A Smart Breeding Evaluation Method and System for Salt-Tolerant Crops in Saline-Alkali Land

Abstract

This invention discloses an intelligent breeding evaluation method and system for salt-tolerant crops in saline-alkali land. The method includes crop data collection, crop data improvement, establishment of a crop breeding evaluation model, and intelligent crop breeding evaluation. This invention belongs to the field of breeding evaluation, specifically referring to an intelligent breeding evaluation method and system for salt-tolerant crops in saline-alkali land. This scheme avoids generating only single salt tolerance level samples by introducing environmental adaptability differential regulation; optimizes the objective function by introducing a stress differential coefficient; and adjusts the feature function of the phenotypic sample generator to ensure that the generated samples are highly adapted to the actual salt stress patterns in saline-alkali land. Through a hierarchical attention mechanism, the contribution of different levels of features in the crop breeding data is evaluated according to the feature importance function, and redundant features are eliminated. The introduction of environmental edge loss improves the distinguishability of adjacent salt tolerance levels in extreme saline-alkali environments; and the distance between the feature centers of adjacent salt tolerance levels is increased, thereby improving the crop breeding evaluation effect.

Description

Technical Field

[0001] This invention relates to the field of breeding evaluation technology, specifically to an intelligent breeding evaluation method and system for salt-tolerant crops in saline-alkali land. Background Technology

[0002] Evaluation methods for salt-tolerant crops in saline-alkali land refer to the means of classifying or judging the salt tolerance of crops by measuring their growth indicators and environmental parameters in a saline-alkali environment. However, general evaluation methods for salt-tolerant crops in saline-alkali land suffer from several problems: unbalanced crop breeding data, leading to the omission of superior varieties; poor environmental adaptability, with crop breeding samples deviating from the phenotypic patterns of crops under actual salt stress, resulting in low reliability of the evaluation; and redundant interference from crop characteristics, with subtle phenotypic differences between adjacent salt tolerance levels, and extreme saline-alkali environments exacerbating characteristic overlap, leading to misjudgment of highly salt-tolerant varieties and poor evaluation results. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent breeding and evaluation method and system for salt-tolerant crops in saline-alkali land. Addressing the problems of unbalanced crop breeding data, missed identification of superior varieties, poor environmental adaptability, and a disconnect between the selected crop samples and actual crop phenotypic patterns under salt stress, leading to low reliability of the breeding evaluation, this solution introduces environmental adaptability differential control and integer feature range control to avoid generating only single salt tolerance level samples. By introducing a stress differential coefficient for objective function optimization, the generated integer data better reflects actual physiological patterns. Finally, through a phenotypic sample generator feature adjustment function combined with a salt tolerance level decay function, the generated samples are highly adapted to the actual salt stress patterns in saline-alkali land. This approach aims to improve the reliability of subsequent crop breeding assessments. Addressing the issues of redundant interference from crop features, subtle phenotypic differences between adjacent salt tolerance levels, and the exacerbation of feature overlap in extreme saline-alkali environments, which can lead to misjudgments of highly salt-tolerant varieties and poor crop breeding assessment results, this solution employs a hierarchical attention mechanism. Based on feature importance functions, it evaluates the contribution of features at different levels of crop breeding data, combines dynamic thresholds to screen high-value features, eliminates redundant features, and strengthens core features related to salt tolerance. Furthermore, it introduces environmental edge loss to enhance the distinguishability between adjacent salt tolerance levels in extreme saline-alkali environments and constructs a salt tolerance level difference loss mechanism to increase the distance between feature centers of adjacent salt tolerance levels, reducing misjudgments of superior varieties. Ultimately, this improves the effectiveness of crop breeding assessments.

[0004] The technical solution adopted by this invention is as follows: This invention provides an intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land, which includes the following steps:

[0005] Step S1: Crop data collection;

[0006] Step S2: Improve crop data;

[0007] Step S3: Establish a crop breeding evaluation model;

[0008] Step S4: Evaluation of intelligent crop breeding.

[0009] Further, in step S1, the crop data collection involves collecting crop breeding data, including crop salt tolerance-related phenotypic data and saline-alkali land environment-related data; and labeling the crop salt tolerance level; and standardizing the acquired data to construct an initial crop breeding dataset.

[0010] Furthermore, in step S2, the crop data improvement specifically includes the following:

[0011] Step S21: Floating-point data augmentation; embed the salt tolerance level label into random noise z, input standardized floating-point data, and generate new floating-point samples through a traditional phenotypic sample generation network, denoted as z. ;in, and These are the newly generated floating-point crop breeding data and their corresponding salt tolerance level labels;

[0012] Step S22: Integer data augmentation; embed salt tolerance level labels into random noise z, input standardized integer data; introduce environmental adaptability differential control, and add integer feature range control, expressed as: ;in, It is the loss of environmental adaptability differential regulation; It is the environmental adaptability differential control coefficient; It is the expectation over a specific distribution; It is the distribution of interpolated samples; It is an interpolated sample between real samples and generated samples, and is an intermediate sample used for environmental adaptability differential regulation. This is the distribution of the interpolated samples; It is the gradient of the phenotypic authenticity discriminator on the interpolated sample; It is the L2 norm;

