An intelligent planting optimization method and system based on artificial intelligence

Abstract

The application relates to the technical field of database construction and artificial intelligence algorithms, and discloses a wisdom planting optimization method and system based on artificial intelligence, which comprises the following steps: establishing a variety database and an environment database, constructing a vector knowledge base, defining the association relationship between entities and relationships, storing the unified vector representation of all entities processed by a joint embedding model in a single vector database according to a pre-trained training multi-modal relationship joint embedding model, creating a unified index, and then inputting a target environment vector and a target trait vector into a prediction model to obtain a target gene vector. The method inversely solves an ideal genotype through a double-branch prediction model based on artificial intelligence algorithms by constructing a variety and environment database. The model analyzes the regulation relationship between the environment and the traits through a cross-attention mechanism, and injects genetic scores as a supervision signal, so that effective optimization directions are provided for variety breeding and planting suggestions.

Description

Technical Field

[0001] This invention relates to the fields of database construction and artificial intelligence algorithm technology, particularly to the field of smart planting technology, and especially to a smart planting optimization method and system based on artificial intelligence. Background Technology

[0002] Currently, significant and increasing investments are being made in the digitalization of citrus, which has, to some extent, accelerated and improved the transformation and efficiency of citrus planting and sales. However, the existing information technology infrastructure is limited by local conditions and needs, with incomplete data and redundant functions. Consequently, the current state of citrus industry informatization in various regions still suffers from insufficient overall planning and a lack of unified standards and specifications.

[0003] Specifically, this manifests as follows: information systems are built and deployed independently by different regions, with system developers from various vendors. This inevitably leads to inconsistencies in data formats, standards, accuracy, and classification. These issues affect data quality, integrity, and reliability, making it difficult to aggregate data and support data exchange.

[0004] Furthermore, most agricultural information applications, including those for citrus, stop at simply displaying the real-time data collected. This involves showing farmers the soil temperature, humidity, EC, and pH values ​​of the plots where the sensors are deployed through icons and numerical values. Such functionality requires citrus farmers to possess a relatively high level of professional knowledge in order to summarize and analyze the collected data and then take relevant measures such as irrigation and fertilization.

[0005] In conclusion, existing citrus planting optimization platforms are limited to local data and have a very low level of intelligence, resulting in very limited planting optimization effects. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a smart planting optimization method and system based on artificial intelligence.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] In a first aspect, this application discloses an intelligent planting optimization method based on artificial intelligence, characterized by comprising: establishing a variety database and an environment database based on historical data, wherein the environment database stores several first environmental feature vectors EA, and the variety database stores basic variety information, a second environmental feature vector EB, gene data G, and a phenotypic vector P corresponding to each variety under the second environmental feature vector EB; generating a corresponding genotype matrix G' based on the gene data G; then constructing a vector knowledge base, defining the association between entities and relations, predicting tail entity vectors based on a pre-trained multimodal relation joint embedding model, wherein the model input includes head entity vectors and relation vectors, storing the unified vector representation of all entities processed by the joint embedding model in a single vector database, and creating a unified index;

[0009] Based on the results of the index, the target environment vector Etarg is retrieved from the environment database. The user constructs the target trait vector Ptarg based on experience, and then inputs the target environment vector Etarg and the target trait vector Ptarg into the prediction model to obtain the target gene vector g'.

[0010] Furthermore, establishing the variety database includes constructing a genotype matrix G' based on gene data G: an initial genotype matrix Graw is constructed based on gene data G, with dimensions of... m × n ;in, m Indicates the number of varieties. n The matrix represents the number of SNP molecular marker sites. Graw [ i , j ] indicates the first i The variety in the first j Genotype coding values ​​at each SNP locus; using singular value decomposition, we can obtain:

[0011] Graw= UFV T

[0012] In the formula, U It is a left singular vector matrix with dimension . m × m In the left singular vector matrix U [ i ,:] is the first i The projection of each sample onto all left singular vectors; F It is a diagonal matrix with dimension 1. m × n Its diagonal elements include: s 1, s 2,…, σr ( r=min( m , n )) is a singular value, and s 1≥ s 2≥…≥ σr ≥0; V T It is a right singular vector matrix with dimension . n × n In the right singular vector matrix V [:, j [For the first] j The weights of each SNP on all right singular vectors;

[0013] Then, dimensionality reduction and truncation are performed, taking the first-order dimensionality. k The principal components include: U k = U [:,0: k ]F k =F[0: k ,0: k ] V k = V [:,0: k ]in, k It is the target dimension number after dimensionality reduction. U k The dimension is m × k , including samples in front k Coordinates along the main genetic structures; F k The dimension is k × k , including the previous k The largest singular value; V k The dimension is n × k Including SNP sites in the front k Loads along the main genetic structural directions; constructing the genotype matrix G 'for: G ′= U k F k Its dimensions are m × k , G Each line of ' yes Representing the i A low-dimensional genetic characteristic representation vector for each variety, with a dimension of 1× k .

[0014] Furthermore, a genome-wide regression model is established for each element pt in the trait vector Pi to estimate the pairing of all SNP markers with the target trait. t joint effect vector Then, obtain the gene vector g corresponding to any variety in the variety database in the genotype matrix G', and calculate the genotype score of the gene vector g for trait t. The genotype scores of all target traits are combined into a single score vector S.

[0015] Furthermore, a genome-wide regression model is established for each element pt in the trait vector Pi to estimate the pair of all SNP markers for the target trait. t joint effect vector This includes: establishing a genome-wide regression model.

