Dioryctria dioryctria propagation and diffusion trend prediction method based on TabNet

Abstract

The invention discloses a Dioryctria diorycis propagation and diffusion trend prediction method based on TabNet, and belongs to the technical field of forestry disease and pest prediction. The problem of improving the accuracy of dioryctria dioryctria occurrence risk prediction is solved. The method comprises the following steps: collecting historical occurrence data, meteorological factor data, forest stand factor data, geographic factor data and human interference factor data of dioryctria dioryctria to obtain a multi-source data set, and preprocessing the multi-source data set to obtain a data set for training a TabNet model; constructing a Dioryctria diorycis propagation and diffusion trend prediction model based on TabNet, wherein the model comprises a feature transformation layer, an attention selector and a multi-step decision mechanism; a weighted binary cross entropy is adopted as a loss function for training, and a trained Dioryctria diorycis propagation and diffusion trend prediction model based on TabNet is obtained; and using the trained TabNet-based dioryctria dioryctria propagation and diffusion trend prediction model to predict future dioryctria dioryctria nationwide county scale occurrence conditions based on historical occurrence conditions.

Description

Technical Field

[0001] This invention belongs to the field of forestry pest and disease prediction technology, specifically involving a method for predicting the spread and diffusion trend of the leafminer moth based on TabNet. Background Technology

[0002] Due to global climate change and frequent human activities, the outbreak trend of forest pests is intensifying, posing a serious threat to the ecological security of my country's forests. The leaf moth (Dioryctria), belonging to the family Pyralidae in the order Lepidoptera, is an important pest of pine trees, distributed in subtropical to subarctic regions of the Northern Hemisphere, with a recorded occurrence in my country for approximately 1200 years. Its larvae feed on needles, causing yellowing and wilting; in severe cases, they bore into the bark, leading to bark peeling and exposed trunks, weakening the tree and increasing its susceptibility to pathogens and other pests. In 2009, due to climate disasters, the leaf moth experienced a devastating outbreak in central Hunan and surrounding provinces, causing severe economic and ecological losses; as of 2023, this pest remained a significant threat in Northeast my country. Taking Korean pine (Pinus koraiensis) as an example, plantations are more vulnerable to damage than natural forests. my country currently has over 300,000 hectares of Korean pine plantations concentrated in the Changbai Mountains and Lesser Khingan Mountains. In some severely affected areas, the damage rate exceeds 70%, seed loss reaches as high as 25%, and even total crop failure occurs, resulting in economic losses of nearly US$12 million annually. Therefore, there is an urgent need to develop a method for predicting the risk of the leafminer borer outbreak that can enhance the identification of key factors, improve the ability to locate high-risk areas, and strengthen predictive performance. Summary of the Invention

[0003] The problem this invention aims to solve is to improve the accuracy of predicting the occurrence risk of the leafminer moth, and proposes a method for predicting the spread and diffusion trend of the leafminer moth based on TabNet.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for predicting the spread and diffusion trend of the leafminer moth based on TabNet includes the following steps:

[0006] S1. Collect historical occurrence data, meteorological factor data, forest stand factor data, geographical factor data, and human disturbance factor data of the leafminer moth to obtain a multi-source dataset;

[0007] S2. Preprocess the data from the multi-source dataset obtained in step S1 to obtain the dataset used for training the TabNet model;

[0008] S3. Construct a TabNet-based model for predicting the spread and diffusion trends of the leafminer moth, including a feature transformation layer, an attention selector, and a multi-step decision-making mechanism;

[0009] S4. Using the dataset obtained in step S2 for training the TabNet model, train the TabNet-based prediction model for the spread and diffusion trend of the leafminer moth from step S3. Use weighted binary cross-entropy as the loss function to obtain the trained TabNet-based prediction model for the spread and diffusion trend of the leafminer moth.

[0010] S5. Using the TabNet-based model for predicting the spread and diffusion trend of the leafminer moth trained in step S4, predict the future occurrence of the leafminer moth at the county-level across the country based on historical occurrences.

