A soybean variety screening method based on nitrogen efficiency index-high yield index two dimensions
By using multimodal image acquisition and deep learning models, combined with nitrogen efficiency and high-yield index, high-efficiency and high-yield soybean varieties were screened, solving the problem of excessive reliance on nitrogen fertilizer in existing technologies and achieving the breeding goals of resource conservation and environmental friendliness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INNER MONGOLIA UNIV FOR THE NATITIES
- Filing Date
- 2026-03-21
- Publication Date
- 2026-06-05
Smart Images
Figure CN122156793A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of soybean variety technology, and in particular to a soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index. Background Technology
[0002] By quantitatively evaluating the comprehensive performance of soybean varieties in terms of nitrogen absorption and utilization efficiency and yield potential, this approach overcomes the limitations of traditional screening methods, such as the singleness of indicators, the one-sidedness of evaluation, and the lack of practical guidance. It realizes the transformation of the screening concept from "high yield priority" to "high yield and high efficiency synergy." Its core purpose is to screen out superior soybean varieties that have both high yield potential and high nitrogen utilization efficiency, so as to reduce the excessive dependence of agricultural production on nitrogen fertilizer, reduce fertilizer application costs and environmental risks, and improve nitrogen resource utilization efficiency. At the same time, the promotion and application of high-quality varieties will increase soybean yield per unit area, ensure national soybean supply security, and promote increased production and income for farmers. In turn, at the agricultural ecosystem level, it will promote the reduction of fertilizer application and the increase of efficiency, reduce greenhouse gas emissions, and protect soil health and biodiversity. It has significant resource-saving value, environmental benefits, economic benefits, and strategic significance.
[0003] Existing technologies either focus solely on variety identity or predict phenotypes, failing to effectively correlate nitrogen efficiency-related indicators with high-yield indices. Therefore, this paper proposes a soybean variety screening method based on a dual-dimensional approach of nitrogen efficiency index and high-yield index. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A soybean variety screening method based on a dual-dimensional nitrogen efficiency index and high-yield index includes the following steps: Step 1: Construct a multimodal image acquisition hardware platform: Integrate an RGB camera, a multispectral camera, a thermal infrared camera, and a chlorophyll fluorometer on a fixed bracket to simultaneously acquire multimodal images of soybean plants under the same field of view, and simultaneously record ambient light and temperature and humidity data. Step 2: Construct a multi-source database: Based on the multimodal image acquisition hardware platform, leaf images, whole-plant images, multispectral data, and corresponding time-series physiological indicators of different soybean varieties at multiple growth stages throughout the entire growth period are collected. Each data set is labeled with variety, growth stage, cultivation conditions, and ecological zone information to establish a multi-source database containing nitrogen efficiency-related indicators and yield components. Step 3: Construct and train a unified deep learning model with multiple tasks, multiple modalities, and multiple temporal phases: Design a deep learning framework that takes the data from the multi-source database as input and outputs nitrogen efficiency-related indicators, yield components, comprehensive high-yield index, and variety identity. The deep learning model extracts common features through a shared encoder and branches out multiple task heads for joint optimization. Step 4: Decision-making and screening: Based on the nitrogen efficiency-related indicators and comprehensive high-yield index output by the deep learning model, and combined with the determined variety identity, calculate the nitrogen efficiency-high-yield synergy index for each variety, and screen the tested soybean varieties based on the nitrogen efficiency-high-yield synergy index to recommend varieties with synergistic advantages to breeders.
[0006] The above further includes: Furthermore, the nitrogen efficiency-related indicators mentioned in step 2 include leaf nitrogen content, nitrogen accumulation, and photosynthetic rate obtained by inverting spectral reflectance and fluorescence parameters acquired through a hyperspectral imaging module and a chlorophyll fluorometer, as well as nitrogen use efficiency and nitrogen harvest index obtained by continuous measurement and calculation at each growth stage.
[0007] Furthermore, the nitrogen use efficiency and nitrogen harvest index are constructed as dynamic feature sequences and stored in a database. These dynamic feature sequences include, but are not limited to, the slope of the nitrogen use efficiency curve over time, rather than relying solely on single-point measurements.