[0016] :

[0017] in, p t It is the phenotypic vector of all varieties on trait t. It is the label effect vector. For the error; where The estimation method is as follows:

[0018] :

[0019] In the formula, λ is the regularization parameter. I It is a with G ′ ⊤ G An identity matrix of the same dimension; the genotype score The calculation method is as follows:

[0020] ,

[0021] This is the transpose of the genotype vector g for a specific variety; the genotype scores corresponding to all traits to be evaluated are combined into a single score vector S = [ s 1, s 2,…, s n ].

[0022] Furthermore, the step of retrieving the target environmental vector Etarg from the environmental database, constructing the target trait vector Ptarg based on user experience, and then inputting the target environmental vector Etarg and the target trait vector Ptarg into the prediction model to obtain the target gene vector g' includes: constructing the prediction model network structure, wherein the prediction model is a two-branch input fusion structure; wherein the target environmental vector Etarg and the target trait vector Ptarg are first processed through independent sub-networks corresponding to the two-branch structure, and the outputs of the two independent sub-networks are fused and then used with the genotype score vector. S Linear layer projection is used as a supervision signal, and the target gene vector of the output result is enhanced by a loss function. g ′ and genotype score S The linear relationship ultimately outputs the target gene vector. g ′ .

[0023] Furthermore, the prediction model network structure includes: an environmental feature processing module and a trait feature processing module. The environmental feature processing module takes the target environmental vector Etarg as input and includes a first hidden layer, a lightweight Transformer layer, and a first fully connected layer. The trait feature processing module takes the target trait vector Ptarg as input and includes a second hidden layer and a second fully connected layer. A fusion module is also included, comprising a cross-attention fusion layer, a linear layer, a random depth residual block, and a channel compression layer.

[0024] Furthermore, the features output by the environmental feature processing module and the trait feature processing module are fused and then input into the cross-attention fusion layer. The query Q, key K, and value V in the cross-attention fusion layer are defined as follows: Q = h env ⋅ Wq , K = h trait ⋅ Wkv , V = h trait ⋅ Wkv ;in, h env For the environmental features in the target environment vector Etarg, Wq The query projection matrix for environmental representation. h trait For the trait features in the target trait vector Pi, Wkv This represents the key / value projection matrix for trait characterization; the environmental attention weights on traits are calculated using scaled dot product attention, reflecting the regulatory strength of environmental factors on different traits, and the genotype score vector is then used. S The linear layer is projected as a monitoring signal.S And concatenate or add it with the attention output, including: S311, calculate the original attention weights:

[0025]

[0026] in d k This refers to the dimension of the key vector, and the scaling factor prevents gradient vanishing; S312, passing the genotype score vector S through a linear layer and a sigmoid activation function to generate a gated vector. ;

[0027]

[0028] in, This is the weight matrix. S313. The attention weight matrix is ​​modulated using a gating vector: where dp is the dimension of the vector, representing the number of target traits;

[0029]

[0030] Here, diag(1-Gate) is a diagonal matrix with (1-Gate) as its diagonal; the stronger the genetic prior of a trait (larger Gate value), the weaker the regulatory attention weight of environmental factors on that trait (smaller 1-Gate value); splicing attention output and S After processing through 3 layers of random depth residual blocks and channel compression layers, the final output is a k-dimensional target gene vector. g ′.

[0031] Furthermore, the target gene vector of the output result is enhanced through a loss function. g ′ and genotype score S The loss function in a linear relationship is set as follows:

[0032]

[0033] In the formula, N The number of training samples, g true( i ) is the first i The true gene vector of each sample S ( i ) is the first i Genotype score vector for each sample, B ∈R k×dp This is the labeling effect matrix, with the following columns: βt , l This is the regularization coefficient.

[0034] Furthermore, the method also includes: increasing the dimensionality of the target gene vector g' to the target gene g'', finding the existing variety closest to the target gene g'' in the variety database as the recommended plant; if the gene sequence consistency is less than a first threshold, then: determining the closest existing variety as the first parent, and storing the genotype of the first parent in the database. yes The similarity to the target gene g'' is calculated. If the similarity is less than a second threshold, proceed to step 5A; if the similarity is greater than or equal to the second threshold, proceed to step 5B. Step 5A includes: obtaining multiple existing varieties closest to the target gene g'' as parents, combining and breeding them to obtain the optimal offspring varieties; planting the offspring varieties in the planting area; and continuously acquiring features from several offspring plants as samples. This feature acquisition can be done by acquiring images of these citrus plants: acquiring features from images uploaded by the planting terminal, and then performing trend prediction for the next stage based on Markov chains according to the current features. The confidence interval of the prediction result is calculated. If the confidence interval reaches a third threshold, it means that a certain growth trend is highly likely to occur, and planting suggestions are made based on this growth trend; otherwise, the prediction is unreliable, and planting suggestions already stored in the existing database are retrieved. Step 5B includes: storing the genotypes of the first parent in the database. yes Distinguishing features are obtained by comparing with the target gene g''. Based on the distinguishing features and the target gene vector g', suggestions are made for the second parent, including: S51B, gene differential quantification, and key site localization. The distinguishing features are obtained by calculating the gene differential vector. get:

[0035] Then, extract the set of SNP sites with significant differences:

[0036] in, t The significance threshold, s Δ is the standard deviation of Δg. For the key differentially expressed loci index set; S52B, establish a second parent screening model, including:

[0037] Based on the aforementioned gene differential vector Using this as input, the output is a set of key differential site indexes. L = { j | |Δg [ j ] | ≥ t·sD}, where τ is the significance threshold and σΔ is the standard deviation of Δg; S522B, Set the parental complementarity scoring function: For each candidate variety in the variety database, calculate its complementarity matching degree with the target gene g'' on the key site set L:

[0038]

[0039] in, The difference weight at position j (the greater the difference, the higher the weight); Gene vectors for each candidate variety, This is an indicator function that returns 1 if the genotypic change direction of the candidate variety at this locus is consistent with the target requirement, and 0 otherwise; it also calculates the hybridization compatibility between the candidate variety and the first parent.