[0011] Furthermore, the specific implementation method of step S1 includes the following steps:

[0012] S1.1. Collect historical occurrence data of the leafminer moth, and label whether the leafminer moth occurred in the county at different times. The label is 1 to indicate occurrence and 0 to indicate no occurrence.

[0013] S1.2. Collect meteorological factor data, including annual average temperature, precipitation, solar radiation, and humidity;

[0014] S1.3. Collect geographic factor data, including altitude and topographic location;

[0015] S1.4. Collect stand factor data, including tree species composition, stand age, density, and canopy closure;

[0016] S1.5. Collect data on human interference factors, including distances to transportation networks and forest distribution patterns;

[0017] S1.6. In ArcGIS software, the collected data is processed using a zoning statistical method, with the spatial zoning results serving as a region mask, to perform zoning statistics on the target raster data; for each spatial zoning... Extract the set of all cell values ​​within it. The average value of each county / district is calculated to obtain a multi-source dataset.

[0018] Furthermore, the preprocessing method in step S2 involves spatial alignment, temporal synchronization, and standardization of the multi-source dataset obtained in step S1, combining them into a feature vector X:

[0019] X = {x1,x2,…,x F}, X ∈ R F

[0020] Where, x F Let F be the feature of dimension F, where F represents the total dimension of the feature and R is the set of real numbers.

[0021] Furthermore, the specific implementation method of step S3 includes the following steps:

[0022] S3.1. Construct the feature transformation layer: It consists of multiple stacked sub-modules, each of which includes a fully connected layer (FC), a batch normalized layer (BN), and a gated linear unit (GLU).

[0023] The feature vector X is first linearly projected to obtain the input feature h0:

[0024] h0=W0X+b0

[0025] Where W0 is the weight matrix of the linear projection, and b0 is the bias term of the linear projection;

[0026] The input feature h0 is mapped through a fully connected layer to obtain the output feature z of the fully connected layer:

[0027] z=Wh0+b

[0028] Where W is the weight matrix of the fully connected layer, and b is the bias term of the fully connected layer;

[0029] Then, batch normalization is performed on the output features of the fully connected layer to obtain batch-normalized features. :

[0030]

[0031] in, It is the average of the current batch. This is the variance of the current batch. To prevent tiny constants with a denominator of zero;

[0032] Then, by controlling the information flow through gated linear units and residual structures, the output of the feature transformation layer is obtained as follows:

[0033] h i = GLU (W a h i-1 W b h i-1 ) + h i-1

[0034] Among them, h i-1 The feature transformation layer is represented by the first... The input feature vector of the layer, Indicates the first The output feature vector of the layer after passing through the GLU gated structure and residual connections Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the linear activation part of the feedforward branch; Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the activation signal for the gated branch; the dimension is... same;

[0035] S3.2. Constructing the Attention Selector: The output of the feature transformation layer is input into the attention selector for decision-making. In the i-th decision step, the attention selector calculates the feature selection weights through linear mapping and sparse activation functions.

[0036] M i = Sparsemax (W i h i +b i )

[0037] Among them, W i For attention selector in the first The linear mapping weight matrix in each decision step is used to map the input feature vector. Projected onto the feature selection space, b i To represent the attention selector in the first... The bias vector used in each decision step is used to adjust the translation term of the linear mapping output, improving the expressiveness and stability of the selector. i This represents the feature selection weight vector, where Sparsemax is the sparse activation function.

[0038] Then, the input feature X for the i-th decision step is generated. i :

[0039] X i =M i ⊙ h i

[0040] Where ⊙ represents element-wise multiplication;

[0041] The attention selector produces two outputs: a decision output used for cumulative prediction, and an intermediate output used to guide feature selection in subsequent steps, resulting in the expression:

[0042] [d i ,h i+1 ] = FeatureTransformer(X i )

[0043] Where, d i h represents the contribution to the decision at step i. i+1 This indicates an intermediate representation passed to the next step;

[0044] S3.3. Constructing a multi-step decision-making mechanism: Setting the model to make decisions step by step within T steps, ultimately predicting... The cumulative contribution of each step yields:

[0045] ;

[0046] During training, a sparsity constraint of KL divergence penalty is applied.

