Dense and reverse-resistant planting breeding strategy optimization system and corn breeding method

Abstract

This invention relates to the field of planting and breeding technology, specifically to an optimization system for planting and breeding strategies that promotes high density and stress tolerance, and a method for maize breeding. In this invention, by running a support vector machine to compare proline concentrations and eliminate low-value entries, the limitations of traditional empirical breeding sample quantity are overcome, enabling precise solution of the targeted gene array for extreme stress resistance potential. By calling the array node coordinates and merging preset ear height and leaf angle parameters into the state tensor, a graph attention network is introduced to calculate the attention weight coefficients between multiple dimensional parameters, completing precise modeling of spatial mutual occlusion and competition constraints among multiple plant individuals under extreme ecological niches, maintaining the peak photosynthetic yield of the population at full environmental density. By analyzing the stomatal conductance sequence within the array to extract the inverse gradient multiplied by the negative penalty coefficient, the shared stress-resistance biological abstraction characteristics across ecological niches are extracted, enhancing the robustness of phenotypic prediction under extreme habitats, and transforming biological pollination risk into a strict mathematical convergence constraint.

Description

Technical Field

[0001] This invention relates to the field of planting and breeding technology, and in particular to a planting and breeding strategy optimization system and a maize breeding method that are tolerant to dense planting and stress. Background Technology

[0002] The field of planting and breeding technology specifically covers plant germplasm resource evaluation, genetic variation analysis, and superior allele aggregation technology. Its actual operation logic is to use multivariate statistical models to analyze the interaction effect between crop genetic variance and environmental variance, solve the narrow heritability parameter, and then screen parental lines with high yield potential and resistance to biological stress. The core of the field is to transform traditional empirical breeding into quantitative engineering based on probability theory and matrix operations through quantitative genetics.

[0003] The optimization system for planting and breeding strategies to tolerate high density and stress, and the specific breeding method for maize, are based on a breeding decision support architecture constructed using a multi-objective constrained mathematical model. The core objectives include, for the first time, analyzing the correlation matrix between the coefficient of variation of the canopy leaf angle and the photosynthetically active radiation interception rate of maize, thereby determining the spatial topology of the high-density plant population, and including secondary terms to quantify the critical threshold of plant osmotic regulatory protein concentration under drought and saline-alkali stress. To achieve the set effect, the scheme aims to output a comprehensive set of operating instructions, including the ratio of male and female parents, the plant row spacing, and the amount of water and fertilizer applied, so that multiple offspring populations can, under the preset full-load environmental density conditions, increase the critical wind speed for root soil stabilization and lodging resistance to the target disaster prevention standard line, and compress the flowering and silking interval time window to within the optimal pollination span, thereby maintaining the photosynthetic yield of the population under extreme ecological niches.

[0004] Existing technologies rely on multivariate statistical models to analyze the interaction effects of crop genetic variance and environmental variance to solve for narrow heritability parameters. Their operational logic heavily depends on linear regression of historical phenotypic data and macroscopic probability matrix operations. Conventional multivariate statistical models presuppose that genotype and environmental effects are independent and exhibit a simple linear additive relationship, making it difficult to capture the microscopic epistatic interactions of multiple alleles within the whole genome and the nonlinear biological mutation characteristics under extreme habitats. Macroscopic quantitative genetic methods often focus on population-average genetic gain, neglecting the microenvironmental competition among individual plants and the transient extreme fluctuations in osmotic regulatory protein concentrations. This results in the parameter prediction generalization ability being fixed in a single experimental habitat, exhibiting severe distortion when facing complex stress scenarios. High-yielding parental lines selected based on narrow genetic parameters of traditional statistical models, when encountering a dual complex stress habitat of sudden drought and extreme dense planting, fail to accurately quantify the microscopic processes of dynamic closure gradient of stomatal conductance and nonlinear decay of photosynthetically effective radiation. This leads to a sharp increase in the time window between flowering and silking of individual plants, causing pollination failure of female ears and large-scale empty stalks and broken ears. Consequently, the photosynthetic yield of the population drops precipitously under full-planting density, completely losing its disaster prevention and stable yield efficacy under extreme ecological niches. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose an optimized system for planting and breeding strategies that are tolerant to dense planting and stress, as well as a method for breeding maize.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a planting and breeding strategy optimization system for dense planting and stress tolerance includes: Genetic effect topology module: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds. After establishing a feature matrix, heterozygous genotype coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate an epistatic interaction weight tensor. Germplasm evolution optimization module: Based on the aforementioned epistatic interaction weight tensor, extreme value coordinates are extracted and preset disaster prevention values ​​are merged, ventilation tissue sequences are projected, latent variables are extracted by superimposing random numbers, nucleotide matrix is ​​reconstructed, and low-value entries are eliminated by comparing concentrations using a support vector machine to obtain a stress-resistant target gene array. Habitat confrontation game module: Based on the stress resistance targeted gene array, analyze the stomatal conductance sequence within the array, retrieve the preset drought stress parameters and extract the reverse gradient multiplied by the negative penalty coefficient, fuse the sequence with the leaf tensor, align the spatial coordinates and merge the overlapping values ​​to obtain the cross-domain shared feature set. Pollination survival prediction module: Based on the stress-resistant target gene array and the cross-domain shared feature set, merge the set values ​​and compare the flowering and silking interval time window, extract the absolute value of the difference and accumulate the weight matrix, calculate the pollen dispersal penalty value and add it to the iteration term, remove the coordinates exceeding the threshold, and construct the habitat tolerance expression array. Spatial layout decision module: Based on the habitat tolerance expression array, after calling the array node coordinates, the preset ear height and leaf angle parameters are merged, the state tensor is filled in, the distribution probability curve is generated through the graph attention network, the extreme value coordinates of the curve and the parent ratio are extracted, and the population adaptation instruction set is output.

