024 dimensions. In the max pooling layer, aggregate the 1024-dimensional features to generate a global feature vector. Step S2004: After receiving the global feature vector, the decoder reduces the 1024-dimensional global feature vector to 2-dimensional in the third multilayer perceptron, and fuses the 2-dimensional global feature vector with the preset number of days after rice transplanting in the fourth multilayer perceptron to perform a nonlinear transformation, so as to output the predicted leaf age and tiller number of the single rice plant. Step S2005: Repeat steps S2001 to S2004 until the structural feature extraction model is trained to a convergent state, so as to determine the structural feature extraction model that has been trained to a convergent state.

[0042] Specifically, the lidar point cloud data of individual rice plants collected by drones at a preset number of days after rice transplanting can be used as the second training sample. The leaf age and tiller number of individual rice plants can be used as the second sample label. The second training sample and the second sample label are matched one-to-one to construct the second sample dataset. First, individual rice plants manually surveyed in each plot are found in the lidar point cloud data, and the center point of each individual rice plant is manually marked. Then, a circle with a radius of 20cm is drawn to extract the point cloud data within the circle. Finally, ground point filtering and outlier filtering are performed on the point cloud samples to obtain pure rice plant point cloud data. All point cloud samples are downsampled to 10,000 points for subsequent processing.

[0043] The structural feature extraction model of this application is based on an end-to-end point cloud deep learning network, including an input layer, an encoder, and a decoder. The LiDAR point cloud data of a single rice plant collected by a drone at a preset number of days after rice transplanting is used as the input to the structural feature extraction model, and the model outputs tillering prediction and leaf age prediction values. The core idea of ​​the structural feature extraction model of this application is to solve the permutation invariance of point clouds through symmetric function-max pooling. In the encoding part, the encoder includes a first T-Net micro-network, a first multilayer perceptron (MLP), a second T-Net micro-network, a second multilayer perceptron (MLP), and a max pooling layer connected in sequence. The main function of the T-Net micro-network is to align and normalize the point cloud data in three-dimensional (3D) space, like straightening a skewed object for subsequent processing. The multilayer perceptron (MLP) performs feature extraction by upscaling the data, increasing the information of each point from the original 3 dimensions (x, y, z) to 1024 dimensions to form a global feature vector.

[0044] The T-Net micro-network can generate an affine transformation matrix to handle spatial transformations such as rotation and translation of point clouds. It achieves this by mapping the point cloud to a k-dimensional space and learning a k×k rotation matrix. Since the rotation matrix is ​​orthogonal, a regularization penalty term needs to be introduced into the loss function to approximate an orthogonal matrix as closely as possible. The expression for this regularization penalty term is: , In the formula, The penalty term for regularization is represented; E is the identity matrix; A is the affine transformation matrix output by the T-Net micronetwork. By adding a Frobenius norm term to the loss function, the output affine transformation matrix A is made as close as possible to an orthogonal matrix. Let A represent the transpose of the affine transformation matrix A; This represents the Frobenius norm.

[0045] In the decoder section, the decoder includes a third multilayer perceptron and a fourth multilayer perceptron connected in sequence. The global features are reduced in dimensionality by the two multilayer perceptrons (MLPs). The third multilayer perceptron reduces the global features to 2 dimensions. The fourth multilayer perceptron has a time feature parameter input, which is the number of days after rice transplanting. This is a time parameter that is highly correlated with the tillering and leaf age of rice. It is fused with the extracted point cloud feature vector to jointly predict tillering and leaf age. Therefore, the fourth multilayer perceptron has 3-dimensional input and 2-dimensional output, which includes leaf age and tillering number.

[0046] In some embodiments, please refer to Figure 2 , Figure 2 This is an exemplary network architecture for a point cloud deep learning network, where Input Points represents the input LiDAR point cloud data; T-Net represents a T-Net micro-network; Transformationmatrix represents an affine transformation matrix; MLP represents a multilayer perceptron; Max pool represents a max pooling layer; Encode represents an encoder; Decoder represents a decoder; Tillers represent the number of tillers; and Leaf age represents the leaf age.

[0047] In some embodiments, please refer to Figure 3 , Figure 3 The horizontal axis represents the number of tillers predicted by the model, and the vertical axis represents the number of tillers measured in the field. The coefficient of determination (R²) between the predicted and measured tillers is 0.91, indicating a strong correlation between the two. The root mean square error (RMSE) is 1.60, reflecting the average error between the predicted and measured tillers, demonstrating that the point cloud deep learning network has good performance in tiller prediction.

