Camellia oleifera seedling drought degree classification method and system based on multi-modal data

Abstract

The invention provides a camellia oleifera seedling drought degree classification method and system based on multi-modal data, and belongs to the technical field of plant detection, the scheme is based on a ConvNeXt architecture, a graph convolutional network (GCN) and a dynamic fusion gate mechanism are combined, image data and root system data of camellia oleifera seedlings are effectively fused, fusion extraction of multi-modal data features is realized, and the method and the system have good application prospects. The classification accuracy of the drought degree of the camellia oleifera seedlings is effectively improved; according to the scheme, the spatial topology modeling capability of a leaf image is enhanced by introducing a graph convolutional network (GCN), the weights of the image and physiological features are adaptively adjusted in combination with a dynamic fusion gate, the weights of the image and root system features are adaptively adjusted through the dynamic fusion gate, the model can accurately capture complementary information among different modals, and the modeling efficiency is improved. And the accuracy of drought stress classification is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of crop drought severity classification technology, and in particular relates to a method and system for classifying the drought severity of Camellia oleifera seedlings based on multimodal data. Background Technology

[0002] Camellia oleifera is an important woody oilseed plant in China. Its tea oil is rich in unsaturated fatty acids and has significant nutritional, medicinal, and industrial value, widely used in food, medicine, and bio-based materials. However, drought stress is a major environmental factor limiting the growth and survival of Camellia oleifera seedlings. Especially against the backdrop of intensified climate change, frequent droughts pose a severe challenge to Camellia oleifera seedling cultivation and management. Traditional assessments of Camellia oleifera health status mainly rely on manual observation or simple physiological indicators. These methods are not only time-consuming and labor-intensive but also have low accuracy and consistency, making it difficult to meet the needs of modern precision and automation in agriculture. Therefore, developing an efficient and accurate method for classifying drought stress in Camellia oleifera is of great significance for optimizing breeding, improving cultivation efficiency, and screening drought-resistant varieties.

[0003] In drought stress studies of Camellia oleifera, plant phenotypic and physiological characteristics are key indicators for assessing health status. Leaf images can visually reflect phenotypic changes under drought stress, such as leaf curling and color changes, while physiological characteristics reveal the plant's adaptive changes to environmental stress. For example, studies have shown that carotenoid content and total chlorophyll content both decrease under drought stress, soluble sugar content first increases and then decreases under mild to moderate drought, and drought significantly reduces the root length, average diameter, and volume of different root systems in Camellia oleifera. Comprehensive analysis of these characteristics provides important evidence for drought stress classification. However, traditional methods face challenges in processing this multidimensional data: image data and physiological characteristic data are heterogeneous modalities, with high data dimensionality, high noise levels, and imbalance issues, making it difficult for classification models to effectively integrate multimodal information and limiting prediction accuracy.

[0004] In recent years, deep learning technology has demonstrated great potential in plant health assessment. Convolutional Neural Networks (CNNs), due to their powerful image feature extraction capabilities, have been widely applied in crop phenotypic analysis. For example, models based on ResNet or EfficientNet can effectively identify the stress state of plant leaves. However, these models are mainly designed for single modalities and struggle to fully utilize structured data such as physiological features, leading to information loss. Graph Convolutional Networks (GCNs), as a deep learning method for processing graph-structured data, have emerged in the agricultural field in recent years. GCNs can effectively capture the spatial correlation of root features by modeling the topological relationships between nodes, but their integration with image features still needs optimization. Furthermore, traditional deep learning models often get stuck in local optima or overfitting when processing multimodal data due to high model complexity or inappropriate hyperparameter selection, affecting classification performance. Summary of the Invention

[0005] This invention provides a method and system for classifying the drought severity of Camellia oleifera seedlings based on multimodal data, which effectively solves the problems existing in the prior art.

[0006] According to a first aspect of the present invention, a method for classifying the drought severity of Camellia oleifera seedlings based on multimodal data is provided, comprising:

[0007] Acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein, the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings;

[0008] Based on the obtained image data and physiological feature data, the drought degree prediction results of camellia seedlings are obtained through a pre-trained deep learning-based drought degree classification model.