[0013] Step S23: Objective function optimization; for integer data augmentation, construct a realism discrimination loss that includes environmental adaptability differential adjustment. , is represented as: ;in, It is the distribution of real integer samples; It is an integer sample; It is the authenticity judgment output of the phenotypic authenticity discriminator; It is the distribution of noise z; It is a phenotypic sample generator based on noise z and salt tolerance level labels. Generate integer samples; construct a salt tolerance level classification loss. , represented as: ; the generation of new integer samples is denoted as ;in, It is the output of the phenotypic authenticity discriminator for classifying the salt tolerance level of real integer samples; It is a phenotypic authenticity discriminator that classifies salt tolerance levels for generated integer samples; it introduces a stress level difference coefficient. Represented as: Define the feature adjustment function for the phenotypic sample generator as follows: ;in, It is a function of salt tolerance level decay. e represents soil electrical conductivity, the value before standardization. and These are the adjusted and unadjusted floating-point phenotypic features, respectively. It represents the crop's salt tolerance level; the final loss function is integer-based data augmentation. Represented as: ; and It is to enhance weight; It is the summation of the differences in electrical conductivity of different soils; It is the mean of floating-point phenotypic characteristics in real samples when the soil electrical conductivity is e;

[0014] Step S24: Data Fusion; Based on the data concatenation operation, merge the generated floating-point sample with the real floating-point sample, denoted as... The generated integer sample is merged with the real integer sample, denoted as . Integrating crop breeding data Input a DNN classifier, using salt tolerance level as the classification target, and optimize the network parameters through backpropagation. If the crop breeding fusion data is not found, the enhancement effect will be discarded; the classification accuracy will be used to verify the enhancement effect; among them, and These are floating-point samples and integer samples from the synthesized crop breeding dataset, respectively. and It corresponds to the label; set a classification accuracy threshold. If the classification accuracy is higher than the classification accuracy threshold, then the crop breeding fusion data obtained at this time will be used as the final crop breeding fusion data.

[0015] Furthermore, in step S3, the establishment of the crop breeding evaluation model is based on a deep learning model; the model updates parameters through gradient descent and adjusts initial parameters through particle swarm optimization; specifically, it includes the following:

[0016] Step S31: Feature extraction network construction; input augmented crop breeding fusion data. Multi-view processing is performed: global average pooling is used to extract the overall feature distribution and generate a global view vector. This includes the overall trend of tiller number in salt-tolerant crops; using a global max-pooling perspective, it extracts local salient features and generates local perspective vectors. This includes extremely salt-tolerant individuals that maintain 80% leaf water content under salt stress; channel weights for two types of features are obtained through fully connected layers and sigmoid activation, highlighting salt-tolerance-related features. The weights are multiplied by the original features and then summed to output the fused feature F(M). i );

[0017] Step S32: Adaptive feature selection; based on the fused features F(M) output from step S31 through a hierarchical attention mechanism. i An importance assessment is performed to select high-contribution features for the fusion stage; a feature importance function I(·) is constructed, using the following formula: ;in, It is F(M) i The low k features of the l-th layer; It is the activation intensity of features within the layer; ; It is the set of the k-th feature values ​​of all samples in the l-th layer; It is the k-th feature value of the i-th sample in the l-th layer; and dynamic thresholding is performed, the dynamic threshold... Represented as: For the k-th feature of the i-th sample in the l-th layer, if the feature importance function value is higher than the dynamic threshold, it is determined to be a high-contribution feature and is retained; otherwise, it is removed. It refers to training rounds;

[0018] Step S33: Layered and refined fusion of salt tolerance features; The high-contribution features selected in Step S32 are fused layer by layer based on salt tolerance features. After unifying the dimensions through a fully connected layer, a progressive fusion is used to allow lower-level basic features to provide detailed support for higher-level derived features, avoiding feature correlation breaks; the formula used to adjust the dimensions of the fully connected layer is: The formula used for progressive fusion is: The formula used for the final feature output is: ;in, is the feature of the i-th sample after adjustment in the l-th layer; ReLU(·) is the ReLU function; FC(·) is a fully connected layer; It is the high contribution feature of the i-th sample in the l-th layer after screening in step S32; It is the fusion feature of the i-th sample at the highest layer; It is global average pooling; It is the final feature vector of the sample;

[0019] Step S34: Network classification; Calculate the feature center of each salt tolerance level using the data-augmented samples, calculate the Euclidean distance between the feature of the sample to be predicted and the feature centers of each class, and the level with the closest distance is the prediction result; the formula used to calculate the feature center of the salt tolerance level is: The formula used to calculate classification probability is: ;in, and These are the salt tolerance feature centers of the j-th and u-th salt tolerance levels, respectively; K is the total number of samples for the j-th salt tolerance level. is the final feature vector of the i-th sample in the j-th class after step S33; x is the sample to be predicted; It is the final feature vector output by the sample to be predicted in step S33; It is the Euclidean distance; It is the probability that the sample to be predicted belongs to the j-th salt tolerance level;

[0020] Step S35: Loss function design; construct the basic loss function, using cross-entropy loss. Based on, it is expressed as: Where Q is the total number of samples participating in training; b is the salt tolerance level index; and These are the true label and the probability of predicting salt tolerance level c for the i-th sample, respectively; construct the salt tolerance level difference loss. The formula used is: Introducing environmental edge loss, we construct adaptive edges based on environmental differences, and for each sample i, we apply these edges to each... Define dynamic edges , represented as: Environmental edge loss Represented as: ;in, It is the salt tolerance characteristic center of the kth salt tolerance level; It is the true level of the i-th sample. The salt tolerance rating characteristic center; It is the basic edge; It is the environment scaling factor; , and It is the environmental average of the sample, the normalized mean; yes The corresponding environmental average; the final loss L is expressed as: ;in, and These are the salt tolerance characteristic centers for Class b and Class b+1 salt tolerance grades, respectively; It is the loss weighting coefficient.