[0048] Please see Figure 4 , Figure 4The horizontal axis represents the leaf ages predicted by the model, and the vertical axis represents the field-measured leaf ages. The coefficient of determination (R²) between the model-predicted leaf ages and the field-measured leaf ages reaches 0.97, indicating a strong correlation between the two. The root mean square error (RMSE) is 0.43, reflecting the average error between the model-predicted leaf ages and the field-measured leaf ages, demonstrating that the point cloud deep learning network has good performance in leaf age prediction.

[0049] Please see Figure 5 , Figure 5 The horizontal axis represents the date (such as 0806, 0810, etc.), and the vertical axis represents the leaf ages. Figure 5 It shows the trend of leaf age change predicted by the model over time. By combining the distribution interval, we can intuitively see the concentration and fluctuation of leaf age prediction values ​​at different periods, which helps to analyze the prediction stability of the model in the process of dynamic changes in leaf age.

[0050] Please see Figure 6 , Figure 6 The horizontal axis represents the date, and the vertical axis represents the number of tillers. Figure 6 The model shows the trend of tiller number changes over time. By observing the distribution interval, the dispersion of the predicted tiller number at different periods can be observed, thereby judging the reliability of the model's prediction in the process of tiller number change over time.

[0051] Step S30: Obtain the first sample dataset, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent the spectral feature time series data constructed from the spectral feature data collected by a single rice plant after a preset number of days following rice transplanting, and the structural feature time series data constructed from the structural feature data. The first sample label represents the early growth vitality score of the rice. The canopy height model of a single rice plant is generated based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to a convergent state, the leaf age and tiller number of the single rice plant are determined according to the lidar point cloud data corresponding to the preset number of days after rice transplanting. After constructing the structural feature data of the single rice plant based on the plant height, leaf age, and tiller number, a first sample dataset is obtained. The first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent the spectral feature time series data constructed from the spectral feature data collected by the single rice plant at the preset number of days after rice transplanting, and the structural feature time series data constructed from the structural feature data. The first sample labels represent the early growth vitality score of the rice. In some embodiments, the step of determining an early growth viability score includes: Step S301: Obtain the number of days required for the target rice variety to reach the effective number of tillers, the maximum number of days required for all rice varieties to reach the effective number of tillers, the minimum number of days required for all rice varieties to reach the effective number of tillers, the number of days required for the target rice variety to reach the time node with the fastest leaf age increase, the maximum number of days required for all rice varieties to reach the time node with the fastest leaf age increase, and the minimum number of days required for all rice varieties to reach the time node with the fastest leaf age increase. Step S302: Calculate the first difference between the number of days required for the target rice variety to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; calculate the second difference between the maximum number of days required for all rice varieties to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; and calculate the first ratio between the first difference and the second difference. Step S303: Calculate the third difference between the number of days required for the target rice variety to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; calculate the fourth difference between the maximum number of days required for all rice varieties to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; and calculate the second ratio between the third difference and the fourth difference. Step S304: Calculate and determine the first sum between the first ratio and the second ratio, calculate and determine the third ratio between the first sum and the second value, and determine the early growth vitality score based on the difference between the first value and the third ratio.

[0052] Specifically, based on the characteristics of "early greening and early tillering" in rice varieties with strong early growth vigor, this application selects two indicators to characterize the timing of greening and tillering in experimental rice varieties. It should be noted that rice tillering is divided into effective tillers and ineffective tillers. Ineffective tillers ultimately cannot develop into panicles normally. Therefore, this application investigates the number of days required to reach the number of effective tillers during manual labeling.

[0053] Please see Figure 7 as well as Figure 8 Taking the rice variety Hemei Simiao as an example, the maximum number of effective tillers at harvest is 7. Based on the results of artificial field surveys, the number of days required for the Hemei Simiao rice variety to reach the number of effective tillers, and the number of days required to reach the time point with the fastest leaf age increase, were calculated. Figure 7 The horizontal axis represents the number of days after transplanting, and the vertical axis represents the number of tillers. Figure 7 The image shows the change in the number of tillers of the rice variety Hemei Simiao after transplanting. The upward trend of the curve reflects the growth process of the number of tillers. The largest dot marks the number of days required for the rice variety to reach the effective number of tillers. By observing the curve, the time point when the number of tillers grows to the effective number can be seen intuitively.