[0040]

[0041] Where γ is the scaling parameter, and GeneticDistance is the genetic distance, which can be calculated using Euclidean distance or an existing kinship database; S523B, Comprehensive screening and sorting;

[0042] A comprehensive score is calculated for each candidate variety, including:

[0043]

[0044] Where α ∈ [0,1] is the weight parameter; finally, sort by comprehensive score in descending order, select the Top M candidate varieties as the second parent recommendations, and output the list of second parents and their complementary site details.

[0045] Its beneficial effects are as follows: The method constructs a variety and environment database, extracts core genetic characteristics using genotype dimensionality reduction, and establishes a genome-wide regression model to quantify the genetic potential score of varieties. Then, based on target environmental characteristics and artificially set target traits, it uses a bi-branch prediction model to inversely solve for the ideal genotype. This model analyzes the regulatory relationship between the environment and traits through a cross-attention mechanism and injects genetic scores as a supervisory signal to ensure that the output genotype conforms to biological laws. Ultimately, it achieves a precise mapping from environmental adaptability and agronomic goals to the optimal genotype, providing effective optimization directions for variety breeding and planting recommendations.

[0046] Secondly, this application discloses an artificial intelligence-based intelligent planting optimization system for implementing the aforementioned artificial intelligence-based planting optimization method. The system includes a cloud platform and a first terminal and a second terminal communicating with the cloud platform. The cloud platform includes: a genotype matrix construction module for receiving gene data G and generating a corresponding genotype matrix G'; and a genome-wide regression modeling module, which is equipped with a genome-wide regression model to estimate the relationship between all SNP markers and the target trait. t joint effect vector The genotype score generation module is used to obtain the gene vector g corresponding to any variety in the genotype matrix G' from the variety database, and calculate the genotype score of the gene vector g for trait t. The system combines the genotype scores of all target traits into a single score vector S; the prediction module retrieves the target environmental vector Etarg and the trait vector Ptarg from the environmental database and outputs the target gene vector g'; the first terminal is for researchers and includes a database management interface and a model training interface; the second terminal is for farmers and includes an image upload interface and a download interface. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the intelligent planting optimization method based on artificial intelligence according to some embodiments of this application;

[0048] Figure 2 This is a simplified schematic diagram of the network results of a prediction model according to some embodiments of this application. Detailed Implementation

[0049] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] See Figure 1-Figure 2 The present invention provides a technical solution:

[0051] Due to varying geographical environments, the impact on crop growth is significant, especially for citrus. Factors such as humidity, temperature, and soil conditions vary considerably across regions, resulting in vastly different citrus varieties suitable for cultivation. For instance, different planting areas exhibit significant differences in key environmental parameters (such as average annual temperature, accumulated temperature (≥10℃), extreme winter temperatures, annual precipitation and seasonal distribution, relative humidity, sunshine duration, soil type (sandy / clay / loam, etc.), soil pH, soil organic matter content, available soil nutrients, altitude, slope, aspect, and water resources).

[0052] On the other hand, different citrus varieties require different optimal fertilization, irrigation, and grafting conditions. Due to farmers' lack of professional knowledge, poor citrus harvests often occur because the varieties are unsuitable for the region or the planting conditions of new varieties are not met. For example, different citrus varieties (and rootstocks) have specific adaptation thresholds and optimal ranges for these environmental parameters. Exceeding these ranges may lead to frost damage, heat damage, drought stress, waterlogging, nutrient imbalance, high incidence of pests and diseases, decreased fruit quality (sugar content, acidity, color, size, flavor, and juiciness), sharp reduction in yield, or even tree death.

[0053] Therefore, according to an embodiment of this application, an artificial intelligence-based smart planting optimization method aims to build a management platform that can be used in various regions and is simultaneously accessible to researchers and farmers. The method can be referenced... Figure 1 Understanding, specifically, includes:

[0054] S1. Establish a variety database and an environmental database based on historical data. This historical data can be data stored from past research and integrated with existing data, or it can be publicly available data from the internet.

[0055] For example, the environmental database can include environmental parameters for each region, such as integrating long-term meteorological station observation data, soil survey data, remote sensing image data (e.g., soil moisture and surface temperature), topographic data (DEM), hydrological data, and field sampling and testing data (supplementing soil nutrients, pH, etc.). The environmental parameters for each region are divided using geographic grids (e.g., 1km x 1km) or administrative regions (counties / townships) as basic units. The variety database includes gene data corresponding to each citrus variety and the phenotypic characteristics of each variety under different environmental parameters. The phenotypic characteristics of each variety under different environmental parameters can be obtained by statistically analyzing and structurally storing existing experimental data. Specifically, the environmental parameters are stored as several first environmental feature vectors EA.

[0056] The variety database stores basic variety information (such as variety name), the second environmental feature vector EB, gene data G, and the corresponding phenotypic vector P for each variety under the second environmental feature vector EB. Gene data G can be constructed based on historical data. For varieties with missing data, molecular marker data (such as SNPs) or information on functional genes / alleles related to the target trait is obtained using high-throughput sequencing technologies (such as whole-genome resequencing and reduced genome sequencing GBS). Each variety corresponds to a gene vector gi and the phenotypic performance of that variety observed and recorded under the second environmental feature vector EBi in a specific environment from past experimental data, which is stored as the corresponding phenotypic vector Pi.

[0057] In some embodiments, establishing the variety database in S1 specifically includes:

[0058] S11. Construct the genotype matrix G'.