[0007] As a further aspect of the present invention, the superordinate interaction weight tensor includes wavelength values, hybrid coding vectors within the feature matrix, and overlapping terms corresponding to truncation operations; the stress-resistant targeted gene array includes preset disaster prevention scalar values, root cortex aeration tissue sequences, reconstructed nucleotide matrices, and surviving node feature parameters; the cross-domain shared feature set includes stomatal conductance sequences, drought stress parameters, inverse gradient penalty tensors, and the mean of overlapping region characterization values; the habitat tolerance expression array includes flowering and silking interval time window values, preset pollen shedding penalty values, a weighted deviation parameter array, and resident safe interval node parameters; and the population adaptation instruction set includes extreme coordinates of the distribution probability curve, parent ratio values, preset population planting spacing parameter scalars, and water and fertilizer application rates.

[0008] As a further aspect of the present invention, the genetic effect topology module includes: Signal mapping submodule: Based on the original sequencing optical signal matrix, extract wavelength numerical parameters, compare the numerical parameters with the pre-set base mapping scalar, assign stem feature numerical matrix to the associated node, extract the heterozygous coding vector inside the feature matrix, merge the coding vector with each element of the feature matrix, and obtain the gene coding parameter matrix; Numerical truncation submodule: Based on the gene coding parameter matrix, it calls the associated feature parameters inside the matrix, extracts the corresponding heterozygous coding vectors for each term, calculates the product of the parameters and coding vectors term by term, truncates multiple negative values ​​inside the product result, extracts the corresponding overlapping terms after each truncation operation, and generates the superordinate interaction weight tensor.

[0009] As a further aspect of the present invention, the germplasm evolution optimization module includes: Weight mapping submodule: Based on the superordinate interaction weight tensor, extract the extreme value coordinate parameters inside the tensor, merge the extreme value coordinates with the preset disaster prevention scalar values, project and map the root cortex ventilation tissue sequence, superimpose the random numerical array to extract the latent variable array parameters, and obtain the variation-driven latent variable features. The mutation reconstruction submodule: Based on the mutation-driven latent variable features, it splices the numerical elements of each item in the condition label matrix, calculates the multidimensional spatial mapping relationship of each matrix element, restores the multidimensional coordinate parameters of the mapped nucleotide sequence, extracts the numerical terms associated with sequence mutation sites, and establishes the reconstructed nucleotide matrix. The stress-resistant screening submodule extracts the numerical parameters of the associated samples based on the reconstructed nucleotide matrix and inputs them into the pre-trained support vector machine. It calls the radial basis function to calculate the geometric distance between the sample feature vector and the classification hyperplane, compares it with the preset proline concentration threshold scalar, calculates the difference between the numerical value and the scalar, removes the sequence entries with a difference lower than the scalar, and aggregates all the feature parameters of the remaining nodes to obtain the stress-resistant targeted gene array.

[0010] As a further aspect of the present invention, the support vector machine specifically involves calling the radial basis kernel function to perform high-dimensional space mapping operations, calculating the geometric distance between multiple associated sample feature vectors and a preset classification hyperplane, obtaining the spatial distribution coordinates of multiple sample feature vectors on both sides of the hyperplane, and determining the spatial polarity of the sample feature vectors relative to the classification hyperplane.

[0011] As a further aspect of the present invention, the habitat adversarial game module includes: The penalty operation submodule: Based on the stress-resistant targeted gene array, the stomatal conductance sequence is analyzed, the preset drought stress parameters are retrieved, the inverse gradient matrix values ​​are extracted and multiplied by the preset negative penalty coefficient, the product terms are calculated and assigned to weight nodes, and the inverse gradient penalty tensor is obtained. Spatial fusion submodule: Based on the inverse gradient penalty tensor, fuse the preset leaf feature tensor, align the spatial physical coordinates of each node, calculate the mean value of the overlapping region representation and replace the original coordinate parameters, extract the merged array, and obtain the cross-domain shared feature set.

[0012] As a further aspect of the present invention, the pollination survival prediction module includes: Risk-weighted submodule: Based on the stress-resistant targeted gene array and the cross-domain shared feature set, merge the associated feature values ​​of the two, compare the flowering and silking interval time window values, calculate the absolute value of the difference, accumulate the absolute value and the preset pollen shedding penalty value to obtain the weighted deviation parameter array; Retention and screening submodule: Based on the weighted deviation parameter array, extract the coordinates of the deviation nodes inside the array, compare the node coordinates with the preset warning threshold values, remove the coordinate records that exceed the warning threshold, aggregate all node parameters in the safe retention range, and obtain the habitat tolerance expression array.

[0013] As a further aspect of the present invention, the spatial layout decision module includes: State tensor construction submodule: Based on the habitat tolerance expression array, extract the coordinate parameters of the convergence nodes inside the array, retrieve the preset ear height and leaf angle parameters, merge the coordinates and the values ​​of the above parameters, load the merged values ​​into the corresponding spatial state dimension, and establish the environmental site state tensor. Probability curve mapping submodule: Based on the environmental site state tensor, extract the values ​​of multiple dimensional parameters inside the tensor, input the graph attention network to calculate the attention weight coefficients between the multiple dimensional parameters, perform the dot product operation between the weight coefficients and the dimensional parameters, map the dot product operation results to generate a probability distribution curve, parse the probability distribution curve to extract the extreme value coordinate parameters, and obtain the set of layout extreme value coordinates. Adaptation instruction output submodule: Based on the set of extreme coordinates of the layout, it parses the extreme coordinate parameters of each item in the set, extracts the parent ratio values ​​corresponding to each coordinate, retrieves the preset population planting spacing parameter scalar, merges the ratio values ​​and the planting spacing parameter scalar, and generates a population adaptation instruction set.

[0014] As a further aspect of the present invention, the graph attention network specifically maps the numerical values ​​of multiple dimensional parameters to feature vectors of graph topology nodes, retrieves a preset learnable weight matrix, multiplies it by the node feature vectors to perform a linear transformation, concatenates multiple adjacent node feature vectors and multiplies them by a preset attention parameter vector, calls a leaky linear rectified activation function to truncate the negative values ​​of the product terms, applies a normalized exponential function to perform numerical scaling, and calculates and outputs an attention weight coefficient matrix among multiple dimensional parameters.