[0054] Figure 8 The horizontal axis represents the number of days after transplanting, and the vertical axis represents the leaf age increase / 4 days. Figure 8 Different bars represent the rate of leaf age increase at different time points after transplanting. The taller the bar, the faster the leaf age increases within that time period. The time points corresponding to the highest bar values ​​represent the periods of fastest leaf age increase. Figure 8 It can clearly determine the number of days required to reach the point where the leaf age increases most rapidly.

[0055] Furthermore, the early growth activity scores of the target rice varieties were comprehensively calculated. Since the 300 varieties planted in this study basically cover a wide range of early growth activity from extremely weak to extremely strong, the early growth activity scores were mapped to a range of 0 to 1 using the maximum-minimum normalization method. Finally, the rice varieties... The formula for calculating the early growth viability score is expressed as follows: , In the formula, Indicates rice varieties Early growth viability score; This indicates the number of days required for the target rice variety to reach the effective number of tillers; This represents the minimum number of days required for all rice varieties to reach the effective number of tillers. This represents the maximum number of days required for all rice varieties to reach the effective number of tillers. Indicates rice varieties The number of days required to reach the point where leaf age increases most rapidly; This represents the maximum number of days required for all rice varieties to reach the point where their leaf age increases at the fastest rate. This represents the minimum number of days required for all rice varieties to reach the point where their leaf age increases at its fastest rate.

[0056] Step S40: Input the current spectral feature data and current structural feature data of a single rice plant of the target variety into the early growth vigor prediction model trained using the first sample dataset, so as to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, and complete the prediction of the early growth vigor of rice.

[0057] After obtaining the first sample dataset, the current spectral and structural feature data of individual rice plants of the target variety are input into the early growth vigor prediction model trained using the first sample dataset. This model predicts the early growth vigor score of individual rice plants of the target variety at a preset number of days after rice transplanting, thus completing the prediction of early growth vigor of rice. The basic network architecture of the early growth vigor prediction model is a Long Short-Term Memory (LSTM) network based on a time attention mechanism. This LSTM network includes an input layer, a dual-branch LSTM module, a feature fusion module, and a prediction layer connected sequentially. The dual-branch LSTM module includes a first LSTM branch module and a second LSTM branch module. The first LSTM branch module processes spectral feature time-series data, and the second LSTM branch module processes structural feature time-series data. Each of the first and second LSTM branch modules includes multiple LSTM processing units corresponding to multiple time steps, and each LSTM processing unit corresponding to a time step is connected to a time attention mechanism module.

[0058] Specifically, a Long Short-Term Memory (LSTM) network based on a time attention mechanism is used as the basic network architecture of the early growth vigor prediction model in this application. The LSTM network based on a time attention mechanism includes an input layer, a dual-branch LSTM module, a feature fusion module, and a prediction layer connected in sequence. The dual-branch LSTM module includes a first LSTM branch module (multispectral module) and a second LSTM branch module (LiDAR point cloud module). The first LSTM branch module is used to process spectral feature time series data, and the second LSTM branch module is used to process structural feature time series data. The feature parameters extracted from the spectral feature time series data and the structural feature time series data are input into the feature fusion module in the LSTM network based on a time attention mechanism. Finally, the feature parameters are connected in the feature fusion module to form a global feature vector. The loss functions of the first LSTM branch module (multispectral module) and the second LSTM branch module (LiDAR point cloud module) are used to train the early growth vigor prediction model through weighted gradient fusion. For each module... loss function In feature parameters The gradient below can be expressed as Then the gradient after fusing the first LSTM branch module (multispectral module) and the second LSTM branch module (LiDAR point cloud module) can be expressed as: , in, This indicates either the first LSTM branch module (multispectral module) or the second LSTM branch module (LiDAR point cloud module). Indicates the characteristic parameters Find the gradient; This represents the objective function of the early growth vigor prediction model. This represents the total number of modules participating in gradient fusion; Indicates the first The weight of each module; Indicates the first The objective function of each module; Indicates the first The objective function of each module has characteristics. Find the gradient, i.e., the gradient of the first step. Gradient information for each module.

[0059] Furthermore, the first LSTM branch module and the second LSTM branch module each include multiple LSTM processing units corresponding to multiple time steps. Each LSTM processing unit corresponding to a time step is connected to a time attention mechanism module. Each LSTM processing unit is followed by a time attention mechanism module, which can adaptively adjust the attention weight of each time step, so that the early growth vitality prediction model gives greater weight to key time steps. At the same time, the weights can be visualized to determine the weights relied upon when using the model for prediction.

