A method and system for intelligent management of maize plant germplasm resources

Abstract

This invention relates to the field of germplasm resource management technology, specifically to an intelligent management method and system for maize plant germplasm resources. The method includes: collecting field images, environmental sensor data, and soil nutrient data from multiple maize growth cycles to construct an original germplasm dataset; performing spatiotemporal alignment on the dataset to construct a multidimensional growth trajectory tensor for a single maize plant, from which morphological and physiological feature time-series sequences are extracted; inputting the two time-series sequences into a pre-trained generative adversarial network (GAN) to output an enhanced growth feature field; associating and encoding the enhanced growth feature field with genotype tag data to generate an enhanced germplasm feature matrix; and then using cluster analysis to classify germplasm resource groups and assign unique group identifiers. This method achieves the regularization and fusion of multi-source heterogeneous germplasm data and feature enhancement, automatically classifying germplasm groups in a data-driven manner, avoiding human subjective errors, and realizing the digital, standardized, and intelligent management and control of maize germplasm resources.

Description

Technical Field

[0001] This invention relates to the field of germplasm resource management technology, and in particular to an intelligent management method and system for maize plant germplasm resources. Background Technology

[0002] In the process of organizing and archiving maize germplasm resources, conventional collection methods often involve separately recording field images, environmental parameters, and soil nutrient data. These data are stored independently and lack a unified spatiotemporal calibration mechanism. Observational data collected across multiple growth cycles are not structured and integrated, making it impossible to establish a systematic data carrier for individual plants. This data cannot fully represent the entire growth process in tensor form, making it difficult to extract continuous temporal sequences of morphological and physiological characteristics.

[0003] Traditional germplasm data analysis workflows lack a processing mode for constructing growth trajectory tensors from multi-source data. Various observational information remains discrete, with low data correlation, resulting in insufficient regularity in subsequent growth feature extraction. Germplasm feature mining lacks a feature enhancement step using generative adversarial networks, leading to single-dimensional extracted features, incomplete feature representation, and an inability to form a complete growth feature field.

[0004] Current technologies struggle to correlate growth characteristics with genotype tag data, making it impossible to construct a germplasm characteristic matrix that integrates phenotypic and genetic information. Germplasm group classification relies heavily on manual experience, lacking quantitative clustering analysis mechanisms, unified classification standards, and standardized identifiers for each group. The industry requirements for digital storage, intelligent classification, and standardized management of germplasm resources cannot be met by existing traditional processing methods for large-scale, long-term management. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent management method and system for maize plant germplasm resources.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent management of maize plant germplasm resources, comprising: Field image data, environmental sensor data, and soil nutrient data of maize plants at multiple growth stages were collected as the original germplasm dataset; Perform spatiotemporal alignment operations on the original germplasm dataset to establish a multidimensional growth trajectory tensor for each individual maize plant; The temporal sequences of morphological features and physiological features of each individual plant are extracted from the multidimensional growth trajectory tensor. The morphological feature time sequence and the physiological feature time sequence are input into a pre-trained generative adversarial network and the generator outputs an enhanced growth feature field. The enhanced growth feature field is associated and encoded with the genotype tag data in the original germplasm dataset to generate an enhanced germplasm feature matrix carrying genotype information. Cluster analysis is performed on the enhanced germplasm feature matrix to divide different germplasm resource groups and assign a unique group identifier to each group.

[0007] As a further aspect of the present invention, field image data, environmental sensor data, and soil nutrient data of maize plants at multiple growth cycles are collected as the original germplasm dataset, specifically including: Multiple image acquisition devices, multiple environmental sensors, and multiple soil nutrient sensors were deployed in a fixed spatial grid within the target corn planting area. When each preset growth cycle is reached, the image acquisition device is triggered to synchronously acquire multi-angle canopy images of each individual plant as the field image data; The air temperature, air humidity, and light intensity at each grid point are continuously recorded as environmental sensor data in a continuous time series. When each irrigation event is triggered, the soil nutrient sensor collects the nitrogen, phosphorus, potassium and organic matter content of soil layers at different depths at each grid point as the soil nutrient data; Field image data, environmental sensor data, and soil nutrient data collected at the same grid point and during the same growth cycle are associated and bound according to the individual plant number; All the associated and bound data is organized into a three-dimensional data cube with plant individual number as the first index, growth cycle number as the second index, and data type as the third index, which serves as the original germplasm dataset.

[0008] As a further aspect of the present invention, a spatiotemporal alignment operation is performed on the original germplasm dataset to establish a multidimensional growth trajectory tensor for each individual maize plant, specifically including: Extract all data slices of the same plant individual number under different growth cycle numbers from the three-dimensional data cube of the original germplasm dataset; Edge detection was performed on the field image data in each data slice to segment the outer contour and leaf vein skeleton of individual plants; The outer contour and the leaf vein skeleton are projected onto a unified spatial coordinate system to eliminate positional offsets between different image acquisition devices; Perform temporal interpolation on the environmental sensor data and soil nutrient data in each data slice to align the sensor data sampling time points between different growth cycles to the same time reference; The aligned outer contour, leaf vein skeleton, environmental sensor data, and soil nutrient data are stacked sequentially according to the time order of the growth cycle. The stacked data structure is defined as a multidimensional tensor with the growth cycle as the time axis, the spatial coordinates as the spatial axis, and the sensor parameters as the feature axis, which is called the multidimensional growth trajectory tensor.

[0009] As a further aspect of the present invention, the morphological feature time sequence and physiological feature time sequence of each individual plant are extracted from the multidimensional growth trajectory tensor, specifically including: For the outer contour in the multidimensional growth trajectory tensor, calculate the projected area, perimeter, and aspect ratio of the outer contour for each growth cycle. The calculated projected area, perimeter, and aspect ratio values ​​are arranged into a morphological feature vector sequence according to the time sequence of the growth cycle, which is used as the morphological feature time sequence. For the leaf vein skeleton in the multidimensional growth trajectory tensor, the total length of leaf veins, the number of leaf vein branch nodes, and the average curvature of leaf veins are extracted for each growth cycle. The extracted total vein length, number of vein branch nodes, and average vein curvature are appended to the corresponding positions in the morphological feature time sequence. For the environmental sensor data and soil nutrient data in the multidimensional growth trajectory tensor, calculate the cumulative value of air temperature, the average value of air humidity, the cumulative value of light intensity, the average value of nitrogen content, the average value of phosphorus content, the average value of potassium content, and the average value of organic matter content for each growth cycle. The calculated cumulative values ​​of air temperature, average air humidity, cumulative light intensity, average nitrogen content, average phosphorus content, average potassium content, and average organic matter content are arranged in chronological order according to the growth cycle to form a physiological feature vector sequence, which is used as the physiological feature time sequence.