[0008] Furthermore, the yield components mentioned in step 2 include the number of pods per plant, the number of seeds per pod, and the estimated weight per hundred seeds, which are automatically obtained by image segmentation and instance segmentation processing of local high-definition images of pods and seeds.
[0009] Furthermore, the unified deep learning framework described in step 3 is specifically as follows: Design a dual-stream Transformer architecture, where one stream processes RGB images and the other stream processes multispectral and thermal infrared data; The system uses cross-attention modules to interact with features at multiple levels, capturing the distribution and translocation characteristics of nitrogen in plants and outputting predicted values of nitrogen efficiency-related indicators.
[0010] Furthermore, the specific method for outputting the comprehensive high-yield index in step 3 is as follows: A branch is set in the model output layer, and a temporal convolutional network or a long short-term memory network is used to dynamically model the characteristics of yield components at multiple growth stages, predict the yield per plant and yield per unit area at maturity, and use actual yield measurement data as a supervision signal for end-to-end training.
[0011] Furthermore, the specific method for outputting the variety identity in step 3 is as follows: By using feature matching or a classifier, the features of the input leaf image and whole plant image are compared with the features of known varieties in the database to determine whether the variety to be tested is a known variety or a new variety.
[0012] Furthermore, the calculation basis for the nitrogen efficiency-high yield synergy index mentioned in step 4 is as follows: Graph neural networks are used to model the coupling relationship between nitrogen efficiency-related indicators and yield components, thereby quantifying the synergistic ability of each soybean variety in terms of both nitrogen efficiency and high yield.
[0013] The present invention has the following beneficial effects: In this invention, the nitrogen efficiency-high yield synergy index is calculated, and the coupling relationship between the two is modeled using a graph neural network. This scientifically quantifies the synergistic performance of each variety in the two dimensions of resource efficiency and high and stable yield. This provides breeders with a new and more valuable screening index, which can accurately recommend breakthrough varieties that are both fertilizer-saving and high-yielding, and truly achieve the unification of breeding goals. Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating the steps of a soybean variety screening method based on a dual-dimensional nitrogen efficiency index and high-yield index proposed in this invention. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] Please see Figure 1 As shown, this invention is a soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index, comprising the following steps: Step 1: Construct a multimodal image acquisition hardware platform: Integrate an RGB camera, a multispectral camera, a thermal infrared camera, and a chlorophyll fluorometer on a fixed bracket to simultaneously acquire multimodal images of soybean plants under the same field of view, and simultaneously record ambient light and temperature and humidity data. The integration method is as follows: a replaceable lens is mounted on a fixed bracket, including a macro lens for capturing leaf images and a wide-angle lens for capturing whole plant images, and an adjustable LED array is equipped as an ambient light compensation device. Compensation is turned on and a calibration plate is placed during standardized imaging, and compensation is turned off during natural light imaging. Step 2: Construct a multi-source database: Based on the multimodal image acquisition hardware platform, leaf images, whole-plant images, multispectral data, and corresponding time-series physiological indicators of different soybean varieties at multiple growth stages throughout the entire growth period are collected. Each data set is labeled with variety, growth stage, cultivation conditions, and ecological zone information to establish a multi-source database containing nitrogen efficiency-related indicators and yield components. Step 3: Construct and train a unified deep learning model with multiple tasks, multiple modalities, and multiple temporal phases: Design a deep learning framework that takes the data from the multi-source database as input and outputs nitrogen efficiency-related indicators, yield components, comprehensive high-yield index, and variety identity. The deep learning model extracts common features through a shared encoder and branches out multiple task heads for joint optimization. Step 4: Decision-making and screening: Based on the nitrogen efficiency-related indicators and comprehensive high-yield index output by the deep learning model, and combined with the determined variety identity, calculate the nitrogen efficiency-high-yield synergy index for each variety, and screen the tested soybean varieties based on the nitrogen efficiency-high-yield synergy index to recommend varieties with synergistic advantages to breeders.
[0017] Step 4 is followed by a visualization step: the key phenotypic change trajectories of each growth stage are visualized to assist breeding decisions.
