Trans-variety camellia oleifera grafting adaptability evaluation method based on transfer learning

Abstract

The invention discloses a transfer learning-based cross-variety camellia oleifera grafting adaptability evaluation method, which comprises the following steps: collecting macroscopic features and microscopic features of target camellia oleifera, the macroscopic features comprising plant morphological indexes of plant height, crown breadth and / or trunk diameter, and the microscopic features comprising gene expression data; carrying out data preprocessing on the collected macroscopic features and microscopic features, and constructing a double-flow neural network; the macroscopic and microscopic feature flow outputs are spliced through an attention mechanism; and inputting the spliced features into a full-connection layer, and obtaining an adaptability score of the target camellia oleifera by using an activation function. Through the scheme of the invention, the feature representation capability of the model can be enhanced, and the generalization level of the model on new varieties can be improved.

Description

Technical Field

[0001] The present application relates to the field of machine learning, and in particular to an adaptability evaluation method for cross-variety tea-oil grafting based on transfer learning. Background Art

[0002] At present, the methods for evaluating the adaptability of cross-variety Camellia oleifera grafting mainly rely on traditional statistical analysis and empirical judgment. The existing technologies in this field have many shortcomings and are in urgent need of improvement and innovation. Specifically, researchers usually collect various physiological indicators of grafted plants, such as survival rate, growth rate, leaf characteristics, etc., and then use simple statistical models (such as linear regression or analysis of variance) to analyze these data. Although this method is simple and intuitive, it is difficult to capture complex nonlinear relationships and potential interactions. The limitations of traditional methods are mainly reflected in the following aspects: First, linear models assume that there is a linear relationship between dependent variables and independent variables, which is often not true in biological systems; second, these methods are difficult to handle high-dimensional data and multi-factor interactions; third, they are sensitive to outliers and noise data, which may lead to biased results. In addition, traditional methods usually ignore time series characteristics and cannot effectively capture dynamic changes during plant growth.

[0003] Another common method is to use machine learning algorithms, such as support vector machines (SVM) or random forests, to build predictive models. These models attempt to predict the adaptability of new grafting combinations by learning known grafting result data. However, these methods often require a large amount of training data, while in practical applications, grafting experimental data between different varieties may be relatively limited. Although machine learning methods are superior to traditional statistical methods in some aspects, there are still some problems: first, the interpretability of the model is poor, and it is difficult to provide researchers with intuitive biological explanations; second, these models are prone to overfitting, especially when training data is limited; third, they are sensitive to feature selection and hyperparameter adjustment, and require a lot of manual intervention.

[0004] Therefore, there is an urgent need for a technical solution that can enhance the feature representation capability of the model and improve the generalization level of the model on new varieties. Summary of the invention

[0005] In order to solve the shortcomings of the prior art, the embodiment of the present application provides an adaptability evaluation method for cross-variety Camellia oleifera grafting based on transfer learning. The present application solves the technical problems that the prior art has a low generalization level on new varieties.

[0006] The embodiment of the present application provides an adaptability assessment method for cross-variety tea oil grafting based on transfer learning, comprising: collecting macroscopic and microscopic characteristics of the target tea oil, the macroscopic characteristics including plant morphological indicators such as plant height, crown width and / or trunk diameter, and the microscopic characteristics including gene expression data; performing data preprocessing on the collected macroscopic and microscopic characteristics, and constructing a two-stream neural network; splicing the macroscopic and microscopic feature stream outputs through an attention mechanism; inputting the spliced ​​features into a fully connected layer, and using an activation function to obtain the adaptability score of the target tea oil.

[0007] In a possible implementation, macroscopic and microscopic characteristics of the target tea oil are collected, wherein the macroscopic characteristics include plant morphological indicators such as plant height, crown width and / or trunk diameter, and the microscopic characteristics include gene expression data, including: selecting plants of the target tea oil to be measured, measuring the plant height using an optical distance measuring device, and measuring the crown width using a length measuring tool, and taking the average value thereof; measuring the trunk diameter at a specific height of the plant, recording the number of leaves and randomly selecting some leaves; using a leaf area meter and a chlorophyll content meter to respectively measure the area and chlorophyll content of the leaves; at a preset time point after the plant is grafted, collecting tissue samples within a specific range around the grafting site, and performing low-temperature rapid freezing treatment; using a nucleic acid extraction reagent to extract total RNA from the tissue sample, and using a spectrophotometer and a bioanalyzer to detect the concentration, purity and integrity of the RNA; performing reverse transcription to obtain cDNA, and using real-time quantitative PCR to detect the expression level of a preselected gene.