[0021] Furthermore, in step S4, the intelligent crop breeding assessment is based on the salt tolerance level assessment model established in step S3, combined with the characteristics of the saline-alkali land environment and breeding objectives, to complete the breeding assessment of highly salt-tolerant crops, specifically including the following:

[0022] Step S41: Salt tolerance potential assessment of candidate breeding materials; collect data from all dimensions of existing breeding materials, input the standardized data into the assessment model of step S3, and output the salt tolerance level prediction results of each material as the basis for salt tolerance potential assessment.

[0023] Step S42: Targeted screening of saline-alkali land environmental adaptability; Based on the environmental data of the target saline-alkali area and the salt tolerance level prediction results of step S41, material-environment recommendations are made.

[0024] This invention provides an intelligent breeding system for salt-tolerant crops in saline-alkali land, comprising a crop data acquisition module, a crop data improvement module, a crop breeding evaluation model establishment module, and a crop intelligent breeding module;

[0025] The crop data acquisition module collects crop breeding data and constructs an initial crop breeding dataset.

[0026] The crop data enhancement module enhances the initial crop breeding dataset by floating-point data enhancement, introduces integer data enhancement for environmental adaptability differential regulation, and introduces stress differential coefficient to optimize the objective function, thus obtaining the final crop breeding fusion data.

[0027] The crop breeding evaluation model building module is based on crop breeding fusion data and deep learning model. It establishes crop breeding evaluation model by extracting salt tolerance features from multiple perspectives, adaptive selection, hierarchical fusion, and combining salt tolerance level difference loss function.

[0028] The intelligent crop breeding module is based on a crop breeding evaluation model and combines the environmental characteristics of the target saline-alkali land to carry out intelligent crop breeding.

[0029] The beneficial effects achieved by the present invention using the above solution are as follows:

[0030] (1) To address the problems of unbalanced crop breeding data and missed identification of superior varieties in general saline-alkali land breeding evaluation methods, poor environmental adaptability, and the disconnect between crop breeding samples and actual crop phenotypic patterns under salt stress, which leads to low reliability of breeding evaluation, this scheme avoids generating only single salt tolerance level samples by introducing environmental adaptability differential control and integer feature range control; optimizes the objective function by introducing stress differential coefficient, making the generated integer data more consistent with actual physiological patterns; and improves the reliability of subsequent crop breeding evaluation by using the phenotypic sample generator feature adjustment function combined with the salt tolerance level decay function to make the generated samples highly compatible with the actual salt stress patterns in saline-alkali land.

[0031] (2) To address the problems of redundant interference of crop characteristics, subtle phenotypic differences between adjacent salt tolerance levels, and the exacerbation of feature overlap in extreme saline-alkali environments, which leads to misjudgment of high salt tolerance varieties and poor crop breeding evaluation results, this scheme uses a hierarchical attention mechanism to evaluate the contribution of different levels of crop breeding data according to the feature importance function. It combines dynamic threshold screening to remove high-value features, eliminate redundant features, and strengthen the core features related to salt tolerance. By introducing environmental edge loss, the scheme improves the distinguishability between adjacent salt tolerance levels in extreme saline-alkali environments. It constructs a salt tolerance level difference loss to increase the distance between the feature centers of adjacent salt tolerance levels and reduce misjudgment of superior varieties. This improves the crop breeding evaluation effect. Attached Figure Description

[0032] Figure 1 A flowchart illustrating an intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land provided by this invention;

[0033] Figure 2 This is a schematic diagram of an intelligent breeding system for salt-tolerant crops in saline-alkali land provided by the present invention.

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0037] Example 1, see Figure 1 This invention provides an intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land, which includes the following steps:

[0038] Step S1: Crop data collection; collect crop breeding data and construct an initial crop breeding dataset;

[0039] Step S2: Crop data enhancement; The initial crop breeding dataset is augmented with floating-point data, augmented with integer data for environmental adaptability differential regulation, and the objective function is optimized by introducing stress differential coefficients to obtain the final crop breeding fusion data;

[0040] Step S3: Establish a crop breeding evaluation model; Based on the fusion data of crop breeding, and using a deep learning model as the foundation, a crop breeding evaluation model is established by extracting salt tolerance features from multiple perspectives, adaptive selection, hierarchical fusion, and combining the salt tolerance level difference loss function.

[0041] Step S4: Intelligent crop breeding assessment; Based on the crop breeding assessment model, intelligent crop breeding is carried out in combination with the environmental characteristics of the target saline-alkali land.