[0009] The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

[0010] Furthermore, the image neural network branch specifically performs the following processing steps: using camellia seedling image data as input, after corresponding preprocessing, it is processed through four sequentially connected feature extraction components, and the image feature vector is obtained through global average pooling; wherein, the ConvNeXt unit of the feature extraction component includes a sequentially connected depthwise separable convolution, layer normalization, a first linear layer, a GELU activation function, a second linear layer, and a GELU activation function, and a graph convolutional network is introduced in the ConvNeXt units of the latter two feature extraction components, the graph convolutional network being set between the layer normalization and the first linear layer.

[0011] Furthermore, the graph convolutional network performs the following processing steps: receiving the feature map from the layer normalization output, treating each pixel position in the feature map as a node, and constructing a graph structure based on the pixel's top, bottom, left, and right adjacency relationships to obtain a new image feature vector representation.

[0012] Furthermore, the image feature vector representation is specifically as follows:

[0013] ;

[0014] ;

[0015] in, Let be the feature vector representation of the i-th node. Let W and b be the input features of the neighbor node j of the i-th node, and W and b be the linear layer weights and biases. , Let the degree of the i-th and j-th nodes be , For cosine similarity, Let T be the input feature of the i-th node, T be the learnable threshold, and N(i) be the set of filtered neighbor nodes.

[0016] Furthermore, the filtered neighbor node set is specifically defined as follows: for the current node's neighbor nodes, if the similarity between the input features of the neighbor node and the input features of the current node is less than a preset threshold, then the neighbor node is filtered out from the current node's neighbor nodes, and a neighbor node set is constructed based on the remaining neighbor nodes.

[0017] Furthermore, the dynamic weighted fusion specifically adopts the following formula:

[0018] ;

[0019] ;

[0020] in, For image feature vectors, For root system eigenvectors, To integrate weights, For activation function, , These are the convolution weights of the first and second convolutional layers, respectively. , These are the biases corresponding to the first and second convolutional layers, respectively. For activation function, This represents the concatenation of image feature vectors and root feature vectors.

[0021] Furthermore, during the training process of the drought severity classification model for camellia seedlings, the sparrow search algorithm is used to optimize the parameters, as specifically shown below:

[0022] ;

[0023] in, Let R2 be the position of the i-th sparrow in the t-th generation, ST be the warning value, rand be the safety threshold, rand be the random number, randn be the normally distributed random number, L be the step size, and T be the maximum number of iterations.

[0024] According to a second aspect of the present invention, a drought severity classification system for Camellia oleifera seedlings based on multimodal data is provided, comprising:

[0025] A data acquisition unit is used to acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings.

[0026] The classification unit is used to obtain the drought level prediction results of camellia seedlings based on the obtained image data and physiological feature data, through a pre-trained deep learning-based drought level classification model for camellia seedlings.

[0027] The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

[0028] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned method for classifying the drought degree of Camellia oleifera seedlings based on multimodal data.

[0029] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the aforementioned method for classifying the drought degree of Camellia oleifera seedlings based on multimodal data.

[0030] The above one or more technical solutions have the following beneficial effects:

[0031] This invention provides a method and system for classifying the drought severity of Camellia oleifera seedlings based on multimodal data. The scheme is based on the ConvNeXt architecture, combined with graph convolutional networks (GCN) and dynamic fusion gate mechanism, which effectively integrates image data and root data of Camellia oleifera seedlings, realizes the fusion and extraction of multimodal data features, and effectively improves the classification accuracy of drought severity of Camellia oleifera seedlings.

[0032] The proposed scheme enhances the spatial topology modeling capability of leaf images by introducing a graph convolutional network (GCN). By combining dynamic fusion gates to adaptively adjust the weights of images and physiological features, the model can accurately capture complementary information between different modalities, significantly improving the accuracy of drought stress classification.

[0033] The proposed solution optimizes the model's learning rate, weight decay, and GCN threshold based on the sparrow search algorithm, effectively improving the model's classification performance.

[0034] The proposed scheme uses data augmentation techniques such as random cropping, flipping, brightness adjustment, and SMOTE oversampling to simulate data distribution under complex environments and improve model robustness.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of the drought severity classification method for Camellia oleifera seedlings based on multimodal data, as described in this embodiment of the invention.

[0038] Figure 2This is a schematic diagram of the overall architecture of the drought severity classification method for Camellia oleifera seedlings based on multimodal data described in this embodiment of the invention.

[0039] Figure 3 This is a schematic diagram of the image neural network branch structure based on the ConvNeXt unit improved by graph convolution, as described in an embodiment of the present invention.

[0040] Figure 4 This is a flowchart illustrating the workflow of the graph convolutional network model described in this embodiment of the invention.