[0008] In a possible implementation, data preprocessing is performed on the collected macroscopic features and microscopic features, including: performing standardization processing on the macroscopic feature data, including obtaining raw data, calculating the mean and standard deviation, and converting the raw data into standardized data using the Z-score method; performing quantitative calculation of relative expression on the microscopic feature data, including: ,in, Indicates the Ct value of the target gene, represents the Ct value of the internal reference gene, Indicates the treatment group value, The control group value, represents the standard deviation, Indicates relative expression.

[0009] In one possible implementation, the two-stream neural network includes: a macro feature processing stream and a micro feature processing stream; wherein, in constructing the two-stream neural network: the macro feature processing stream includes an input layer, multiple fully connected layers and an activation layer connected in sequence; the micro feature processing stream includes an input layer, multiple one-dimensional convolutional layers, a pooling layer, a fully connected layer and an activation layer connected in sequence, wherein the output of the one-dimensional convolutional layer is subjected to a pooling operation after activation.

[0010] In a possible implementation, the macro and micro feature stream outputs are spliced ​​through an attention mechanism, including: splicing the macro and micro feature stream output vectors to obtain a spliced ​​feature vector; performing a nonlinear transformation on the spliced ​​feature vector using trainable parameters, and calculating the attention weight through a softmax function; and performing weighted fusion of the macro feature stream output vector and the micro feature stream output vector according to the attention weight to obtain a fused feature vector.

[0011] In a possible implementation, before the macro and micro feature flow outputs are spliced ​​through the attention mechanism, the method includes: dividing the existing tea oil variety data into a source domain data set and a target domain data set, wherein the source domain data set contains common varieties with sufficient data, and the target domain data set contains new varieties with sparse data; calculating the maximum average difference between the source domain data set and the target domain data set, using the Gaussian kernel function as the feature mapping function, and determining the value range of the kernel width parameter through cross-validation; constructing a loss function, including the loss terms of the source domain data set and the target domain data set, and the maximum average difference weighted term between the source domain data set and the target domain data set, and setting the initial value and variation range of the balance parameter; using the adaptive moment estimation optimization algorithm to train the model, setting the initial learning rate, learning rate decay strategy, batch size and number of training iterations; during the training process, regularly calculating the performance indicators on the validation set, and saving the model with the best performance.

[0012] In a possible implementation, constructing a loss function includes: ,in, represents the total loss function, represents the source domain loss, represents the equilibrium parameter, with an initial value of 0.1. represents the maximum average difference between the source domain and the target domain, represents the target domain loss.

[0013] In a possible implementation, the maximum average difference includes: ,in, represents the number of source domain samples, represents the i-th sample in the source domain, represents the number of samples in the target domain, represents the jth sample in the target domain, represents the Gaussian kernel function, represents the Hilbert square norm.

[0014] In the adaptability evaluation method of cross-variety Camellia oleifera grafting based on transfer learning provided above, the embodiment of the present application can simultaneously utilize information at the macro and micro levels through a multi-scale transfer learning framework, thereby improving the feature representation ability of the model. Furthermore, in some embodiments, the weights of the source domain and the target domain are adjusted through a dynamic weight adaptive algorithm, which can improve the generalization ability of the model on new varieties. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 A schematic flow chart of an adaptability evaluation method for cross-variety Camellia oleifera grafting based on transfer learning provided in an embodiment of the present application; Figure 2 A flowchart of a method for training a two-stream neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0018] Those skilled in the art can understand that the terms "first", "second" and the like in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor represent the necessary logical order between them. It should also be understood that in the embodiments of the present disclosure, "multiple" can refer to two or more, and "at least one" can refer to one, two or more. It should also be understood that for any component, data or structure mentioned in the embodiments of the present disclosure, in the absence of explicit limitation or contrary revelation given in the context, it can generally be understood as one or more. In addition, the term "and / or" in the present disclosure is only a description of the association relationship of the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are an "or" relationship. It should also be understood that the description of each embodiment in the present disclosure emphasizes the differences between the embodiments, and the same or similar parts can refer to each other. For the sake of brevity, they will not be repeated one by one.