[0010] As a further aspect of the present invention, the average curvature of the leaf vein is obtained by averaging the curvature values ​​of the Frenet frame at each pixel point in the leaf vein skeleton.

[0011] As a further aspect of the present invention, the morphological feature time sequence and the physiological feature time sequence are input into a pre-trained generative adversarial network, and the generator outputs an enhanced growth feature field, specifically including: The morphological feature time sequence and the physiological feature time sequence are concatenated along the feature dimension to form a fused feature matrix; The fused feature matrix is ​​input into the generator of the generative adversarial network; The generator of the generative adversarial network performs multi-layer deconvolution operations on the fused feature matrix to expand the feature dimension and the temporal dimension; After each deconvolution operation, batch normalization and activation function mapping are performed on the intermediate feature map output by that layer. Spatial domain smoothing filtering is performed on the feature map output from the last deconvolution operation to eliminate high-frequency noise artifacts; The smoothed and filtered feature map is output as the enhanced growth feature field.

[0012] As a further aspect of the present invention, the spatial domain smoothing filter employs an anisotropic Gaussian filter, the standard deviation of which is adaptively adjusted along the time dimension of the feature field.

[0013] As a further aspect of the present invention, the enhanced growth feature field is associated with and encoded with the genotype tag data in the original germplasm dataset to generate an enhanced germplasm feature matrix carrying genotype information, specifically including: Extract the genotype tag data corresponding to the individual plant number from the original germplasm dataset; The genotype tag data includes the coding sequence of the single nucleotide polymorphism site and the coding sequence of the molecular marker banding for the individual plant. The enhanced growth feature field is divided into feature sub-blocks of the same number as the genotype tag data according to the plant individual number; Perform average pooling on each feature sub-block to compress the feature sub-block into a fixed-length feature vector; Each fixed-length feature vector is concatenated with the corresponding single nucleotide polymorphism site coding sequence and molecular marker banding coding sequence. All the spliced ​​vectors are stacked into a two-dimensional matrix according to the individual plant number, which serves as the enhanced germplasm feature matrix carrying genotype information.

[0014] As a further aspect of the present invention, cluster analysis is performed on the enhanced germplasm feature matrix to classify different germplasm resource groups and assign a unique group identifier to each group, specifically including: Perform principal component dimensionality reduction on the enhanced germplasm feature matrix to extract principal component feature vectors whose cumulative contribution rate exceeds a preset threshold; The extracted principal component feature vectors are projected into a low-dimensional space to obtain the low-dimensional embedding points of each plant individual. Initialize a preset number of candidate cluster centers and continuously update the position of each candidate cluster center using an iterative optimization algorithm; After each iteration update, each low-dimensional embedding point is assigned to the cluster of the nearest candidate cluster center; When the rate of change of the candidate cluster center is lower than the preset convergence threshold, the iteration stops and the cluster division result at this time is taken as the final cluster division. Assign a unique numerical group identifier to each final group and establish a mapping table between the group identifier and the individual plant numbers within the group.

[0015] As a further aspect of the present invention, the present invention also includes an intelligent management system for maize plant germplasm resources. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the intelligent management method for maize plant germplasm resources described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Spatiotemporal alignment processing was performed on the original maize germplasm dataset from multiple sources and periods to eliminate misalignment biases in time nodes and spatial sampling locations among different data categories, and to construct a multidimensional growth trajectory tensor corresponding to a single maize plant. This tensor structure uniformly carries multidimensional growth information from images, environment, and soil. Morphological and physiological feature time-series sequences were extracted from the structured tensor, allowing the external morphological evolution and internal physiological changes of the plant growth process to form a continuous and traceable temporal representation, thus standardizing the organization of multi-source discrete data.

[0017] Morphological and physiological time-series sequences are input into a generative adversarial network (GAN) to enhance feature generation. The enhanced growth feature field output by the generator enriches the expression dimensions of the original features and improves the representation ability of latent features in the growth process. The enhanced growth feature field is combined with genotype tag data for association coding, enabling deep integration of growth phenotypic information and genetic information, and constructing an enhanced germplasm feature matrix with multi-layered association information.

[0018] Cluster analysis is conducted based on an enhanced germplasm feature matrix, automatically distinguishing germplasm resource groups according to the data's own distribution structure, thus eliminating the bias caused by subjective human judgment. Each germplasm resource is assigned a unique and fixed group identifier, giving each type of germplasm resource a unique and identifiable code, forming a standardized germplasm resource classification and numbering system that is compatible with the operational process of large-scale systematic archiving and intelligent classification of maize germplasm resources. Attached Figure Description

[0019] Figure 1 This is a state diagram of an intelligent management method for maize plant germplasm resources according to the present invention; Figure 2 A flowchart for constructing the original germplasm dataset; Figure 3 This is a flowchart for extracting time-series morphological and physiological features. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] See Figure 1 A method for intelligent management of maize plant germplasm resources is implemented as follows: Field image data, environmental sensor data, and soil nutrient data of maize plants at multiple growth cycles are collected as the original germplasm dataset. Spatiotemporal alignment is performed on the original germplasm dataset to establish a multidimensional growth trajectory tensor for each individual maize plant. The morphological feature time series and physiological feature time series of each individual plant are extracted from the multidimensional growth trajectory tensor. The morphological feature time series and physiological feature time series are input into a pre-trained generative adversarial network, and the generator outputs an enhanced growth feature field. The enhanced growth feature field is associated with the genotype label data in the original germplasm dataset to generate an enhanced germplasm feature matrix carrying genotype information. Cluster analysis is performed on the enhanced germplasm feature matrix to classify different germplasm resource groups and assign a unique group identifier to each group.

[0022] In one embodiment of the present invention, see [reference] Figure 2 Multiple image acquisition devices, environmental sensors, and soil nutrient sensors were deployed in a fixed spatial grid within the target maize planting area. At the arrival of each preset growth cycle, the image acquisition devices were triggered to synchronously acquire multi-angle canopy images of each individual plant as field image data. In a continuous time series, the environmental sensors continuously recorded the air temperature, air humidity, and light intensity at each grid point as environmental sensor data. At each irrigation event, the soil nutrient sensors collected the nitrogen, phosphorus, potassium, and organic matter content of soil layers at different depths at each grid point as soil nutrient data. The field image data, environmental sensor data, and soil nutrient data collected at the same grid point and in the same growth cycle were associated and bound according to the individual plant number. All the associated and bound data were organized into a three-dimensional data cube with the individual plant number as the first index, the growth cycle number as the second index, and the data type as the third index as the original germplasm dataset.