[0042] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, crop data collection involves collecting crop breeding data, including crop salt tolerance-related phenotypic data and saline-alkali land environment-related data; and labeling the crop with salt tolerance level tags; the crop salt tolerance-related phenotypic data includes integer phenotypic data: plant height, effective tiller number, seed setting rate, survival days, and floating-point phenotypic data: leaf relative water content, chlorophyll content, root length, and salt tolerance index; the saline-alkali land environment-related data includes integer environmental data: soil salinity level, irrigation frequency, and floating-point environmental data: soil electrical conductivity and soil pH value; the crop salt tolerance level tags are 1-5, with higher levels indicating better salt tolerance; the acquired data are standardized to construct an initial crop breeding dataset;

[0043] Data acquisition: Plant height, random sampling, measuring the vertical distance from the ground to the top of the main stem, and removing deformed stems caused by salt damage during measurement;

[0044] Tillers: Randomly select 10 clumps of plants along an S-shaped route, retain 3 main stems in each clump, and count the number of tillers that can produce ears and seeds in each individual plant.

[0045] To calculate the seed setting rate, random samples were taken from each plot, and after threshing, plump seeds and empty or shriveled seeds were separated.

[0046] Chlorophyll content was measured on 20 days after salt stress, with 3 healthy leaves selected from each plant and 3 sites measured on each leaf, avoiding the veins, and the average value was taken.

[0047] Root length: Random samples were taken from each plot, the entire root system was dug up, rinsed clean with running water, and laid flat on a transparent glass plate to obtain the total root length;

[0048] Salt tolerance index: The soil EC value was set at 8 mS / cm, and the biomass ratio of the treatment group to the control group (soil EC value of 2 mS / cm) was calculated to obtain the salt tolerance index.

[0049] Soil electrical conductivity was measured by inserting the probe into the soil to a depth of 20 cm, with one measuring point every 10 m. Temperature was recorded and corrected to the standard value of 25℃.

[0050] To measure soil pH, insert a strong alkali resistant electrode 5 cm into the soil and take the reading after stabilization. Repeat the measurement three times at each point and take the average value.

[0051] Soil salinity levels are determined by assigning soil EC values ​​as follows: Level 1, EC < 2 mS / cm; Level 2, 2-4 mS / cm; Level 3, 4-6 mS / cm; Level 4, 6-8 mS / cm; Level 5, EC > 8 mS / cm. A quincunx five-point sampling method is used, and the EC values ​​are measured after mixing to represent the regional level.

[0052] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, crop data improvement is necessary because in high-salt environments like saline-alkali land, most crops have weak salt tolerance, and varieties with strong salt tolerance are few in number and have a long artificial cultivation cycle, resulting in such samples accounting for a very low proportion in the dataset. Therefore, by expanding the number of small samples, the omission of superior varieties can be avoided. Specifically, this includes the following:

[0053] Step S21: Floating-point data augmentation; embed the salt tolerance level label into random noise z, input standardized floating-point data, and generate new floating-point samples through a traditional phenotypic sample generation network, denoted as z. ;in, and These are the newly generated floating-point crop breeding data and their corresponding salt tolerance level labels;

[0054] Step S22: Integer Data Augmentation; Addressing the issues of high dispersion in integer data and the distorted generation of traditional phenotypic sample generation networks in crop breeding data, an improved phenotypic sample generation network with environmental adaptability differential control is adopted; salt tolerance level labels are embedded with random noise z, and standardized integer data is input; environmental adaptability differential control is introduced to avoid generating only single salt tolerance level samples, and integer feature range control is added, expressed as: ;in, It is the loss of environmental adaptability differential regulation; It is the environmental adaptability differential control coefficient; It is the expectation over a specific distribution; It is the distribution of interpolated samples; It is an interpolated sample between real samples and generated samples, and is an intermediate sample used for environmental adaptability differential regulation. This is the distribution of the interpolated samples; It is the gradient of the phenotypic authenticity discriminator on the interpolated sample; It is the L2 norm;

[0055] Step S23: Objective function optimization; for integer data augmentation, construct a realism discrimination loss that includes environmental adaptability differential adjustment. , represented as: ;in, It is the distribution of real integer samples; It is an integer sample; It is the authenticity judgment output of the phenotypic authenticity discriminator; It is the distribution of noise z; It is a phenotypic sample generator based on noise z and salt tolerance level labels. Generate integer samples; construct a salt tolerance level classification loss. , represented as: ; the generation of new integer samples is denoted as ;in, It is the output of the phenotypic authenticity discriminator for classifying the salt tolerance level of real integer samples; The phenotypic authenticity discriminator outputs the salt tolerance level classification of generated integer samples; a stress grade difference coefficient is introduced to adjust the feature distribution of the phenotypic sample generator based on the soil EC value, so that the generated samples conform to the phenotypic pattern of specific salt grade differences. Represented as: Define the feature adjustment function for the phenotypic sample generator as follows: ;in, It is a function of salt tolerance level decay. e represents soil electrical conductivity, the value before standardization. and These are the adjusted and unadjusted floating-point phenotypic features, respectively. It represents the crop's salt tolerance level; the final loss function is integer-based data augmentation. Represented as: ; and It enhances the weights; the phenotypic authenticity discriminator is a single-network dual-output structure: the same phenotypic authenticity discriminator outputs two results simultaneously. and ; It is the summation of the differences in electrical conductivity of different soils; It is the mean of floating-point phenotypic characteristics in real samples when the soil electrical conductivity is e;

[0056] By strengthening the matching between floating-point phenotypic features and salinity grade differences in saline-alkali land, a physiologically reasonable basis for the generation of integer features is provided; this prevents the generated samples from becoming disconnected from the actual stress patterns of saline-alkali land.