[0019] At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered part of the specification. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] Figure 1A flow chart of a method for evaluating the adaptability of cross-variety Camellia oleifera grafting based on transfer learning provided in an embodiment of the present application. It should be noted that, at present, the methods for evaluating the adaptability of cross-variety Camellia oleifera grafting often rely on traditional statistical analysis and empirical judgment. Researchers usually collect various physiological indicators of grafted plants, such as survival rate, growth rate, leaf characteristics, etc., and then use simple statistical models (such as linear regression or analysis of variance) to analyze these data. Although this method is simple and intuitive, it is difficult to capture complex nonlinear relationships and potential interactions.

[0022] Another common approach is to use machine learning algorithms, such as support vector machines (SVM) or random forests, to build predictive models. These models attempt to predict the adaptability of new grafting combinations by learning from known grafting results data. However, these methods often require a large amount of training data, while in practical applications, grafting experimental data between different varieties may be relatively limited.

[0023] The existing technology faces two main problems: 1. Data sparsity and variety specificity: The data of grafting experiments between different oil-tea camellia varieties are often sparse, because conducting large-scale grafting experiments is time-consuming and expensive. Traditional machine learning methods are prone to overfitting problems when faced with such sparse data, resulting in insufficient model generalization ability. Specifically, the model may overfit to the characteristics of certain specific varieties and fail to accurately predict the grafting adaptability of new varieties. This phenomenon can be mathematically expressed as a high variance problem of the model: ,in, is the prediction function of the model, and x is the input feature. When the available training data is sparse, will become larger, causing the model to be unstable on new data.

[0024] 2. Limitations of feature representation: Existing methods usually use artificially extracted features, such as morphological characteristics or basic physiological indicators of plants. This feature extraction method may ignore some potentially important information, especially interactions at the molecular level. Traditional feature extraction methods have difficulty capturing the dynamic changes in gene expression levels, which are crucial for evaluating grafting adaptability. From the perspective of information theory, this problem can be described as information loss in the feature extraction process: I(X; Y) - I(f(X); Y) > ε, where I(X; Y) represents the mutual information between the original data X and the target variable Y, I(f(X); Y) represents the mutual information between the extracted features f(X) and Y, and ε is a positive number representing an acceptable threshold of information loss. When this difference is large, it means that a lot of valuable information has been lost in the feature extraction process.

[0025] In order to solve the above technical problems, the embodiment of the present application discloses a multi-scale transfer learning framework, which can simultaneously utilize information at the macroscopic (such as morphological features) and microscopic (such as gene expression data) levels. Specifically, a two-stream neural network structure is constructed, in which one stream processes macroscopic features and the other stream processes microscopic features. The two streams are fused through the attention mechanism to achieve cross-scale information integration. At the same time, in order to solve the problem of data sparsity, a dynamic weight adaptive algorithm is disclosed. The algorithm can dynamically adjust the weights in the transfer learning process according to the similarity between the source domain (variety with sufficient data) and the target domain (new variety).

[0026] The two work together to form a complete cross-variety Camellia oleifera grafting adaptability assessment method. The multi-scale transfer learning framework provides a richer and more comprehensive feature representation, while the dynamic weight adaptation algorithm ensures that the model can effectively use limited data for learning and generalization.

[0027] Specifically, if Figure 1 As shown, in step S101, macroscopic characteristics and microscopic characteristics of the target tea oil plant are collected, wherein the macroscopic characteristics include plant morphological indicators such as plant height, crown width and / or trunk diameter, and the microscopic characteristics include gene expression data.

[0028] It should be understood that "collection" here is a broad concept, including but not limited to field measurement, instrument detection, gene sequencing and other data acquisition methods. Macro characteristics mainly refer to the morphological indicators of oil-tea camellia plants, including plant height, crown width and / or trunk diameter. These indicators are important parameters for evaluating the growth status and adaptability of oil-tea camellia. Among them, plant height reflects the longitudinal growth of oil-tea camellia, which is usually measured using a height meter or laser rangefinder; crown width indicates the horizontal expansion of the crown of the oil-tea camellia, which is usually obtained by measuring the maximum diameter of the crown in the east-west and north-south directions and taking the average value; the trunk diameter reflects the thickness of the oil-tea camellia, which is usually measured at 1.3 meters above the ground using a breast diameter ruler. These macro characteristics can intuitively reflect the growth status and environmental adaptability of oil-tea camellia.