[0023] In practice, multiple image acquisition devices, environmental sensors, and soil nutrient sensors are deployed in a fixed spatial grid within the target corn planting area. The grid spacing is set to 2 meters. At each grid point, one image acquisition device, one environmental sensor, and one soil nutrient sensor are installed. The image acquisition device is a multispectral camera, the environmental sensor is an integrated temperature, humidity, and light sensor, and the soil nutrient sensor is an ion-selective electrode array sensor. At the arrival of each preset growth cycle, the image acquisition device is triggered to synchronously acquire multi-angle canopy images of each individual plant as field image data. The growth cycle includes five stages: seedling stage, jointing stage, tasseling stage, silking stage, and grain-filling stage. Canopy images are acquired from five different angles at the arrival of each growth cycle. In a continuous time series, the environmental sensors continuously record air temperature, air humidity, and light intensity at each grid point as environmental sensor data. The sampling interval for the continuous time series is 30 minutes, and the recorded time range covers the period from sowing... Throughout the entire growth period from planting to maturity, soil nutrient sensors collect nitrogen, phosphorus, potassium, and organic matter content at different soil depths at each grid point during each irrigation event. These different soil depths include 0-20 cm, 20-40 cm, and 40-60 cm. The irrigation control system automatically records the trigger time of each irrigation event. Soil nutrient data is collected after a 2-hour delay following each irrigation event. Field image data, environmental sensor data, and soil nutrient data collected at the same grid point and during the same growth cycle are associated and bound according to the plant individual number. The plant individual number is a unique identifier assigned to each corn plant at the time of planting. The binding operation performs a database join by reading the spatial coordinate information and timestamp information recorded by each sensor. All the associated and bound data is organized into a three-dimensional data cube as the original germplasm dataset, with the plant individual number as the first index, the growth cycle number as the second index, and the data type as the third index. The definition of the three-dimensional data cube is shown in the formula:

[0024] in: Represents a data entry in a 3D data cube, subscript Number each plant individually, with a subscript. Number the growth cycle, subscript Index for data types, Indicates the first The individual plant was in the first Field image data at each growth cycle Indicates the first The individual plant was in the first Environmental sensor data during each growth cycle Indicates the first The individual plant was in the first Soil nutrient data for each growth cycle.

[0025] In some embodiments, the target corn planting area is divided into a fixed spatial grid of 20 meters × 20 meters, with each grid point corresponding to a unique grid number. The image acquisition device, the environmental sensor, and the soil nutrient sensor are all bound to the grid number. The storage record of the acquired data simultaneously writes the timestamp of the acquisition time and the grid number. When associating and binding according to the plant individual number, the field image data, environmental sensor data, and soil nutrient data of the same grid point and the same growth cycle are matched to the corresponding plant individual number based on the grid number and the timestamp. Each plant individual number is recorded to its corresponding grid number through a geographic information system at the time of sowing.

[0026] In some embodiments, the preset growth cycle is determined by an accumulated temperature model. When the effective accumulated temperature accumulated since the sowing date reaches the first threshold, the seedling stage is determined to be entered; when it reaches the second threshold, the jointing stage is determined to be entered; when it reaches the third threshold, the tasseling stage is determined to be entered; when it reaches the fourth threshold, the silking stage is determined to be entered; and when it reaches the fifth threshold, the grain-filling stage is determined to be entered. When each growth cycle is reached, the image acquisition device is triggered to perform an acquisition task. The acquisition task includes taking canopy images from five angles within 2 seconds, with an interval of 0.4 seconds between each angle.

[0027] Optionally, the multi-angle canopy images include five images acquired from a 90-degree overhead view directly above the plant, a 45-degree eastward tilt angle, a 45-degree westward tilt angle, a 45-degree southward tilt angle, and a 45-degree northward tilt angle. Each image has a resolution of 1920×1080 pixels and is stored in RGB three-channel color image format. Optionally, during the recording of the continuous time series, if environmental sensor data is missing at a certain sampling moment, linear interpolation is performed using data from two adjacent sampling moments to complete the data. The linear interpolation formula is as follows: ,in For the time points where data is missing, and These are the two most recent valid sampling times before and after the missing time, and they satisfy... < < , for Real-time recorded environmental sensor data values for The actual environmental sensor data values ​​recorded at any given time. The interpolation formula contains... The output value type should be consistent with the input value type; that is, when the original data for interpolation is an air temperature value, The unit is degrees Celsius; when the original data for interpolation is air humidity value... The unit is relative humidity as a percentage; when the original data for interpolation is light intensity value, The unit is lux. The interpolated and completed data, along with the original acquired data, is stored as environmental sensor data and aligned with the time reference in subsequent spatiotemporal alignment operations.

[0028] It is understood that the irrigation event is automatically triggered when the soil moisture sensor reading falls below 60% of field capacity. The trigger time of each irrigation event is recorded as an event timestamp, along with the irrigation duration and volume. The soil nutrient sensor collects data two hours after the irrigation event is triggered, at which point the moisture distribution is uniform and the ion concentration is relatively stable. The three soil depths collected correspond to the topsoil, subsoil, and tillage layers, respectively. It is also understood that when the three-dimensional data cube uses the plant individual number as the first index, it allows direct access to complete data slices of that plant across all growth cycles. When the growth cycle number is used as the second index, it allows direct access to all data types of all plant individuals within that growth cycle. When the data type is used as the third index, it allows direct access to all plant individuals and a specific data type across all growth cycles.

[0029] In one embodiment of the present invention, all data slices of the same plant individual number under different growth cycle numbers are extracted from the three-dimensional data cube of the original germplasm dataset. Edge detection is performed on the field image data in each data slice to segment the outer contour and leaf vein skeleton of a single plant individual. The outer contour and leaf vein skeleton are projected onto a unified spatial coordinate system to eliminate the positional offset of different image acquisition devices. Temporal interpolation is performed on the environmental sensor data and soil nutrient data in each data slice to align the sensor data sampling time points between different growth cycles to the same time reference. The aligned outer contour, leaf vein skeleton, environmental sensor data and soil nutrient data are stacked sequentially according to the time order of the growth cycle. The stacked data structure is defined as a multidimensional tensor with the growth cycle as the time axis, the spatial coordinate as the spatial axis and the sensor parameters as the feature axis as the multidimensional growth trajectory tensor.