[0041] Figure 5 This is the ROC curve of the drought severity classification method for Camellia oleifera seedlings based on multimodal data described in this embodiment of the invention;

[0042] Figure 6 This is a structural diagram of the drought severity classification system for Camellia oleifera seedlings based on multimodal data, as described in this embodiment of the invention. Detailed Implementation

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0045] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0046] To address the above problems, in one or more implementations, such as Figure 1 and Figure 2 As shown in the figure, this embodiment provides a method for classifying the drought severity of Camellia oleifera seedlings based on multimodal data, including:

[0047] Acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein, the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings;

[0048] Based on the obtained image data and physiological feature data, the drought degree prediction results of camellia seedlings are obtained through a pre-trained deep learning-based drought degree classification model.

[0049] The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

[0050] In specific implementation, such as Figure 3 As shown, the drought severity classification model for camellia seedlings includes an input layer, a hidden layer, and an output layer, specifically:

[0051] The input layer accepts multimodal data, including image data (shape [batchsize,3,224,224]) and root feature data (shape [batchsize,3], including root diameter, number of root tips and root surface area).

[0052] The hidden layer is the core module of the entire model, specifically including the image neural network branch and the root neural network branch, wherein:

[0053] The image neural network branch performs the following processing steps: using camellia seedling image data as input, after corresponding preprocessing, it is processed by a series of sequentially connected feature extraction components consisting of downsampling units and ConvNeXt units, and the image feature vector is obtained by global average pooling.

[0054] Specifically, the image neural network branch is constructed based on the ConvNeXt architecture, which includes four stages (i.e., four feature extraction components). Each stage includes a downsampling layer and a ConvNeXt unit. The ConvNeXt unit includes a sequentially connected depthwise separable convolution, layer normalization, a first linear layer, a GELU activation function, a second linear layer, and a GELU activation function. In the third and fourth stages, a graph convolutional network is introduced between the layer normalization and the first linear layer in the ConvNeXt unit.

[0055] In specific implementation, such as Figure 4As shown, the graph convolutional network performs the following processing steps: receiving a feature map from the layer normalization output (i.e., its input is the feature map processed by depthwise separable convolution and layer normalization), treating each pixel position in the feature map as a node, and constructing a graph structure based on the pixel's vertical, horizontal, and vertical adjacency relationships to obtain a new image feature vector. .

[0056] In a specific implementation, the image feature vector is represented as follows:

[0057] ;

[0058] ;

[0059] in, Let be the feature vector representation of the i-th node. Let W and b be the input features of the neighbor node j of the i-th node, and W and b be the linear layer weights and biases. , Let the degree of the i-th and j-th nodes be , For cosine similarity, Let T be the input feature of the i-th node, T be the learnable threshold, and N(i) be the set of filtered neighbor nodes.

[0060] In a specific implementation, the construction of the filtered neighbor node set N(i) is as follows: for the neighbor nodes of the current node, if the similarity between the input features of the neighbor nodes and the input features of the current node is less than a preset threshold, then the neighbor nodes of the current node are filtered out, and a neighbor node set is constructed based on the remaining neighbor nodes.

[0061] The root neural network branch employs a multilayer perceptron, which projects root feature data onto the same dimension as the image feature vector to obtain the root feature vector. ;

[0062] Based on the obtained image feature vector and root system eigenvectors The fused feature vector is obtained through dynamic weighted fusion. Specifically, the dynamic weighted fusion generates fusion weights through a two-layer convolutional network, using the following formula:

[0063] ;

[0064] ;

[0065] in, For image feature vectors, For root system eigenvectors, To integrate weights, For activation function, , These are the convolution weights of the first and second convolutional layers, respectively. , These are the biases corresponding to the first and second convolutional layers, respectively. For activation function, This represents the concatenation of image feature vectors and root feature vectors.