[0029] The microscopic features focus on the gene expression data of Camellia oleifera. Gene expression refers to the process of transcribing genetic information from DNA into RNA and then translating it into protein. Gene expression data are usually obtained through high-throughput sequencing technologies, such as RNA sequencing (RNA-seq) or microarray analysis. These data can reflect the gene activity state of Camellia oleifera under specific environmental conditions and provide molecular-level evidence for evaluating its adaptability. The collection process of gene expression data usually includes multiple steps such as sample collection, RNA extraction, sequencing or chip hybridization, and data processing. This microscopic feature analysis can reveal the molecular mechanism of Camellia oleifera's adaptation to the environment and provide a deeper basis for adaptability assessment.

[0030] Furthermore, the method includes: selecting target oil-tea plants to be tested, measuring plant heights using an optical distance measuring device, and measuring crown widths using a length measuring tool, and taking an average value thereof; measuring trunk diameters at a specific height of the plants, recording the number of leaves and randomly selecting some leaves; using a leaf area meter and a chlorophyll content meter to respectively measure the area and chlorophyll content of the leaves; collecting tissue samples within a specific range around the grafting site at a preset time point after the plants are grafted, and performing low-temperature rapid freezing treatment; extracting total RNA from the tissue samples using a nucleic acid extraction reagent, and detecting the concentration, purity and integrity of the RNA using a spectrophotometer and a bioanalyzer; performing reverse transcription to obtain cDNA, and using real-time quantitative PCR to detect the expression level of a preselected gene.

[0031] In one implementation scenario, the steps for collecting macroscopic features are as follows: Select healthy tea oil plants and use a laser rangefinder (such as Leica DISTO D2, with a measurement accuracy of ±1.5mm) to measure the plant height, with a value accurate to 0.1cm. Use a tape measure to measure the crown width of the plant in the east-west and north-south directions, and take the average value, accurate to 1cm. Use a vernier caliper (such as Mitutoyo 500-196-30, with an accuracy of 0.01mm) to measure the trunk diameter 30cm above the ground, and take the average value of the two vertical directions. Record the number of leaves, and randomly select 30 fully expanded leaves to measure the leaf area using a portable leaf area meter (such as LI-COR LI-3000C, with an accuracy of ±1%). Use a portable chlorophyll meter (such as Konica Minolta SPAD-502Plus, with an accuracy of ±1.0SPAD unit) to measure the leaf SPAD value, measure 3 points on each leaf, and take the average value.

[0032] The steps for collecting microscopic features were as follows: leaf and phloem tissue samples within 5 cm above and below the grafting site were collected 7, 14, and 21 days after grafting. The samples were quickly frozen using liquid nitrogen and ground into powder using a mortar. Total RNA was extracted using TRIzol reagent, and RNA concentration and purity were detected using a NanoDrop 2000 spectrophotometer. RNA integrity was checked using an Agilent 2100 bioanalyzer. Reverse transcription was performed using the PrimeScript RT reagent Kit to obtain cDNA. Real-time fluorescence quantitative PCR (qRT-PCR) technology was used to detect the expression levels of key genes. Genes related to graft healing, such as WRKY transcription factors, β-1,3-glucanase, and chitinase, were selected. The Bio-Rad CFX96 real-time PCR system was used for detection, and 3 biological replicates and 3 technical replicates were set for each sample.

[0033] At step S102, the collected macroscopic and microscopic features are subjected to data preprocessing, and a two-stream neural network is constructed. Data preprocessing is a key step before deep learning model training, and its purpose is to improve data quality and enhance the learning effect of the model. For macroscopic features, preprocessing may include data normalization, outlier processing, missing value filling, etc. For example, plant height, crown width, and trunk diameter may have different dimensions, and they need to be converted to the same scale through normalization. For microscopic features, i.e., gene expression data, preprocessing may involve data standardization, dimensionality reduction, batch effect correction, etc. For example, RNA-seq data can be standardized using tools such as DESeq2 or edgeR, or dimensionality reduction can be performed using methods such as principal component analysis (PCA) to extract the most representative gene expression features.

