Interpretable breeding recommendation method based on crop trait and gene mapping analysis

Abstract

The present invention provides an interpretable breeding recommendation method based on crop trait and gene mapping analysis, which belongs to the field of breeding recommendation technology based on computer data processing. First, multi-angle images of crops are collected, and their corresponding SNP genomic markers are collected at the same time; secondly, an entity-level crop trait feature extraction model is designed to realize the identification of key traits and spatial position output of each crop. Subsequently, models are constructed for continuous traits and discrete traits respectively, and a contribution score table of SNP sites to each trait is obtained, and the trait target evaluation baseline data is generated by combining the reference genotype and the unit contribution change value. Finally, combined with the target trait configuration input by the user, a multi-objective loss function is constructed and optimized to generate the optimal SNP gene combination scheme. An intelligent breeding method with the ability of trait identification, gene contribution quantification and interpretable gene combination optimization is provided, which meets the comprehensive technical needs of modern breeding engineering.

Description

Technical Field

[0001] The present invention belongs to the technical field of breeding recommendation based on computer data processing, and in particular relates to an interpretable breeding recommendation method based on crop trait and gene mapping analysis. Background Art

[0002] In modern agricultural breeding systems, the precise identification of structural traits and the effective modeling of associations between genomes have become the key technical foundation for achieving complex trait improvements in crops, such as high quality, high yield, and wide adaptability. This is especially true in typical multi-trait crops such as sorghum, where trait expression is regulated by multiple genes and often exhibits significant spatiotemporal heterogeneity and environmental sensitivity. Therefore, how to efficiently and accurately obtain plant structural trait information and analyze its genetic control mechanisms at the molecular level has become a major bottleneck restricting the improvement of breeding efficiency. With the development of single nucleotide polymorphism (SNP) molecular markers, a series of new trait assessment and gene analysis methods have emerged in the breeding field, attempting to alleviate the problems of low efficiency and high subjectivity of traditional methods. However, existing methods still face challenges such as limited trait identification accuracy and weak genetic explanatory power, making it difficult to support the needs of targeted breeding for practical application scenarios.

[0003] At present, the main methods of crop intelligent breeding include the following:

[0004] Trait-gene association analysis methods based on statistical models: Traditional crop breeding widely uses statistical model methods such as genome-wide association analysis (GWAS) or quantitative trait loci (QTL) positioning to establish the correlation between trait phenotypes and genotypes. This type of method relies on a large amount of measured phenotypic data and high-density genotypic data to identify genetic loci that are significantly associated with traits through linear mixed models or regression models. However, this type of method is usually only applicable to data sets with stable phenotypes and large sample sizes, and it is difficult to handle complex interactions between multiple traits and the effects of environmental interference. At the same time, its output is mostly a list of statistically significant sites, lacking the ability to reversely reason towards trait targets, making it difficult to directly guide actual genotype configuration and breeding combination optimization;

[0005] Image recognition-based phenotypic trait extraction methods: In recent years, with the development of remote sensing and computer vision technologies, some studies have attempted to use image recognition models (such as CNN and UNet) to automatically extract crop morphological traits, such as plant height, leaf area, and number of ears, from drone images to replace traditional manual measurement methods. This type of method significantly improves the efficiency and scale of phenotypic collection and is suitable for large-scale crop monitoring and rapid phenotypic screening. However, this type of method usually focuses on the identification and extraction of traits and lacks the ability to jointly model with genomic information. It is unable to complete causal relationship analysis between traits and genes, let alone provide locus-level intervention recommendations or genetic combination recommendations for breeding goals;

[0006] Gene prediction and material screening methods based on machine learning: Some methods attempt to apply deep neural networks to crop breeding scenarios, such as predicting the mapping relationship between genotypes and target traits through multi-layer perceptrons, graph neural networks, or Transformer structures. These methods are able to model complex relationships between nonlinear and multivariable variables, and are superior to traditional linear models in prediction accuracy. However, most of these methods can only provide "black box" prediction results, lack support for the interpretability of the prediction results, do not have the ability to identify key control sites, and are difficult to meet the traceability requirements in variety improvement. In addition, some methods fail to effectively integrate user trait goals or breeding scenario requirements, and cannot achieve target-oriented customized gene combination design.

[0007] In summary, existing crop intelligent breeding methods either lack the ability to reverse model traits to genes or are insufficiently interpretable, making it difficult to provide recommendation solutions with causal reasoning and target optimization capabilities under complex breeding needs. Summary of the Invention

[0008] To address the above issues, the present invention proposes an interpretable breeding recommendation method based on crop trait and gene mapping analysis, comprising the following steps:

[0009] S1, collect multi-angle remote sensing images of crop fields and simultaneously collect SNP genomic marker information of corresponding plant leaf tissues;

[0010] S2 inputs the multi-angle remote sensing images collected by S1 into the trained entity-level crop trait feature extraction model, and outputs the spatial position results and trait recognition results of each plant;

[0011] S3, combining the spatial location results of the plants to match the SNP genomic marker information of each plant with the corresponding trait identification results to construct a gene-trait pairing dataset;

[0012] S4, based on the trait recognition results, a Lasso regression model for continuous traits and a random forest classification model for discrete traits are constructed respectively, and the model training is performed based on the paired data set obtained in S3 to obtain the parameters of the two models after training;

[0013] S5: Calculate the contribution score of each SNP site to the trait based on the two model parameters, concatenate the contribution scores of all traits and construct a structured contribution output table; at the same time, count the reference SNP combinations and reference trait results in the current scenario, and derive the estimated trait change value corresponding to the unit contribution based on the contribution, thus obtaining the baseline data for the trait target assessment;

[0014] S6, based on the obtained structured contribution output table, trait target evaluation baseline data and user target trait configuration, constructs a multi-objective loss function and performs optimization search to output the optimal SNP gene combination solution.