[0057] Step S24: Data Merging; Based on the data concatenation operation, use the pd.concat() function of the pandas library to merge the generated floating-point samples with the real floating-point samples, denoted as... The generated integer sample is merged with the real integer sample, denoted as . Integrating crop breeding data Input a DNN classifier with salt tolerance level as the classification target, and optimize the network parameters through backpropagation. If the crop breeding fusion data is not found, the enhancement effect will be discarded; the classification accuracy will be used to verify the enhancement effect; among them, and These are floating-point samples and integer samples from the synthesized crop breeding dataset, respectively. and It corresponds to the label; set a classification accuracy threshold. If the classification accuracy is higher than the classification accuracy threshold, then the crop breeding fusion data obtained at this time will be used as the final crop breeding fusion data.

[0058] By performing the above operations, this scheme addresses the problems of unbalanced crop breeding data, missed identification of superior varieties, poor environmental adaptability, and a disconnect between the crop breeding samples and the actual phenotypic patterns of crops under salt stress, which leads to low reliability of breeding evaluation in general saline-alkali land breeding evaluation methods. This is achieved by introducing environmental adaptability differential control and integer feature range control to avoid generating only single salt tolerance level samples; by introducing a stress differential coefficient for objective function optimization, the generated integer data is made more consistent with actual physiological patterns; and by using a phenotypic sample generator feature adjustment function combined with a salt tolerance level decay function, the generated samples are highly adapted to the actual salt stress patterns in saline-alkali land, thereby improving the reliability of subsequent crop breeding evaluation.

[0059] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the crop breeding evaluation model is established based on a deep learning model. It captures salt-tolerance-related features through multi-view extraction of salt-tolerance features, adaptive selection, and hierarchical fusion of salt-tolerance features. Combined with the salt-tolerance level difference loss function, it includes two stages: training and testing. Finally, it outputs the crop salt-tolerance evaluation level and establishes the crop breeding evaluation model. This solves the problems of complex phenotypic-environmental characteristics of salt-tolerant crops in saline-alkali land, few salt-tolerance level samples, and difficulty in classifying adjacent salt-tolerance levels. The model updates parameters through gradient descent algorithm and adjusts initial parameters through particle swarm optimization algorithm. Specifically, it includes the following:

[0060] Step S31: Feature extraction network construction; input augmented crop breeding fusion data. Multi-view processing is performed: global average pooling is used to extract the overall feature distribution and generate a global view vector. This includes the overall trend of tiller number in salt-tolerant crops; using a global max-pooling perspective, it extracts local salient features and generates local perspective vectors. This includes extremely salt-tolerant individuals that maintain 80% leaf water content under salt stress; channel weights for two types of features are obtained through fully connected layers and sigmoid activation, highlighting salt-tolerance-related features. The weights are multiplied by the original features and then summed to output the fused feature F(M). i ); By using multi-view pooling, the overall distribution pattern and local key features are captured respectively, and the weights are dynamically strengthened to enhance the salt-tolerant core features and filter out irrelevant features;

[0061] Global view vector Local view vector Where N is the number of features in the c-th channel, and k is the feature index; It is the k-th feature value of the c-th channel; and These are the 1st and Nth eigenvalues ​​of the c-th channel, respectively; the channel weight calculation formula is expressed as: ; ; Fusion features ;in, It is the Sigmoid function; and These are learnable network parameters from a global perspective; and These are the channel weights for the global view and the local view, respectively; and These are learnable network parameters from a local perspective;

[0062] Step S32: Adaptive feature selection; based on the fused features F(M) output from step S31 through a hierarchical attention mechanism. iAn importance assessment is performed to select high-contribution features for the fusion stage; a feature importance function I(·) is constructed, using the following formula: ;in, It is F(M) i The low k features of the l-th layer; It is the activation intensity of features within the layer; ; It is the set of the k-th feature values ​​of all samples in the l-th layer; It is the k-th feature value of the i-th sample in the l-th layer; and a dynamic threshold is applied, with the threshold adjusted with each training round to retain high-contribution features; dynamic threshold Represented as: For the k-th feature of the i-th sample in the l-th layer, if the feature importance function value is higher than the dynamic threshold, it is determined to be a high-contribution feature and is retained; otherwise, it is removed. It refers to training rounds;

[0063] Different levels of features have different values ​​for salt tolerance classification; the shallowest feature l=1 is the original phenotypic layer, corresponding to basic phenotypic features including plant height; l=2 is the environmental feature layer, which is a single type of original environmental feature, corresponding to basic environmental features including soil salinity; l=3 is the primary fusion layer, which is a preliminary phenotypic-environmental fusion feature, corresponding to preliminary phenotypic-environmental interaction features including tiller number × pH value; the deepest feature l=4 is the advanced fusion layer, corresponding to deep phenotypic-environmental interaction features including salinity × leaf water content;