[0030] In practice, all data slices of the same plant individual number under different growth cycle numbers are extracted from the three-dimensional data cube of the original germplasm dataset. Edge detection is performed on the field image data in each data slice to segment the outer contour and leaf vein skeleton of a single plant individual. The edge detection uses the Canny operator to project the outer contour and leaf vein skeleton onto a unified spatial coordinate system to eliminate the positional offset of different image acquisition devices. The projection transformation formula is as follows:

[0031] in: This represents the three-dimensional coordinates of pixels on the outer contour or leaf vein skeleton in a unified spatial coordinate system. This represents the three-dimensional coordinates of the same pixel in the camera coordinate system of the image acquisition device. This represents the rotation matrix from the camera coordinate system to the unified spatial coordinate system. This represents the translation vector from the camera coordinate system to a unified spatial coordinate system. Temporal interpolation is performed on the environmental sensor data and soil nutrient data in each data slice to align the sampling time points of the sensor data across different growth cycles to the same time reference. This time reference is based on midnight on the day of sowing. A cubic spline interpolation function is used to resample the data along the time axis at a density of one time point per day. The aligned outer contour, leaf vein skeleton, environmental sensor data, and soil nutrient data are then stacked sequentially according to the time order of the growth cycle. The stacked data structure is defined as a multidimensional tensor with the growth cycle as the time axis, the spatial coordinates as the spatial axis, and the sensor parameters as the feature axes, serving as a multidimensional growth trajectory tensor. The expression is ,in For the time axis index of the growth cycle, , , For spatial coordinate axes, For the characteristic axis index of the sensor parameters, Indicates the outer contour value. Indicates the leaf vein skeleton value. This indicates the sensor parameter values.

[0032] In some embodiments, the dual-threshold method in edge detection sets a low threshold of 50 and a high threshold of 150. Pixels with gradient magnitudes higher than the high threshold are marked as strong edge points. Pixels with gradient magnitudes between the low and high thresholds are only retained as edge points if they are connected to strong edge points. The resulting closed outer contour curve is simplified using the Douglas-Puk algorithm, retaining inflection points and deleting collinear intermediate points. In some embodiments, when the sensor data sampling time points between different growth cycles are aligned to the same time base, the start and end times of each growth cycle are determined by the accumulated temperature model. Within each growth cycle, the entire growth cycle is divided into equal intervals of days, with the start time as day 0 and the end time as the last day. The sampling time point for each day is set to 12 noon, and the sensor data values ​​at that time are calculated using a cubic spline interpolation function.

[0033] Optionally, after the outer contour and leaf vein skeleton are projected onto a unified spatial coordinate system, for the outer contour point clouds of the same plant individual acquired from different image acquisition devices, an iterative nearest-point algorithm is used for registration and fusion. This process is iterated until the root mean square error is less than 0.001 meters, at which point the fused outer contour point cloud density is no less than 100 points per square centimeter. Optionally, the cubic spline interpolation function used in the temporal interpolation operation constructs a cubic polynomial over the interval between two adjacent original sampling points. The boundary conditions are set to natural boundary conditions, i.e., the second derivative at the endpoints is zero. The interpolated data sequence is completely consistent with the original data sequence at the original sampling points.

[0034] It can be understood that the length of the time axis in the multidimensional growth trajectory tensor is equal to the total number of growth cycles, the three dimensions of the spatial axis correspond to the X-axis, Y-axis and Z-axis in a unified spatial coordinate system, and the length of the feature axis is equal to the total number of sensor parameters in the environmental sensor data and soil nutrient data.

[0035] It is understandable that when the aligned outer contour, leaf vein skeleton, environmental sensor data and soil nutrient data are stacked sequentially according to the time order of the growth cycle, the time interval between different growth cycles in the stacked data structure is not uniform, and each position on the time axis directly uses the start time of the corresponding growth cycle as the timestamp.

[0036] In one embodiment of the present invention, see [reference] Figure 3For the outer contour in the multidimensional growth trajectory tensor, the projected area, perimeter, and aspect ratio of the outer contour are calculated for each growth cycle. These calculated values ​​are then arranged in chronological order according to the growth cycle to form a morphological feature vector sequence, which serves as the morphological feature temporal sequence. For the leaf vein skeleton in the multidimensional growth trajectory tensor, the total vein length, number of vein branch nodes, and average vein curvature are extracted for each growth cycle. These extracted values ​​are then appended to the corresponding positions in the morphological feature temporal sequence. The average vein curvature is calculated using the leaf vein skeleton... The average value of the curvature value of the Frenet frame at each pixel is obtained. For the environmental sensor data and soil nutrient data in the multidimensional growth trajectory tensor, the cumulative value of air temperature, the average value of air humidity, the cumulative value of light intensity, the average value of nitrogen content, the average value of phosphorus content, the average value of potassium content, and the average value of organic matter content are calculated for each growth cycle. The calculated cumulative values ​​of air temperature, the average value of air humidity, the cumulative value of light intensity, the average value of nitrogen content, the average value of phosphorus content, the average value of potassium content, and the average value of organic matter content are arranged in the time order of the growth cycle to form a physiological feature vector sequence as the physiological feature time sequence.

[0037] In specific implementation, the projected area, perimeter, and aspect ratio of the outer contour in the multidimensional growth trajectory tensor are calculated for each growth cycle. The outer contour is a closed curve extracted from the multidimensional growth trajectory tensor. The projected area is obtained by projecting the outer contour curve onto the XOY plane of a unified spatial coordinate system and multiplying the number of pixels in the projected area by the actual area corresponding to each pixel. The perimeter is obtained by summing the Euclidean distances between adjacent pixels on the outer contour curve. The aspect ratio is obtained by dividing the length of the smallest bounding rectangle of the outer contour by its width. The calculated projected area, perimeter, and aspect ratio are arranged in chronological order according to the growth cycle to form a morphological feature vector sequence as a morphological feature temporal sequence. Each growth cycle corresponds to a three-dimensional morphological feature vector, and the five three-dimensional morphological feature vectors corresponding to five growth cycles are concatenated in chronological order to form a sequence of length 15.