[0015] Preferably, the multi-angle remote sensing image uses the top view as the main image perspective, and the oblique view and side view as auxiliary image perspectives. During data collection, the structural characteristics of each sorghum plant in multiple spatial dimensions are recorded, including plant height, , average ear length , spike type category , crown width , effective branch number , stem thickness , coloring properties .

[0016] Preferably, when constructing the dataset, for each sorghum plant that has completed spatial annotation and trait annotation, the corresponding genomic information collection and standardized coding process is carried out, specifically including:

[0017] First, leaf tissue samples were collected from each labeled plant. Then, genomic DNA was extracted using standard plant DNA extraction methods, and raw molecular marker data was obtained through SNP chip typing. After that, a high-quality SNP locus set significantly associated with the target trait was screened through population variability analysis. ;in, Indicates the SNP sites, and ;in is the number of pre-selected SNPs;

[0018] Then, for each SNP site Using the unified 0 / 1 / 2 numerical code, we get Genotype representation of SNP loci , , and 0 represents inferior homozygous aa, 1 represents heterozygous Aa, and 2 represents superior homozygous AA; therefore, the All gene combinations of the plants are ,and .

[0019] Preferably, the entity-level crop trait feature extraction model includes a plant space recognition channel, a structural trait perception channel, and a trait fusion recognition channel;

[0020] The plant space recognition channel is based on the main viewing angle image As input, it is used to realize the spatial position perception of individual plants and the extraction of coarse-grained structural contours, which is suitable for the identification of dominant traits of spatial structure;

[0021] The structural feature perception channel takes an oblique view With side view The auxiliary image is used as input to model the three-dimensional structural characteristics of the plant, supplementing the missing structural characteristics information caused by occlusion or angle limitation in the main view, and is suitable for extracting fine-grained continuous characteristics;

[0022] The trait fusion recognition channel is used to jointly model the output features of the plant space recognition channel and the structural trait perception channel. First, the output features of the two channels are spliced ​​together as the input of the fusion recognition channel to obtain the channel joint features. ; Then input into the cross-view fusion convolution layer to generate a unified feature representation structural feature discriminant feature ; Then, feature compression is performed through two layers of convolution to obtain the compressed features , The data are sent to the spatial position recognition head and the structural trait recognition head respectively. The spatial position recognition head outputs the spatial position result of each plant, and the structural trait recognition head outputs the trait recognition result.

[0023] Preferably, the plant space recognition channel first extracts the initial plane features through the pre-trained ResNet34 network to obtain the initial texture features of the plant main view , then the characteristics After five positions-enhancement-pooling structure groups, the spatial position convolution layer and the saliency enhancement convolution layer realize the enhanced extraction of plant boundaries, contours and salient areas; each output is compressed by the maximum pooling layer, and the position-aware enhanced features are output. ; will eventually Unify the dimensions through the Conv1 convolution layer and output the main feature results of the channel .

[0024] Preferably, the structural feature perception channel first extracts multi-angle structural features through the pre-trained Swin-Tiny model to obtain multi-view structural encoding features Then, the trait feature selection layer is introduced to construct multiple trait channel branches corresponding to multiple types of structural traits, and the trait specific channel responses are extracted in parallel to generate trait significant features. ; Afterwards, features The image is fed into the cross-view fusion convolution layer to align and fusion model the structural information in the side view and oblique image, enhance the three-dimensional consistency representation and output the view fusion structural features. The fusion result is compressed by average pooling to form a fusion compression feature The above trait selection-fusion-pooling combination is repeated 5 times, and the final feature is unified in dimension through Conv1 convolution to output the auxiliary features of the structural trait perception channel .

[0025] Preferably, the S4 is specifically:

[0026] Trait identification results include plant height prediction results , average ear length prediction results , spike type category prediction results , crown width prediction results , prediction results of effective branch number , stem thickness prediction results , coloring trait prediction results ;in, , , , , It is a continuous trait. , It is a discrete trait;

[0027] For continuous traits, gene annotation As input, continuous traits are parallel output, and a Lasso linear regression model is constructed; this model can output multi-path prediction results of the target continuous traits in parallel;

[0028] For discrete traits, gene annotation As input, discrete traits are parallel outputs, and a random forest classifier is constructed; this model can output multi-path classification results of target discrete traits in parallel;

[0029] Based on the constructed gene-trait pairing dataset, the Lasso and random forest classifier models were trained respectively to obtain the trained Lasso and random forest classifier model parameters.

[0030] Preferably, the S5 includes:

[0031] For continuous traits, the regression coefficient vector in the trained Lasso model is used as the linear contribution of each site, and the regression coefficient of all SNP sites is the contribution score of the trait;

[0032] For discrete traits, the trained random forest classifier outputs a feature importance score as the information gain contribution of each SNP in the classification task. The score is based on the frequency of the SNP being selected as a node splitting variable in all decision trees, as well as the sum of the Gini impurity reduction values ​​brought about by the split operation. After global accumulation and normalization, the contribution score of each SNP site to the trait is obtained.

[0033] The complete SNP site contribution score corresponding to each trait is obtained, that is, , , , , , , , respectively, plant height SNP contribution score, average ear length SNP contribution score, crown width SNP contribution score, effective branch number SNP contribution score, stem diameter SNP contribution score, ear type category SNP contribution score, coloring trait SNP contribution score; and the contribution scores of all traits are spliced ​​to obtain the SNP contribution score of the complete trait ;

[0034] SNP contribution score based on the complete traits obtained , construct a standardized contribution structured output table; each table corresponds to a trait, and the table records all SNP site numbers corresponding to the trait and their contribution scores.