[0059] The gene data G was obtained through genome-wide selection. A genotype matrix G' was constructed based on the gene data G. To avoid the curse of dimensionality, effective dimensionality reduction and feature extraction were performed to obtain the genotype matrix G', including:

[0060] For each variety, represent all its genotype data (via SNP markers) as a vector. Assume there are m varieties (or samples) and n SNP marker loci. Construct an initial genotype matrix Graw with dimensions [missing information]. m × n .in: m Indicates the number of varieties or individual samples; n This represents the number of SNP molecular marker sites (i.e., the original high-dimensional feature number); in the matrix Graw [ i , j ] indicates the first i The variety in the first j Genotype coding values ​​(e.g., 0, 1, 2 or standardized values) at each SNP locus.

[0061] Using singular value decomposition (SVD), we can obtain:

[0062] Graw= UFV T

[0063] In the formula, U It is a left singular vector matrix with dimension . m × m . U The column vectors are orthogonal, that is, they satisfy... U T U = I (Orthogonal matrix). U The column vectors describe the orientation of the sample (variety) in the genetic structure space, and the left singular vector matrix... U [ i ,:] can be seen as the first i The projection of each sample onto all left singular vectors; F It is a diagonal matrix with dimension 1. m × n Its diagonal elements include: s 1, s 2,…, σr ( r =min( m , n )) represents singular values, sorted from largest to smallest, i.e. s 1≥ s 2≥…≥ σr ≥0. The magnitude of the singular value reflects the intensity (importance) of the variation in the direction of the corresponding singular vector; V T ( V The transpose of (x, y) is a right singular vector matrix with dimension 1. n × n. V The column vectors are orthogonal, that is, they satisfy... V T V = I . V The column vectors describe the loading directions of SNP sites (primary traits) in the genetic structure space, and the right singular vector matrix... V [:, j ] indicates the first j The weights of each SNP on all right singular vectors.

[0064] Based on this, we perform dimensionality reduction and truncation, taking the first digit. k One principal component. The operation is as follows:

[0065] U k = U [:,0: k ](Pick U The former k (column), F k =F[0: k ,0: k (Take the first part of F) k (The diagonal elements form a diagonal matrix). V k = V [:,0: k ](Pick V The former k (List). Among them, k It is the target dimension number after dimensionality reduction ( k ≪ n ), used to determine how much major genetic variation information to retain. k The selection method can be based on any one or more methods such as scree plot, cumulative variance rate, cross-validation, etc. The specific means are well known to those in the art and will not be elaborated here. U k The dimension is m × k , including samples in front k Coordinates along the main genetic structures; F k The dimension is k × k , including the previous k The largest singular value; V k The dimension is n × k Including SNP sites in the front k Loads in the directions of the main genetic structures.

[0066] Finally, a genotype matrix is ​​constructed through dimensionality reduction. G 'for: G ′= U k F k Its dimensions are m × k .in, G Each line of ' yes Representing the i A low-dimensional genetic feature representation vector for each variety (sample), with a dimension of 1× k ; yes ′[ l ]( l =1,…, k ) indicates the first i The variety in the first l The primary genetic direction (from the first) l The score or coordinates on a singular vector (defined by a singular vector). This value synthesizes the variety's scores or coordinates on a large number of original SNP sites ( n Genetic information is expressed in a more compact and information-concentrated way, rather than in a single genetic unit.

[0067] S12. Constructing the trait vector Pi includes:

[0068] Based on the phenotypic performance of each variety under various environmental parameters obtained from historical data, a phenotypic vector Pi is constructed, where each phenotypic vector Pi corresponds to a second environmental vector. EBi For example, EBi = [latitude, longitude, average annual temperature 18.5℃, soil pH 5.8, ...] and Pi = [yield 32t / ha, sugar content 12.5°Brix, ...].

[0069] Furthermore, since the variety database needs to be retrieved and updated in subsequent steps after its construction, and the various types of data in the database have non-linear mapping relationships, some embodiments also include S13, which constructs a vector knowledge base. S13 specifically includes:

[0070] S131. Define the association between entities and relations;

[0071] Entities are mapped to relations. For example, entities include varieties, environmental conditions, phenotypic traits, gene loci, etc. Relationship types are defined, including adapting to, exhibiting, benefiting from, etc. Training triples are constructed: triples in the format of (head entity, relation, tail entity) are extracted from historical data as training supervision signals.

[0072] S132, Train a multimodal relation joint embedding model;

[0073] The model input includes a head entity vector; for example, a variety vector = f(gene vector gi'); an input relation vector, which is a learnable vector representing adaptation to or expression of an equire relationship; and then predicts tail entity vectors. The training objective of the model is to output tail entity vectors that correspond to the head entity vectors and relation vectors. For example:

[0074] f(Wogan gene vector 1) + V_ adapts to ≈ environmental vector_EB123; f(Wogan gene vector 2) + V_ exhibits ≈ high sugar content phenotype vector;

[0075] The triples constructed in S131 can be used for training, and positive and negative samples can be constructed separately for training.

[0076] Specifically, the multimodal relationship joint embedding model includes a variety encoder, an environment encoder, and a phenotypic encoder; it is trained based on the input positive and negative samples, the scores of all positive and negative samples are calculated according to the scoring function, the total loss is calculated according to the loss function L, and all parameters are updated through the backpropagation algorithm and gradient descent optimizer (such as Adam), and this process is repeated until the model converges.

[0077] S133. Construct a unified vector knowledge base and index it;

[0078] All entities, after being processed by the joint embedding model, are represented by a unified vector and stored in a single vector database (such as Pgvector, Weaviate). The stored data includes: variety vectors, obtained from gene vectors via encoder f; second environmental feature vectors EBi and phenotypic vectors Pi, either phenotypic vectors Pi or encoded vectors; and relation vectors, relation embedding vectors obtained from model training (such as V_adapted to).

[0079] Then, a uniform ANN index is created for all types of entity vectors, enabling cross-modal similarity searches to be performed in this uniform vector space.