[0066] In specific implementation, the training of the drought severity classification model for camellia seedlings is carried out through the following process:

[0067] First, images, substrate moisture content, and root characteristic data (including root diameter, number of root tips, and root surface area) of camellia seedlings under different drought conditions were collected to form a drought severity classification dataset for camellia seedlings; specifically:

[0068] In this embodiment, healthy, disease-free, and uniform two-year-old grafted seedlings of Camellia oleifera Changlin No. 4, 40, and 53 were selected as experimental materials. Holes were drilled in the outer wall of the container at depths of 5cm, 10cm, and 15cm below the substrate surface for measuring substrate moisture content. These holes were sealed with tape and the container was thoroughly watered until fully saturated. Soil temperature and humidity sensors were used to measure substrate moisture content. Seedling leaf images were photographed, and the root system was scanned using a root analysis system (WinRHINO) to obtain root diameter, root tip number, and root surface area. The experiment continued until large-scale seedling mortality or stable survival occurred. Based on substrate moisture content and plant phenotype, drought severity was classified as severe drought (less than 3%), moderate drought (3%–5%), mild drought (5%–10%), and adequate moisture (less than 10%). Among them, the substrate moisture content is only used during the training phase, and there is no need to measure it again in actual applications after the model is trained; the root features are obtained through the laboratory scanning system during the training phase, while in actual applications, the model can automatically infer the root features based on the images to complete the drought level identification. If there is a portable root observation device, it can also be collected directly on site, but it is not a necessary condition.

[0069] Secondly, data preprocessing is performed: the collected data is preprocessed, including random cropping and scaling of images, random flipping, brightness and contrast adjustment, and normalization of pixel values ​​to [0,1]; as well as SMOTE oversampling, outlier removal, and standardization of root features to optimize the dataset. Outliers are detected using the Z-score method, with the formula: , where x i Let μ be the eigenvalue of the i-th root system, μ be the mean of the eigenvalue, and σ be the standard deviation. Finally, the set is divided into training, validation, and test sets in an 8:1:1 ratio.

[0070] Finally, the drought severity classification model for the tea oleifera seedlings is trained and validated based on the constructed training set, validation set, and test set. Specifically:

[0071] (1) Set the image size of the input data to [batchsize,3,224,224], and the root feature dimension to [batchsize,3], including root diameter, number of root tips, and root surface area. Set the maximum number of iterations to T. (2) Randomly select samples from the enhanced training dataset for batch training, with a batch size of 8.

[0072] (3) Input the selected training samples and their labels into the model for training. Use weighted cross-entropy loss and mean squared error (MSE) loss as the loss function, the formula is:

[0073] ;

[0074] Among them, classification loss for: Where N is the number of samples, and C=4 is the number of categories. For category weights, For real labels, To predict probabilities. Regression loss. for: ,in, These are the true root system eigenvalues. This is the predicted value. The total loss function is:

[0075] ;

[0076] Where λ is the regression loss scaling factor.

[0077] (4) Image data is used to extract spatial features by improving ConvNeXt units based on GCN layers, and root features are projected to the same dimension by multilayer perceptron.

[0078] (5) Optimize hyperparameters (learning rate, weight decay, initial threshold, etc.) using the Sparrow Search Algorithm (SSA), the formula is:

[0079] ;

[0080] in, Let R2 be the position of the i-th sparrow in the t-th generation, ST be the warning value, rand be the safety threshold, rand be the random number, randn be the normally distributed random number, L be the step size, and T be the maximum number of iterations.

[0081] (6) Update the network parameters according to the total loss function using the backpropagation algorithm and the AdamW optimizer.

[0082] (7) Repeat the above steps to traverse all enhanced training data samples until the maximum number of iterations T is reached.

[0083] Finally, the trained model was used to classify drought levels in the test set and to verify the accuracy of the model's predictions by comparing them with the actual drought levels and root system characteristics, thus ensuring the reliability of the model.

[0084] In one or more embodiments, such as Figure 6 As shown, a drought severity classification system for Camellia oleifera seedlings based on multimodal data is provided corresponding to the above method, including:

[0085] A data acquisition unit is used to acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings.

[0086] The classification unit is used to obtain the drought level prediction results of camellia seedlings based on the obtained image data and physiological feature data, through a pre-trained deep learning-based drought level classification model for camellia seedlings.

[0087] The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

[0088] It is understood that the system described in this embodiment corresponds one-to-one with the method embodiment described above, and its technical details have been explained in Embodiment 1, so they will not be repeated here.

[0089] In further embodiments, the following is also provided:

[0090] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the methods described in the above embodiments. For brevity, further details are omitted here.

[0091] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0092] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0093] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the methods described in the above embodiments.