[0035] Preferably, the specific process of obtaining the trait target assessment baseline data is:

[0036] Based on the trained Lasso regression and random forest classification models, statistical analysis was performed on all plant samples to extract the most common SNP combinations. As the basic reference genotype configuration in the current scenario, and calculate its corresponding reference traits As a reference trait; at the same time, combined with the model mapping relationship between contribution and measured trait increment, a set of trait change estimates corresponding to unit contribution is constructed for all traits ; Reference genotype configuration , reference traits , and the set of trait change estimates Together constitute the baseline data for trait target assessment .

[0037] Preferably, the S6 specifically includes:

[0038] Users set multiple target trait requirements , and specify the priorities of different traits at the same time, and parse them into target trait configuration matrix; based on the set target trait requirements configuration , output structured contribution table and trait target assessment baseline data , construct the evaluation function between site and trait:

[0039] First, the reference SNP genotype configuration and its corresponding reference traits As the initial state of optimization; then, according to the estimated value of the trait change corresponding to the unit contribution , and the contribution value of each site Calculate the genotype of each SNP site in different states The actual physical increment of the target trait is quantified as , in order to obtain its physical fitting value for the target trait and the target value of the trait set by the user Perform MSE error calculation and sum the preset weights of different traits to obtain the objective function result ;

[0040] Use simulated annealing algorithm to solve nonlinear optimization problems and output the optimal SNP combination solution that meets trait requirements .

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) Entity-level crop trait feature extraction model: Combining multi-view image collaborative modeling with channel branching structure, a crop trait recognition model with a three-channel architecture is proposed. The plant spatial recognition channel is used to perceive the target boundary and distribution. The structural trait perception channel is used to achieve saliency enhancement and semantic decoupling of different traits. Finally, structural modeling is unified in the trait fusion recognition channel to achieve high-precision trait structured recognition in complex occlusion and angle-varying environments.

[0043] (2) Trait-gene modeling method with interpretability and quantitative contribution analysis capabilities: For continuous and discrete trait characteristics, Lasso regression and random forest modeling strategies are introduced to construct multi-path prediction models between SNP sites and target traits, and the numerical contribution of each SNP site is extracted by combining regression coefficients, feature importance, etc., to achieve causal and interpretable mapping from genotype to trait phenotype, and at the same time, a contribution-trait change unit mapping is constructed to provide data support for quantitative breeding;

[0044] (3) SNP gene combination optimization and recommendation mechanism for target trait configuration: Combining the target trait values ​​and priority configurations input by the user, a multi-objective loss function is constructed, and a simulated annealing algorithm is introduced to achieve joint optimization of SNP genotype combinations. Finally, an interpretable and controllable genetic locus combination scheme is output, and the optimal candidate materials are screened by comparing with the existing germplasm database, forming a closed-loop target-oriented breeding recommendation path, and realizing intelligent reverse deduction from breeding goals to gene configuration;

[0045] (4) It has a wide range of applications and can be used for a variety of crops such as sorghum, wheat, and rice. It only requires adaptive selection of the corresponding crop trait structure, such as plant height, average ear length, ear type, crown width, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1This is a flow chart of the overall technical route of the present invention.

[0047] Figure 2 This is the network structure diagram of the entity-level crop trait feature extraction model of the present invention.

[0048] Figure 3 This is a flowchart of the crop trait and genome interpretability modeling of the present invention.

[0049] Figure 4 The MAPE (%) test results for the prediction results of the five traits in the embodiment are shown in FIG.

[0050] Figure 5 : The R² test results for the prediction results of the five traits in the embodiment.

[0051] Figure 6 This is a heat map of the contribution of ten representative SNP sites to seven traits in the example. DETAILED DESCRIPTION

[0052] The present invention proposes an interpretable breeding recommendation method based on quantitative analysis of crop traits and gene mapping. This method constructs a full-process breeding intelligent modeling from image perception to gene recommendation. First, a multi-angle remote sensing image acquisition strategy is used to obtain the top view, oblique view and side view of each crop in the crop field, and the spatial position and trait annotation of each crop are performed. At the same time, the corresponding SNP genomic molecular markers are collected to construct a standardized image-trait-gene joint data set; secondly, an entity-level crop trait feature extraction model is designed, which includes a plant spatial recognition channel, a structural trait perception channel and a trait fusion recognition channel to achieve the key trait identification and structured output of each crop. Subsequently, a multi-trait interpretable modeling module is constructed with the structural trait recognition results and genomic features as input. Lasso regression models and random forest models are constructed for continuous traits and discrete traits respectively. The contribution scores of SNP sites to each trait are obtained to form a structured output table, and the baseline data for trait target evaluation is generated by combining the reference genotype and the unit contribution change value. Finally, based on the user-entered target trait configuration, a multi-objective loss function is constructed and optimized using a simulated annealing algorithm to generate the optimal SNP gene combination scheme. This is then compared with the material database to screen recommended germplasms, enabling intelligent breeding recommendations based on interpretable genetic contributions. This system has a wide range of applications and can be used in a variety of crops, including sorghum, wheat, and rice, requiring only adaptive selection of the corresponding crop trait structure.

[0053] The specific implementation process of the present invention is described in detail below with reference to specific embodiments.

[0054] This embodiment takes sorghum crops as an example. The overall process is as follows Figure 1 shown.

[0055] Multi-angle remote sensing image data acquisition and dataset construction: Multi-angle remote sensing image acquisition was performed on sorghum fields, with the top view as the primary image perspective and the oblique view and side view as auxiliary image perspectives. The structural traits of each sorghum plant in multiple spatial dimensions were recorded. Subsequently, the spatial position and dominant structural traits of each plant were annotated, and the corresponding leaf tissue was collected to extract SNP genomic marker information. An image-trait dataset was then constructed.