[0080] Therefore, relational reasoning and querying based on vector space can be implemented, as shown in the following example:

[0081] Example 1

[0082] User input: An environment vector Enew for a new environment.

[0083] The system executes the following steps: Calculate Enew - V_fit; in the vector database, search for the variety vector most similar to (Enew - V_fit).

[0084] Return the Top K varieties. This is equivalent to performing a query on the graph to "find all varieties whose 'adapted to' relation points to environment Enew", but done entirely through vector operations.

[0085] S2. For each element pt in the trait vector Pi (such as yield p1, sugar content p2, acidity p3, etc.), establish a genome-wide regression model:

[0086] in, p t G is the phenotypic vector of all varieties on trait t, and G' is the aforementioned genotype matrix. It is a labeled effect vector; for example, p t It is the joint effect vector of all SNP markers on the target trait t (such as sugar content). This is the error.

[0087] in The estimation method is as follows:

[0088]

[0089] In the formula, λ is the regularization parameter. I It is a with G ′ ⊤ G An identity matrix of the same dimension.

[0090] Understandably, genome-wide regression models aim to estimate the impact of all SNP markers on the target trait. t joint effect vector ， However, the genetic potential of a single variety cannot be evaluated. Therefore, further research is needed:

[0091] S21. For any variety in the variety database, its corresponding gene vector in the genotype matrix G' is g. Then, its genotype score for trait t is... for:

[0092]

[0093] Therefore, the group effect βt The quantitative score is converted into a genotype vector g specific to a particular variety. This is the transpose of the genotype vector g. Transposing it facilitates... Calculations are made to determine the phenotypic characteristics of the variety. t The genetic potential (independent of the environment).

[0094] Then, the genotype scores corresponding to all traits that need to be evaluated are combined into a score vector S=[ s 1, s 2,…, s n ].

[0095] Furthermore, since environmental feature vectors are obtained from existing experimental data and cannot cover all situations, it is not possible to directly match local environmental conditions based on environmental feature vectors in the variety database to obtain corresponding variety recommendations. Therefore, in some embodiments, the method further includes:

[0096] S3. Based on the farmer's geographical location, retrieve the corresponding target environmental vector Etarg from the environmental database. Then, based on past experience, i.e., the citrus varieties previously grown in the area, and combined with the index results in S133, breeders select possible citrus optimization directions to construct a target trait vector Ptarg. Based on this, input the target environmental vector Etarg and the target trait vector Ptarg into the prediction model to obtain the target gene vector g'.

[0097] Specifically, step S3 includes:

[0098] S31. Construct a prediction model network structure. The prediction model is a two-branch input fusion structure. The target environment vector Etarg and the target trait vector Ptarg are first processed through independent subnetworks (two branches) and then fused. This is to better capture the heterogeneity of environmental characteristics (such as temperature and soil pH) and trait characteristics (such as sugar content and yield). Furthermore, genotype scoring is used during training. S The linear layer projection is used as a supervision signal, and the target gene vector of the output result is enhanced by a loss function. g ′ and genotype score S The linear relationship is established. A Dropout layer is added to prevent overfitting and to accommodate potential missing or noisy historical data. The output layer uses linear activation to ensure... g ′ is a real number vector that can be directly used for matching in variety databases.

[0099] Specifically, such as Figure 2 As shown, the network structure includes an environment feature processing module, a trait feature processing module, and a fusion module. The environment feature processing module takes the target environment vector Etarg as input, which is 18-dimensional in this example. The environment feature processing module includes at least a first hidden layer, a lightweight Transformer layer, and a first fully connected layer. The trait feature processing module takes the target trait vector Ptarg as input, which is 8-dimensional in this example. The trait feature processing module includes at least a second hidden layer and a second fully connected layer.

[0100] Then, the environmental feature processing module and the trait feature processing module output to the fusion module, which includes a cross-attention fusion layer, a linear layer, a random depth residual block, and a channel compression layer.

[0101] The features output from the environmental feature processing module and the trait feature processing module are fused and then input into the cross-attention fusion layer. In the cross-attention fusion layer, the query (Q), key (K), and value (V) are defined as follows:

[0102] Q = h env ⋅ Wq , K = h trait ⋅ Wkv , V = h trait ⋅ Wkv ;

[0103] in, h env The environmental features (such as climate and soil) in the target environment vector Etarg. Wq The query projection matrix for environmental representation. h trait For the trait features in the target trait vector Pi, Wkv This is the key / value projection matrix representing the traits.

[0104] The environmental attention weights on traits are calculated by scaling the dot product attention, reflecting the regulatory strength of environmental factors on different traits. Then, biological priors (such as known contributions of genotypes to traits) are introduced, and the genotype score vector is... S The linear layer is projected as a monitoring signal. S This is then concatenated or added to the attention output. Specifically, this includes:

[0105] S311. Calculate the original attention weights:

[0106]

[0107] in d k It is the dimension of the key vector, and the scaling factor prevents gradient vanishing.

[0108] S312. Pass the genotype score vector S through a linear layer and a sigmoid activation function to generate a gated vector. .

[0109]

[0110] in, This is the weight matrix. Here, dp is the bias vector, and dp is the dimension of the vector, representing the number of target traits. In essence, this gating vector encodes which traits are more easily influenced / determined by genetic potential. A value closer to 1 indicates stronger genetic control over the trait, and a correspondingly weaker attentional weight from environmental regulation; a value closer to 0 indicates greater susceptibility to environmental influences.

[0111] S313. Modulate the attention weight matrix using a gating vector:

[0112]

[0113] Here, diag(1-Gate) is a diagonal matrix with (1-Gate) as the diagonal; the stronger the genetic prior of a trait (larger Gate value), the weaker the regulatory attention weight of environmental factors on that trait (smaller 1-Gate value).