[0015] An optimization method for planting and breeding strategies that are tolerant to high density and stress, and a method for breeding maize, wherein the optimization method for planting and breeding strategies that are tolerant to high density and stress, and a method for breeding maize, are executed based on the above-mentioned optimization system for planting and breeding strategies that are tolerant to high density and stress, and include the following steps: S1: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds to determine the stalk feature matrix assigned to maize gene loci, heterozygous coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate a superordinate interaction weight tensor. S2: Based on the superordinate interaction weight tensor, extract the extreme value coordinates and merge the disaster prevention values, project the maize aeration tissue sequence, superimpose random numbers to extract latent variables, reconstruct the nucleotide matrix, and use a support vector machine to compare concentrations and remove low-value entries to obtain the stress-resistant target gene array. S3: Based on the stress-resistant targeted gene array, analyze the maize stomatal conductance sequence, retrieve drought stress parameters to extract the reverse gradient multiplied by the negative penalty coefficient, fuse the sequence with the maize leaf tensor, align the spatial coordinates and merge the overlapping values ​​to obtain a cross-domain shared feature set. S4: Based on the stress-resistant targeted gene array and cross-domain shared feature set, the time window of maize flowering and silking interval was merged by numerical comparison, the weight matrix of the cumulative weight of the difference was extracted, the pollen penalty value was calculated and added to the iteration term, the coordinates exceeding the threshold were removed, and the habitat tolerance expression array was constructed. S5: Based on the habitat tolerance expression array, after calling the array node coordinates, merge the parameters of maize ear height and leaf angle, fill in the state tensor, generate the distribution probability curve by graph attention network, extract the extreme value coordinates and parent ratio, and output the maize population adaptation instruction set.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by running a support vector machine to compare proline concentration and remove low-value entries, the extreme values ​​of the distribution of high stress-resistant genotypes in the multidimensional potential space are accurately quantified, breaking through the limitations of the number of samples in traditional empirical breeding, and realizing the accurate solution of the extreme value stress-resistant potential targeted gene array. In this invention, the state tensor is filled by calling the array node coordinates and merging the preset ear height and leaf angle parameters. A graph attention network is introduced to calculate the attention weight coefficients between multiple dimensional parameters, so as to complete the accurate modeling of the spatial mutual occlusion and competition constraints of multiple individual plants under the extreme ecological niche and maintain the peak photosynthetic yield of the full-load environmental density population. In this invention, the reverse gradient multiplier is extracted by analyzing the stomatal conductance sequence within the array, and the overlapping values ​​of the fused sequence and leaf tensor are merged to extract the shared stress-resistance biological abstract features across ecological niches. This enhances the robustness of phenotypic prediction under extreme habitats and transforms biological pollination risks into strict mathematical convergence constraints. Attached Figure Description

[0017] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1 Please see Figure 1 The present invention provides a technical solution: a planting and breeding strategy optimization system for dense planting and stress resistance includes.

[0020] Genetic effect topology module: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds. After establishing a feature matrix, heterozygous genotype coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate an epistatic interaction weight tensor. Germplasm evolution optimization module: Based on the epistatic interaction weight tensor, extreme value coordinates are extracted and pre-set disaster prevention values ​​are merged, ventilation tissue sequences are projected, latent variables are extracted by superimposing random numbers, nucleotide matrix is ​​reconstructed, and low-value entries are eliminated by comparing concentrations through support vector machine to obtain stress resistance target gene array; Habitat confrontation game module: Based on the stress resistance targeted gene array, the stomatal conductance sequence within the array is analyzed, the preset drought stress parameters are retrieved and the negative gradient multiplied by the negative penalty coefficient is extracted, the sequence and leaf tensor are fused, the spatial coordinates are aligned and the overlapping values ​​are merged to obtain the cross-domain shared feature set. Pollination survival prediction module: Based on the stress-resistant targeted gene array and cross-domain shared feature set, the set values ​​are merged and compared with the flowering and silking interval time window. After extracting the absolute value of the difference, the weight matrix is ​​accumulated, the pollen dispersal penalty value is calculated and added to the iteration term, the coordinates exceeding the threshold are removed, and the habitat tolerance expression array is constructed. Spatial layout decision module: Based on habitat tolerance expression array, after calling array node coordinates, the preset ear height and leaf angle parameters are merged, the state tensor is filled in, the distribution probability curve is generated through graph attention network, the extreme value coordinates of the curve and the parent ratio are extracted, and the population adaptation instruction set is output.

[0021] The superordinate interaction weight tensor includes wavelength values, hybrid coding vectors within the feature matrix, and overlapping terms corresponding to truncation operations. The stress-resistant targeted gene array includes preset disaster prevention scalar values, root cortex aeration tissue sequences, reconstructed nucleotide matrices, and surviving node feature parameters. The cross-domain shared feature set includes stomatal conductance sequences, drought stress parameters, inverse gradient penalty tensors, and the mean values ​​of overlapping region characterization values. The habitat tolerance expression array includes flowering and silking interval time window values, preset pollen shedding penalty values, weighted deviation parameter arrays, and resident safe interval node parameters. The population adaptation instruction set includes extreme coordinates of the distribution probability curve, parent ratio values, preset population planting spacing parameter scalars, and water and fertilizer application rates.