[0056] Entity-level crop trait feature extraction model construction: A three-channel crop trait recognition model was designed. This model includes a plant spatial recognition channel, a structural trait perception channel, and a trait fusion recognition channel. Ultimately, it predicts the structural traits of each plant. End-to-end training is performed using the constructed image-trait dataset, outputting structured trait recognition results.

[0057] Modeling of crop traits and genome interpretability: Using the output trait identification results and the extracted standardized SNP genotypes as input, a Lasso regression model for continuous traits and a random forest classification model for discrete traits are constructed to achieve modeling and prediction of multiple traits; and a structured output table is further generated. On this basis, the reference SNP combinations and reference trait results in the current scenario are statistically analyzed, and the estimated trait change corresponding to the unit contribution is derived in combination with the contribution, forming the baseline data for trait target evaluation for subsequent breeding optimization.

[0058] Trait-target-oriented SNP gene combination generation and breeding recommendations: Users set target trait requirements and weight preferences. The system constructs a multi-objective loss function based on the output contribution structure, trait change estimates and reference combinations, and uses a simulated annealing algorithm for joint optimization to search for the optimal SNP genotype combination scheme; finally, the optimized gene combination is compared and matched with the sorghum material database, and the closest germplasm individual is output as the recommended breeding gene reference scheme.

[0059] 1. Multi-angle remote sensing image data acquisition and dataset construction

[0060] To reduce the limitations of single-perspective recognition of sorghum field crops due to problems such as plant shape obstruction, uneven lighting, and missing local perspectives, the present invention adopts a multi-angle remote sensing image acquisition strategy, using a top-down view as the primary image perspective and combining auxiliary angle images to achieve a complete record of crop structural traits and establish a foundation for identification. At the same time, combined with manual labeling, spatial numbering, and genotype acquisition processes, a standardized, structured image-trait-gene joint dataset for training is constructed.

[0061] 1. Multi-angle remote sensing image acquisition: A multi-angle camera system equipped with a drone is used to systematically acquire remote sensing images of sorghum fields. Image perspectives are divided into two categories: primary image perspective and auxiliary image perspective:

[0062] (1) The top view is the main perspective image, which is collected by a camera installed vertically downward. It is used for spatial position calibration and structural feature annotation of the plant and serves as the reference image for all subsequent annotation tasks.

[0063] (2) Oblique views and side views are used as auxiliary image perspectives to record crop side structure, plant outline, and ear morphology. Specifically, the oblique view is captured by a camera at a certain angle (30 to 60 degrees) relative to the ground, and the side view is captured by a camera approximately parallel to the ground from the side of the ridge, thereby enhancing the model's perception of spatial structure.

[0064] The above collected images must ensure that the same crop can be identified in images from different angles; finally, the main perspective image from the top is collected as , the auxiliary images of the oblique view and the side view are and , the three perspective images together constitute a complete multi-angle remote sensing image combination .

[0065] 2. Plant spatial position and shape information annotation: Using the main view image as the standard view, and the oblique view and side view as auxiliary images, organize staff to use auxiliary annotation tools to annotate all sorghum plants in the image; specifically, the spatial position of each plant (Reflected by the center coordinates, length and height of the rectangle), plant height , average ear length , spike type category (0 for compact, 1 for loose), crown width , effective number of branches , stem thickness , coloring properties ;

[0066] Therefore, the trait information corresponding to each plant is ; The spatial position and trait annotation results are ; Mark all plants in the top view to get complete marking results ,in Indicates the The spatial location and trait annotation information of each plant, and ; The total number of plants.

[0067] 3. Plant genome molecular markers: For each sorghum plant that has completed spatial annotation and trait annotation, the corresponding genome information collection and standardized coding process is carried out. The specific steps are as follows:

[0068] First, leaf tissue samples were collected from each labeled plant. Then, genomic DNA was extracted using standard plant DNA extraction methods, and raw molecular marker data was obtained through SNP chip typing. After that, a high-quality SNP locus set significantly associated with the target trait was screened through population variability analysis. ;in, Indicates the SNP sites, and ;in is the number of pre-selected SNPs;

[0069] Then, for each SNP site Using the unified 0 / 1 / 2 numerical code, we get Genotype representation of SNP loci , , and 0 represents inferior homozygous aa, 1 represents heterozygous Aa, and 2 represents superior homozygous AA; therefore, the All gene combinations of the plants are ,and ;

[0070] Based on this method, we can obtain a complete The gene annotation results of the plants are ;

[0071] Image-trait dataset construction: Multi-angle remote sensing image combination For input data, complete spatial position and trait annotation results are constructed To output data, a set of image-property data sets is obtained. Based on this method, comprehensive data collection is performed to obtain a complete image-property data set.

[0072] 2. Design of entity-level crop trait feature extraction model

[0073] In order to realize automatic sorghum plant recognition and trait recognition, and ensure the adaptability under plant occlusion, angle change and trait significant difference in complex field scenes, the present invention constructs a multi-view collaborative perception entity-level crop trait feature extraction model. As input, a multi-channel structure design is adopted, which includes plant space recognition channel, structural trait perception channel and trait fusion recognition channel, to complete the automatic recognition from image to structural trait. The model architecture is as follows Figure 2 shown.