[0060] In some embodiments, the structural feature extraction model and early growth viability prediction model of this application can be built using the PyTorch framework and trained on an NVIDIA RTX 4090 with 24G of video memory. Both the structural feature extraction model and the early growth viability prediction model in this application are implemented using PyTorch, with stochastic gradient descent (SGD) as the optimizer and L2 regularization to prevent overfitting. For the final model prediction evaluation, the coefficient of determination (COP) was calculated on the test set. ), root mean square error ( ) and relative root mean square error ( ), of which the coefficient of determination ( The formula for calculating ) is expressed as: , The formula for calculating the root mean square error (RMSE) is as follows: , Relative root mean square error ( The formula for calculating ) is expressed as: , in, The coefficient of determination measures the degree of fit between the model's predicted values ​​and the actual values. It ranges from 0 to 1, with values ​​closer to 1 indicating a better fit. The model represents the first Predicted values ​​for each sample; Represents actual value The average value; Indicates the first The actual value of each sample; Indicates the number of samples; Root mean square error (RMSE) measures the average error between the model's predicted value and the actual value. A smaller value indicates higher model prediction accuracy. Relative root mean square error is an indicator that relativizes the root mean square error. It is used to more intuitively reflect the proportion of the error relative to the average level of the actual value, and is presented in the form of a percentage.

[0061] The early growth vigor prediction model of this application is trained using the first sample dataset mentioned above. After the early growth vigor prediction model is trained to convergence, the early growth vigor score of a single rice plant of the target variety can be predicted based on the current spectral feature data and current structural feature data of the single rice plant of the target variety at a preset number of days after rice transplanting.

[0062] In a specific embodiment, the step of inputting the current spectral feature data and current structural feature data of a single rice plant of the target variety into an early growth vigor prediction model trained using the first sample dataset to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting includes: Step S401: Input the current spectral feature time series data and the current structural feature time series data of the target variety of single rice plant into the input layer of the early growth vigor prediction model; Step S402: In the first LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the spectral feature time series data, and in the second LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the structural feature time series data. Step S403: In the time attention mechanism module corresponding to each time step, calculate the weights of the spectral feature data or structural feature data output by the corresponding LSTM processing unit to obtain the spectral feature weighted vector or structural feature weighted vector corresponding to the LSTM processing unit at each time step. Step S404: After calculating the weights of the spectral feature data of all time steps through their respective time attention mechanism modules, the data are weighted and aggregated into a global weighted vector of spectral features. After calculating the weights of the structural feature data of all time steps through their respective time attention mechanism modules, the data are weighted and aggregated into a global weighted vector of structural features. Step S405: In the feature fusion module, the global weighted vector of the spectral features and the global weighted vector of the structural features are concatenated and fused to obtain a global feature vector. In the prediction layer, the global feature vector is nonlinearly mapped by a fully connected layer and a Sigmoid activation function to output the early growth vitality score of a single rice plant of the target variety after a preset number of days after rice transplanting.

[0063] In some embodiments, please refer to Figure 9 , Figure 9This is an exemplary network architecture for a Long Short-Term Memory (LSTM) network based on a temporal attention mechanism. Here, Input represents the input, LSTM represents the Long Short-Term Memory network, MS Module represents the multispectral module, Dense represents the densely connected layer, Softmax represents the Softmax activation function, Feature Vector represents the feature vector, Full Connect represents the fully connected layer, Sigmoid represents the Sigmoid activation function, Fusion Model represents the fusion model, LiDAR Module represents the LiDAR point cloud module, Attention Mechanism represents the temporal attention mechanism module, Early growth vigor score (MS) represents the early growth vigor score output by the multispectral module, Early growth vigor score (MS+LiDAR) represents the early growth vigor score output by the fusion model constructed from the multispectral module and the LiDAR point cloud module, and Early growth vigor score (LiDAR) represents the early growth vigor score output by the LiDAR point cloud module.

[0064] In some embodiments, please refer to Figure 10 , Figure 10 This is a scatter plot of predicted and true values ​​from a multispectral (MS) and LiDAR data fusion model, with the horizontal axis representing predicted values ​​and the vertical axis representing true values. The plot is labeled with... This indicates that the predicted values ​​of the fusion model are highly correlated with the actual values; the figure is marked with... as well as The small average error between the predicted and actual values ​​indicates that the fusion model combining multispectral (MS) and lidar (LiDAR) data performs well in predicting early growth vitality.