[0038] In specific implementation, the total length of the leaf veins, the number of leaf vein branch nodes, and the average curvature of the leaf veins are extracted from the leaf vein skeleton in the multidimensional growth trajectory tensor according to each growth cycle. The leaf vein skeleton is a set of connected lines with a width of a single pixel extracted from the interior of the outer contour. The total length of the leaf veins is obtained by summing the number of pixels of all connected lines on the leaf vein skeleton and multiplying it by the actual length corresponding to each pixel. The number of leaf vein branch nodes is obtained by detecting the number of pixels on the leaf vein skeleton whose neighboring pixels are greater than 2. The average curvature of the leaf veins is obtained by calculating the curvature value of the Frenet frame at each pixel in the leaf vein skeleton and taking the average value. The curvature value calculation formula is:

[0039] in: Indicates the arc length parameter on the leaf vein skeleton The curvature value at that point, The three-dimensional vector parametric equation representing the leaf vein skeleton curve. Indicates the curve with respect to arc length The first derivative vector is the tangent vector. Indicates the curve with respect to arc length The second derivative vector is the curvature vector. This represents the cross product operation of vectors. The Euclidean norm of the vector is used to append the extracted total vein length, number of vein branch nodes, and average vein curvature to the corresponding positions in the morphological feature time sequence. The appending operation concatenates the three values ​​of total vein length, number of vein branch nodes, and average vein curvature corresponding to each growth cycle to the end of the morphological feature vector of that growth cycle, so that the morphological feature vector of each growth cycle is expanded from three dimensions to six dimensions. The five growth cycles form an extended morphological feature time sequence with a length of 30.

[0040] In specific implementation, the environmental sensor data and soil nutrient data in the multidimensional growth trajectory tensor are used to calculate the cumulative air temperature, average air humidity, cumulative light intensity, average nitrogen content, average phosphorus content, average potassium content, and average organic matter content for each growth cycle. The environmental sensor data includes the air temperature, air humidity, and light intensity values ​​recorded daily within each growth cycle. The cumulative air temperature is the sum of the air temperature values ​​for all days within that growth cycle. The average air humidity is the arithmetic mean of the air humidity values ​​for all days within that growth cycle. The cumulative light intensity is the sum of the light intensity values ​​for all days within that growth cycle. The soil nutrient data includes the cumulative values ​​for each growth cycle. The nitrogen, phosphorus, potassium, and organic matter content values ​​recorded daily over a long period are used. The average nitrogen content is the arithmetic mean of the nitrogen content values ​​for all days within that growth cycle. The average phosphorus, potassium, and organic matter content values ​​are calculated in the same way. The calculated cumulative air temperature, average air humidity, cumulative light intensity, average nitrogen, average phosphorus, average potassium, and average organic matter content values ​​are arranged in chronological order according to the growth cycle to form a physiological feature vector sequence, which is used as the physiological feature time sequence. Each growth cycle corresponds to a seven-dimensional physiological feature vector. The five seven-dimensional physiological feature vectors corresponding to the five growth cycles are concatenated in chronological order to form a sequence of length 35.

[0041] In some embodiments, the minimum bounding rectangle of the outer contour is obtained by a rotating caliper algorithm. The specific steps are to calculate the convex hull of the outer contour, then rotate the coordinate system for each edge of the convex hull so that the edge is parallel to the coordinate axis and calculate the area of ​​the bounding rectangle at this time. After traversing all edges, the bounding rectangle with the smallest area is selected as the minimum bounding rectangle. Its length is defined as the length of the longer side of the bounding rectangle, and its width is defined as the length of the shorter side of the bounding rectangle.

[0042] In some embodiments, during the detection of the number of leaf vein branch nodes, each pixel on the leaf vein skeleton is traversed and the number of skeleton pixels in its eight neighborhoods is counted. If the number of skeleton pixels in the neighborhood is greater than 2, the pixel is marked as a branch node. If the number of skeleton pixels in the neighborhood is equal to 1, the pixel is marked as an endpoint. If the number of skeleton pixels in the neighborhood is equal to 2, the pixel is marked as a normal connection point. The total number of pixels marked as branch nodes is the number of leaf vein branch nodes.

[0043] Optionally, before calculating the Frenet frame curvature value at each pixel on the leaf vein skeleton, a cubic B-spline fit is performed on the leaf vein skeleton curve to obtain a continuously differentiable parameter curve. Cubic B-spline fitting uses uniform node vectors, and the fitted curve exhibits second-order continuity at the nodes. The first and second derivatives are then calculated using numerical differencing. The approximate formula for the derivative is: and ,in Take the average arc length interval between adjacent pixels.

[0044] Optionally, the data for each growth cycle in the morphological feature time series and the physiological feature time series are normalized, and the normalization formula is as follows: ,in These are the original eigenvalues. This represents the minimum value for this feature dimension across all individual plants. The normalization process is performed independently for the morphological and physiological feature time series, representing the maximum value for all individual plants in this feature dimension. It can be understood that the projected area, perimeter, and aspect ratio of the outer contour are calculated for the same individual plant number in each growth cycle. The calculation results for different individual plant numbers are independent of each other, and the calculation results for the same individual plant number across different growth cycles are strictly arranged in chronological order, consistent with the order of growth cycle numbers G1 to G5.

[0045] It is understood that when calculating the cumulative air temperature, average air humidity, cumulative light intensity, average nitrogen content, average phosphorus content, average potassium content, and average organic matter content, the number of days in each growth cycle is determined by the accumulated temperature model. The number of days varies for different growth cycles, and the cumulative air temperature is calculated by directly using the original summation without dividing by the number of days.

[0046] In one embodiment of the present invention, a fusion feature matrix is ​​formed by concatenating morphological feature time series and physiological feature time series along the feature dimension. The fusion feature matrix is ​​input into the generator of a generative adversarial network. The generator of the generative adversarial network performs multi-layer deconvolution operations on the fusion feature matrix to expand the feature dimension and time dimension. After each deconvolution operation, batch normalization and activation function mapping operations are performed on the intermediate feature map output by the layer. Spatial domain smoothing filtering is performed on the feature map output by the last deconvolution operation to eliminate high-frequency noise artifacts. The spatial domain smoothing filtering adopts an anisotropic Gaussian filter, and the standard deviation of its filter kernel is adaptively adjusted along the time dimension of the feature field. The smoothed feature map is output as an enhanced growth feature field.

