A land property classification method based on dual knowledge self-distillation multilayer perceptron

Abstract

The present invention discloses a land property classification method based on a dual knowledge self-distillation multilayer perceptron, comprising: step 1, using soil properties obtained by sensors as node features, using paths connecting adjacent sensors as edges, and obtaining node topology information; step 2, constructing a training sample set, which includes: node features, node topology information and corresponding soil classification labels; constructing a soil classification prediction model; inputting the node features in the training sample set into the soil classification prediction model, performing dual knowledge self-distillation in combination with the node topology information and calculating the loss, updating all parameters through a back propagation algorithm, and obtaining an optimized soil classification prediction model; step 3, inputting the node features of the soil to be classified into the optimized soil classification prediction model, and the output layer outputs the probability corresponding to each category; step 4, taking the category corresponding to the highest prediction probability as the final result of soil category prediction.

Description

Technical Field

[0001] The present invention belongs to the field of smart agriculture technology, and in particular relates to a land property classification method based on a dual knowledge self-distillation multi-layer perceptron. Background Art

[0002] Soil classification is a critical task in fields such as agriculture, environmental science, and geology. It aims to systematically categorize soils based on their physical, chemical, and biological properties. Effective soil classification can help formulate appropriate land use policies, improve crop yields, manage water resources, and predict ecological changes. With the development of remote sensing technology and geographic information systems (GIS), soil classification using graph data has become increasingly common. Graph data can effectively represent the spatial relationships and similarities between soil samples. Traditional methods typically rely on manual feature extraction and empirical models, which are inefficient and susceptible to human bias.

[0003] In recent years, the introduction of deep learning technology has brought new opportunities for soil classification. By automatically learning soil characteristics, neural networks can improve classification accuracy. Graph data is becoming increasingly important in many fields, including social network analysis, bioinformatics, and the Internet of Things. Soil information, as a graph with inherent spatial structure, belongs to the category of homomorphic graphs where node properties (soil) and edge properties (adjacency) are identical. Therefore, neural network frameworks are best suited for processing graph data. Traditional graph convolutional neural networks (GCNs) are widely used due to their ability to effectively capture inter-node relationships. However, their slow training speed and large model size limit their application to large datasets. Other methods also fail to consider soil spatial topology, reducing model accuracy.

[0004] In the existing technology, the multimodal neural network classification method sets detectors in the soil to monitor and obtain multimodal data of the soil in real time, which is used to train the network to give classification results. However, it fails to fully consider the spatial structure of the soil, resulting in the neglect of the soil's topological information, thereby affecting the classification accuracy. In addition, the multimodal framework has disadvantages such as high data noise and easy overfitting.

[0005] Another technology, graph convolutional neural network (GCN), can process the spatial topological information of soil, but it relies on the feature aggregation and local neighborhood information of the graph structure, which leads to computational bottlenecks when processing large-scale graph data. Although some improved solutions such as GraphSAGE and ChebNet attempt to improve efficiency through sample selection or frequency domain methods, they still fail to effectively solve the problems of model complexity and training speed. Summary of the Invention

[0006] The purpose of the present invention is to provide a land property classification method based on dual knowledge self-distillation multi-layer perceptron, which can improve training speed, reduce model complexity, integrate soil topology information, reduce overfitting and noise effects, and accurately classify soil properties of different scales and types.

[0007] The technical solution provided by the present invention is:

[0008] A land property classification method based on dual knowledge self-distillation multilayer perceptron includes the following steps:

[0009] Step 1: The sensors placed in the soil are regarded as nodes, the soil properties obtained by the sensors are regarded as node features, and the paths connecting adjacent sensors are regarded as edges, thereby obtaining the soil graph topology structure and node topology information;

[0010] Step 2: Construct a training sample set, which includes: node features, node topology information and corresponding soil classification labels;

[0011] A soil classification prediction model based on a pure multi-layer perceptron network is constructed, comprising: an input layer, two hidden layers, two inference layers, and an output layer; node features in the training sample set are input into the soil classification prediction model, processed by the two hidden layers, and then respectively input into the two inference layers; dual knowledge self-distillation is performed and loss is calculated by combining the node topology information and the corresponding soil classification labels; all parameters are updated through a back-propagation algorithm to obtain an optimized soil classification prediction model;

[0012] Step 3: Input the node features of the soil data to be classified into the optimized soil classification prediction model, and the output layer outputs the probability corresponding to each category;

[0013] Step 4: From the probabilities corresponding to each category output by the output layer, the category corresponding to the highest prediction probability is taken as the final result of soil category prediction.

[0014] Preferably, the dual knowledge self-distillation includes: feature-level self-distillation and label-level self-distillation.

[0015] Preferably, the overall loss function of the dual knowledge self-distillation is:

[0016] ;

[0017] Where, is the overall loss function of dual knowledge self-distillation; is the loss function for label-level self-distillation; is the feature-level self-distillation loss function; An important hyperparameter for balancing feature-level self-distillation and label-level self-distillation.

[0018] Preferably, the loss function of the feature-level self-distillation is:

[0019] ;

[0020] Where, is the target node; Target node Neighbor nodes of For the target node Non-adjacent nodes, i.e. negative samples; is the first simulation point constructed by the attention mechanism; is the number of nodes; is the number of hidden layers; Target node Neighbor nodes A collection of For the The target node of the hidden layer output Implicit representation of node features; For the Negative samples of the hidden layer output Implicit representation of node features; Target node Its neighboring nodes The node feature implicitly represents the first simulation point of weighted fitting Implicit representation of node features; is the target node after linear transformation in the second inference layer The prediction results; is the target node after SoftMax conversion in the second inference layer The predicted probability distribution of is the simulated point after linear transformation in the first inference layer The prediction results; is the negative sample after SoftMax conversion in the first inference layer The predicted probability distribution of Negative samples the number of Target node With negative samples The edge between is the set of edges between nodes.