[0065] Please see Figure 11 , Figure 11 This is a histogram of the prediction error of the multispectral (MS) and lidar (LiDAR) data fusion model, with the horizontal axis representing error and the vertical axis representing count. The figure shows the distribution of sample numbers within different error intervals. The overall error follows an approximately normal distribution trend, with most errors concentrated between -0.2 and 0.2, indicating that the fusion model has high prediction stability.

[0066] In some embodiments, the relationship between early growth vigor and final yield in rice was investigated, showing a positive correlation, but this relationship was not absolutely linear. Varieties with higher early growth vigor were more likely to achieve higher yields; however, not all high-vigor varieties achieved high yields, while almost all high-yielding varieties possessed high early growth vigor. This result indicates that early and rapid growth is a necessary condition for high rice yields, but not a sufficient condition.

[0067] Combining the scatter plot and percentage distribution, it can be seen that most varieties are concentrated in two quadrants: "low yield + low biomass" (50.9%) and "high yield + high biomass" (24.9%), indicating that the growth vigor of most varieties is consistent with their yield trend. Varieties in the "high yield + high biomass" quadrant exhibit strong early growth ability and ultimately achieve higher yields per plant, which is consistent with the physiological laws of rice growth, namely, early growth advantage usually lays the foundation for high yields later. Conversely, "low yield + low biomass" varieties show weak growth vigor throughout their growth period, making it difficult to accumulate sufficient biomass, thus resulting in lower final yields.

[0068] It is worth noting that some varieties fell into the "high yield + low score" (2.3%) and "low yield + high score" (22.0%) quadrants, revealing the multidimensionality of rice yield formation. Among them, the "high yield + low score" varieties, despite having lower early growth vigor scores, still achieved higher yields. This may be due to their strong late-stage compensatory growth capabilities, such as higher photosynthetic efficiency, stronger grain filling ability, or better tillering and panicle formation rate. Conversely, the "low yield + high score" varieties exhibited strong early growth vigor but ultimately lower yields, possibly due to factors such as late-stage tiller degeneration, nutrient competition, limited panicle grain number, or environmental stress.

[0069] These results suggest that in rice breeding and precision management, focusing solely on early growth vigor is insufficient to fully predict yield; it is also necessary to consider later growth performance and physiological characteristics. For example, the existence of some "high-yielding + low-tillering" varieties indicates that despite lower early growth vigor, higher yields can still be achieved by compensating for these disadvantages through later management practices (such as appropriate fertilization and improved grain-filling efficiency). Therefore, in precision rice management strategies, differentiated management plans can be developed for varieties with different growth types. For example, for "high-tillering + low-yielding" varieties, strengthening later-stage tillering control and water and fertilizer management can improve final yield, while for "low-tillering + high-yielding" varieties, optimizing early growth control can promote better growth balance.

[0070] As can be seen from the above embodiments, compared with the prior art, this application addresses the problems in the prior art where multispectral indices can indirectly reflect the tillering trend of the population but cannot achieve accurate analysis at the single plant scale, and where traditional rasterization methods cause the loss of three-dimensional features, while segmentation networks based on regular voxels are difficult to adapt to the topological heterogeneity of rice tillers. This has led to the fact that current LiDAR technology has not yet achieved high-throughput analysis of tiller numbers. This application includes, but is not limited to, the following beneficial effects: Firstly, the rice early growth vigor prediction method proposed in this application addresses the difficulty in extracting fine structural features such as tiller number and leaf age. It designs an end-to-end point cloud deep learning network, using a T-Net micro-network to achieve spatial alignment and normalization of the point cloud, combined with a multilayer perceptron for feature dimensionality upscaling and downscaling, and incorporating time parameters. This overcomes the bottleneck in extracting tiller number and leaf age at the single-plant scale. The network achieves a determination coefficient (R²) of 0.91 for tiller number prediction and 0.97 for leaf age prediction, resolving the contradiction of insufficient population inversion accuracy and low efficiency in single-plant measurement in traditional methods, and achieving high-throughput and accurate analysis of key growth parameters. Simultaneously, it innovatively employs a two-level feature fusion structure LSTM network to integrate multimodal time series data, combining an attention mechanism to quantify the importance of data at different growth stages, overcoming the shortcomings of insufficient stability and interpretability in traditional static phenotypic analysis.