[0047] In the specific implementation, the morphological feature time series and the physiological feature time series are concatenated along the feature dimension to form a fusion feature matrix. The morphological feature time series is an extended morphological feature time series of length 30, and the physiological feature time series is a physiological feature vector sequence of length 35. The dimension of the fusion feature matrix formed after concatenation along the feature dimension is the number of growth cycles multiplied by the total number of features in a single cycle, i.e., 5 rows multiplied by 65 columns. Here, 5 represents the five growth cycles: seedling stage, jointing stage, tasseling stage, silking stage, and grain filling stage, and 65 represents the sum of the number of morphological features (30) and the number of physiological features (35) corresponding to each growth cycle. The fusion feature matrix is ​​input into the generator of the generative adversarial network. The generative adversarial network is pre-trained using labeled data collected in the same maize planting area in historical years. The generator adopts a deep convolutional generative adversarial network structure and uses the fusion feature matrix as a conditional input variable.

[0048] In practical implementation, the generator of the generative adversarial network performs multiple deconvolution operations on the fused feature matrix to expand the feature dimension and time dimension. This deconvolution operation is also called a transposed convolution operation. The first deconvolution layer receives the input fused feature matrix and expands the number of feature channels from 1 to 512, while simultaneously expanding the length of the time dimension from 5 to 10. The second deconvolution layer reduces the number of feature channels from 512 to 256 and expands the length of the time dimension from 10 to 20. The third deconvolution layer reduces the number of feature channels from 256 to 128 and expands the length of the time dimension from 20 to 40. The fourth deconvolution layer reduces the number of feature channels from 128 to 40. The number of feature channels is reduced to 64, and the time dimension is expanded from 40 to 80. The fifth deconvolution layer reduces the number of feature channels from 64 to 32 and expands the time dimension from 80 to 160. Each deconvolution layer uses a 4×4 kernel with a stride of 2 and uses the same padding method to ensure that the output size is exactly twice the input size. After each deconvolution operation, batch normalization and activation function mapping are performed on the intermediate feature maps output by that layer. The batch normalization operation is performed independently on each feature channel. After calculating the mean and variance of the intermediate feature maps in the current batch, each element is normalized. The normalization formula is as follows: ,in For the input values ​​of the batch normalization operation, This represents the average value of the current batch across the corresponding channel. This represents the variance of the current batch in the corresponding channel. The value is a numerical stability constant. The normalized value is transformed by a learnable scaling parameter and a translation parameter. The activation function mapping operation uses a linear rectified function with leakage and a negative half-axis slope of 0.2.

[0049] In practice, spatial domain smoothing filtering is performed on the feature map output by the last deconvolution operation to eliminate high-frequency noise artifacts. The spatial domain smoothing filtering uses an anisotropic Gaussian filter, and the filter kernel expression is as follows:

[0050] in: This represents the spatial offset of the anisotropic Gaussian filter. , , and time offset The filter kernel weights at that location, , , These represent the offsets in three spatial directions relative to the center of the filter kernel in the feature map spatial coordinate system. This represents the time offset of the feature map relative to the center of the filter kernel along the time axis. , , These represent the standard deviations of the filter in the three spatial directions, all of which are fixed at 1.2. The standard deviation of the filter in the time direction is adaptively adjusted along the time dimension of the feature field. ,in The baseline time standard deviation is set to 2.0. The adaptive coefficient is set to 0.1. The absolute value of the time offset is represented by the spatial window size of the filter kernel, which is 5×5×5. The temporal window size adapts to the temporal position. For each position in the feature map, the values ​​of all pixels in the spatial and temporal neighborhoods centered on that position are multiplied by the corresponding filter kernel weights and then summed to obtain the smoothed value at that position. The smoothed feature map is output as an enhanced growing feature field. The dimension of the enhanced growing feature field is 160×32, where 160 represents the length of the expanded temporal dimension and 32 represents the number of feature channels.

[0051] In some embodiments, the numerical stability constant in the batch normalization operation The value is 10 to the power of -5. The initial value of the learnable scaling parameter is 1, and the initial value of the learnable translation parameter is 0. The batch normalization operation uses the mean and variance of the current batch data during the training phase and the global mean and global variance recorded during the training phase during the inference phase.

[0052] In some embodiments, the spatial direction standard deviation of the anisotropic Gaussian filter , , The fixed value is determined based on the spatial pixel spacing in the feature map. When the spatial pixel spacing is 0.5 cm per pixel, the standard deviation of 1.2 corresponds to a spatial filtering range of a spherical region with a diameter of approximately 2.4 cm. The time-direction adaptive coefficient... Setting it to 0.1 ensures that the filter weights decay faster at time points farther from the center, thus preserving the temporal dynamics of the growth trajectory. Optionally, the generator of the generative adversarial network adds a random deactivation operation after each deconvolution operation and before the batch normalization operation, with a random deactivation ratio of 0.3. That is, during each forward propagation, the output value of neurons in the intermediate feature map is set to zero with a 30% probability, thereby enhancing the generator's generalization ability.

[0053] Optionally, the leaky linear rectifier function is defined as follows: ,in The input value is 0.2, and the slope of the negative half-axis is 0.2. This function outputs a small non-zero value instead of directly outputting zero when the input value is negative. It can be understood that the fused feature matrix maintains the temporal order of the growth cycle when concatenating along the feature dimensions. That is, the first row corresponds to all feature values ​​of the seedling stage, the second row to all feature values ​​of the jointing stage, the third row to all feature values ​​of the tasseling stage, the fourth row to all feature values ​​of the silking stage, and the fifth row to all feature values ​​of the grain-filling stage. Within each row, the morphological feature time sequence comes first, followed by the physiological feature time sequence.

[0054] It can be understood that the sum of the filter kernel weights of the anisotropic Gaussian filter is normalized to 1, that is, the sum of the filter kernel weight values ​​at all offset positions is equal to 1, to ensure that the filtering operation does not change the overall brightness range and time standard deviation of the feature map. Follow The increase in size makes the filtering effect smoother for time points far from the center.