[0114] Thus, by injecting supervisory signals, the model can explicitly utilize known genotype-trait association priors to enhance the biological plausibility of fused features.

[0115] splicing attention output and S After processing through 3 layers of random depth residual blocks (DropPath rate 0.2) and channel compression layers, the final output is a k-dimensional target gene vector. g ′.

[0116] S32. Update all weight parameters (branch weights are updated independently) until convergence, and set the loss function as follows:

[0117]

[0118] In the formula, N The number of training samples, g true( i ) is the first i The true gene vectors of each sample (from a variety database). S ( i ) is the first i Genotype score vector for each sample, B ∈R k×dp This is the labeling effect matrix, with the following columns: βt (Characteristics) t The effect vector was obtained from the genome-wide regression model with fixed parameters (not updated). l Regularization coefficient (e.g.) l =0.1), controlling the strength of the genotype score constraint to ensure prediction g ′ and genetic potential S Consistent.

[0119] For example, input target environment vector Etarg ∈R de ; dimension de This includes environmental characteristics (e.g., average annual temperature, accumulated temperature, extreme low winter temperature, annual precipitation, relative humidity, sunshine duration, soil type code, soil pH, soil organic matter content, soil nitrogen / phosphorus content, altitude, slope, aspect, water source conditions, etc.). Typical de ≈15-20 (depending on data granularity). Trait vector Ptarg ∈R dp Dimension dp Quantitative characteristics (e.g., yield, sugar content, acidity, color, size, flavor, slag-forming properties, etc.); typical dp ≈7-10. Furthermore, the output is defined as the target gene vector. g ′∈R k Dimension k The cumulative variance rate is chosen in this example to represent the length of the gene vector after dimensionality reduction.

[0120] S4. Upscale the target gene vector g' to the target gene g''. Find the existing variety in the variety database that is closest to the target gene g'' as the recommended plant. If the gene sequence similarity is less than the first threshold, then: determine the closest existing variety as the first parent, and store the genotype of the first parent in the database. yes The proximity to the target gene g'' is calculated. If the proximity is less than the second threshold, proceed to step 5A. If the proximity is greater than or equal to the second threshold, proceed to step 5B. The proximity can be calculated based on Euclidean distance or Mahalanobis distance.

[0121] S5A. Several existing varieties closest to the target gene g'' are obtained as parents and combined. Then, breeders carry out breeding work and finally select the best offspring varieties. Then, the offspring varieties are planted in the planting area, and several offspring plants are used as samples for continuous feature acquisition. This feature acquisition may be to obtain images of these citrus plants.

[0122] Because the growth of citrus trees is strongly correlated with the period in front of them, and has virtually no correlation with earlier periods (e.g., the growth of citrus trees only shows a strong correlation with the citrus plants of the previous few days, and has virtually no correlation with earlier periods), it is possible to base our analysis on the various traits of the corresponding varieties stored in the database, as well as past experience data:

[0123] Feature extraction is performed on images uploaded from planting terminals. Then, based on the current features, a Markov chain is used to predict the next stage of trend and calculate the confidence interval of the prediction result. If the confidence interval reaches the third threshold, it means that a certain growth trend is highly likely to occur, and planting suggestions are made based on this growth trend; otherwise, the prediction is unreliable, and planting suggestions already stored in the existing database are retrieved.

[0124] S5B, Store the genotypes of the first parent in the database. yes Distinguishing features are obtained by comparing with the target gene g'', and suggestions are made for the second parent based on the distinguishing features and the target gene vector g'; specifically including:

[0125] S51B, gene differential quantification and key site localization:

[0126] The distinguishing features are calculated by measuring gene difference vectors. get:

[0127]

[0128] Then, extract the set of SNP sites with significant differences:

[0129]

[0130] in, t This is the significance threshold (default 3.0). s Δ is the standard deviation of Δg. This is a set of indexes of key differential sites.

[0131] Understandably, the index set of key differential loci represents regions where there are significant differences between the first parent and the target genotype.

[0132] S52B, Establishing a second parental screening model, including:

[0133] Based on the aforementioned gene differential vector Using this as input, the output is a set of key differential site indexes. L = { j | |Δg [ j ] | ≥ t·sD}, where τ is the significance threshold (e.g., 3.0) and σΔ is the standard deviation of Δg.

[0134] S522B, Setting a parental complementarity scoring function:

[0135] For each candidate variety in the variety database, calculate its complementarity matching degree with the target gene g'' on the key site set L:

[0136]

[0137] in, The difference weight at position j (the greater the difference, the higher the weight); Gene vectors for each candidate variety, This is an indicator function that returns 1 if the genotypic change direction of the candidate variety at this locus is consistent with the target requirement (i.e., has the same sign as Δg), otherwise returns 0; it also calculates the hybridization compatibility between the candidate variety and the first parent (based on a known compatibility database or genetic distance):

[0138]

[0139] Where γ is the scaling parameter and GeneticDistance is the genetic distance, which can be calculated using Euclidean distance or an existing kinship database.

[0140] S523B, Comprehensive Filtering and Sorting;

[0141] A comprehensive score is calculated for each candidate variety, including:

[0142]

[0143] Where α ∈ [0,1] is the weight parameter (default 0.7, which emphasizes genetic complementarity).

[0144] Finally, sort the candidates in descending order of their comprehensive scores and select the top M candidate varieties as second parent recommendations (M is usually 3-5). Output the list of second parents and their complementary site details (i.e., a visualized difference compensation report).

[0145] According to an embodiment of this application, an artificial intelligence-based smart planting optimization system is used in the artificial intelligence-based planting optimization method. The system includes a cloud platform and a first terminal and a second terminal communicating with the cloud platform.