[0071] Secondly, the rice early growth vigor prediction method of this application, which integrates LiDAR point cloud deep learning features and multispectral data, achieves an R² of 0.79 and an RMSE of only 0.07 for early growth vigor prediction, significantly outperforming models using LiDAR alone (R²=0.72, RMSE=0.08) or multispectral data alone (R²=0.54, RMSE=0.11). The three-dimensional structural parameters provided by LiDAR compensate for the shortcomings of spectral indices in the three-dimensional representation of the canopy, while the physiological parameters of multispectral data enhance the model's understanding of biochemical processes. Furthermore, the two are dynamically complementary at different growth stages, with structural parameters contributing more significantly within 12 days after transplanting, and spectral indices contributing more prominently after canopy closure. Further optimization of the input parameter combination revealed that the highest prediction accuracy (R²=0.84) was achieved when "tiller number + leaf age + plant height" was fused with multispectral indices, fully validating the complementary effect of multimodal data.

[0072] Third, the method for predicting early growth vigor in rice, as described in this application, visualizes the process through a time attention mechanism, accurately pinpointing 12 to 20 days after transplanting as the critical period influencing early growth vigor prediction. During this stage, canopy phenotypic data contributes the most weight, shortening the window of opportunity for traditional empirical judgment. It was also found that leaf age has the highest predictive value among single parameters, while the dynamic synergy between structural parameters and spectral indices is the core influencing factor on early growth vigor. Furthermore, the relationship between early growth vigor and yield was clarified. High-yielding varieties generally possess high early growth vigor, but high-vigor varieties do not necessarily have high yields, confirming that early growth and rapid development are a necessary but not sufficient condition for high yield. Moreover, varieties exhibiting "high yield + high phenotypic" or "low yield + low phenotypic" characteristics account for over 75%, providing a new perspective for yield formation mechanism research.

[0073] Fourth, the method for predicting early growth vigor of rice in this application achieves quantitative evaluation of early growth vigor of rice at the single-plant scale in field conditions, breaking the traditional model that relies on manual judgment and providing a standardized technical means for breeding rice varieties with strong early growth vigor. By comparing the early growth vigor scores of different rice varieties, data support is provided for variety adaptability assessment. Simultaneously, the study clarifies the management priorities for varieties with different growth types, such as strengthening late-stage regulation for "high-scoring + low-yielding" varieties and optimizing early-stage management for "low-scoring + high-yielding" varieties, providing personalized solutions for precision agriculture management. Furthermore, the constructed technical system can be extended to the field of crop phenotyping, laying a solid foundation for the application of multimodal remote sensing technology in intelligent agriculture.

[0074] Please see Figure 12This application provides a device for predicting the early growth vigor of rice, comprising a spectral feature determination module 1100, a structural feature determination module 1200, a dataset acquisition module 1300, and a growth vigor prediction module 1400. The spectral feature determination module 1100 is configured to acquire multispectral image data and lidar point cloud data of multiple individual rice plants of a target variety corresponding to a preset number of days after rice transplanting, and determine the spectral feature data of the individual rice plants based on the multispectral image data. The structural feature determination module 1200 is configured to generate a canopy height model of the individual rice plants based on the lidar point cloud data to extract the plant height of the individual rice plants, and determine the leaf age and tiller number of the individual rice plants based on the lidar point cloud data corresponding to the preset number of days after rice transplanting using a structural feature extraction model trained to convergence, and construct the structural feature data of the individual rice plants based on the plant height, leaf age, and tiller number. The dataset acquisition module 1300 is configured to acquire... A first sample dataset is taken, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent spectral feature time series data constructed from spectral feature data collected from a single rice plant at a preset number of days after rice transplanting, and structural feature time series data constructed from structural feature data. The first sample labels represent the early growth vigor score of the rice. The growth vigor prediction module 1400 is configured to input the current spectral feature data and current structural feature data of a single rice plant of the target variety into the early growth vigor prediction model trained using the first sample dataset, so as to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, thereby completing the prediction of the early growth vigor of rice.

[0075] Based on any embodiment of this application, please refer to Figure 13 Another embodiment of this application also provides an electronic device, which can be implemented by a computer device, such as... Figure 13 The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store control information sequences. When the computer-readable instructions are executed by the processor, the processor can implement a method for predicting early growth vigor of rice. The processor of the computer device provides computing and control capabilities, supporting the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the method for predicting early growth vigor of rice according to this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0076] In this embodiment, the processor is used to execute... Figure 12 The specific functions of each module are described, and the memory stores the program code and various data required to execute these modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules in the rice early growth vigor prediction device of this application, and the server can call the server's program code and data to execute the functions of all modules.

[0077] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the rice early growth vigor prediction method described in any embodiment of this application.

[0078] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the rice early growth vigor prediction method described in any embodiment of this application.