[0055] In one embodiment of the present invention, genotype tag data corresponding to each plant individual number is extracted from the original germplasm dataset. The genotype tag data includes the single nucleotide polymorphism (SNP) coding sequence and molecular marker banding coding sequence of that plant individual. The enhanced growth feature field is divided into feature sub-blocks of the same number as the genotype tag data according to the plant individual number. Average pooling is performed on each feature sub-block to compress it into a fixed-length feature vector. Each fixed-length feature vector is concatenated with its corresponding SNP coding sequence and molecular marker banding coding sequence. All concatenated vectors are stacked according to the plant individual number to form a two-dimensional matrix as the enhanced germplasm feature matrix carrying genotype information. To enhance the germplasm feature matrix, principal component dimensionality reduction is performed to extract principal component feature vectors whose cumulative contribution rate exceeds a preset threshold. The extracted principal component feature vectors are projected into a low-dimensional space to obtain the low-dimensional embedding points of each plant individual. A preset number of candidate cluster centers are initialized, and an iterative optimization algorithm is used to continuously update the position of each candidate cluster center. After each iteration, each low-dimensional embedding point is assigned to the cluster to which the nearest candidate cluster center belongs. When the rate of change of the position of the candidate cluster center is lower than a preset convergence threshold, the iteration stops, and the cluster division result at this time is taken as the final cluster division. A unique numerical cluster identifier is assigned to each final cluster, and a mapping table between the cluster identifier and the individual plant numbers in the cluster is established.

[0056] In the specific implementation, genotype tag data corresponding to each plant individual number is extracted from the original germplasm dataset. The genotype tag data includes the single nucleotide polymorphism (SNP) coding sequence and the molecular marker banding coding sequence for that plant individual. The SNP coding sequence has a length of 2048 sites, with each site having a value of 0, 1, or 2. The molecular marker banding coding sequence has a length of 256 sites, with each site having a value of 0 or 1. The enhanced growth feature field is divided into feature sub-blocks of the same number as the genotype tag data according to the plant individual number. The dimension of the enhanced growth feature field is the number of plant individuals multiplied by the time dimension length of 160 multiplied by the number of feature channels of 32. After segmentation, each plant individual number corresponds to a three-dimensional feature sub-block with a dimension of 160 multiplied by 32. Average pooling is performed on each feature sub-block to compress it into a fixed-length feature vector. The average pooling formula is:

[0057] in: This represents the eigenvector after average pooling. The values ​​of each feature channel This indicates that the total length of the feature sub-block in the time dimension is 160. Indicates the index number in the time dimension. The time index in the feature sub-block is Feature channel index is The element values ​​are used to concatenate each fixed-length feature vector with the corresponding single nucleotide polymorphism site coding sequence and molecular marker banding coding sequence. After concatenation, a new vector of length 2336 is formed. All concatenated vectors are stacked into a two-dimensional matrix according to the plant individual number as an enhanced germplasm feature matrix carrying genotype information.

[0058] In specific implementation, principal component dimensionality reduction is performed on the enhanced germplasm feature matrix to extract principal component feature vectors with a cumulative contribution rate exceeding a preset threshold, which is set to 85%. The extracted principal component feature vectors are projected into a low-dimensional space to obtain the low-dimensional embedding points of each plant individual. A preset number of candidate cluster centers are initialized, and an iterative optimization algorithm is used to continuously update the position of each candidate cluster center. After each iteration, each low-dimensional embedding point is assigned to the cluster to which the nearest candidate cluster center belongs. When the rate of change of the position of the candidate cluster center is lower than a preset convergence threshold, the iteration stops, and the cluster division result at this time is taken as the final cluster division, which is set to 0.001. A unique numerical cluster identifier is assigned to each final cluster, and a mapping table between the cluster identifier and the individual plant numbers within the cluster is established.

[0059] In some embodiments, the preset number of candidate cluster centers is determined by the elbow rule, that is, clustering is performed in the range of K values ​​from 2 to 15 and the silhouette coefficient corresponding to each K value is calculated. The K value with the largest silhouette coefficient is selected as the preset number. The candidate cluster centers are initialized using the K-means++ algorithm. The first candidate cluster center is randomly selected from the low-dimensional embedding points. Each subsequent candidate cluster center is selected from the remaining low-dimensional embedding points with a probability proportional to the square of the distance between it and the existing candidate cluster centers.

[0060] In some embodiments, the iterative optimization algorithm is a K-means clustering algorithm. When updating the position of the candidate cluster center in each iteration, for each cluster, the arithmetic mean of the coordinates of all low-dimensional embedding points in each dimension is calculated as the new candidate cluster center coordinates of that cluster. The position change rate is defined as the ratio of the average position offset of all candidate cluster centers after this iteration relative to the previous iteration to the average position magnitude of the candidate cluster centers in the previous iteration.

[0061] Optionally, before the principal component dimensionality reduction operation, standard deviation normalization is performed on each column of the enhanced germplasm feature matrix, and the normalization formula is as follows: ,in The first element in the original matrix Line number The element values ​​of the column, For the first The arithmetic mean of the column, For the first The standard deviation of the column. Optionally, the average pooling operation is performed in the time dimension. For a feature sub-block with a dimension of 160 x 32, the arithmetic mean of the values ​​at all time positions in each feature channel is calculated along the time dimension to obtain a feature vector of length 32. It can be understood that each row in the enhanced germplasm feature matrix carrying genotype information corresponds to a plant individual number, and the order of the feature vectors within a row is: feature vector first, single nucleotide polymorphism site coding sequence in the middle, and molecular marker banding coding sequence last.

[0062] It can be understood that each group identifier in the mapping table corresponds to only one group, and the union of the set of plant individual numbers corresponding to all group identifiers is equal to the complete set of all plant individual numbers.

[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent management of maize plant germplasm resources, characterized in that, include: Field image data, environmental sensor data, and soil nutrient data of maize plants at multiple growth stages were collected as the original germplasm dataset; Perform spatiotemporal alignment operations on the original germplasm dataset to establish a multidimensional growth trajectory tensor for each individual maize plant; The temporal sequences of morphological features and physiological features of each individual plant are extracted from the multidimensional growth trajectory tensor. The morphological feature time sequence and the physiological feature time sequence are input into a pre-trained generative adversarial network and the generator outputs an enhanced growth feature field. The enhanced growth feature field is associated and encoded with the genotype tag data in the original germplasm dataset to generate an enhanced germplasm feature matrix carrying genotype information. Cluster analysis is performed on the enhanced germplasm feature matrix to divide different germplasm resource groups and assign a unique group identifier to each group.