[0018] In one embodiment, the nitrogen efficiency-related indicators mentioned in step 2 include leaf nitrogen content, nitrogen accumulation, and photosynthetic rate obtained by inverting spectral reflectance and fluorescence parameters acquired through a hyperspectral imaging module and a chlorophyll fluorometer, as well as nitrogen use efficiency and nitrogen harvest index obtained by continuous measurement and calculation at each growth stage.
[0019] In one embodiment, the nitrogen use efficiency and nitrogen harvest index are constructed as dynamic feature sequences and stored in a database. These dynamic feature sequences include, but are not limited to, the slope of the nitrogen use efficiency curve over time, rather than relying solely on single-point measurements.
[0020] In one embodiment, the yield components mentioned in step 2 include the number of pods per plant, the number of seeds per pod, and the estimated weight per hundred seeds, which are automatically calculated by performing image segmentation and instance segmentation on local high-definition images of pods and seeds.
[0021] It should be noted that the specific analysis process for constructing a multi-source database is as follows: Determine the collection plan and growth period division: For the multiple soybean varieties tested, their entire growth period was divided into five key growth periods: seedling stage, flowering stage, pod-setting stage, grain-filling stage, and maturity stage, and the specific collection time points for each growth period were determined. Multimodal image synchronous acquisition: Using the multimodal image acquisition hardware platform, leaf images, whole plant images, multispectral data and chlorophyll fluorescence parameters of each soybean variety are synchronously acquired for each fixed sample plant at each acquisition time point. During the acquisition process, light intensity, ambient temperature and air humidity data are synchronously recorded by environmental sensors. Measurement and calculation of time-series physiological indicators: At each growth stage, the fixed sample plants were sampled destructively or non-destructively to measure the total nitrogen accumulation and grain nitrogen content of the plants. Nitrogen use efficiency and nitrogen harvest index were calculated according to the formula as the core components of nitrogen efficiency-related indicators. Nitrogen use efficiency is specifically defined as the ratio of grain yield to total nitrogen content of the plant, and nitrogen harvest index is specifically defined as the ratio of grain nitrogen content to total nitrogen content of the plant. The values of these two indicators at each growth stage were constructed into a dynamic feature sequence, including the slope of their change curve over time. Extraction of yield components: During the pod-filling and ripening stages, high-resolution images of pods and seeds collected are processed for image segmentation and instance segmentation. The number of pods per plant, the number of seeds per pod, and the weight of 100 seeds are automatically counted and estimated. As the core components of yield, the image segmentation and instance segmentation algorithm adopts the Mask R-CNN deep learning model. By automatically identifying and segmenting high-resolution images of pods and seeds collected during the pod-filling and ripening stages, the algorithm achieves accurate counting of the number of pods per plant and the number of seeds per pod, as well as automatic estimation of the weight of 100 seeds. Spectral index inversion and feature generation: The collected multispectral data and chlorophyll fluorescence parameters are processed, and leaf nitrogen content, nitrogen accumulation, photosynthetic rate, normalized vegetation index, and chlorophyll content index are obtained through the preset spectral index model. These serve as a supplementary component of nitrogen efficiency-related indicators. Specifically, the leaf nitrogen content and photosynthetic rate are obtained through the preset spectral index model by using the spectral reflectance of leaves in the visible to near-infrared bands obtained by the hyperspectral imaging module, combined with chlorophyll fluorescence parameters, and inverted through the calculation model of nitrogen balance index and SPAD value. Multi-dimensional data annotation and association: For all the original and derived data obtained in the above steps, label each one with the variety name, growth period name, specific collection date, cultivation condition type and ecological zone source, and bind the leaf images, whole plant images, spectral data and physiological index data of the same plant and the same period through cross-modal association. Constructing a multi-source database: All labeled and linked data, including raw image data, nitrogen efficiency-related indicators obtained from inversion, extracted yield components, environmental parameters, and labeling information, are stored in a database system in a unified format to form a multi-source database for model training and variety selection.
[0022] In one embodiment, the unified deep learning framework described in step 3 is specifically: Design a dual-stream Transformer architecture, where one stream processes RGB images and the other stream processes multispectral and thermal infrared data; The system uses cross-attention modules to interact with features at multiple levels, capturing the distribution and translocation characteristics of nitrogen in plants and outputting predicted values of nitrogen efficiency-related indicators.