[0146] The cloud platform includes: a genotype matrix construction module, used to receive gene data G and generate the corresponding genotype matrix G'; and a genome-wide regression modeling module, which is equipped with a genome-wide regression model to estimate the relationship between all SNP markers and the target trait. t joint effect vector The genotype score generation module is used to obtain the gene vector g corresponding to any variety in the genotype matrix G' from the variety database, and calculate the genotype score of the gene vector g for trait t. The genotype scores of all target traits are combined into a score vector S; the prediction module retrieves the target environmental vector Etarg and the trait vector Ptarg from the environmental database and outputs the target gene vector g'.

[0147] The first terminal is for researchers, including a database management interface for maintaining variety and environmental databases, and a model training interface for regulating hyperparameters (such as the regularization coefficient λ) and other preset parameters of the genome-wide regression modeling module, as well as the bi-branch prediction model structure. The second terminal is for farmers, including an image upload interface for real-time acquisition of field environmental data and generation of target environmental vectors, and a download interface for receiving variety recommendations and growth warning suggestions from the cloud platform.

[0148] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A smart planting optimization method based on artificial intelligence, characterized in that, include: A variety database and an environment database are established based on historical data. The environment database stores several first environmental feature vectors EA, and the variety database stores basic variety information, second environmental feature vectors EB, gene data G, and the phenotypic vector P corresponding to each variety under the second environmental feature vector EB. Generate the corresponding genotype matrix G' based on the gene data G; For each element pt in the trait vector Pi, a genome-wide regression model is built to estimate the relationship between all SNP markers and the target trait. t joint effect vector Then, obtain the gene vector g corresponding to any variety in the variety database in the genotype matrix G', and calculate the genotype score of the gene vector g for trait t. And combine the genotype scores of all target traits into a single score vector S; Then, a vector knowledge base is constructed, defining the association between entities and relations. The tail entity vector is predicted based on the pre-trained multimodal relation joint embedding model. The model input includes the head entity vector and relation vector. All entities are stored in a single vector database after being processed by the joint embedding model, and a unified index is created. Based on the index results, the target environment vector Etarg is retrieved from the environment database. The user constructs the target trait vector Ptarg based on experience. Then, the target environment vector Etarg and the target trait vector Ptarg are input into the prediction model to obtain the target gene vector g', which includes: A prediction model network structure is constructed, wherein the prediction model is a two-branch input fusion structure; wherein the target environment vector Etarg and the target trait vector Ptarg are first processed through independent sub-networks corresponding to the two-branch structure, and the outputs of the two independent sub-networks are fused and then used with the genotype score vector. S Linear layer projection is used as a supervision signal, and the target gene vector of the output result is enhanced by a loss function. g ′ and genotype score S The linear relationship ultimately outputs the target gene vector. g ′; The prediction model network structure includes: The system includes an environmental feature processing module and a trait feature processing module. The environmental feature processing module takes the target environmental vector Etarg as input and includes a first hidden layer, a lightweight Transformer layer, and a first fully connected layer. The trait feature processing module takes the target trait vector Ptarg as input and includes a second hidden layer and a second fully connected layer. The system also includes a fusion module, which comprises a cross-attention fusion layer, a linear layer, a random depth residual block, and a channel compression layer.

2. The intelligent planting optimization method based on artificial intelligence according to claim 1, characterized in that, The establishment of the variety database includes constructing a genotype matrix G' based on gene data G: An initial genotype matrix Graw is constructed based on the gene data G, with dimensions of . m × n ;in, m Indicates the number of varieties. n The matrix represents the number of SNP molecular marker sites. Graw [ i , j ] indicates the first i The first variety in j Genotype coding values ​​at each SNP locus; Using singular value decomposition, we can obtain: Graw= UFV T ； In the formula, U It is a left singular vector matrix with dimension . m × m In the left singular vector matrix U [ i ,:] is the first i The projection of each sample onto all left singular vectors; F It is a diagonal matrix with dimension 1. m × n Its diagonal elements include: σ 1, σ 2,…, σr ( r =min( m , n )) is a singular value, and σ 1≥ σ 2≥…≥ σr ≥0; V T It is a right singular vector matrix with dimension . n × n In the right singular vector matrix V [:, j ] is the first j The weights of each SNP on all right singular vectors; Then, dimensionality reduction and truncation are performed, taking the first-order dimensionality. k The principal components include: U k = U [:,0: k ] F k =F[0: k ,0: k ] V k = V [:,0: k ] in, k It is the target dimension number after dimensionality reduction. U k The dimension is m × k , including samples in front k Coordinates along the directions of the main genetic structures; F k The dimension is k × k , including the previous k The largest singular value; V k The dimension is n × k Including SNP sites in the front k Loads in the main genetic structural directions; Constructing a genotype matrix G 'for: G ′= U k F k Its dimensions are m × k , G Each line of ' gi Representing the i A low-dimensional genetic trait representation vector for each variety, with a dimension of 1× k .

3. The intelligent planting optimization method based on artificial intelligence according to claim 1, characterized in that, A genome-wide regression model is established for each element pt in the trait vector Pi to estimate the relationship between all SNP markers and the target trait. t joint effect vector include: Establish a genome-wide regression model: ； in, p t It is the phenotypic vector of all varieties on trait t. It is the label effect vector. For error; in The estimation method is as follows: ； In the formula, λ is the regularization parameter. I It is a with G ′ ⊤ G An identity matrix of the same dimension; The genotype score The calculation method is as follows: , This is the transpose of the genotype vector g for a specific variety; Combine the genotype scores corresponding to all traits that need to be evaluated into a single score vector S=[ s 1, s 2,…, s n ].