[0034] The two-stream neural network is a special deep learning architecture that can process two different types of input data at the same time. In an embodiment of the present application, one stream of the two-stream neural network is used to process macro features, and the other stream is used to process micro features. This architectural design allows the model to learn and extract representations of the two features separately, and then fuse them in subsequent steps. The macro feature stream may consist of multiple fully connected layers, each of which uses a nonlinear activation function (such as ReLU) to increase the expressive power of the model. The micro feature stream may contain convolutional layers or recurrent neural network layers to capture complex patterns and timing information in gene expression data.

[0035] Specifically, it includes: performing standardization processing on the macroscopic feature data, including obtaining raw data, calculating the mean and standard deviation, and converting the raw data into standardized data using the Z-score method; performing quantitative calculation of the relative expression amount on the microscopic feature data, including: ,in, Indicates the Ct value of the target gene, represents the Ct value of the internal reference gene, Indicates the treatment group value, The control group value, represents the standard deviation, Indicates relative expression.

[0036] The two-stream neural network includes: a macro feature processing stream and a micro feature processing stream; wherein, in constructing the two-stream neural network: the macro feature processing stream includes an input layer, multiple fully connected layers and an activation layer connected in sequence; the micro feature processing stream includes an input layer, multiple one-dimensional convolutional layers, a pooling layer, a fully connected layer and an activation layer connected in sequence, wherein the output of the one-dimensional convolutional layer is subjected to a pooling operation after activation.

[0037] In one embodiment, a two-stream neural network can be constructed. The network structure is as follows: Macro feature processing flow: input layer → fully connected layer (128 nodes) → ReLU activation → Dropout (0.3) → fully connected layer (64 nodes) → ReLU activation; Micro-feature processing flow: input layer → one-dimensional convolution layer (32 filters, kernel size 3) → ReLU activation → maximum pooling → one-dimensional convolution layer (64 filters, kernel size 3) → ReLU activation → global average pooling → fully connected layer (64 nodes) → ReLU activation.

[0038] At step S103, the macroscopic and microscopic feature stream outputs are spliced ​​together through the attention mechanism. The attention mechanism allows the model to "focus on" the most relevant parts when processing the input. In the embodiment of the present application, the introduction of the attention mechanism enables the model to dynamically adjust the importance weights of macroscopic features and microscopic features. For example, for some oil-tea camellia varieties, gene expression may better reflect its adaptability than morphological features. At this time, the attention mechanism will give microscopic features a higher weight. Vice versa. This adaptive weight allocation mechanism greatly improves the flexibility and generalization ability of the model.

[0039] Specifically, the attention mechanism can be implemented by calculating the attention weights. First, the outputs of the macro feature stream and the micro feature stream are transformed through a fully connected layer to obtain the query vector (query), the key vector (key), and the value vector (value). Then, the attention weights are obtained by calculating the similarity between the query vector and the key vector and normalizing them with the softmax function. Finally, these weights are multiplied by the value vector and summed to obtain the weighted feature representation. This process can be expressed as: , where Q, K, and V represent query, key, and value matrices, respectively. is the dimension of the key vector. In this way, the model is able to automatically learn how to optimally combine macro and micro features.

[0040] Feature concatenation is the process of merging the macro and micro feature vectors processed by the attention mechanism into a larger vector. This process can be achieved through simple vector concatenation operations, or more complex fusion methods such as weighted summation or gating mechanisms can be used. The concatenated feature vector contains comprehensive information about Camellia oleifera, including both macroscopic morphological features and microscopic gene expression patterns, providing rich input for subsequent fitness scoring.

[0041] Furthermore, it includes: splicing the macro and micro feature flow output vectors to obtain a spliced ​​feature vector; performing a nonlinear transformation on the spliced ​​feature vector using trainable parameters, and calculating the attention weight through a softmax function; and performing weighted fusion of the macro feature flow output vector and the micro feature flow output vector according to the attention weight to obtain a fused feature vector.

[0042] For example, in the feature fusion layer, the outputs of the macro and micro feature streams are concatenated and then fused through the attention mechanism. The specific implementation of the attention mechanism is as follows: the concatenated feature vector is recorded as . Calculate the attention weights: , where v, W and b are learnable parameters, and the Xavier initialization method is used for initialization. Weighted fusion: .