[0023] In one embodiment, the specific method for outputting the comprehensive high-yield index in step 3 is as follows: A branch is set in the model output layer, and a temporal convolutional network or a long short-term memory network is used to dynamically model the characteristics of yield components at multiple growth stages, predict the yield per plant and yield per unit area at maturity, and use actual yield measurement data as a supervision signal for end-to-end training.
[0024] In one embodiment, the specific method for outputting the variety identity in step 3 is as follows: By using feature matching or a classifier, the features of the input leaf image and whole plant image are compared with the features of known varieties in the database to determine whether the variety to be tested is a known variety or a new variety.
[0025] It should be noted that the specific analytical process for building and training a deep learning model is as follows: Define multimodal input data: For each soybean variety sample in the database, the model input consists of three parts: RGB image features : Represents the original pixel matrix extracted from leaf and whole-plant images. For time-series data, it is represented as , where T represents different growth stages; Multispectral and thermal infrared characteristics : Represents multi-channel data acquired from multispectral and thermal infrared cameras, including red-edge band, near-infrared band, and thermal infrared information, denoted as ; Time-series physiological index sequence This represents time-series physiological indicators recorded at fixed time intervals from seedling stage to maturity, such as yield components like the number of pods per plant and the weight of 100 seeds at each growth stage, as well as nitrogen efficiency-related indicators like leaf nitrogen content and photosynthetic rate. It is expressed as... , where N is the total number of time points; Constructing a dual-stream Transformer encoder (shared encoder): First stream: RGB image encoding stream, this stream is responsible for processing RGB image features. First, By performing patch embedding through a convolutional layer, a sequence of feature vectors for a series of image patches is obtained. Then, these sequences are fed into a standard Transformer encoder to capture the spatial structure and global contextual information within the image. This process is represented as follows: ,in, Positional encoding is used to preserve the spatial location information of image patches. These are depth features extracted from an RGB image; The second stream: the multispectral and thermal infrared data encoding stream, which is responsible for processing multispectral and thermal infrared features. Similar to RGB streams, firstly... Embedding is performed to obtain a sequence of feature vectors. This is then fed into another Transformer encoder to extract spectral and thermal infrared features, represented as: ,in, Depth features extracted from multispectral and thermal infrared data; Cross-modal feature interaction: A cross-attention module is introduced between multiple layers of two streams. Taking the l-th layer as an example, the output of the RGB stream... As a query (Query), the output of the multispectral stream As a key (Key) and value (Value) performs cross-attention calculations, allowing RGB features to learn nitrogen-related information from spectral features. Conversely, it allows spectral features to learn spatial structure information from RGB features. The calculation formula is as follows: ; ; Through multi-layered cross-interactions, a joint feature representation that integrates spatial structure and spectral information is ultimately obtained. This feature fully captures information on the distribution and transport of nitrogen; Constructing a multi-task learning decoder (multiple task heads): In obtaining joint feature representation The model then feeds this data into multiple parallel task headers to output multiple predictions simultaneously: Task Header 1: Prediction of Nitrogen Efficiency Related Indicators This task header is based on joint features By using a multilayer perceptron (MLP) regression network, predicted values of nitrogen efficiency-related indicators are directly output, including but not limited to leaf nitrogen content. Nitrogen utilization efficiency and nitrogen harvest index , is represented as: ; Task Header 2: Comprehensive High-Yield Index Prediction Header (Introducing Time Series Model): This task head not only depends on the current joint features It also incorporates historical time-series data. We will With time-series physiological index sequences The extracted output component features are spliced together to form a new time-series feature sequence. Then, this temporal sequence is input into a temporal convolutional network. Alternatively, in a Long Short-Term Memory (LSTM) network, dynamic modeling is performed on the yield components across multiple growth stages to learn their temporal variations and their weights in relation to the final yield. Finally, a fully connected layer outputs a predicted comprehensive high-yield index. (i.e., predicted yield per plant or yield per unit area at maturity), where the TCN network effectively expands the receptive field through dilated convolution, enabling it to capture dependencies over a longer time span. Task Head 3: Variety Identification Head: It is responsible for determining the variety of the sample being tested. It combines features. Input into a classifier or feature matching module; Using a classifier approach: a softmax layer is used to output the probability that the sample belongs to each known variety in the database. ; Feature matching method: Calculate Template characteristics of all known varieties in the database The similarity between the varieties (e.g., cosine similarity) is used to select the variety with the highest similarity as the prediction result. If the highest similarity score is lower than a