4. The intelligent planting optimization method based on artificial intelligence according to claim 3, characterized in that, The features output by the environmental feature processing module and the trait feature processing module are fused and then input into the cross-attention fusion layer. The query Q, key K, and value V in the cross-attention fusion layer are defined as follows: Q = h env ⋅ Wq, K = h trait ⋅ Wkv, V = h trait ⋅ Wkv ； in, h env For the environmental features in the target environment vector Etarg, Wq The query projection matrix for environmental representation. h trait For the trait features in the target trait vector Pi, Wkv The key / value projection matrix represents the traits; The environmental attention weights on traits are calculated by scaling the dot product attention, reflecting the regulatory strength of environmental factors on different traits, and the genotype score vector is then used. S The linear layer is projected as a monitoring signal. S And concatenate or add it to the attention output, including: S311. Calculate the original attention weights: ； in d k It is the dimension of the key vector, and the scaling factor prevents gradient vanishing; S312. Pass the genotype score vector S through a linear layer and a sigmoid activation function to generate a gated vector. ; ； in, This is the weight matrix. Here, dp is the bias vector, and dp is the dimension of the vector, which represents the number of target traits. S313. Modulate the attention weight matrix using a gating vector: ； Wherein, diag(1-Gate) is a diagonal matrix with (1-Gate) as the diagonal; the stronger the genetic prior of a certain trait, the larger the Gate value, which correspondingly weakens the regulatory attention weight of environmental factors on that trait, and the smaller the 1-Gate value. splicing attention output and S After processing through 3 layers of random depth residual blocks and channel compression layers, the final output is a k-dimensional target gene vector. g ′.

5. The intelligent planting optimization method based on artificial intelligence according to claim 4, characterized in that, The target gene vector of the output result is enhanced by the loss function. g ′ and genotype score S The loss function in a linear relationship is set as follows: ； In the formula, N The number of training samples, g true( i ) is the first i The true gene vector of each sample S ( i ) is the first i Genotype score vector for each sample, B ∈R k×dp This is the labeling effect matrix, with the following columns: βt , λ This is the regularization coefficient.

6. The intelligent planting optimization method based on artificial intelligence as described in claim 5, characterized in that, Also includes: The target gene vector g' is increased in dimension to the target gene g'', and the existing varieties that are closest to the target gene g'' are found in the variety database as recommended plants; If the gene sequence similarity is less than the first threshold, then: the closest existing variety is identified as the first parent, and the genotype of the first parent is stored in the database. gi Calculate the proximity to the target gene g''. If the proximity is less than the second threshold, proceed to step 5A. If the proximity is greater than or equal to the second threshold, proceed to step 5B. Step 5A includes: obtaining multiple existing varieties that are closest to the target gene g'' as parents, combining them, and breeding and screening to obtain the optimal progeny varieties; planting the progeny varieties in the planting area; and continuously acquiring features from several progeny plants as samples, which involves acquiring images of these citrus plants. The system extracts features from images uploaded by the planting terminal, then uses Markov chains to predict the next stage of trends based on the current features. It calculates the confidence interval of the prediction results. If the confidence interval reaches the third threshold, it means that a certain growth trend is very likely to occur, and planting suggestions are made based on this growth trend. Otherwise, if the prediction is unreliable, the system retrieves the planting suggestions already stored in the existing database. Step 5B includes: The genotypes of the first parent are stored in the database. gi Distinguishing features are obtained by comparing with the target gene g''. Based on the distinguishing features and the target gene vector g', suggestions are made for the second parent, including: S51B, gene differential quantification and key site localization: The distinguishing features are calculated by measuring gene difference vectors. get: ； Then, extract the set of SNP sites with significant differences: ； in, τ The significance threshold, σ Δ is the standard deviation of Δg. A set of indexes of key differentially expressed sites; S52B, Establishing a second parental screening model, including: Based on the aforementioned gene differential vector Using this as input, the output is a set of key differential site indexes. L = { j | |Δg [ j ] | ≥ τ·σΔ }, where τ is the significance threshold and σΔ is the standard deviation of Δg; S522B, Setting a parental complementarity scoring function: For each candidate variety in the variety database, calculate its complementarity matching degree with the target gene g'' on the key site set L: ； in, The difference weight at position j is given by the fact that the greater the difference, the higher the weight. Gene vectors for each candidate variety, This is an indicator function that returns 1 if the genotypic change direction of the candidate variety at this locus is consistent with the target requirement, and 0 otherwise; it also calculates the hybridization compatibility between the candidate variety and the first parent. ； Where γ is the scaling parameter and GeneticDistance is the genetic distance, which can be calculated using Euclidean distance or an existing kinship database; S523B, Comprehensive Filtering and Sorting; A comprehensive score is calculated for each candidate variety, including: ； Where α ∈ [0,1] are weight parameters; Finally, sort the candidates in descending order of their comprehensive scores, select the top M candidate varieties as second parent recommendations, and output the list of second parents and details of their complementary sites.

7. A smart planting optimization system based on artificial intelligence, characterized in that, The system for implementing the AI-based smart planting optimization method according to any one of claims 1-6 includes a cloud platform and a first terminal and a second terminal communicating with the cloud platform. The cloud platform includes: a genotype matrix construction module, used to receive gene data G and generate the corresponding genotype matrix G'; and a genome-wide regression modeling module, which is equipped with a genome-wide regression model to estimate the relationship between all SNP markers and the target trait. t joint effect vector The genotype score generation module is used to obtain the gene vector g corresponding to any variety in the genotype matrix G' from the variety database, and calculate the genotype score of the gene vector g for trait t. The genotype scores of all target traits are combined into a score vector S; the prediction module retrieves the target environmental vector Etarg and the trait vector Ptarg from the environmental database and outputs the target gene vector g'. The first terminal is for researchers and includes a database management interface and a model training interface; the second terminal is for farmers and includes an image upload interface and a download interface.

Images