[0022] The genetic effects topology module includes: Signal mapping submodule: Based on the original sequencing optical signal matrix, extract wavelength numerical parameters, compare the numerical parameters with the pre-set base mapping scalar, assign stem feature numerical matrix to the associated node, extract the heterozygous coding vector inside the feature matrix, merge the coding vector with each element of the feature matrix, and obtain the gene coding parameter matrix; Numerical truncation submodule: Based on the gene coding parameter matrix, it calls the associated feature parameters inside the matrix, extracts the corresponding heterozygous coding vectors for each term, calculates the product of the parameter and the coding vector term by term, truncates the multiple negative values ​​inside the product result, extracts the corresponding overlapping terms after each truncation operation, and generates the superordinate interaction weight tensor. Signal mapping submodule: Based on the original sequencing optical signal matrix, a multilayer perceptron feedforward neural network algorithm is adopted. The network architecture is set with 256 input layer nodes, 128 first hidden layer nodes, and 64 second hidden layer nodes. The weight tensors within the network are initialized to follow a normal distribution with a mean of 0 and a variance of 0.05. The global learning rate parameter for backpropagation is set to a constant 0.01. The batch processing single input sample size is configured to 32. Multinomial wavelength numerical parameters are extracted along the diagonal of the two-dimensional tensor. The arithmetic mean of the multinomial wavelength numerical parameters is calculated and compared with the pre-set base mapping standard. The preset base mapping scalar is read manually from the historical control group peak constant in the external read-only memory during the power-on phase of the device and set as a scalar in the threshold range of 530 nm to 680 nm. When the wavelength value parameter falls within the threshold range of the band, a binary activation level signal is output to assign a stem feature value matrix to the topology node of the association graph. The bottom layer slice index addressing instruction is called to extract the single column hybrid coding vector inside the stem feature value matrix along the first dimension direction. The hybrid coding vector is spliced ​​and merged with the multiple floating-point elements inside the stem feature value matrix along the row-major memory physical arrangement direction to generate a gene coding parameter matrix. The numerical truncation submodule, based on the gene-encoding parameter matrix, employs a leaky linear rectified activation algorithm. It calls the physical address of the processor's internal cache register, reads the associated multinomial floating-point feature parameters within the gene-encoding parameter matrix, sets the continuous memory addressing step size to 4 bytes, extracts the corresponding hybrid encoding vectors, executes a dot-multiplication instruction sequence using the single-instruction multiple-data extended instruction set, calculates the product of the associated floating-point feature parameters and the hybrid encoding vectors term by term, and obtains a continuous sequence containing 1024 double-precision floating-point product results. The leaky linear rectified activation algorithm's internal negative slope penalty hyperparameter is configured to a constant value. 0.01, sequentially traverse the continuous sequence of all product results to identify negative values ​​below the scalar limit of 0, multiply the identified negative values ​​by the negative slope penalty hyperparameter 0.01 to perform a proportional decay operation, truncate multiple negative values ​​within the product results, configure a memory value comparison mask and extract the corresponding overlapping terms after multiple truncation operations, set the memory extraction step size for the corresponding overlapping terms to 2 consecutive step size units, load all extracted values ​​sequentially into the physical structure of the three-dimensional tensor space, configure the three-dimensional tensor space dimension scale parameter to 64 x 64 x 128 data element nodes, and generate the final upper-level interaction weight tensor.

[0023] The germplasm evolution optimization module includes: The weight mapping submodule extracts the extreme value coordinate parameters inside the tensor based on the upper-level interaction weight tensor, merges the extreme value coordinates with the preset disaster prevention scalar values, projects and maps the root cortical ventilation tissue sequence, and superimposes a random numerical array to extract the latent variable array parameters to obtain the mutation-driven latent variable features. The mutation reconstruction submodule: Based on the mutation-driven latent variable features, it splices the numerical elements of the condition label matrix, calculates the multidimensional spatial mapping relationship of each matrix element, restores the multidimensional coordinate parameters of the mapped nucleotide sequence, extracts the numerical terms associated with sequence mutation sites, and establishes the reconstructed nucleotide matrix. The stress-resistant screening submodule: Based on the reconstructed nucleotide matrix, it extracts the numerical parameters of the associated samples and inputs them into the pre-trained support vector machine. It calls the radial basis kernel function to calculate the geometric distance between the sample feature vector and the classification hyperplane, compares it with the preset proline concentration threshold scalar, calculates the difference between the numerical value and the scalar, removes the sequence entries with a difference lower than the scalar, and aggregates the feature parameters of all remaining nodes to obtain the stress-resistant targeted gene array. The weight mapping submodule, based on the upper-level interaction weight tensor, employs a conditional mutation autoencoder algorithm. It extracts the peak wind speed constant from historical wind speed data logs from hardware sensors and writes it to a read-only register, setting it to a physical value of 25 meters per second as a preset disaster prevention scalar value. The conditional mutation autoencoder algorithm is configured with an input layer receiving 512 floating-point dimensions, a first fully connected hidden layer with 256 computation nodes, and both the mean and log-variance output layers with 128 neurons each. The learning rate is configured to be 0.005. It extracts the floating-point extremum coordinate parameters from the upper-level interaction weight tensor and processes them sequentially. The extreme value coordinates are merged with the preset disaster prevention scalar value. The matrix multiplication instruction is called to multiply the merged value by the weight matrix of the first fully connected hidden layer and add the corresponding bias vector. The product result is processed by calling the activation instruction of the rectified linear unit with leakage. The root cortex ventilation tissue sequence is projected and mapped. The pseudo-random number generator hardware interface is called to read the system clock nanosecond-level timestamp as a random number seed. A random numerical array following a normal distribution with a mean of 0 and a standard deviation of 1 is generated. The reparameterized sampling operation is performed to calculate the product of the mean and the corresponding variance and superimposes the random numerical array to extract the latent variable array parameters and obtain the mutation-driven latent variable features. The mutation reconstruction submodule, based on mutation-driven latent variable features, employs a conditional mutation autoencoder decoding algorithm. It reads the fixed values ​​from the drought stress level comparison table stored in the main control chip's serial peripheral interface, setting the label dimension to 64 single-precision floating-point numbers as a preset conditional label matrix. Multiple numerical elements of the preset conditional label matrix are then tiled and stitched along the tensor channel dimension with the mutation-driven latent variable features. The first deconvolutional layer of the conditional mutation autoencoder decoding algorithm is configured with a kernel size of 3 x 3 pixels and a stride of 2 pixels. The second deconvolutional layer is configured with 1024 output channels. The diagram shows the process of extracting and splicing tensors, inputting them into the first deconvolutional layer to perform multidimensional feature upscaling and mapping calculations, calling batch normalization operation instructions to calibrate the multidimensional spatial mapping relationship of multiple matrix elements, calculating the multidimensional logistic activation values ​​of the output layer, calling memory slicing instructions to extract the first 512 feature vectors along the spatial coordinate axes to restore the mapping of nucleotide sequence multidimensional coordinate parameters, setting an index addressing mask to traverse the multidimensional coordinate parameters to locate base sites with a mutation frequency exceeding 5%, calling register read instructions to extract sequence variant site-related numerical items, and pushing them into a two-dimensional memory block structure in row-major order to establish a reconstructed nucleotide matrix. The stress-resistance screening submodule, based on a reconstructed nucleotide matrix, employs a support vector machine (SVM) algorithm. The parameter 150 mg / g is manually input into the initialization control panel as a preset proline concentration threshold scalar. The SVM algorithm error penalty parameter is configured to a constant of 100, and the internal gamma hyperparameter of the radial basis function is configured to 0.025. The module reads multinomial floating-point feature data from the reconstructed nucleotide matrix to extract associated sample numerical parameters and transforms them into a high-dimensional tensor space. The extracted associated sample numerical parameters are then compared with the pre-loaded support vector set, and the squared Euclidean distance is calculated. The squared distance is then multiplied by a negative gamma. The hyperparameter values ​​are calculated and the natural exponential operation instruction is called to generate the kernel matrix space mapping value. The linear combination product adder instruction is called to calculate the geometric distance between the sample feature vector and the classification hyperplane. The arithmetic logic unit is called to execute the subtraction microinstruction to compare the geometric distance with the preset proline concentration threshold scalar. The difference between the returned floating-point value and the corresponding scalar is calculated. The status comparison register is configured to identify and mark the sequence index with a value less than 0. The memory reclamation instruction is called to remove the sequence entries with a difference lower than the scalar based on the sequence index. The continuous dynamic memory space is allocated to copy and aggregate all the feature parameters of the remaining nodes byte by byte. The anti-stress targeted gene array is output.