[0043] In addition (such as Figure 2 As shown in the figure, before the macro and micro feature flow outputs are spliced ​​through the attention mechanism, the method includes: dividing the existing Camellia oleifera variety data into a source domain dataset and a target domain dataset, wherein the source domain dataset contains common varieties with sufficient data, and the target domain dataset contains new varieties with sparse data; calculating the maximum average difference between the source domain dataset and the target domain dataset, using the Gaussian kernel function as the feature mapping function, and determining the value range of the kernel width parameter through cross-validation; constructing a loss function, including the loss terms of the source domain dataset and the target domain dataset, and the weighted term of the maximum average difference between the source domain dataset and the target domain dataset, and setting the initial value and variation range of the balance parameter; using the adaptive moment estimation optimization algorithm to train the model, setting the initial learning rate, learning rate decay strategy, batch size and number of training iterations; during the training process, regularly calculating the performance indicators on the validation set, and saving the model with the best performance.

[0044] At step S104, the concatenated features are input into the fully connected layer, and the activation function is used to obtain the adaptability score of the target tea oil. The fully connected layer is a type of layer in a neural network, which connects all neurons in the previous layer to each neuron in the current layer. In the embodiment of the present application, the function of the fully connected layer is to map the concatenated high-dimensional features to the adaptability score space. This process can be expressed as: , where x is the input feature vector, W is the weight matrix, b is the bias vector, f is the activation function, and y is the output. The choice of activation function is crucial to the performance of the model. Common activation functions include ReLU (Rectified Linear Unit), Sigmoid, and Tanh. In an embodiment of the present application, a suitable activation function can be selected according to the specific needs of the adaptability score. For example, if you want the score range to be between 0 and 1, you can choose the Sigmoid function; if you want the score range to be between -1 and 1, you can choose the Tanh function.

[0045] In one embodiment, the fused features are passed through the last fully connected layer (32 nodes) and ReLU activation, and finally the grafting fitness score (a continuous value between 0 and 1) is output through the sigmoid function.

[0046] Figure 2 A schematic diagram of a flow chart of a method for training a two-stream neural network provided in an embodiment of the present application. During the model training process, a dynamic weight adaptive algorithm can be used to solve the problem of data sparsity.

[0047] like Figure 2 As shown, at step S201, the existing oil-tea variety data is divided into a source domain data set and a target domain data set, wherein the source domain data set contains common varieties with sufficient data, and the target domain data set contains new varieties with sparse data. That is, the existing oil-tea variety data is divided into a source domain (common varieties with sufficient data) and a target domain (new varieties with sparse data).

[0048] In step S202, the maximum average difference between the source domain data set and the target domain data set is calculated, a Gaussian kernel function is used as a feature mapping function, and a value range of a kernel width parameter is determined by cross-validation. The maximum average difference includes: ,in, represents the number of source domain samples, represents the i-th sample in the source domain, represents the number of samples in the target domain, represents the jth sample in the target domain, Represents the feature mapping function, using the Gaussian kernel function: , σ is the kernel width parameter, which can be determined by cross-validation and usually takes a value between 0.1 and 10. represents the Hilbert square norm.

[0049] In step S203, a loss function is constructed, including loss terms of the source domain dataset and the target domain dataset, and a weighted term of the maximum average difference between the source domain dataset and the target domain dataset, and an initial value and a variation range of the balance parameter are set. Constructing the loss function includes: ,in, represents the total loss function, represents the source domain loss, represents the balance parameter, with an initial value of 0.1 and gradually increasing to 1.0 as training progresses. represents the maximum average difference between the source domain and the target domain, represents the target domain loss.

[0050] At step S204, the model is trained using the adaptive moment estimation optimization algorithm, and the initial learning rate, learning rate decay strategy, batch size, and number of training iterations are set. The Adam optimizer can be used for model training, with the initial learning rate set to 0.001 and decayed by 10% every 50 epochs. The batch size is set to 32, and 300 epochs are trained.

[0051] At step S205, during the training process, the performance index on the validation set is calculated regularly, and the model with the best performance is saved. For example, the performance on the validation set is calculated once every 10 epochs, and the best model is saved.

[0052] It should be noted that the model performance can be evaluated using leave-one-out cross-validation to calculate the mean absolute error (MAE) and Pearson correlation coefficient. For new grafting combinations of Camellia oleifera varieties, macroscopic and microscopic feature data are collected and input into the trained model to obtain the grafting adaptability score. The score results are divided into three levels: high (>0.8), medium (0.5-0.8) and low (<0.5) to provide guidance for actual grafting operations.