[0047]

[0048] in, It is a uniform distribution, and λ is the sparsity regularization coefficient. This is a sparsity constraint term;

[0049] S3.4. Constructing the output layer: After undergoing a multi-step decision-making mechanism, the TabNet network completes multiple sparse selections and nonlinear combinations of features to obtain a global latent space representation. Then, the output layer maps this global latent space representation to the result space required for the prediction task, and uses a fully connected layer + Sigmoid activation function to obtain the result of the risk prediction of the occurrence of the leafminer moth.

[0050] Furthermore, the weighted binary cross-entropy in step S4 The expression is:

[0051]

[0052] in, This is a real label; To predict probabilities; , These are the weights of the positive and negative samples, respectively, and N is the number of training samples.

[0053] Furthermore, in step S5, the county codes are matched and connected in ArcGIS software to generate a visual map, including the future spread range of the leafminer moth and the risk level classification.

[0054] The beneficial effects of this invention are:

[0055] This invention presents a TabNet-based method for predicting the spread of the leafminer moth. Addressing issues such as insufficient utilization of spatial features, severe data imbalance interference, and inadequate identification of high-risk areas in leafminer spread prediction, this invention constructs a unified spatiotemporal feature input framework that effectively integrates meteorological, environmental, and social factors, improving the comprehensiveness and consistency of data utilization. Utilizing TabNet's stepwise feature selection and sparse attention mechanism, the model can automatically filter key spread factors during multi-step decision-making, significantly improving the identification of high-risk areas. Simultaneously, TabNet's self-attention mechanism dynamically adjusts the weight allocation of different features, better reflecting the complexity and variability of leafminer spread. The multi-step decision-making mechanism enables the model to capture both sudden spread characteristics and model long-term evolutionary trends, thereby enhancing the stability and robustness of the prediction. Overall, this invention outperforms existing methods in prediction accuracy, model interpretability, and key area identification capabilities.

[0056] This invention presents a TabNet-based method for predicting the spread and diffusion trends of the leafminer moth, enabling the visualization of its spread range, risk level, and potential transmission routes. This provides forestry departments with an intuitive decision support tool. Based on the prediction results, managers can implement targeted monitoring and precise control measures in high-risk areas, optimize resource allocation, improve control efficiency, and reduce economic losses and ecological risks. This invention is not only applicable to the spread prediction of the leafminer moth but can also be extended to the spatiotemporal spread modeling and risk assessment of other forest pests, demonstrating significant potential for widespread application and industrialization value. Attached Figure Description

[0057] Figure 1 This is a flowchart of a method for predicting the spread and diffusion trend of the leafminer moth based on TabNet, as described in this invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0059] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0060] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 Detailed explanation is as follows:

[0061] Example 1:

[0062] A method for predicting the spread and diffusion trend of the leafminer moth based on TabNet includes the following steps:

[0063] S1. Collect historical occurrence data, meteorological factor data, forest stand factor data, geographical factor data, and human disturbance factor data of the leafminer moth to obtain a multi-source dataset;

[0064] Furthermore, the specific implementation method of step S1 includes the following steps:

[0065] S1.1. Collect historical occurrence data of the leafminer moth, and label whether the leafminer moth occurred in the county at different times. The label is 1 to indicate occurrence and 0 to indicate no occurrence.

[0066] S1.2. Collect meteorological factor data, including annual average temperature, precipitation, solar radiation, and humidity;

[0067] S1.3. Collect geographic factor data, including altitude and topographic location;

[0068] S1.4. Collect stand factor data, including tree species composition, stand age, density, and canopy closure;

[0069] S1.5. Collect data on human interference factors, including distances to transportation networks and forest distribution patterns;

[0070] S1.6. In ArcGIS software, the collected data is processed using a zoning statistical method, with the spatial zoning results serving as a region mask, to perform zoning statistics on the target raster data; for each spatial zoning... Extract the set of all cell values ​​within it. The average value of each county / district is calculated to obtain a multi-source dataset.