[0094] The methods described in the above embodiments can be directly executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0095] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0096] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for classifying drought severity in Camellia oleifera seedlings based on multimodal data, characterized in that, include: Acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein, the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings; Based on the obtained image data and physiological feature data, the drought degree prediction results of camellia seedlings are obtained through a pre-trained deep learning-based drought degree classification model. The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

2. The drought severity classification method for Camellia oleifera seedlings based on multimodal data as described in claim 1, characterized in that, The image neural network branch specifically performs the following processing steps: using camellia seedling image data as input, after corresponding preprocessing, it is processed through four sequentially connected feature extraction components, and the image feature vector is obtained through global average pooling; wherein, the ConvNeXt unit of the feature extraction component includes a sequentially connected depthwise separable convolution, layer normalization, a first linear layer, a GELU activation function, a second linear layer, and a GELU activation function, and a graph convolutional network is introduced in the ConvNeXt units of the latter two feature extraction components, and the graph convolutional network is set between the layer normalization and the first linear layer.

3. The drought severity classification method for Camellia oleifera seedlings based on multimodal data as described in claim 2, characterized in that, The graph convolutional network performs the following processing steps: it receives the feature map from the layer normalization output, treats each pixel position in the feature map as a node, and constructs a graph structure based on the pixel's top, bottom, left, and right adjacency relationships to obtain a new image feature vector representation.

4. The drought severity classification method for Camellia oleifera seedlings based on multimodal data as described in claim 3, characterized in that, The image feature vector is specifically represented as follows: ； ； in, Let be the feature vector representation of the i-th node. Let W and b be the input features of the neighbor node j of the i-th node, and W and b be the linear layer weights and biases. , Let the degree of the i-th and j-th nodes be , For cosine similarity, Let T be the input feature of the i-th node, T be the learnable threshold, and N(i) be the set of filtered neighbor nodes.

5. The drought severity classification method for Camellia oleifera seedlings based on multimodal data as described in claim 4, characterized in that, The filtered neighbor node set is specifically defined as follows: for the current node's neighbor nodes, if the similarity between the input features of the neighbor node and the input features of the current node is less than a preset threshold, then the neighbor node is filtered out from the current node's neighbor nodes, and a neighbor node set is constructed based on the remaining neighbor nodes.

6. The method for classifying drought severity of Camellia oleifera seedlings based on multimodal data as described in claim 1, characterized in that, The dynamic weighted fusion specifically adopts the following formula: ； ； in, For image feature vectors, For root system eigenvectors, To integrate weights, For activation function, , These are the convolution weights of the first and second convolutional layers, respectively. , These are the biases corresponding to the first and second convolutional layers, respectively. For activation function, This represents the concatenation of image feature vectors and root feature vectors.

7. The method for classifying drought severity of Camellia oleifera seedlings based on multimodal data as described in claim 1, characterized in that, During the training process of the drought severity classification model for camellia seedlings, the sparrow search algorithm is used to optimize the parameters, as detailed below: ； in, Let R2 be the position of the i-th sparrow in the t-th generation, ST be the warning value, rand be the safety threshold, rand be the random number, randn be the normally distributed random number, L be the step size, and T be the maximum number of iterations.

8. A drought severity classification system for Camellia oleifera seedlings based on multimodal data, characterized in that, include: A data acquisition unit is used to acquire image data and physiological characteristic data of Camellia oleifera seedlings; wherein the physiological characteristic data is root system characteristic data of Camellia oleifera seedlings. The classification unit is used to obtain the drought level prediction results of camellia seedlings based on the obtained image data and physiological feature data, through a pre-trained deep learning-based drought level classification model for camellia seedlings. The drought severity classification model for camellia seedlings specifically performs the following processing steps: Using acquired camellia seedling image data and root feature data as input, the camellia seedling image data is processed through an image neural network branch based on a ConvNeXt architecture to obtain image feature vectors, and the root feature data is processed through a root neural network branch based on a multilayer perceptron to obtain root feature vectors. The image neural network branch includes several sequentially connected feature extraction components composed of downsampling units and ConvNeXt units. In the residual structure of the last two feature extraction components, the ConvNeXt units, a graph convolutional network is introduced to perform spatial topological modeling of the camellia seedling leaf data in the image. Based on the obtained image feature vectors and root feature vectors, a fused feature vector is obtained through dynamic weighted fusion. Based on the obtained fused feature vector, a preset classifier is used to obtain the drought severity classification result for the camellia seedlings.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the drought classification method for Camellia oleifera seedlings based on multimodal data as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the drought classification method for Camellia oleifera seedlings based on multimodal data as described in any one of claims 1-7.

Images