[0024] Support Vector Machine (SVM) specifically calls the radial basis function to perform high-dimensional space mapping operations, calculates the geometric distance between multiple associated sample feature vectors and a pre-defined classification hyperplane, obtains the spatial distribution coordinates of multiple sample feature vectors on both sides of the hyperplane, and determines the spatial polarity of the sample feature vectors relative to the classification hyperplane.

[0025] The habitat adversarial game module includes: Penalty operation submodule: Based on the stress-resistant targeted gene array, the stomatal conductance sequence is analyzed, the preset drought stress parameters are retrieved, the inverse gradient matrix values ​​are extracted and multiplied by the preset negative penalty coefficient, the product terms are calculated and assigned to weight nodes, and the inverse gradient penalty tensor is obtained. Spatial Fusion Submodule: Based on the inverse gradient penalty tensor, it fuses the pre-set leaf feature tensor, aligns the spatial physical coordinates of each node, calculates the mean value of the overlapping area representation and replaces the original coordinate parameters, extracts the merged array, and obtains the cross-domain shared feature set. The penalty operation submodule, based on the stress-resistance targeted gene array, employs a domain adversarial neural network gradient inversion algorithm. It configures the network with 512 fully connected layer nodes, an optimizer learning rate of 0.001, and a batch processing capacity of 64 independent computational samples. It directly reads instructions from memory to parse the continuous stomatal conductance sequence within the stress-resistance targeted gene array. It reads historical soil moisture sensor logs via the motherboard's integrated circuit bus, setting a 15% volumetric water content constant as a preset drought stress parameter. It then extracts the stomatal conductance sequence and performs multidimensional matrix multiplication with the preset drought stress parameter. The returned values ​​are extracted to construct the inverse gradient matrix. The microprocessor's floating-point unit reads the constant -0.05 stored in the device's flash memory read-only memory area as a preset negative penalty coefficient. The single instruction multiple data stream multiplication extension instruction is called to perform multiplication of the multiple floating-point values ​​inside the inverse gradient matrix with the preset negative penalty coefficient term by term, obtaining a continuous sequence of 256 double-precision floating-point product terms. The product terms are extracted in row-major order and written to the weight nodes of the multiple hidden layers of the neural network. The tensor concatenation instruction is called to combine the physical memory blocks of the multiple weight nodes to generate the inverse gradient penalty tensor. The spatial fusion submodule, based on the inverse gradient penalty tensor, employs bilinear interpolation and average pooling algorithms. It imports a mature maize 3D point cloud scanning dataset via an external universal serial bus interface and uses a compression algorithm to reduce the dimensionality to a 128x128 two-dimensional floating-point array, which is then set as a preset leaf feature tensor. The module calls direct memory access controller microinstructions to extract the preset leaf feature tensor and the inverse gradient penalty tensor, loading them into the main control chip's L2 cache. The feature extraction step size for the bilinear interpolation and average pooling algorithms is configured to be 2 pixel units, and the pooling window size is set to a 4x4 matrix grid. Finally, the module calls the affine transformation matrix operation unit to execute two-dimensional spatial rotation and translation floating-point instructions. Align the spatial physical coordinates of each node within the inverse gradient penalty tensor and the preset leaf feature tensor, retrieve the overlapping region matrix grid within the multidimensional space, call the hardware floating-point accumulator to accumulate all representation values ​​within the matrix grid in parallel and perform a division operation by the total number of grid nodes (16) to obtain the arithmetic mean of the representation values ​​in the overlapping region, call the memory overwrite instruction to write the arithmetic mean and replace the old coordinate parameters corresponding to the original spatial coordinates, call the continuous memory allocation operation instruction to concatenate and replace the data, extract and merge the two-dimensional floating-point array, extract all high-frequency feature vectors within the merged array and load them sequentially into the target register to obtain the cross-domain shared feature set.