[0074] 1. Plant space identification channel design: This channel uses the top-down main perspective image The input is used to realize the spatial position perception and coarse-grained structural contour extraction of individual plants, which is suitable for the identification of dominant spatial structural traits such as spike type and effective branch number. Specifically, the channel first extracts the initial plane features through the pre-trained ResNet34 network to obtain the initial texture features of the main view of the plant. , then the characteristics After five positions-enhancement-pooling structure groups, the spatial position convolution layer and the saliency enhancement convolution layer realize the enhanced extraction of plant boundaries, contours and salient areas; each output is compressed by the maximum pooling layer, and the position-aware enhanced features are output. ; will eventually Unify the dimensions through the Conv1 convolution layer and output the main feature results of the channel ;

[0075] The spatial position convolution layer is used to extract spatial visual feature information. Specifically, the horizontal and vertical coordinates of each pixel are first normalized and encoded into a position vector, which is then channel-concatenated with the original features to be processed. Then, a Conv3 convolution layer and a ReLU activation function are used for joint modeling to improve the model's ability to perceive relative positions and separate plants.

[0076] The saliency enhancement convolutional layer introduces two branches: spatial attention and pixel attention. First, features are extracted through the Conv3 convolutional layer. Then, the attention weight map is calculated through the spatial attention and pixel attention modules respectively. After splicing in the channel dimension, the fusion result is obtained through the ReLU activation function, thereby enhancing the salient plant area in the feature space and suppressing background interference.

[0077] 2. Design of structural property perception channel: This channel uses an oblique view With side view The auxiliary image is used as input to model the three-dimensional structural features of the plant and supplement the missing information of structural traits caused by occlusion or angle limitation in the main view. It is mainly suitable for extracting fine-grained continuous traits such as plant height, crown width, ear length, and stem thickness. This channel first extracts multi-angle structural features through the pre-trained Swin-Tiny model to obtain multi-angle structural encoding features. Then, the trait feature selection layer is introduced to construct 7 trait channel branches (corresponding to 7 types of structural traits) to extract trait-specific channel responses in parallel and generate trait significant features. ; Afterwards, features The image is fed into the cross-view fusion convolution layer to align and fusion model the structural information in the side view and oblique image, enhance the three-dimensional consistency representation and output the view fusion structural features. The fusion result is compressed by average pooling to form a fusion compression feature The above trait selection-fusion-pooling combination is repeated 5 times, and the final feature is unified in dimension through Conv1 convolution to output the auxiliary features of the structural trait perception channel ;

[0078] The trait feature selection layer constructs a channel attention structure based on trait semantics. Each trait is filtered through an independent channel attention layer, and the seven trait features are concatenated and normalized for output. This mechanism enables explicit differentiation of different trait modeling paths and enhances interpretability.

[0079] The cross-view fusion convolution layer first locally enhances the input features through the Conv3 convolution layer, then performs saliency matching and channel weighting on the same plant structure in different viewpoints through the cross-view attention mechanism, and finally outputs the fused features through the ReLU activation function, effectively capturing structural continuity and contour consistency.

[0080] 3. Trait fusion recognition channel design: This channel is used to jointly model the output features of the plant space recognition channel and the structural trait perception channel to achieve spatial position recognition and structural trait prediction of each plant. First, the output features of the two channels are combined into a and Perform channel splicing as the input of the fusion recognition channel to obtain the channel joint feature ; Then input into the cross-view fusion convolution layer to generate a unified feature representation structural feature discriminant feature ; Then, feature compression is performed through two layers of convolution to obtain the compressed features ;

[0081] In the prediction output stage, The data are fed into two sets of recognition head modules respectively: (1) Spatial position recognition head, which outputs the spatial position of each plant through the fully connected layer + Dropout + ReLU structure , i.e. the length and height of the coordinates and the area occupied by the plant; (2) Structural trait recognition head, which includes a continuous trait prediction branch and a discrete trait classification branch. Continuous trait prediction uses a full connection + Dropout + ReLU structure to output numerical traits such as plant height, ear length, and crown width; discrete trait classification uses a full connection + BatchNorm + Softmax activation function to predict discrete attributes such as ear type category and color type (coloring trait); and finally outputs the complete structural trait recognition results for each plant. ;

[0082] Finally, the spatial position of each plant Complete identification results of structural traits Together they constitute the spatial position and trait identification results of all plants. .

[0083] 4. Model training: The image-trait dataset constructed in the S1 stage is used as training samples, and the Adam optimizer is used for end-to-end training. The loss function is composed of the spatial position error based on the mean squared error (MSE), the continuous trait regression error, and the classification cross-entropy loss. Training is performed until the preset maximum number of iterations is reached to obtain the final trained trait recognition model.

[0084] 3. Crop Traits and Genome Interpretability Modeling

[0085] In order to establish a quantitative mapping relationship between crop traits and genomes, and thereby achieve trait modeling and key gene identification with causal explanation capabilities, the present invention designs an interpretable modeling module for crop traits and genomes. This module uses the predicted position and structural trait characteristics of each plant to SNP genomic characteristics of each plant collected in As module input, the output is a quantitative contribution relationship table between each trait and all SNP sites, as well as trait target evaluation baseline data, thus providing a decision basis for target-oriented SNP gene combination generation and breeding recommendations. The specific process is as follows Figure 3 shown.

[0086] Plant gene-trait data alignment: based on prediction results , including the spatial position of each plant and structural trait recognition results ; Through the spatial position of the plant Matching genetic annotations for each plant and trait identification results ; and use this to construct a gene-trait paired dataset.