[0071] S2. Preprocess the data from the multi-source dataset obtained in step S1 to obtain the dataset used for training the TabNet model;

[0072] Furthermore, the preprocessing method in step S2 involves spatial alignment, temporal synchronization, and standardization of the multi-source dataset obtained in step S1, combining them into a feature vector X:

[0073] X = {x1,x2,…,x F}, X ∈ R F

[0074] Where, x F Let F be the feature of dimension F, where F represents the total dimension of the feature and R is the set of real numbers.

[0075] Furthermore, for multi-source data, spatial alignment, temporal synchronization, and standardization are performed to construct a unified input matrix, combining historical records and environmental factors into structured tabular data suitable for TabNet model training. Unlike traditional neural networks that directly feed input into fully connected layers, TabNet uses a FeatureTransformer for non-linear mapping. In the input layer, linear projection is first performed:

[0076] h0=W0X+b0

[0077] Where W0∈R d×F It is a weight matrix, b0∈R d Here, is the bias term, and d is the dimension of the transformed latent space. In this way, the input features are mapped to a higher-dimensional representation space that is better suited for learning:

[0078] h0∈R d ;

[0079] S3. Construct a TabNet-based model for predicting the spread and diffusion trends of the leafminer moth, including a feature transformation layer, an attention selector, and a multi-step decision-making mechanism;

[0080] Furthermore, the specific implementation method of step S3 includes the following steps:

[0081] S3.1. Construct the feature transformation layer: It consists of multiple stacked sub-modules, each of which includes a fully connected layer (FC), a batch normalized layer (BN), and a gated linear unit (GLU).

[0082] The feature vector X is first linearly projected to obtain the input feature h0:

[0083] h0=W0X+b0

[0084] Where W0 is the weight matrix of the linear projection, and b0 is the bias term of the linear projection;

[0085] The input feature h0 is mapped through a fully connected layer to obtain the output feature z of the fully connected layer:

[0086] z=Wh0+b

[0087] Where W is the weight matrix of the fully connected layer, and b is the bias term of the fully connected layer;

[0088] Then, batch normalization is performed on the output features of the fully connected layer to obtain batch-normalized features. :

[0089]

[0090] in, It is the average of the current batch. This is the variance of the current batch. To prevent tiny constants with a denominator of zero;

[0091] Then, by controlling the information flow through gated linear units and residual structures, the output of the feature transformation layer is obtained as follows:

[0092] h i = GLU (W a h i-1 W b h i-1 ) + h i-1

[0093] Among them, h i-1 The feature transformation layer is represented by the first... The input feature vector of the layer, Indicates the first The output feature vector of the layer after passing through the GLU gated structure and residual connections Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the linear activation part of the feedforward branch; Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the activation signal for the gated branch; the dimension is... same;

[0094] This approach preserves the original information while enhancing the model's non-linear expressive power. Training stability is ensured through BN + GLU + residuals; the results of each feature transformation layer provide candidate representations for the subsequent Feature Selector; compared to traditional MLPs, TabNet's feature transformation layers have stronger sparse selection capabilities, automatically focusing on the most critical factors for predicting the leafminer moth.

[0095] Furthermore, GLU is a key innovation of TabNet, controlling the flow of information through a gating mechanism:

[0096] GLU (a,b) = a⊙σ (b)

[0097] Where a,b∈R d They come from two independent linear transformations, It is the sigmoid activation function, and ⊙ represents element-wise multiplication.

[0098] In other words, some features are "suppressed" and some features are "amplified", achieving sparse feature transformation.

[0099] S3.2. Constructing the Attention Selector: The output of the feature transformation layer is input into the attention selector for decision-making. In the i-th decision step, the attention selector calculates the feature selection weights through linear mapping and sparse activation functions.