[0026] The pollination survival prediction module includes: Risk-weighted submodule: Based on the stress-resistant targeted gene array and the cross-domain shared feature set, the associated feature values ​​of the two are merged, the flowering and silking interval time window values ​​are compared, the absolute value of the difference is calculated, and the absolute value is accumulated with the preset pollen shedding penalty value to obtain the weighted deviation parameter array. The retention and screening submodule: Based on the weighted deviation parameter array, the coordinates of the deviation nodes inside the array are extracted, the node coordinates are compared with the preset warning threshold values, the coordinate records that exceed the warning threshold are removed, and all node parameters in the safe retention range are aggregated to obtain the habitat tolerance expression array. Risk-weighted submodule: Based on the stress-resistant targeted gene array and the cross-domain shared feature set, it adopts the L1 norm distance calculation algorithm, calls the main control chip memory block to transmit micro-instructions to merge the stress-resistant targeted gene array and the cross-domain shared feature set containing 512 double-precision floating-point related feature values, reads the flowering and silking interval time window value in the external read-only memory, configures the microprocessor arithmetic logic unit to execute parallel subtraction instructions to compare the flowering and silking interval time window value with the merged related feature value, calls the hardware absolute value operation instruction to calculate and return a 1024-dimensional difference absolute value tensor, reads the pollination failure rate statistics table of the drought test period in the historical meteorological database to solidify the constant setting value of 0.20 as the preset pollen dispersal penalty value, configures the tensor accumulator to execute the preset pollen dispersal penalty value and the difference absolute value tensor item by item scalar addition operation, extracts all continuous floating-point memory blocks in the internal part of the addition operation output register, and generates a weighted deviation parameter array; The retention and filtering submodule, based on a weighted deviation parameter array, employs a dynamic threshold mask filtering algorithm. It sets the direct memory access step size parameter to 8 bytes to extract the coordinates of 2048 two-dimensional physical spatial deviation nodes within the weighted deviation parameter array. A preset warning threshold value (3.14) is set by reading a constant input manually via the serial communication interface during the device system initialization phase. A single-instruction multiple-data-stream comparison instruction set is invoked to compare the deviation corresponding to the deviation node coordinates with the preset warning threshold value in parallel. Based on the comparison result, a Boolean mask tensor composed of 0s and 1s is generated. A logical NOT instruction is invoked to invert the polarity of the Boolean mask tensor value. A bitwise AND operation is performed between the Boolean mask tensor and the original deviation node coordinates. The memory reclamation management interface is invoked to remove physical memory address blocks corresponding to coordinate records exceeding the warning threshold. A continuous memory allocation instruction is configured to aggregate all node parameters residing in the safe zone and copy them to the target physical cache, generating a habitat tolerance expression array.

[0027] The spatial layout decision module includes: State tensor construction submodule: Based on habitat tolerance expression array, extract the coordinate parameters of convergence nodes inside the array, retrieve the preset ear height and leaf angle parameters, merge the coordinates and the values ​​of the above parameters, load the merged values ​​into the corresponding spatial state dimension, and establish the environmental site state tensor. The probability curve mapping submodule extracts the values ​​of multiple dimensional parameters within the environment site state tensor, inputs them into the graph attention network to calculate the attention weight coefficients between the multiple dimensional parameters, performs a dot product operation between the weight coefficients and the dimensional parameters, maps the dot product operation results to generate a probability distribution curve, parses the probability distribution curve to extract the extreme value coordinate parameters, and obtains the set of layout extreme value coordinates. Adaptation instruction output submodule: Based on the set of layout extreme coordinates, it parses the extreme coordinate parameters of each item in the set, extracts the parent ratio values ​​corresponding to each coordinate, retrieves the preset population planting spacing parameter scalar, merges the ratio values ​​and the planting spacing parameter scalar, and generates a population adaptation instruction set. The state tensor construction submodule, based on the habitat tolerance expression array, employs a multidimensional tensor mapping and physical memory allocation algorithm. It calls the direct memory access controller microinstruction to set the read step size to 16 bytes, extracts the coordinate parameters of the convergent nodes within the habitat tolerance expression array, and reads the physical constants burned into the external electrically erasable read-only memory during system initialization via the serial peripheral interface bus. It sets a specific value of 120 cm as the preset ear height parameter and a specific value of 25 degrees as the preset leaf angle parameter. It then calls the single instruction multiple data stream extended instruction set to perform floating-point data concatenation, merging multiple coordinates with the aforementioned multiple parameter values. Finally, it calls the dynamic memory allocation function to allocate a continuous 64-megabyte physical cache space, configuring the physical cache space dimension scale parameter to 1024 x 1024 x 3 floating-point element nodes. Finally, it calls the cache write microinstruction to load the merged values ​​into the corresponding spatial state dimension physical memory block in row-major order, thus establishing the environmental site state tensor. The probability curve mapping submodule, based on the environment site state tensor, employs a multi-head graph attention network and probability density estimation algorithm. It calls tensor slice addressing instructions to extract multi-dimensional parameter values ​​from the environment site state tensor. The multi-head attention mechanism within the multi-head graph attention network and probability density estimation algorithm is configured with 8 concurrent threads. The network's single-node hidden layer feature dimension parameters are configured to be 256 single-precision floating-point numbers. The negative slope parameter with leaky linear rectified activation constant is configured to be constant at 0.2. The global learning rate fine-tuning step size is configured to be 0.001. The input to the multi-head graph attention network is processed by the multiply-accumulate operation unit within the field-programmable gate array to calculate the attention weights between the multi-dimensional parameters. The matrix is ​​calculated by calling floating-point tensor multiplication microinstructions to perform dot product feature aggregation operations on the attention weight coefficient matrix and multi-dimensional parameter numerical tensors. The natural logarithm base physical constant is read from the motherboard's read-only memory, and the natural exponential operation instruction is called to perform normalization numerical scaling operations to map the dot product feature aggregation operation results. The product results generate a floating-point distribution probability curve containing 4096 data sampling nodes. The first-order forward difference microinstruction is called to calculate the first-order derivative values ​​of multiple data sampling nodes and compare them to determine the boundaries of small positive extreme points. The distribution probability curve is analyzed to extract extreme value coordinate parameters. The data push stack microinstruction is called to push all extracted extreme value coordinate parameters into the underlying continuous physical address data stack structure in order to obtain the set of layout extreme value coordinates. The adaptation instruction output submodule, based on the layout extreme coordinate set, employs heuristic rule matching and binary instruction encoding algorithms. It calls the data stack to pop up operation micro-instructions, parsing multiple extreme coordinate parameters within the layout extreme coordinate set in a last-in-first-out order. It configures the two-dimensional memory addressing offset, reads the floating-point ratio constant stored in the register, and extracts the corresponding parent ratio values ​​for multiple coordinates. It reads the analog electrical signal generated by the operator turning the band switch on the physical control panel during the device's power-on self-test phase, converts it into a digital constant, and sets 60 cm as the preset population planting spacing parameter scalar. It calls logical shift operation instructions to align the data bit width, merge the parent ratio values ​​with the preset population planting spacing parameter scalar, configures the serial communication data frame format parameters including a 1-bit start bit, 8-bit data bits, and a 1-bit stop bit, and combines even parity rules. It calls the data frame encapsulation micro-instruction to convert the merged data into a hexadecimal machine code byte stream, generating a population adaptation instruction set.