preset threshold, it is determined to be a new variety, as shown below: ; Define multi-task loss function and joint optimization: Jointly optimize the entire model and define a multi-task loss function. , which is the weighted sum of the loss functions of each task head: ,in, This refers to the loss in the prediction head of nitrogen efficiency-related indicators, typically measured using the mean squared error (MSE) loss, which is used to measure the predicted value. With real labels The differences between them This is the loss of the head predicted by the comprehensive high-yield index, also using MSE loss to measure the predicted yield. Compared with actual production data The difference lies in using actual yield measurement data as a monitoring signal for end-to-end training. The loss function is used for determining the variety identity. For classification tasks, cross-entropy loss is used; for feature matching tasks, triplet loss can be used to bring similar samples closer together and push away features of dissimilar samples. It is a weighted hyperparameter that balances the importance of each task; Using the backpropagation algorithm, gradient descent is employed to iteratively optimize the parameters of the entire network (including the parameters of the shared encoder and each task head) end-to-end until... convergence.
[0026] In one embodiment, the calculation basis for the nitrogen efficiency-high yield synergy index in step 4 is: Graph neural networks are used to model the coupling relationship between nitrogen efficiency-related indicators and yield components, thereby quantifying the synergistic ability of each soybean variety in terms of nitrogen efficiency and high yield.
[0027] It should be noted that the specific analytical process for calculating the nitrogen efficiency-high yield synergy index is as follows: Construct the feature matrix of the variety node: Each soybean variety is considered as a node in the graph. Nitrogen efficiency-related indicators and yield components for this variety are extracted from multi-source databases and concatenated to form the initial feature vector of this node. ; The nitrogen efficiency-related indicators include, but are not limited to: leaf nitrogen content. Nitrogen accumulation Photosynthetic rate Nitrogen utilization efficiency and nitrogen harvest index ; The yield components include, but are not limited to: number of pods per plant. Number of grains per pod and weight per 100 grains ,node The initial feature vector is represented as: For all N known varieties, construct the initial feature matrix. ; Constructing a relationship diagram structure between varieties: Definition diagram ,in, For variety node set, This is a set of edges, constructed based on the physiological similarity and genetic distance between varieties, specifically determined through the following two methods: Edge connections based on phenotypic similarity: Calculate edge connections between any two variety nodes. and If the Euclidean distance between nitrogen efficiency-related indicators and yield components is less than a preset threshold, an edge is added between the two nodes, indicating that they are similar in terms of nitrogen efficiency and high yield phenotypes. Edge connections based on genetic distance: If two varieties have similar genetic backgrounds, edge connections are added based on prior knowledge to strengthen the information transfer between varieties with a common genetic basis.
[0028] The final adjacency matrix A records the connection relationships between nodes; Message passing and feature aggregation in neural networks: Feature learning is performed using a multi-layer graph convolutional network (GCN). In each layer l, nodes update their own feature representation by aggregating the features of their neighboring nodes, thereby capturing the coupling relationship between nitrogen efficiency-related indicators and yield components. The propagation rules of graph convolutional layers are as follows: ,in, For adjacency matrices with added self-connections, for The degree matrix, , Let l be the node feature matrix of the l-th layer. As the initial input, Let be the learnable weight matrix of the l-th layer. It is a non-linear activation function; After L-layer graph convolution, each variety node The final feature representation It has deeply integrated nitrogen efficiency-related indicators and yield component information of its own and structurally related varieties, reflecting the coupling relationship between nitrogen efficiency and high yield indicators; Calculate the nitrogen efficiency-high yield synergy index: The final feature representation of the nodes output by the graph neural network The input is fed into a synergistic index calculation module composed of fully connected layers to obtain the nitrogen efficiency-high yield synergistic index for this variety. , represented as ,in, and As learnable parameters, the Sigmoid function maps its output to the [0,1] interval. The higher the value, the stronger the synergistic effect of the variety in both nitrogen efficiency and high yield. That is, the more prominent the ability to absorb and utilize nitrogen under the condition of maintaining high yield, or the higher the yield can be achieved while utilizing nitrogen efficiently. Model training and optimization: Using known variety data from a multi-source database, and with actual measured yield and nitrogen efficiency data as supervisory signals, the aforementioned graph neural network and synergy index calculation module are jointly trained end-to-end. Loss function. Defined as the mean squared error (MSE) between the synergy index and the actual overall performance of the variety: ,in, This provides authentic synergistic labels for varieties based on historical breeding data or expert experience. Backpropagation is used to optimize network parameters, enabling the model to accurately quantify the nitrogen efficiency-high yield synergistic index for each soybean variety.