[0100] M i = Sparsemax (W i h i +b i )

[0101] Among them, W i For attention selector in the first The linear mapping weight matrix in each decision step is used to map the input feature vector. Projected onto the feature selection space, b i To represent the attention selector in the first... The bias vector used in each decision step is used to adjust the translation term of the linear mapping output, improving the expressiveness and stability of the selector. i This represents the feature selection weight vector, where Sparsemax is the sparse activation function.

[0102] Sparsemax produces a sparse probability distribution:

[0103]

[0104] Where τ(z) is a threshold, ensuring that ∑ j Sparsemax (z) j = 1, so only the most relevant few features will be selected, and the weights of other features will become 0. The weight vector M generated by the selector i Multiplying the input features element-wise yields the filtered input:

[0105] Then, the input feature X for the i-th decision step is generated. i :

[0106] X i =M i ⊙ h i

[0107] Here, ⊙ represents element-wise multiplication; this step ensures that each layer focuses only on key features, improving the sparsity and interpretability of the model. Sparsity: Through Sparsemax, only a small number of key features are activated, reducing redundant computation; Adaptability: The feature selection results differ for different samples and decision steps, exhibiting dynamic characteristics; Interpretability: The final feature weight distribution Mi can directly explain "which factors are most important for predicting the risk of the leafminer borer." The attention selector enables TabNet to have dynamic feature selection capabilities at both the sample and step levels, automatically focusing on the factors that have the greatest impact on the risk prediction of the leafminer borer. This mechanism improves both the model's prediction accuracy and its interpretability.

[0108] The attention selector produces two outputs: a decision output used for cumulative prediction, and an intermediate output used to guide feature selection in subsequent steps, resulting in the expression:

[0109] [d i ,h i+1 ] = FeatureTransformer(X i )

[0110] Where, d i h represents the contribution to the decision at step i. i+1 This indicates an intermediate representation passed to the next step;

[0111] In traditional deep learning models (such as MLP and CNN), all features are involved in the computation, leading to redundancy and overfitting. TabNet's attention selector, however, dynamically selects the most useful features based on different input samples and decision steps, making the model more interpretable and improving computational efficiency. In the task of predicting the risk of leafminer outbreaks, this mechanism means the model can automatically "pick out" the most critical meteorological, stand, or geographical factors, such as temperature, forest age, and precipitation, while weakening irrelevant or redundant features.

[0112] S3.3. Constructing a multi-step decision-making mechanism: Setting the model to make decisions step by step within T steps, ultimately predicting... The cumulative contribution of each step yields:

[0113] ;

[0114] During training, a sparsity constraint of KL divergence penalty is applied.

[0115]

[0116] in, It is a uniform distribution, and λ is the sparsity regularization coefficient. The sparsity constraint ensures the model focuses on only a few key features, improving interpretability. By simulating the multi-step human thought process, different features are selected at each step; strong interpretability: the feature selection and decision contribution at each step can be analyzed independently; robustness: sparse regularization avoids the model relying on too many redundant features; cumulative prediction: the final result is the sum of contributions from multiple steps, making it more stable and reliable. The multi-step decision-making mechanism is TabNet's core innovation, enabling the model to progressively and dynamically select features and accumulate decision contributions, thus more accurately identifying key factors, simulating complex occurrence mechanisms, and maintaining strong interpretability in the task of predicting the risk of leafminer moth outbreaks.

[0117] S3.4. Constructing the output layer: After undergoing a multi-step decision-making mechanism, the TabNet network completes multiple sparse selections and nonlinear combinations of features to obtain a global latent space representation. Then, the output layer maps this global latent space representation to the result space required for the prediction task, and uses a fully connected layer + Sigmoid activation function to obtain the result of the risk prediction of the occurrence of the leafminer moth.

[0118] Furthermore, the output layer typically employs a fully connected layer with a sigmoid activation function.

[0119]

[0120] in, The final implicit representation obtained by the multi-step decision-making mechanism; The weights and biases of the output layer; For the Sigmoid function:

[0121] ;

[0122] Received , representing the probability of the occurrence of the leafminer moth.