[0087] Multi-trait modeling structure design: trait identification results All trait prediction and identification results are included, including plant height prediction results , average ear length prediction results , spike type category prediction results , crown width prediction results , prediction results of effective branch number , stem thickness prediction results , coloring trait prediction results ;in, , , , , It is a continuous trait. , It is a discrete trait;

[0088] For continuous traits, gene annotation As input, continuous traits ( , , , , ) is a parallel output, and a Lasso linear regression model is constructed; this model can output multi-path prediction results of the target continuous traits in parallel;

[0089] For discrete traits, gene annotation As input, discrete traits ( , ) is a parallel output to build a random forest classifier; this model can output multi-path classification results of target discrete traits in parallel;

[0090] Based on the constructed gene-trait pairing dataset, the Lasso and random forest classifier models were trained respectively to obtain the trained Lasso and random forest classifier model parameters.

[0091] Site contribution extraction: Based on the trained model structure, the numerical contribution of all SNP sites is extracted for each trait. Specifically, it includes:

[0092] (1) For continuous traits, the regression coefficient vector in the trained Lasso model is used as the linear contribution of each site, and the regression coefficient of all SNP sites is the contribution score of the trait;

[0093] (2) For discrete traits, the trained random forest classifier outputs a feature importance score as the information gain contribution of each SNP in the classification task; the score is based on the frequency of the SNP being selected as a node splitting variable in all decision trees and the sum of the Gini impurity reduction values ​​brought about by the split operation. After global accumulation and standardization, the contribution score of each SNP site in the trait is obtained.

[0094] Therefore, the complete SNP site contribution score corresponding to each trait is obtained, that is, , , , , , , , respectively, plant height SNP contribution score, average ear length SNP contribution score, crown width SNP contribution score, effective branch number SNP contribution score, stem diameter SNP contribution score, ear type category SNP contribution score, coloring trait SNP contribution score; and the contribution scores of all traits are spliced ​​to obtain the SNP contribution score of the complete trait .

[0095] Trait-site contribution table construction and structured output: SNP contribution scores based on the obtained complete traits , construct a standardized contribution structured output table; each table corresponds to a trait, and the table records all the SNP site numbers corresponding to the trait and their contribution scores. The output format is: trait name-SNP number-contribution value triplet, and form structured contribution data .

[0096] On this basis, the system further performs statistical analysis on all plant samples based on the trained Lasso regression and random forest classification models to extract the most common SNP combinations. As the basic reference genotype configuration in the current scenario, and calculate its corresponding reference traits At the same time, combined with the model mapping relationship between contribution and measured trait increment, a set of trait change estimates corresponding to unit contribution is constructed for all traits. (For example, when the contribution is 1.0, the average plant height increases by about 8.6 cm and the ear length increases by about 3.2 cm); Reference genotype configuration , reference traits , and the set of trait change estimates Together constitute the baseline data for trait target assessment ;

[0097] Therefore, the output of this module is structured contribution data , and baseline data for trait target assessment .

[0098] 4. Trait-oriented SNP gene combination generation and breeding recommendations

[0099] In order to achieve the optimal selection of genes under the user-input trait target and generate an interpretable SNP genetic gene combination that meets the trait expectation, the present invention designs a gene combination recommendation module based on joint contribution optimization. , trait target assessment baseline data , and user-specified trait requirements As input, the multi-trait joint control mechanism, the regulatory conflicts of different sites and the combination trade-offs are comprehensively considered, and the SNP genotype combination generation path is constructed through a multi-objective optimization algorithm to output the optimal breeding gene recommendation scheme.

[0100] User target trait configuration input: The user sets multiple target trait requirements (For example: plant height , unit cm; ear length , spike type category ), and specify the priorities of different traits (for example, setting the optimization weights of "plant height" and "ear length" higher than other traits). The system parses it into a target trait configuration matrix to guide the construction of the joint optimization objective function to ensure that the excellent traits of the overall morphological structure are realized first.

[0101] Initial state and objective function design: configuration based on the set target trait requirements , output structured contribution table and trait target assessment baseline data , construct the evaluation function between site and trait;

[0102] First, the reference SNP genotype configuration is provided and its corresponding reference traits As the initial state of optimization;

[0103] In addition, the estimated value of the trait change corresponding to the unit contribution , and the contribution value of each site (Indicates the SNP sites on the The normalized contribution value of the target traits, and , ), calculate the different genotype states of each SNP site The actual physical increment of the target trait is quantified as , in order to obtain its physical fitting value for the target trait and the target value of the trait set by the user Perform MSE error calculation and sum the preset weights of different traits to obtain the objective function result ;

[0104] Gene combination search and optimization modeling: The simulated annealing algorithm is used to solve the nonlinear optimization problem; the following steps are included:

[0105] First, to determine the reference SNP genotype configuration As the initial solution, the objective function constructed Calculate its fitness score;

[0106] Then, in each iteration, the system performs neighborhood perturbations on the current gene combination (randomly selects several sites and changes their genotypes) to form new candidate solutions. , and based on the Metropolis criterion, it is determined whether to accept the solution to escape the local optimum. The temperature parameter is gradually reduced using an exponential decay strategy.

[0107] The entire optimization process continues until the maximum number of iterations is reached, and the optimal SNP combination solution that meets the trait requirements is finally output. ;

[0108] Breeding recommendation generation: optimize the SNP genotype combination plan It is used as a genetic recommendation reference for the target traits, and is compared and matched with the sorghum material database to screen out the candidate material individuals closest to the optimized combination and output as the final recommended germplasm, thereby providing a decision-making basis for breeding material screening.

[0109] 5. Analysis of experimental results

[0110] To verify the effectiveness and practicality of the explainable breeding recommendation method proposed in this paper, two types of experiments were designed to evaluate its performance, focusing on: (1) the recognition accuracy of the entity-level crop trait feature extraction model; and (2) the quantitative modeling relationship between traits and genomes.