[0029] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A soybean variety screening method based on a dual-dimensional nitrogen efficiency index and high-yield index, characterized in that, Includes the following steps: Step 1: Construct a multimodal image acquisition hardware platform: Simultaneously acquire multimodal images of soybean plants under the same field of view, and simultaneously record ambient light and temperature and humidity data; Step 2: Construct a multi-source database: Based on the multimodal image acquisition hardware platform, leaf images, whole-plant images, multispectral data, and corresponding time-series physiological indicators of different soybean varieties at multiple growth stages throughout the entire growth period are collected. Each data set is labeled with variety, growth stage, cultivation conditions, and ecological zone information to establish a multi-source database containing nitrogen efficiency-related indicators and yield components. Step 3: Construct and train a deep learning model: Design a deep learning framework that takes the data from the multi-source database as input and outputs nitrogen efficiency-related indicators, yield components, comprehensive high-yield index, and variety identity. The deep learning model extracts common features through a shared encoder and branches out multiple task heads for joint optimization. Step 4: Decision-making and screening: Based on the nitrogen efficiency-related indicators and comprehensive high-yield index output by the deep learning model, and combined with the determined variety identity, calculate the nitrogen efficiency-high-yield synergy index for each variety, and screen the tested soybean varieties based on the nitrogen efficiency-high-yield synergy index to recommend varieties to breeders.
2. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The nitrogen efficiency-related indicators mentioned in step 2 include leaf nitrogen content, nitrogen accumulation, and photosynthetic rate obtained by inverting spectral reflectance and fluorescence parameters acquired through a hyperspectral imaging module and a chlorophyll fluorometer, as well as nitrogen use efficiency and nitrogen harvest index obtained by continuous measurement and calculation at each growth stage.
3. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 2, characterized in that, The nitrogen utilization efficiency and nitrogen harvest index are constructed as dynamic feature sequences and stored in the database.
4. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The yield components mentioned in step 2 include the number of pods per plant, the number of seeds per pod, and the estimated weight per hundred seeds, which are automatically obtained by image segmentation and instance segmentation of local high-definition images of pods and seeds.
5. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The unified deep learning framework mentioned in step 3 is specifically as follows: Design a dual-stream Transformer architecture, where one stream processes RGB images and the other stream processes multispectral and thermal infrared data; The system uses cross-attention modules to interact with features at multiple levels, capturing the distribution and translocation characteristics of nitrogen in plants and outputting predicted values of nitrogen efficiency-related indicators.
6. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The specific method for outputting the comprehensive high-yield index in step 3 is as follows: A branch is set in the model output layer, and a temporal convolutional network or a long short-term memory network is used to dynamically model the characteristics of yield components at multiple growth stages, predict the yield per plant and yield per unit area at maturity, and use actual yield measurement data as a supervision signal for end-to-end training.
7. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The specific method for outputting the variety identity in step 3 is as follows: By using feature matching or a classifier, the features of the input leaf image and whole plant image are compared with the features of known varieties in the database to determine whether the variety to be tested is a known variety or a new variety.
8. The soybean variety screening method based on a dual dimension of nitrogen efficiency index and high yield index as described in claim 1, characterized in that, The calculation basis for the nitrogen efficiency-high yield synergy index mentioned in step 4 is as follows: A graph neural network was used to model the coupling relationship between nitrogen efficiency-related indicators and yield components.