[0123] S4. Using the dataset obtained in step S2 for training the TabNet model, train the TabNet-based prediction model for the spread and diffusion trend of the leafminer moth from step S3. Use weighted binary cross-entropy as the loss function to obtain the trained TabNet-based prediction model for the spread and diffusion trend of the leafminer moth.

[0124] Furthermore, the weighted binary cross-entropy in step S4 The expression is:

[0125]

[0126] in, This is a real label; To predict probabilities; , The weights for positive and negative samples are respectively, and N is the number of training samples. The high-dimensional latent representation h obtained from the multi-step decision mechanism is converted into a predicted probability / risk value. Appropriate activation functions (Sigmoid / Softmax / Linear) are used to meet the needs of different tasks. Combined with weighted BCE / FocalLoss and other methods, the imbalance of data is effectively solved and the ability to identify high-risk small sample areas is improved.

[0127] S5. Using the TabNet-based model for predicting the spread and diffusion trend of the leafminer moth trained in step S4, predict the future occurrence of the leafminer moth at the county-level across the country based on historical occurrences.

[0128] Furthermore, step S5 involves matching and connecting county / district codes in ArcGIS software to generate a visual map, including the future spread range and risk level classification of the leafminer moth. The comparison results are shown in Table 1.

[0129] Table 1

[0130] This embodiment describes a TabNet-based method for predicting the spread and diffusion trends of the leafminer moth. It addresses existing methods for predicting leafminer occurrence risk by addressing issues such as insufficient feature utilization, severe data imbalance, and weak model identification of high-risk areas. This proposed TabNet-based method offers higher factor modeling accuracy and generalization ability, effectively improving the scientific validity and practicality of the prediction. The method combines deep learning and structured data modeling, utilizing TabNet's stepwise feature selection mechanism and sparse attention strategy to automatically filter and deeply characterize multi-source features such as historical data on leafminer occurrence, meteorological factors, and forest stand structure. A weighted loss function is used to mitigate sample imbalance, thereby achieving accurate identification of high-risk areas and prediction of leafminer occurrence trends.

[0131] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0132] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting the spread and diffusion trend of the leafminer moth based on TabNet, characterized in that, Includes the following steps: S1. Collect historical occurrence data, meteorological factor data, forest stand factor data, geographical factor data, and human disturbance factor data of the leafminer moth to obtain a multi-source dataset; S2. Preprocess the data from the multi-source dataset obtained in step S1 to obtain the dataset used for training the TabNet model; S3. Construct a TabNet-based model for predicting the spread and diffusion trends of the leafminer moth, including a feature transformation layer, an attention selector, and a multi-step decision-making mechanism; S4. Using the dataset obtained in step S2 for training the TabNet model, train the TabNet-based prediction model for the spread and diffusion trend of the leafminer moth from step S3. Use weighted binary cross-entropy as the loss function to obtain the trained TabNet-based prediction model for the spread and diffusion trend of the leafminer moth. S5. Using the TabNet-based model for predicting the spread and diffusion trend of the leafminer moth trained in step S4, predict the future occurrence of the leafminer moth at the county-level across the country based on historical occurrences.

2. The method for predicting the spread and diffusion trend of the leafminer moth based on TabNet according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Collect historical occurrence data of the leafminer moth, and label whether the leafminer moth occurred in the county at different times. The label is 1 to indicate occurrence and 0 to indicate no occurrence. S1.

2. Collect meteorological factor data, including annual average temperature, precipitation, solar radiation, and humidity; S1.

3. Collect geographic factor data, including altitude and topographic location; S1.

4. Collect stand factor data, including tree species composition, stand age, density, and canopy closure; S1.

5. Collect data on human interference factors, including distances to transportation networks and forest distribution patterns; S1.

6. In ArcGIS software, the collected data is processed using a zoning statistical method, with the spatial zoning results serving as a regional mask, to perform zoning statistics on the target raster data; For each spatial partition Extract the set of all cell values ​​within it. The average value of each county / district is calculated to obtain a multi-source dataset.