[0053] In addition, when collecting plant samples, individuals with good growth conditions and no obvious diseases and pests should be selected to ensure the representativeness of the data. During RNA extraction, RNase contamination should be strictly controlled, and DEPC-treated water and autoclaved utensils should be used. When performing qRT-PCR, it is necessary to select appropriate internal reference genes (such as Actin or GAPDH) and verify their expression stability. During model training, close attention should be paid to overfitting problems. This can be alleviated by adding L2 regularization (weight decay parameter set to 0.0001) or early stopping strategy. Since there may be systematic errors in experiments of different batches, it is recommended to correct batch effects in the data preprocessing stage. The ComBat algorithm can be used or batch information can be introduced into the model as a covariate.

[0054] Next, the overall method will be described in conjunction with an embodiment. A new tea oil variety A is intended to be grafted onto an existing rootstock variety B. The following steps are followed to perform the evaluation: Collect macroscopic characteristic data: The average plant height of variety A was measured to be 220.3 cm, the crown width was 157.8 cm, and the trunk diameter was 3.27 cm. Thirty leaves were randomly selected, and the average leaf area was measured to be 31.2 cm², and the average SPAD value was 39.7.

[0055] Collecting microscopic feature data: Samples were collected 7 days, 14 days and 21 days after grafting, RNA was extracted and qRT-PCR analysis was performed. Taking WRKY1 transcription factor as an example, the relative expression levels at the three time points were 1.32, 2.87 and 1.95, respectively.

[0056] Data preprocessing: Z-score standardization was performed on macroscopic characteristics. For example, the standardized plant height was: (220.3 -215.6) / 12.4 = 0.379 (assuming that the mean of historical data was 215.6 cm and the standard deviation was 12.4 cm). Relative quantitative analysis was performed on microscopic characteristic data. For example, the expression level of WRKY1 at 14 days was 2.87 times that of the control.

[0057] The processed data was fed into the trained model: the model output grafting fitness score was 0.72.

[0058] Interpretation of results: The score of 0.72 belongs to the medium adaptability level, indicating that the grafting combination of varieties A and B has a certain possibility of success, but still needs to be operated with caution and closely monitored.

[0059] In summary, the embodiments of the present application bring the following beneficial effects at the technical level: Improved feature representation capabilities: The multi-scale transfer learning framework can simultaneously utilize information at both the macro and micro levels, greatly improving the model's feature representation capabilities. This enables the model to capture more potential factors, such as changes in gene expression levels, and thus more accurately assess grafting fitness.

[0060] Enhanced model generalization ability: The dynamic weight adaptation algorithm significantly improves the model's generalization ability on new varieties by adjusting the weights of the source and target domains. This solves the problem of data sparsity and enables the model to better cope with unseen grafting combinations.

[0061] Reduced data requirements: By effectively utilizing existing knowledge and cross-domain migration, the embodiments of the present application reduce the demand for large-scale experimental data. This not only saves time and resources, but also enables small-scale experiments to obtain reliable evaluation results.

[0062] Providing interpretable insights: The introduction of the attention mechanism not only improves model performance, but also provides a certain degree of interpretability for the results. By analyzing the attention weights of different features, researchers can understand which factors have a greater impact on grafting adaptability.

[0063] Furthermore, an embodiment of the present application also provides a device for evaluating the adaptability of cross-variety tea oil grafting, comprising: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any of the above methods.

[0064] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the above-mentioned processing methods.

[0065] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0066] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0067] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements limited by the statement "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment including the elements.

[0068] The above description of the disclosed embodiments enables professionals and technicians in the field to implement or use the present application. Various modifications to these embodiments will be apparent to professionals and technicians in the field, and the general principles defined in the embodiments of the present application can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown in the embodiments of the present application, but will conform to the widest range consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A method for evaluating the adaptability of cross-variety Camellia oleifera grafting based on transfer learning, characterized in that: include: Collecting macroscopic and microscopic characteristics of the target Camellia oleifera, wherein the macroscopic characteristics include plant morphological indicators such as plant height, crown width and / or trunk diameter, and the microscopic characteristics include gene expression data; Perform data preprocessing on the collected macroscopic and microscopic features, and construct a two-stream neural network; The macro and micro feature flow outputs are spliced ​​together through the attention mechanism; The concatenated features are input into the fully connected layer, and the activation function is used to obtain the adaptability score of the target tea oil tree.