[0028] The graph attention network specifically maps multi-dimensional parameter values ​​to graph topology node feature vectors, retrieves a pre-set learnable weight matrix, multiplies it by the node feature vectors to perform a linear transformation, concatenates multiple adjacent node feature vectors and multiplies them by a pre-set attention parameter vector, calls a leaky linear rectified activation function to truncate the negative values ​​of the product terms, applies a normalized exponential function to perform numerical scaling, and calculates and outputs an attention weight coefficient matrix between multi-dimensional parameters.

[0029] Please see Figure 2The optimization of planting and breeding strategies for density tolerance and stress tolerance, as well as maize breeding methods, include the following steps: S1: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds to determine the stalk feature matrix assigned to maize gene loci, heterozygous coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate a superordinate interaction weight tensor. S2: Based on the upper-level interaction weight tensor, extreme value coordinates are extracted and disaster prevention values ​​are merged. The maize aeration tissue sequence is projected, and latent variables are extracted by random number extraction. The nucleotide matrix is ​​reconstructed, and low-value entries are removed by comparison of concentration using a support vector machine to obtain the stress-resistant target gene array. S3: Based on the stress-resistant targeted gene array, the stomatal conductance sequence of maize was analyzed, the drought stress parameters were retrieved to extract the inverse gradient multiplied by the negative penalty coefficient, the sequence was fused with the maize leaf tensor, the spatial coordinates were aligned and the overlapping values ​​were merged to obtain the cross-domain shared feature set. S4: Based on the stress-resistant targeted gene array and cross-domain shared feature set, the time window of maize flowering and silking interval was merged by numerical comparison, the weight matrix of the cumulative weight of the difference was extracted, the pollen penalty value was calculated and added to the iteration term, the coordinates exceeding the threshold were removed, and the habitat tolerance expression array was constructed. S5: Based on the habitat tolerance expression array, after calling the array node coordinates, the parameters of corn ear height and leaf angle are merged, the state tensor is filled in, the distribution probability curve is generated by the graph attention network, the extreme value coordinates and parent ratio are extracted, and the corn population adaptation instruction set is output.

[0030] 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 system for optimizing planting and breeding strategies to tolerate high density and stress, characterized in that, The system includes: Genetic effect topology module: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds. After establishing a feature matrix, heterozygous genotype coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate an epistatic interaction weight tensor. Germplasm evolution optimization module: Based on the aforementioned epistatic interaction weight tensor, extreme value coordinates are extracted and preset disaster prevention values ​​are merged, ventilation tissue sequences are projected, latent variables are extracted by superimposing random numbers, nucleotide matrix is ​​reconstructed, and low-value entries are eliminated by comparing concentrations using a support vector machine to obtain a stress-resistant target gene array. Habitat confrontation game module: Based on the stress resistance targeted gene array, analyze the stomatal conductance sequence within the array, retrieve the preset drought stress parameters and extract the reverse gradient multiplied by the negative penalty coefficient, fuse the sequence with the leaf tensor, align the spatial coordinates and merge the overlapping values ​​to obtain the cross-domain shared feature set. Pollination survival prediction module: Based on the stress-resistant target gene array and the cross-domain shared feature set, merge the set values ​​and compare the flowering and silking interval time window, extract the absolute value of the difference and accumulate the weight matrix, calculate the pollen dispersal penalty value and add it to the iteration term, remove the coordinates exceeding the threshold, and construct the habitat tolerance expression array. Spatial layout decision module: Based on the habitat tolerance expression array, after calling the array node coordinates, the preset ear height and leaf angle parameters are merged, the state tensor is filled in, the distribution probability curve is generated through the graph attention network, the extreme value coordinates of the curve and the parent ratio are extracted, and the population adaptation instruction set is output.

2. The optimized system for planting and breeding strategies to tolerate high density and stress, as described in claim 1, is characterized in that... The superordinate interaction weight tensor includes wavelength values, hybrid coding vectors within the feature matrix, and overlapping terms corresponding to truncation operations. The stress-resistant targeted gene array includes preset disaster prevention scalar values, root cortex aeration tissue sequences, reconstructed nucleotide matrices, and surviving node feature parameters. The cross-domain shared feature set includes stomatal conductance sequences, drought stress parameters, inverse gradient penalty tensors, and the mean of overlapping region characterization values. The habitat tolerance expression array includes flowering and silking interval time window values, preset pollen shedding penalty values, a weighted deviation parameter array, and resident safe interval node parameters. The population adaptation instruction set includes extreme coordinates of the distribution probability curve, parent ratio values, preset population planting spacing parameter scalars, and water and fertilizer application rates.

3. The optimized system for planting and breeding strategies to tolerate high density and stress as described in claim 1, characterized in that, The genetic effect topology module includes: Signal mapping submodule: Based on the original sequencing optical signal matrix, extract wavelength numerical parameters, compare the numerical parameters with the pre-set base mapping scalar, assign stem feature numerical matrix to the associated node, extract the heterozygous coding vector inside the feature matrix, merge the coding vector with each element of the feature matrix, and obtain the gene coding parameter matrix; Numerical truncation submodule: Based on the gene coding parameter matrix, it calls the associated feature parameters inside the matrix, extracts the corresponding heterozygous coding vectors for each term, calculates the product of the parameters and coding vectors term by term, truncates multiple negative values ​​inside the product result, extracts the corresponding overlapping terms after each truncation operation, and generates the superordinate interaction weight tensor.

4. The optimized system for planting and breeding strategies to tolerate high density and stress as described in claim 1, characterized in that, The germplasm evolution optimization module includes: Weight mapping submodule: Based on the superordinate interaction weight tensor, extract the extreme value coordinate parameters inside the tensor, merge the extreme value coordinates with the preset disaster prevention scalar values, project and map the root cortex ventilation tissue sequence, superimpose the random numerical array to extract the latent variable array parameters, and obtain the variation-driven latent variable features. The mutation reconstruction submodule: Based on the mutation-driven latent variable features, it splices the numerical elements of each item in the condition label matrix, calculates the multidimensional spatial mapping relationship of each matrix element, restores the multidimensional coordinate parameters of the mapped nucleotide sequence, extracts the numerical terms associated with sequence mutation sites, and establishes the reconstructed nucleotide matrix. The stress-resistant screening submodule extracts the numerical parameters of the associated samples based on the reconstructed nucleotide matrix and inputs them into the pre-trained support vector machine. It calls the radial basis function to calculate the geometric distance between the sample feature vector and the classification hyperplane, compares it with the preset proline concentration threshold scalar, calculates the difference between the numerical value and the scalar, removes the sequence entries with a difference lower than the scalar, and aggregates all the feature parameters of the remaining nodes to obtain the stress-resistant targeted gene array.