3. The method for predicting the spread and diffusion trend of the leafminer moth based on TabNet according to claim 2, characterized in that, The preprocessing method in step S2 involves spatial alignment, temporal synchronization, and standardization of the multi-source dataset obtained in step S1, combined to form a feature vector X: X = {x1,x2,…,x F }, X ∈ R F 4 Where, x F Let F be the feature of dimension F, where F represents the total dimension of the feature and R is the set of real numbers.

4. The method for predicting the spread and diffusion trend of the leafminer moth based on TabNet according to claim 3, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Construct the feature transformation layer: It consists of multiple stacked sub-modules, each of which includes a fully connected layer (FC), a batch normalized layer (BN), and a gated linear unit (GLU). The feature vector X is first linearly projected to obtain the input feature h0: h0 = W0X + b0; Where W0 is the weight matrix of the linear projection, and b0 is the bias term of the linear projection; The input feature h0 is mapped through a fully connected layer to obtain the output feature z of the fully connected layer: z = Wh0 + b; Where W is the weight matrix of the fully connected layer, and b is the bias term of the fully connected layer; Then, batch normalization is performed on the output features of the fully connected layer to obtain batch-normalized features. : ； in, It is the average of the current batch. It is the variance of the current batch. To prevent tiny constants with a denominator of zero; Then, by controlling the information flow through gated linear units and residual structures, the output of the feature transformation layer is obtained as follows: h i = GLU (W a h i-1 ,W b h i-1 ) + h i-1 ; Among them, h i-1 The feature transformation layer is represented by the first... The input feature vector of the layer, Indicates the first The output feature vector of the layer after passing through the GLU gated structure and residual connections Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the linear activation part of the feedforward branch; Indicates the effect on input features The linear transformation weight matrix is ​​used to generate the activation signal for the gated branch; the dimension is... same; S3.

2. Constructing the Attention Selector: The output of the feature transformation layer is input into the attention selector for decision-making. In the i-th decision step, the attention selector calculates the feature selection weights through linear mapping and sparse activation functions. M i = Sparsemax (W i h i +b i ); Among them, W i For attention selector in the first The linear mapping weight matrix in each decision step is used to map the input feature vector. Projected onto the feature selection space, b i To represent the attention selector in the first... The bias vector used in each decision step is used to adjust the translation term of the linear mapping output, improving the expressiveness and stability of the selector. i This represents the feature selection weight vector, and Sparsemax is the sparse activation function. Then, the input feature X for the i-th decision step is generated. i : X i =M i ⊙ h i ; Where ⊙ represents element-wise multiplication; The attention selector produces two outputs: a decision output used for cumulative prediction, and an intermediate output used to guide feature selection in subsequent steps, resulting in the expression: [d i ,h i+1 ]= FeatureTransformer (X i ); Where, d i h represents the contribution to the decision at step i. i+1 This indicates an intermediate representation passed to the next step; S3.

3. Constructing a multi-step decision-making mechanism: Setting the model to make decisions step by step within T steps, ultimately predicting... The cumulative contribution of each step yields: ； During training, a sparsity constraint of KL divergence penalty is applied. ； in, It is a uniform distribution, and λ is the sparsity regularization coefficient. This is a sparsity constraint term; S3.

4. Constructing the output layer: After undergoing a multi-step decision-making mechanism, the TabNet network completes multiple sparse selections and nonlinear combinations of features to obtain a global latent space representation. Then, the output layer maps this global latent space representation to the result space required for the prediction task, and uses a fully connected layer + Sigmoid activation function to obtain the result of the risk prediction of the occurrence of the leafminer moth.

5. The method for predicting the spread and diffusion trend of the leafminer moth based on TabNet according to claim 4, characterized in that, Weighted binary cross-entropy in step S4 The expression is: ； in, This is a real label; To predict probabilities; , These are the weights of the positive and negative samples, respectively, and N is the number of training samples.

6. The method for predicting the spread and diffusion trend of the leafminer moth based on TabNet according to claim 5, characterized in that, Step S5 involves matching and connecting county / district codes in ArcGIS software to generate a visual map, including the future spread range and risk level classification of the leafminer moth.

Images