2. The intelligent management method for maize plant germplasm resources according to claim 1, characterized in that, Field image data, environmental sensor data, and soil nutrient data of maize plants at multiple growth stages were collected as the original germplasm dataset, specifically including: Multiple image acquisition devices, multiple environmental sensors, and multiple soil nutrient sensors were deployed in a fixed spatial grid within the target corn planting area. When each preset growth cycle is reached, the image acquisition device is triggered to synchronously acquire multi-angle canopy images of each individual plant as the field image data; The air temperature, air humidity, and light intensity at each grid point are continuously recorded as environmental sensor data in a continuous time series. When each irrigation event is triggered, the soil nutrient sensor collects the nitrogen, phosphorus, potassium and organic matter content of soil layers at different depths at each grid point as the soil nutrient data; Field image data, environmental sensor data, and soil nutrient data collected at the same grid point and during the same growth cycle are associated and bound according to the individual plant number; All the associated and bound data is organized into a three-dimensional data cube with plant individual number as the first index, growth cycle number as the second index, and data type as the third index, which serves as the original germplasm dataset.

3. The intelligent management method for maize plant germplasm resources according to claim 2, characterized in that, Perform spatiotemporal alignment operations on the original germplasm dataset to establish a multidimensional growth trajectory tensor for each individual maize plant, specifically including: Extract all data slices of the same plant individual number under different growth cycle numbers from the three-dimensional data cube of the original germplasm dataset; Edge detection was performed on the field image data in each data slice to segment the outer contour and leaf vein skeleton of individual plants; The outer contour and the leaf vein skeleton are projected onto a unified spatial coordinate system to eliminate positional offsets between different image acquisition devices; Perform temporal interpolation on the environmental sensor data and soil nutrient data in each data slice to align the sensor data sampling time points between different growth cycles to the same time reference; The aligned outer contour, leaf vein skeleton, environmental sensor data, and soil nutrient data are stacked sequentially according to the time order of the growth cycle. The stacked data structure is defined as a multidimensional tensor with the growth cycle as the time axis, the spatial coordinates as the spatial axis, and the sensor parameters as the feature axis, which is called the multidimensional growth trajectory tensor.

4. The intelligent management method for maize plant germplasm resources according to claim 3, characterized in that, The temporal sequences of morphological features and physiological features of each individual plant are extracted from the multidimensional growth trajectory tensor, specifically including: For the outer contour in the multidimensional growth trajectory tensor, calculate the projected area, perimeter, and aspect ratio of the outer contour for each growth cycle. The calculated projected area, perimeter, and aspect ratio values ​​are arranged into a morphological feature vector sequence according to the time sequence of the growth cycle, which is used as the morphological feature time sequence. For the leaf vein skeleton in the multidimensional growth trajectory tensor, the total length of leaf veins, the number of leaf vein branch nodes, and the average curvature of leaf veins are extracted for each growth cycle. The extracted total vein length, number of vein branch nodes, and average vein curvature are appended to the corresponding positions in the morphological feature time sequence. For the environmental sensor data and soil nutrient data in the multidimensional growth trajectory tensor, calculate the cumulative value of air temperature, the average value of air humidity, the cumulative value of light intensity, the average value of nitrogen content, the average value of phosphorus content, the average value of potassium content, and the average value of organic matter content for each growth cycle. The calculated cumulative values ​​of air temperature, average air humidity, cumulative light intensity, average nitrogen content, average phosphorus content, average potassium content, and average organic matter content are arranged in chronological order according to the growth cycle to form a physiological feature vector sequence, which is used as the physiological feature time sequence.

5. The intelligent management method for maize plant germplasm resources according to claim 4, characterized in that, The average curvature of the leaf veins is obtained by averaging the curvature values ​​of the Frenet frame at each pixel in the leaf vein skeleton.

6. The intelligent management method for maize plant germplasm resources according to claim 1, characterized in that, The morphological feature time sequence and the physiological feature time sequence are input into a pre-trained generative adversarial network, and the generator outputs an enhanced growth feature field, specifically including: The morphological feature time sequence and the physiological feature time sequence are concatenated along the feature dimension to form a fusion feature matrix; The fused feature matrix is ​​input into the generator of the generative adversarial network; The generator of the generative adversarial network performs multi-layer deconvolution operations on the fused feature matrix to expand the feature dimension and the temporal dimension; After each deconvolution operation, batch normalization and activation function mapping are performed on the intermediate feature map output by that layer. Spatial domain smoothing filtering is performed on the feature map output from the last deconvolution operation to eliminate high-frequency noise artifacts; The smoothed and filtered feature map is output as the enhanced growth feature field.

7. The intelligent management method for maize plant germplasm resources according to claim 6, characterized in that, The spatial domain smoothing filter employs an anisotropic Gaussian filter, whose standard deviation of the filter kernel is adaptively adjusted along the time dimension of the feature field.

8. The intelligent management method for maize plant germplasm resources according to claim 6, characterized in that, The enhanced growth feature field is associated and encoded with the genotype tag data in the original germplasm dataset to generate an enhanced germplasm feature matrix carrying genotype information, specifically including: Extract the genotype tag data corresponding to the individual plant number from the original germplasm dataset; The genotype tag data includes the coding sequence of the single nucleotide polymorphism site and the coding sequence of the molecular marker banding for the individual plant. The enhanced growth feature field is divided into feature sub-blocks of the same number as the genotype tag data according to the plant individual number; Perform average pooling on each feature sub-block to compress the feature sub-block into a fixed-length feature vector; Each fixed-length feature vector is concatenated with the corresponding single nucleotide polymorphism site coding sequence and molecular marker banding coding sequence. All the spliced ​​vectors are stacked into a two-dimensional matrix according to the individual plant number, which serves as the enhanced germplasm feature matrix carrying genotype information.

9. The intelligent management method for maize plant germplasm resources according to claim 8, characterized in that, Cluster analysis is performed on the enhanced germplasm feature matrix to classify different germplasm resource groups and assign a unique group identifier to each group, specifically including: Perform principal component dimensionality reduction on the enhanced germplasm feature matrix to extract principal component feature vectors whose cumulative contribution rate exceeds a preset threshold; The extracted principal component feature vectors are projected into a low-dimensional space to obtain the low-dimensional embedding points of each plant individual. Initialize a preset number of candidate cluster centers and continuously update the position of each candidate cluster center using an iterative optimization algorithm; After each iteration update, each low-dimensional embedding point is assigned to the cluster of the nearest candidate cluster center; When the rate of change of the candidate cluster center is lower than the preset convergence threshold, the iteration stops and the cluster division result at this time is taken as the final cluster division. Assign a unique numerical group identifier to each final group and establish a mapping table between the group identifier and the individual plant numbers within the group.

10. A smart management system for maize plant germplasm resources, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent management method for maize plant germplasm resources as described in any one of claims 1 to 9.

Images