[0064] Step S33: Layered and refined fusion of salt tolerance features; The high-contribution features selected in Step S32 are fused layer by layer using salt tolerance features. This is achieved through three steps: dimension unification, hierarchical association, and feature condensation, outputting a final feature vector that balances the details of basic indicators with the patterns of derived indicators. Salt-tolerant crop classification relies on both individual plant phenotypic and population phenotypic-environment-derived patterns, but the dimensional differences between different levels are significant. After unifying the dimensions through a fully connected layer, a progressive fusion method is used to provide detailed support for higher-level derived features from lower-level basic features, avoiding feature association breaks. The formula used to adjust the dimensions in the fully connected layer is: The formula used for progressive fusion is: The formula used for the final feature output is: ;in, is the feature of the i-th sample after adjustment in the l-th layer; ReLU(·) is the ReLU function; FC(·) is a fully connected layer; It is the high contribution feature of the i-th sample in the l-th layer after screening in step S32; It is the fusion feature of the i-th sample at the highest layer; It is global average pooling; It is the final feature vector of the sample;

[0065] Step S34: Network classification; Calculate the feature center of each salt tolerance level using the data-augmented samples, calculate the Euclidean distance between the feature of the sample to be predicted and the feature centers of each class, and the level with the closest distance is the prediction result; the formula used to calculate the feature center of the salt tolerance level is: The formula used to calculate classification probability is: ;in, and These are the salt tolerance feature centers of the j-th and u-th salt tolerance levels, respectively; K is the total number of samples for the j-th salt tolerance level. is the final feature vector of the i-th sample in the j-th class after step S33; x is the sample to be predicted; It is the final feature vector output by the sample to be predicted in step S33; It is the Euclidean distance; It is the probability that the sample to be predicted belongs to the j-th salt tolerance level; in the breeding of salt-tolerant crops, there are few samples of highly salt-tolerant varieties. Salt tolerance level information is concentrated by salt tolerance level feature centers to construct stable centers; and by projecting the samples to a feature space with controllable distance, the dependence on large-scale labeled data is reduced.

[0066] Step S35: Loss function design; construct the basic loss function, using cross-entropy loss. Based on, it is expressed as: Where Q is the total number of samples participating in training; b is the salt tolerance level index; and These are the true label and the probability of predicting the i-th sample as grade c, respectively; the true label is one-hot encoded, and when sample i is grade 5, then... , ; Constructing a loss due to differences in salt tolerance levels This is used to enhance spatial separation, increasing the distance between the centers of adjacent salt tolerance grades. The formula used is: Introducing environmental edge loss, we construct adaptive edges based on environmental differences, and for each sample i, we apply these edges to each... Define dynamic edges , represented as: Environmental edge loss Represented as: ;in, It is the salt tolerance characteristic center of the kth salt tolerance level; It is the true level of the i-th sample. The salt tolerance rating characteristic center; It is the basic edge; It is the environment scaling factor; , and It is the environmental average of the sample, the normalized mean; yes The corresponding environmental average value; adjacent salt tolerance levels are difficult to distinguish in extreme environments, through Automatically enlarging edges significantly improves resolution in high-salinity areas; the final loss L is represented as: ;in, and These are the salt tolerance characteristic centers for Class b and Class b+1 salt tolerance grades, respectively; It is the loss weighting coefficient.

[0067] By performing the above operations, this scheme addresses the problems of redundant interference from crop features, subtle phenotypic differences between adjacent salt tolerance levels, and the exacerbation of feature overlap in extreme saline-alkali environments, which can lead to misjudgment of highly salt-tolerant varieties and poor crop breeding evaluation results. Instead, it addresses these issues by employing a hierarchical attention mechanism. This mechanism evaluates the contribution of features at different levels of crop breeding data according to a feature importance function, combines dynamic thresholding to screen high-value features, eliminates redundant features, and strengthens core features related to salt tolerance. Furthermore, it introduces environmental edge loss to improve the distinguishability between adjacent salt tolerance levels in extreme saline-alkali environments and constructs a salt tolerance level difference loss mechanism to increase the distance between the feature centers of adjacent salt tolerance levels, reducing misjudgment of superior varieties. Ultimately, this improves the effectiveness of crop breeding evaluation.

[0068] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the intelligent crop breeding evaluation is based on the salt tolerance level evaluation model established in step S3. Combining the characteristics of the saline-alkali land environment and the breeding objectives, the breeding evaluation of highly salt-tolerant crops is completed, specifically including the following:

[0069] Step S41: Salt tolerance potential assessment of candidate breeding materials; collect data from all dimensions of existing breeding materials, input the standardized data into the assessment model of step S3, and output the salt tolerance level prediction results of each material as the basis for salt tolerance potential assessment.

[0070] Step S42: Targeted screening for adaptability to saline-alkali land environment; Based on the environmental data of the target saline-alkali area and the salt tolerance level prediction results of step S41, material-environment recommendations are made; When the target environment is a moderate to severe stress area with a salt concentration of 4-6 mS / cm, breeding materials of level 4 are selected as candidates for main varieties; When the target environment is a severe stress area with a salt concentration >6 mS / cm, breeding materials of level 5 are selected as candidates for main varieties and as parents for salt tolerance gene donation in hybridization breeding.

[0071] Example 6, see Figure 2Based on the above embodiments, this embodiment provides an intelligent breeding system for salt-tolerant crops in saline-alkali land, including a crop data acquisition module, a crop data improvement module, a crop breeding evaluation model establishment module, and a crop intelligent breeding module.

[0072] The crop data acquisition module collects crop breeding data and constructs an initial crop breeding dataset.