2. The adaptability evaluation method according to claim 1, characterized in that: in, Collecting macroscopic and microscopic characteristics of the target Camellia oleifera, wherein the macroscopic characteristics include plant morphological indicators of plant height, crown width and / or trunk diameter, and the microscopic characteristics include gene expression data, including: Select the target tea oil plants to be tested, use an optical distance measuring device to measure the plant height, and use a length measuring tool to measure the crown width, and take the average value; The trunk diameter was measured at a specific height of the plant, the number of leaves was recorded and some leaves were randomly selected; Using a leaf area meter and a chlorophyll content meter to measure the area and chlorophyll content of the leaves respectively; At a preset time point after the plant is grafted, tissue samples within a specific range around the grafted site are collected and rapidly frozen at low temperature; Extracting total RNA from the tissue sample using a nucleic acid extraction reagent, and detecting the concentration, purity and integrity of the RNA using a spectrophotometer and a bioanalyzer; Reverse transcription was performed to obtain cDNA, and real-time quantitative PCR was used to detect the expression level of the preselected gene.

3. The adaptability evaluation method according to claim 1, characterized in that: in, Data preprocessing is performed on the collected macroscopic and microscopic features, including: Performing standardization processing on the macro characteristic data, including obtaining raw data, calculating the mean and standard deviation, and converting the raw data into standardized data using a Z-score method; The relative expression amount of the microscopic feature data is quantitatively calculated, including: ， in, Indicates the Ct value of the target gene, represents the Ct value of the internal reference gene, Indicates the treatment group value, The control group value, represents the standard deviation, Indicates relative expression.

4. The adaptability evaluation method according to claim 1, characterized in that: The two-stream neural network includes: a macro feature processing stream and a micro feature processing stream; Among them, in building a two-stream neural network: The macro feature processing flow includes an input layer, a plurality of fully connected layers and an activation layer connected in sequence; The micro-feature processing flow includes an input layer, multiple one-dimensional convolutional layers, a pooling layer, a fully connected layer and an activation layer connected in sequence, wherein the output of the one-dimensional convolutional layer is subjected to a pooling operation after being activated.

5. The adaptability evaluation method according to claim 1, characterized in that: in, The macro and micro feature flow outputs are spliced ​​together through the attention mechanism, including: The macroscopic and microscopic feature flow output vectors are concatenated to obtain a concatenated feature vector; Performing a nonlinear transformation on the concatenated feature vector using trainable parameters, and calculating the attention weight using a softmax function; The macro feature stream output vector and the micro feature stream output vector are weightedly fused according to the attention weight to obtain a fused feature vector.

6. The adaptability evaluation method according to claim 1, characterized in that: in, Before the macro and micro feature flow outputs are concatenated through the attention mechanism, it includes: The existing Camellia oleifera variety data is divided into a source domain dataset and a target domain dataset, where the source domain dataset contains common varieties with sufficient data, and the target domain dataset contains new varieties with sparse data. The maximum average difference between the source domain dataset and the target domain dataset is calculated, a Gaussian kernel function is used as the feature mapping function, and the value range of the kernel width parameter is determined by cross-validation; Construct a loss function, including the loss terms of the source domain dataset and the target domain dataset, and the maximum average difference weighted term between the source domain dataset and the target domain dataset, and set the initial value and variation range of the balance parameter; The model is trained using the adaptive moment estimation optimization algorithm, and the initial learning rate, learning rate decay strategy, batch size, and number of training iterations are set; During training, performance metrics on the validation set are calculated periodically and the best performing model is saved.

7. The adaptability evaluation method according to claim 6, characterized in that: in, Construct a loss function, including: ， in, represents the total loss function, represents the source domain loss, represents the equilibrium parameter, with an initial value of 0.

1. represents the maximum average difference between the source domain and the target domain, represents the target domain loss.

8. The adaptability evaluation method according to claim 7, characterized in that: in, Maximum mean differences, including: ， in, represents the number of source domain samples, represents the i-th sample in the source domain, represents the number of samples in the target domain, represents the jth sample in the target domain, represents the Gaussian kernel function, represents the Hilbert square norm.

Images