[0111] 1. Comparative Analysis of Entity-Level Feature Extraction Models

[0112] On the constructed image-trait dataset, the prediction accuracy of the proposed three-channel trait recognition model and two typical comparison models for five continuous traits (plant height, average ear length, crown width, number of effective branches, and stem diameter) was evaluated. The comparison models include:

[0113] (1) ResNet34 model: takes a single-view image as input, extracts deep features and then directly performs trait regression;

[0114] (2) EfficientNet-B0 model: has a lightweight structure, but does not include multi-view perception and saliency guidance mechanisms;

[0115] The evaluation indicators used are mean absolute percentage error (MAPE) and coefficient of determination (R²). MAPE reflects the relative average deviation between the predicted value and the true value, and the smaller the MAPE, the smaller the error; R² measures the degree of fit of the model to the trait fluctuation, and the larger the R², the higher the fitting accuracy. Figure 4 and Figure 5 As shown;

[0116] Experimental results demonstrate that the proposed model outperforms the comparison models in both MAPE and R² metrics. Specifically, for difficult-to-identify traits such as "number of effective branches" and "stem thickness," the proposed architecture significantly reduces error and improves prediction consistency, demonstrating the significant enhancement of the stability and generalization capabilities of the convolution, trait feature selection module, and view fusion strategy in complex field environments. Furthermore, compared to ResNet34 and EfficientNet-B0, the proposed model exhibits superior adaptability across different trait dimensions, validating its robust performance in multi-view collaborative recognition tasks.

[0117] 2. Interpretability Analysis of Trait-Gene Modeling Contribution

[0118] In order to evaluate the trait modeling and SNP site contribution quantification capabilities, 10 representative SNP sites were selected from the modeling results, and their normalized contributions to seven structural traits (plant height, average ear length, crown width, number of effective branches, stem diameter, ear type category, and coloring traits) were analyzed, and a heat map was drawn as shown in the figure. Figure 6 shown.

[0119] The results of the heat map analysis show that some SNP sites (such as SNP_003 and SNP_005) have a stable positive contribution to multiple traits and have obvious multi-trait regulatory capabilities; some sites (such as SNP_008 and SNP_010) show a negative regulatory trend on specific traits (such as the effective branch number or coloring traits) and have a potential inhibitory effect; there is a cross-influence mechanism between multiple traits, that is, the same SNP site contributes to different traits in different directions, verifying the complexity of the genetic control of crop traits. The results of this heat map experiment illustrate the advantages of interpretable modeling: on the one hand, it can explicitly identify key regulatory sites, and on the other hand, it can provide a quantitative reference for multi-objective optimization, effectively supporting the decision-making of subsequent breeding strategies.

[0120] Therefore, the interpretable breeding recommendation method proposed in the present invention can achieve high-precision and robust trait prediction in complex field environments. In terms of quantitative modeling of crop traits and genomes, the present invention verifies the combined action mechanism of SNP sites on multiple traits through contribution structure modeling and interpretable heat map analysis, providing a clear modeling basis for goal-oriented breeding optimization. The two sets of experimental results fully verified the performance advantages of the present invention in terms of recognition accuracy and modeling interpretability, providing interpretable quantitative results for subsequent breeding optimization, and having good breeding practice value.

[0121] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

[0122] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An interpretable breeding recommendation method based on crop trait and gene mapping analysis, characterized in that: The following steps are involved: S1, collect multi-angle remote sensing images of crop fields and simultaneously collect SNP genomic marker information of corresponding plant leaf tissues; S2 inputs the multi-angle remote sensing images collected by S1 into the trained entity-level crop trait feature extraction model, and outputs the spatial position results and trait recognition results of each plant; The entity-level crop trait feature extraction model includes a plant space recognition channel, a structural trait perception channel, and a trait fusion recognition channel; The plant space recognition channel is based on the main viewing angle image As input, it is used to realize the spatial position perception of individual plants and the extraction of coarse-grained structural contours, which is suitable for the identification of dominant traits of spatial structure; The structural feature perception channel takes an oblique view With side view The auxiliary image is used as input to model the three-dimensional structural characteristics of the plant, supplementing the missing structural characteristics information caused by occlusion or angle limitation in the main view, and is suitable for extracting fine-grained continuous characteristics; The trait fusion recognition channel is used to jointly model the output features of the plant space recognition channel and the structural trait perception channel. First, the output features of the two channels are spliced ​​together as the input of the fusion recognition channel to obtain the channel joint features. ; Then input into the cross-view fusion convolution layer to generate a unified feature representation structural feature discriminant feature ; Then, feature compression is performed through two layers of convolution to obtain the compressed features , The data are sent to the spatial position recognition head and the structural trait recognition head respectively. The spatial position recognition head outputs the spatial position result of each plant, and the structural trait recognition head outputs the trait recognition result. S3, combining the spatial location results of the plants to match the SNP genomic marker information of each plant with the corresponding trait identification results to construct a gene-trait pairing dataset; S4, based on the trait recognition results, a Lasso regression model for continuous traits and a random forest classification model for discrete traits are constructed respectively, and the model training is performed based on the paired data set obtained in S3 to obtain the parameters of the two models after training; S5: Calculate the contribution score of each SNP site to the trait based on the two model parameters, concatenate the contribution scores of all traits and construct a structured contribution output table; at the same time, count the reference SNP combinations and reference trait results in the current scenario, and derive the estimated trait change value corresponding to the unit contribution based on the contribution, thus obtaining the baseline data for the trait target assessment; S6, based on the obtained structured contribution output table, trait target evaluation baseline data and user target trait configuration, constructs a multi-objective loss function and performs optimization search to output the optimal SNP gene combination solution.

2. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 1, characterized in that: The multi-angle remote sensing image uses the top view as the main image perspective, and the oblique view and side view as auxiliary image perspectives. During data collection, the structural characteristics of each sorghum plant in multiple spatial dimensions are recorded, including plant height. , average ear length , spike type category , crown width , effective branch number , stem thickness , coloring properties .

3. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 1, characterized in that: When constructing the dataset, the corresponding genomic information collection and standardized coding process was carried out for each sorghum plant that had completed spatial and trait annotation. Specifically, the process included: First, leaf tissue samples were collected from each labeled plant. Then, genomic DNA was extracted using standard plant DNA extraction methods, and raw molecular marker data was obtained through SNP chip typing. After that, a high-quality SNP locus set significantly associated with the target trait was screened through population variability analysis. ;in, Indicates the SNP sites, and ;in is the number of pre-selected SNPs; Then, for each SNP site Using the unified 0 / 1 / 2 numerical code, we get Genotype representation of SNP loci , , and 0 represents inferior homozygous aa, 1 represents heterozygous Aa, and 2 represents superior homozygous AA; therefore, the All gene combinations of the plants are ,and .

4. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 1, characterized in that: The plant space recognition channel first extracts the initial plane features through the pre-trained ResNet34 network to obtain the initial texture features of the plant main view. , then the characteristics After five positions-enhancement-pooling structure groups, the spatial position convolution layer and the saliency enhancement convolution layer realize the enhanced extraction of plant boundaries, contours and salient areas; each output is compressed by the maximum pooling layer, and the position-aware enhanced features are output. ; will eventually Unify the dimensions through the Conv1 convolution layer and output the main feature results of the channel .

5. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 1, characterized in that: The structural feature perception channel first extracts multi-angle structural features through the pre-trained Swin-Tiny model to obtain multi-view structural encoding features Then, the trait feature selection layer is introduced to construct multiple trait channel branches corresponding to multiple types of structural traits, and the trait specific channel responses are extracted in parallel to generate trait significant features. ; Afterwards, features The image is fed into the cross-view fusion convolution layer to align and fusion model the structural information in the side view and oblique image, enhance the three-dimensional consistency representation and output the view fusion structural features. ; The fusion result is compressed by average pooling to form a fusion compression feature The above trait selection-fusion-pooling combination is repeated 5 times, and the final feature is unified in dimension through Conv1 convolution to output the auxiliary features of the structural trait perception channel .

6. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 1, characterized in that: The S4 is specifically: Trait identification results include plant height prediction results , average ear length prediction results , spike type category prediction results , crown width prediction results , prediction results of effective branch number , stem thickness prediction results , coloring trait prediction results ;in, , , , , It is a continuous trait. , It is a discrete trait; For continuous traits, gene annotation As input, continuous traits are parallel output, and a Lasso linear regression model is constructed; this model can output multi-path prediction results of the target continuous traits in parallel; For discrete traits, gene annotation As input, discrete traits are parallel outputs, and a random forest classifier is constructed; this model can output multi-path classification results of target discrete traits in parallel; Based on the constructed gene-trait pairing dataset, the Lasso and random forest classifier models were trained respectively to obtain the trained Lasso and random forest classifier model parameters.

7. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 6, characterized in that: The S5 includes: For continuous traits, the regression coefficient vector in the trained Lasso model is used as the linear contribution of each site, and the regression coefficient of all SNP sites is the contribution score of the trait; For discrete traits, the trained random forest classifier outputs a feature importance score as the information gain contribution of each SNP in the classification task. The score is based on the frequency of the SNP being selected as a node splitting variable in all decision trees, as well as the sum of the Gini impurity reduction values ​​brought about by the split operation. After global accumulation and normalization, the contribution score of each SNP site to the trait is obtained. The complete SNP site contribution score corresponding to each trait is obtained, that is, , , , , , , , respectively, plant height SNP contribution score, average ear length SNP contribution score, crown width SNP contribution score, effective branch number SNP contribution score, stem diameter SNP contribution score, ear type category SNP contribution score, coloring trait SNP contribution score; and the contribution scores of all traits are spliced ​​to obtain the SNP contribution score of the complete trait ; SNP contribution score based on the complete traits obtained , construct a standardized contribution structured output table; each table corresponds to a trait, and the table records all SNP site numbers corresponding to the trait and their contribution scores.

8. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 7, characterized in that: The specific process of obtaining the trait target assessment baseline data is as follows: Based on the trained Lasso regression and random forest classification models, statistical analysis was performed on all plant samples to extract the most common SNP combinations. As the basic reference genotype configuration in the current scenario, and calculate its corresponding reference traits As a reference trait; at the same time, combined with the model mapping relationship between contribution and measured trait increment, a set of trait change estimates corresponding to unit contribution is constructed for all traits ; Reference genotype configuration , reference traits , and the set of trait change estimates Together constitute the baseline data for trait target assessment .

9. The interpretable breeding recommendation method based on crop trait and gene mapping analysis according to claim 7, characterized in that: The S6 specifically includes: Users set multiple target trait requirements , and specify the priorities of different traits at the same time, and parse them into target trait configuration matrix; based on the set target trait requirements configuration , output structured contribution table and trait target assessment baseline data , construct the evaluation function between site and trait: First, the reference SNP genotype configuration and its corresponding reference traits As the initial state of optimization; then, according to the estimated value of the trait change corresponding to the unit contribution , and the contribution value of each site Calculate the genotype of each SNP site in different states The actual physical increment of the target trait is quantified as , in order to obtain its physical fitting value for the target trait and the target value of the trait set by the user Perform MSE error calculation and sum the preset weights of different traits to obtain the objective function result ; Use simulated annealing algorithm to solve nonlinear optimization problems and output the optimal SNP combination solution that meets trait requirements .

Images