[0073] The crop data enhancement module enhances the initial crop breeding dataset by floating-point data enhancement, introduces integer data enhancement for environmental adaptability differential regulation, and introduces stress differential coefficient to optimize the objective function, thus obtaining the final crop breeding fusion data.

[0074] The crop breeding evaluation model building module is based on crop breeding fusion data and deep learning model. It establishes crop breeding evaluation model by extracting salt tolerance features from multiple perspectives, adaptive selection, hierarchical fusion, and combining salt tolerance level difference loss function.

[0075] The intelligent crop breeding module is based on a crop breeding evaluation model and combines the environmental characteristics of the target saline-alkali land to carry out intelligent crop breeding.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0078] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intelligent breeding and evaluation of salt-tolerant crops in saline-alkali land, characterized in that: The method includes the following steps: Step S1: Crop data collection; collect crop breeding data and construct an initial crop breeding dataset; Step S2: Crop data enhancement; The initial crop breeding dataset is augmented with floating-point data, augmented with integer data for environmental adaptability differential regulation, and the objective function is optimized by introducing stress differential coefficients to obtain the final crop breeding fusion data; Step S3: Establish a crop breeding evaluation model; Based on the fusion data of crop breeding, and using a deep learning model as a foundation, establish a crop breeding evaluation model; Step S4: Intelligent crop breeding assessment; Intelligent crop breeding is carried out based on the crop breeding assessment model and combined with the environmental characteristics of the target saline-alkali land. In step S2, the objective function optimization is to construct a realism discrimination loss that includes environmental adaptability differential adjustment for integer data augmentation. , represented as: ;in, It is the distribution of real integer samples; It is an integer sample; It is the authenticity judgment output of the phenotypic authenticity discriminator; It is the expectation over a specific distribution; It is the loss of environmental adaptability differential regulation; It is the distribution of noise z; It is a phenotypic sample generator based on noise z and salt tolerance level labels. Generate integer samples; construct a salt tolerance level classification loss. , represented as: ; the generation of new integer samples is denoted as ;in, It is the output of the phenotypic authenticity discriminator for classifying the salt tolerance level of real integer samples; It is a phenotypic authenticity discriminator that classifies salt tolerance levels for generated integer samples; it introduces a stress level difference coefficient. Represented as: Define the feature adjustment function for the phenotypic sample generator as follows: ;in, It is a function of salt tolerance level decay. e represents soil electrical conductivity, the value before standardization. and These are the adjusted and unadjusted floating-point phenotypic features, respectively. It represents the crop's salt tolerance level; the final loss function is integer-based data augmentation. Represented as: ; and It is to enhance weight; It is the summation of the differences in electrical conductivity of different soils; It is the mean of floating-point phenotypic characteristics in real samples when the soil electrical conductivity is e.

2. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 1, characterized in that: In step S2, the crop data improvement specifically includes the following: Step S21: Floating-point data augmentation; embed the salt tolerance level label into random noise z, input standardized floating-point data, and generate new floating-point samples through a traditional phenotypic sample generation network, denoted as z. ;in, and These are the newly generated floating-point crop breeding data and their corresponding salt tolerance level labels; Step S22: Integer data augmentation; embed salt tolerance level labels into random noise z, input standardized integer data; introduce environmental adaptability differential control, and add integer feature range control, expressed as: ;in, It is the environmental adaptability differential control coefficient; It is the distribution of interpolated samples; It is an interpolated sample between real samples and generated samples, and is an intermediate sample used for environmental adaptability differential regulation. This is the distribution of the interpolated samples; It is the gradient of the phenotypic authenticity discriminator on the interpolated sample; It is the L2 norm; Step S23: Objective function optimization; Step S24: Data Fusion; Based on the data concatenation operation, merge the generated floating-point sample with the real floating-point sample, denoted as... The generated integer sample is merged with the real integer sample, denoted as . Integrating crop breeding data Input a DNN classifier with salt tolerance level as the classification target, and optimize the network parameters through backpropagation. If the crop breeding fusion data is not found, the enhancement effect will be discarded; the classification accuracy will be used to verify the enhancement effect; among them, and These are floating-point samples and integer samples from the synthesized crop breeding dataset, respectively. and It corresponds to the label; set a classification accuracy threshold. If the classification accuracy is higher than the classification accuracy threshold, then the crop breeding fusion data obtained at this time will be used as the final crop breeding fusion data.

3. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 2, characterized in that: In step S3, the establishment of the crop breeding evaluation model is based on a deep learning model; the model updates parameters through gradient descent and adjusts initial parameters through particle swarm optimization; specifically, it includes the following: Step S31: Feature extraction network construction; input augmented crop breeding fusion data. Multi-view processing is performed: global average pooling is used to extract the overall feature distribution and generate a global view vector. This includes the overall trend of tiller number in salt-tolerant crops; using a global max-pooling perspective, it extracts local salient features and generates local perspective vectors. This includes extremely salt-tolerant individuals that maintain 80% leaf water content under salt stress; channel weights for two types of features are obtained through fully connected layers and sigmoid activation, highlighting salt-tolerance-related features. The weights are multiplied by the original features and then summed to output the fused feature F(M). i ); Step S32: Adaptive feature selection; based on the fused features F(M) output from step S31 through a hierarchical attention mechanism. i An importance assessment is performed to select high-contribution features for the fusion stage; a feature importance function I(·) is constructed, using the following formula: ;in, It is F(M) i The low k features of the l-th layer; It is the activation intensity of features within the layer; ; It is the set of the k-th feature values ​​of all samples in the l-th layer; It is the k-th feature value of the i-th sample in the l-th layer; and dynamic thresholding is performed, with the dynamic threshold... Represented as: For the k-th feature of the i-th sample in the l-th layer, if the feature importance function value is higher than the dynamic threshold, it is determined to be a high-contribution feature and is retained; otherwise, it is removed. It refers to training rounds; Step S33: Layered and refined fusion of salt tolerance features; The high-contribution features selected in Step S32 are fused layer by layer based on salt tolerance features. After unifying the dimensions through a fully connected layer, a progressive fusion is used to allow lower-level basic features to provide detailed support for higher-level derived features, avoiding feature correlation breaks; the formula used to adjust the dimensions of the fully connected layer is: The formula used for progressive fusion is: The formula used for the final feature output is: ;in, is the feature of the i-th sample after adjustment in the l-th layer; ReLU(·) is the ReLU function; FC(·) is a fully connected layer; It is the high contribution feature of the i-th sample in the l-th layer after screening in step S32; It is the fusion feature of the i-th sample at the highest layer; It is global average pooling; It is the final feature vector of the sample; Step S34: Network classification; Step S35: Loss function design.

4. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 3, characterized in that: In step S3, the network classification calculates the feature center of each salt tolerance level using the data-augmented samples, calculates the Euclidean distance between the feature of the sample to be predicted and the feature centers of each class, and the level with the closest distance is the prediction result; the formula used to calculate the salt tolerance level feature center is: ; The formula used to calculate classification probability is: ;in, and These are the salt tolerance feature centers of the j-th and u-th salt tolerance levels, respectively; K is the total number of samples for the j-th salt tolerance level. is the final feature vector of the i-th sample in the j-th class after step S33; x is the sample to be predicted; It is the final feature vector output by the sample to be predicted in step S33; It is the Euclidean distance; It is the probability that the sample to be predicted belongs to the j-th salt tolerance level.

5. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 4, characterized in that: In step S3, the loss function design involves constructing a base loss, using cross-entropy loss. Based on, it is expressed as: Where Q is the total number of samples participating in training; b is the salt tolerance level index; and These are the true label and the probability of predicting salt tolerance level c for the i-th sample, respectively; construct the salt tolerance level difference loss. The formula used is: Introducing environmental edge loss, we construct adaptive edges based on environmental differences, and for each sample i, we apply these edges to each... Define dynamic edges , represented as: Environmental edge loss Represented as: ;in, It is the salt tolerance characteristic center of the kth salt tolerance level; It is the true level of the i-th sample. The salt tolerance rating characteristic center; It is the basic edge; It is the environment scaling factor; , and It is the environmental average of the sample, the normalized mean; yes The corresponding environmental average; the final loss L is expressed as: ;in, and These are the salt tolerance characteristic centers for Class b and Class b+1 salt tolerance grades, respectively; It is the loss weighting coefficient.

6. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 5, characterized in that: In step S4, the intelligent crop breeding assessment is based on the salt tolerance level assessment model established in step S3, combined with the characteristics of the saline-alkali land environment and breeding objectives, to complete the breeding assessment of highly salt-tolerant crops, specifically including the following: Step S41: Salt tolerance potential assessment of candidate breeding materials; Full-dimensional data collection was conducted on existing breeding materials, and after standardization, the data was input into the evaluation model in step S3. The predicted salt tolerance level of each material was output as the basis for salt tolerance potential assessment. Step S42: Targeted screening of saline-alkali land environmental adaptability; Based on the environmental data of the target saline-alkali area and the salt tolerance level prediction results of step S41, material-environment recommendations are made.

7. The intelligent breeding and evaluation method for salt-tolerant crops in saline-alkali land according to claim 6, characterized in that: In step S1, the crop data collection involves collecting crop breeding data, including crop salt tolerance-related phenotypic data and saline-alkali land environment-related data. The crops were labeled with salt tolerance levels; the acquired data were standardized to construct an initial crop breeding dataset.

8. A smart breeding and evaluation system for salt-tolerant crops in saline-alkali land, used to implement the smart breeding and evaluation method for salt-tolerant crops in saline-alkali land as described in any one of claims 1-7, characterized in that: It includes a crop data acquisition module, a crop data improvement module, a crop breeding evaluation model establishment module, and a crop intelligent breeding module; The crop data acquisition module collects crop breeding data and constructs an initial crop breeding dataset. The crop data enhancement module enhances the initial crop breeding dataset by floating-point data enhancement, introduces integer data enhancement for environmental adaptability differential regulation, and introduces stress differential coefficient to optimize the objective function, thus obtaining the final crop breeding fusion data. The crop breeding evaluation model building module is based on crop breeding fusion data and deep learning model. It establishes crop breeding evaluation model by extracting salt tolerance features from multiple perspectives, adaptive selection, hierarchical fusion, and combining salt tolerance level difference loss function. The intelligent crop breeding module is based on a crop breeding evaluation model and combines the environmental characteristics of the target saline-alkali land to carry out intelligent crop breeding.

Images