[0079] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0080] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting early growth vigor of rice, characterized in that, include: Acquire multispectral image data and lidar point cloud data of multiple individual rice plants of the target variety at a preset number of days after rice transplanting, and determine the spectral feature data of the individual rice plants based on the multispectral image data. The canopy height model of a single rice plant is generated based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to a convergent state, the leaf age and number of tillers of the single rice plant are determined according to the lidar point cloud data corresponding to the preset number of days after rice transplanting. The structural feature data of the single rice plant is constructed based on the plant height, the leaf age, and the number of tillers. Obtain a first sample dataset, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent spectral feature time series data constructed from spectral feature data collected from a single rice plant at a preset number of days after rice transplanting, and structural feature time series data constructed from structural feature data. The first sample labels represent the early growth vigor score of the rice. The current spectral and structural feature data of a single rice plant of the target variety are input into the early growth vigor prediction model trained using the first sample dataset to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, thus completing the prediction of the early growth vigor of rice.

2. The method for predicting early growth vigor of rice according to claim 1, characterized in that, The spectral feature data include normalized vegetation index, chlorophyll index, green normalized vegetation index, soil-optimized vegetation index, normalized red edge difference index, and canopy coverage. The basic network architecture of the structural feature extraction model is a point cloud deep learning network, which includes an input layer, an encoder, and a decoder. The encoder includes two T-Net mini-networks and two multilayer perceptrons. The encoder includes a first T-Net mini-network, a first multilayer perceptron, a second T-Net mini-network, a second multilayer perceptron, and a max pooling layer connected in sequence. The decoder includes a third multilayer perceptron and a fourth multilayer perceptron connected in sequence. The basic network architecture of the early growth vigor prediction model is a Long Short-Term Memory (LSTM) network based on a time attention mechanism. The LSTM network based on the time attention mechanism includes an input layer, a dual-branch LSTM module, a feature fusion module, and a prediction layer connected in sequence. The dual-branch LSTM module includes a first LSTM branch module and a second LSTM branch module. The first LSTM branch module is used to process spectral feature time series data, and the second LSTM branch module is used to process structural feature time series data. The first LSTM branch module and the second LSTM branch module each include multiple LSTM processing units corresponding to multiple time steps. Each LSTM processing unit corresponding to a time step is connected to a time attention mechanism module.

3. The method for predicting early growth vigor of rice according to claim 2, characterized in that, The steps for training a structural feature extraction model include: Obtain a second sample dataset, wherein the second sample dataset includes multiple second training samples and their corresponding second sample labels. The second training samples represent the lidar point cloud data of a single rice plant corresponding to a preset number of days after rice transplanting. The second sample labels represent the leaf age and tiller number of a single rice plant. The multiple second training samples and their corresponding second sample labels are input into a preset structural feature extraction model. A 3×3 affine transformation matrix is ​​generated in the first T-Net micro-network to perform three-dimensional spatial alignment on the lidar point cloud data. In the first multilayer perceptron, the three-dimensionally aligned lidar point cloud data is upsized from 3D to 64D to extract local features. In the second T-Net micro-network, a 64×64 affine transformation matrix is ​​generated to align the 64-dimensional local features in the feature space. In the second multilayer perceptron, the features after feature space alignment are increased to 1024 dimensions. In the max pooling layer, the 1024-dimensional features are aggregated to generate a global feature vector. After receiving the global feature vector, the decoder reduces the 1024-dimensional global feature vector to 2-dimensional in the third multilayer perceptron, and then fuses the 2-dimensional global feature vector with the preset number of days after rice transplanting in the fourth multilayer perceptron to perform a nonlinear transformation, so as to output the predicted leaf age and tiller number of the single rice plant. Repeat the above steps until the structural feature extraction model is trained to a convergent state, in order to determine the structural feature extraction model that has been trained to a convergent state.