5. The planting and breeding strategy optimization system for density tolerance and stress tolerance according to claim 1, characterized in that, The support vector machine specifically calls the radial basis kernel function to perform high-dimensional space mapping operation, calculates the geometric distance between multiple associated sample feature vectors and the preset classification hyperplane, obtains the spatial distribution coordinates of multiple sample feature vectors on both sides of the hyperplane, and determines the spatial polarity of the sample feature vectors relative to the classification hyperplane.

6. The planting and breeding strategy optimization system for density tolerance and stress tolerance according to claim 1, characterized in that, The habitat adversarial game module includes: The penalty operation submodule: Based on the stress-resistant targeted gene array, the stomatal conductance sequence is analyzed, the preset drought stress parameters are retrieved, the inverse gradient matrix values ​​are extracted and multiplied by the preset negative penalty coefficient, the product terms are calculated and assigned to weight nodes, and the inverse gradient penalty tensor is obtained. Spatial fusion submodule: Based on the inverse gradient penalty tensor, fuse the preset leaf feature tensor, align the spatial physical coordinates of each node, calculate the mean value of the overlapping region representation and replace the original coordinate parameters, extract the merged array, and obtain the cross-domain shared feature set.

7. The planting and breeding strategy optimization system for density tolerance and stress tolerance according to claim 1, characterized in that, The pollination survival prediction module includes: Risk-weighted submodule: Based on the stress-resistant targeted gene array and the cross-domain shared feature set, merge the associated feature values ​​of the two, compare the flowering and silking interval time window values, calculate the absolute value of the difference, accumulate the absolute value and the preset pollen shedding penalty value to obtain the weighted deviation parameter array; Retention and screening submodule: Based on the weighted deviation parameter array, extract the coordinates of the deviation nodes inside the array, compare the node coordinates with the preset warning threshold values, remove the coordinate records that exceed the warning threshold, aggregate all node parameters in the safe retention range, and obtain the habitat tolerance expression array.

8. The optimized system for planting and breeding strategies to tolerate high density and stress as described in claim 1, characterized in that, The spatial layout decision module includes: State tensor construction submodule: Based on the habitat tolerance expression array, extract the coordinate parameters of the convergence nodes inside the array, retrieve the preset ear height and leaf angle parameters, merge the coordinates and the values ​​of the above parameters, load the merged values ​​into the corresponding spatial state dimension, and establish the environmental site state tensor. Probability curve mapping submodule: Based on the environmental site state tensor, extract the values ​​of multiple dimensional parameters inside the tensor, input the graph attention network to calculate the attention weight coefficients between the multiple dimensional parameters, perform the dot product operation between the weight coefficients and the dimensional parameters, map the dot product operation results to generate a probability distribution curve, parse the probability distribution curve to extract the extreme value coordinate parameters, and obtain the set of layout extreme value coordinates. Adaptation instruction output submodule: Based on the set of extreme coordinates of the layout, it parses the extreme coordinate parameters of each item in the set, extracts the parent ratio values ​​corresponding to each coordinate, retrieves the preset population planting spacing parameter scalar, merges the ratio values ​​and the planting spacing parameter scalar, and generates a population adaptation instruction set.

9. The optimized system for planting and breeding strategies to tolerate high density and stress, as described in claim 1, is characterized in that... The graph attention network specifically maps the numerical values ​​of multiple dimensional parameters to feature vectors of graph topology nodes, retrieves a preset learnable weight matrix, multiplies it by the node feature vectors to perform a linear transformation, concatenates multiple adjacent node feature vectors and multiplies them by a preset attention parameter vector, calls a leaky linear rectified activation function to truncate the negative values ​​of the product terms, applies a normalized exponential function to perform numerical scaling, and calculates and outputs an attention weight coefficient matrix among multiple dimensional parameters.

10. Optimization of planting and breeding strategies for maize that are tolerant to high density and stress, and methods for maize breeding, characterized in that, The implementation of the planting and breeding strategy optimization system for tolerance to high density and stress as described in any one of claims 1-9 includes the following steps: S1: Based on the original sequencing optical signal matrix, wavelength values ​​are extracted and compared with preset base thresholds to determine the stalk feature matrix assigned to maize gene loci, heterozygous coding vectors are extracted, vectors and matrices are multiplied, negative values ​​are truncated and overlapping terms are extracted to generate a superordinate interaction weight tensor. S2: Based on the superordinate interaction weight tensor, extract the extreme value coordinates and merge the disaster prevention values, project the maize aeration tissue sequence, superimpose random numbers to extract latent variables, reconstruct the nucleotide matrix, and use a support vector machine to compare concentrations and remove low-value entries to obtain the stress-resistant target gene array. S3: Based on the stress-resistant targeted gene array, analyze the maize stomatal conductance sequence, retrieve drought stress parameters to extract the reverse gradient multiplied by the negative penalty coefficient, fuse the sequence with the maize leaf tensor, align the spatial coordinates and merge the overlapping values ​​to obtain a cross-domain shared feature set. S4: Based on the stress-resistant targeted gene array and cross-domain shared feature set, the time window of maize flowering and silking interval was merged by numerical comparison, the weight matrix of the cumulative weight of the difference was extracted, the pollen penalty value was calculated and added to the iteration term, the coordinates exceeding the threshold were removed, and the habitat tolerance expression array was constructed. S5: Based on the habitat tolerance expression array, after calling the array node coordinates, merge the parameters of maize ear height and leaf angle, fill in the state tensor, generate the distribution probability curve by graph attention network, extract the extreme value coordinates and parent ratio, and output the maize population adaptation instruction set.

Images