4. The method for predicting early growth vigor of rice according to claim 2, characterized in that, The steps of inputting the current spectral and structural feature data of a single rice plant of the target variety into an early growth vigor prediction model trained using the first sample dataset to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting include: The current spectral feature time series data and current structural feature time series data of a single rice plant of the target variety are input into the input layer of the early growth vigor prediction model; In the first LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the spectral feature time series data, and in the second LSTM branch module, the LSTM processing unit at each time step performs time dynamic feature extraction on the structural feature time series data. In the temporal attention mechanism module corresponding to each time step, the weights of the spectral feature data or structural feature data output by the corresponding LSTM processing unit are calculated to obtain the spectral feature weighted vector or structural feature weighted vector corresponding to the LSTM processing unit at each time step. The spectral feature data of all time steps are weighted and aggregated into a global weighted vector of spectral features after the weights are calculated by their respective time attention mechanism modules. The structural feature data of all time steps are weighted and aggregated into a global weighted vector of structural features after the weights are calculated by their respective time attention mechanism modules. In the feature fusion module, the global weighted vector of the spectral features and the global weighted vector of the structural features are concatenated and fused to obtain a global feature vector. In the prediction layer, the global feature vector is nonlinearly mapped by a fully connected layer and a Sigmoid activation function to output the early growth vitality score of a single rice plant of the target variety at a preset number of days after rice transplanting.

5. The method for predicting early growth vigor of rice according to claim 1, characterized in that, The steps to determine an early growth viability score include: The goal is to obtain the number of days required for the target rice variety to reach the effective number of tillers, the maximum number of days required for all rice varieties to reach the effective number of tillers, the minimum number of days required for all rice varieties to reach the effective number of tillers, the number of days required for the target rice variety to reach the time node with the fastest leaf age increase, the maximum number of days required for all rice varieties to reach the time node with the fastest leaf age increase, and the minimum number of days required for all rice varieties to reach the time node with the fastest leaf age increase. Calculate a first difference between the number of days required for the target rice variety to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; calculate a second difference between the maximum number of days required for all rice varieties to reach the effective number of tillers and the minimum number of days required for all rice varieties to reach the effective number of tillers; and calculate a first ratio between the first difference and the second difference. Calculate the third difference between the number of days required for the target rice variety to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; calculate the fourth difference between the maximum number of days required for all rice varieties to reach the fastest leaf age increase point and the minimum number of days required for all rice varieties to reach the fastest leaf age increase point; and calculate the second ratio between the third difference and the fourth difference. A first sum between the first ratio and the second ratio is calculated, a third ratio between the first sum and the second value is calculated, and the early growth viability score is determined based on the difference between the first value and the third ratio.

6. The method for predicting early growth vigor of rice according to claim 1, characterized in that, The step of generating a canopy height model of a single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant includes: Acquire lidar point cloud data of a single rice plant of the target variety at a preset number of days after rice transplanting; A digital surface model and a digital elevation model of a single rice plant are generated based on the lidar point cloud data. The canopy height model of the single rice plant is determined based on the difference between the digital surface model and the digital elevation model. The plant height of a single rice plant is extracted based on the canopy height model of the single rice plant.

7. The method for predicting early growth vigor of rice according to claim 2, characterized in that, The step of determining the spectral characteristic data of the single rice plant based on the multispectral image data includes: Acquire multispectral image data of a single rice plant of the target variety at a preset number of days after rice transplanting. Preprocess the multispectral image data to determine the multispectral reflectance of the single rice plant at the preset number of days after rice transplanting. The multispectral reflectance includes green band reflectance, red band reflectance, red edge band reflectance, and near-infrared band reflectance. The spectral characteristic data of a single rice plant corresponding to a preset number of days after rice transplanting are determined based on the multispectral reflectance.

8. A device for predicting early growth vigor of rice, characterized in that, include: The spectral feature determination module is configured to acquire multispectral image data and lidar point cloud data of multiple individual rice plants of the target variety corresponding to a preset number of days after rice transplanting, and determine the spectral feature data of the individual rice plants based on the multispectral image data. The structural feature determination module is configured to generate a canopy height model of a single rice plant based on the lidar point cloud data to extract the plant height of the single rice plant. Based on the structural feature extraction model trained to a convergent state, the module determines the leaf age and tiller number of the single rice plant based on the lidar point cloud data corresponding to a preset number of days after rice transplanting. The module then constructs the structural feature data of the single rice plant based on the plant height, leaf age, and tiller number. The dataset acquisition module is configured to acquire a first sample dataset, wherein the first sample dataset includes multiple first training samples and their corresponding first sample labels. The first training samples represent spectral feature time series data constructed from spectral feature data collected from a single rice plant after a preset number of days following rice transplanting, and structural feature time series data constructed from structural feature data. The first sample labels represent the early growth vigor score of the rice. The growth vigor prediction module is configured to input the current spectral feature data and current structural feature data of a single rice plant of the target variety into the early growth vigor prediction model trained using the first sample dataset, so as to predict the early growth vigor score of a single rice plant of the target variety at a preset number of days after rice transplanting, thereby completing the prediction of the early growth vigor of rice.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images