[0021] Preferably, the loss function of the label-level self-distillation is:

[0022] ;

[0023] Where, is the cross loss entropy function; is a set of nodes with known corresponding soil classification labels; Target node Known soil classification labels; For the The target node of the hidden layer output Neighbor nodes Implicit representation of node features; is the target node after SoftMax conversion in the first inference layer Neighbor nodes The predicted probability distribution of .

[0024] Preferably, each of the hidden layers is connected in a fully connected neuron manner.

[0025] Preferably, each of the hidden layers includes: batch normalization, activation function, random dropout and linear transformation.

[0026] Preferably, the representation formula of the input layer is:

[0027] ;

[0028] Where, It is the implicit representation of node features output by the input layer; is the node feature of soil data.

[0029] Preferably, the hidden layer is represented by:

[0030] ;

[0031] Where, For the The node features of the hidden layer output are implicitly represented; For the The node features of the hidden layer input are implicitly represented; is the parameter matrix of a specific layer; is the random dropout function; is batch normalization; is the activation function Leaky Relu.

[0032] Preferably, an edge sampling strategy is used to optimize the loss calculation of dual knowledge self-distillation. The optimized loss functions of feature-level self-distillation and label-level self-distillation are:

[0033] ;

[0034] Where, is the batch size, is a small batch of edge sets; are all sampling points in the mini-batch edge set, ; Target node A collection of Target node Its neighboring nodes edge; The negative sample nodes are obtained after optimizing the negative sample selection through the edge sampling strategy; is a probability distribution, and the default is uniform distribution. ; is the target node after linear transformation in the second inference layer Neighbor nodes The prediction results; is the target node after SoftMax conversion in the second inference layer Neighbor nodes The predicted probability distribution of is the second simulation point constructed by the attention mechanism; Target node Its neighboring nodes The node feature implicitly represents the second simulation point obtained by weighted fitting Implicit representation of node features is the second simulated point after linear transformation in the first inference layer prediction results.

[0035] The beneficial effects of the present invention are:

[0036] The land property classification method based on dual knowledge self-distillation multi-layer perceptron provided by the present invention can improve training speed, reduce model complexity, integrate soil topology information, reduce overfitting and noise effects, and accurately classify soil properties of different scales and types. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the prediction model training of the land property classification method based on the dual knowledge self-distillation multi-layer perceptron described in the present invention.

[0038] Figure 2 This is a prediction flow chart of the prediction model of the land property classification method based on dual knowledge self-distillation multi-layer perceptron described in the present invention.

[0039] Figure 3 Schematic diagram of characteristic-level self-distillation according to the present invention.

[0040] Figure 4 Schematic diagram of label-grade self-distillation according to the present invention. DETAILED DESCRIPTION

[0041] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0042] like Figure 1-2 As shown, the present invention provides a land property classification method based on dual knowledge self-distillation multilayer perceptron, and the specific steps are as follows:

[0043] Step 1: Arrange sensors in the soil, use the sensors arranged in the soil as nodes, and use the soil properties obtained by the sensors as the node features of each node. The soil properties include soil temperature and soil moisture obtained by the sensors; the paths connecting adjacent sensors are used as edges to form the soil graph topology structure, thereby obtaining node topology information; the node features of each node and the node topology information together constitute the soil data.

[0044] In this embodiment, it is assumed that represents a soil map, where is a collection of nodes; is the set of edges between nodes, i.e. the soil graph topology; let is a node feature set; where is the number of nodes; each node There is one dimensional feature vector associated.

[0045] Step 2, soil classification prediction model training phase: Construct a training sample set, which includes: nodes with known soil classification labels, node features, node topology information and known corresponding soil classification labels. In this embodiment, the common semi-supervised node settings are followed, and a part of the nodes are With corresponding and known soil classification labels , as the training sample set data.

[0046] A soil classification prediction model based on a pure multi-layer perceptron network is constructed, which includes: an input layer, two hidden layers, two inference layers and an output layer. The two hidden layers are the first hidden layer and the second hidden layer, and the two inference layers are the first inference layer. and the second inference layer Inputting the node features in the training sample set into the soil classification prediction model, and processing them in sequence through the input layer, the first hidden layer, and the second hidden layer to obtain the implicit representation of the node features of the first hidden layer and the implicit representation of the node features of the second hidden layer;

[0047] Among them, the representation formula of the input layer is:

[0048] ;

[0049] Where, It is the implicit representation of node features output by the input layer; is the node feature of soil data;

[0050] The main function of the input layer is to receive the initial node features of each node; the node features are the attribute representations of each node in the soil data; the input layer processes the node features, usually including linear transformation (i.e., matrix multiplication through a weight matrix), and outputs the implicit feature representation of each node. This implicit representation is a high-dimensional encoding of the input data and is usually used in subsequent feature optimization and inference processes;

[0051] Each hidden layer adopts a fully connected neuron connection mode; each hidden layer includes: batch normalization, activation function, random dropout and linear transformation;

[0052] The hidden layer is represented by:

[0053] ;

[0054] Where, For the The node features of the hidden layer output are implicitly represented; For the The node features of the hidden layer input are implicitly represented; is the parameter matrix of a specific layer; is the random dropout function; is batch normalization; is the activation function Leaky Relu.

[0055] Each of the inference layers includes: linear transformation and SoftMax normalization, wherein the first inference layer only participates in the soil classification prediction model training stage, and the second inference layer participates in the soil classification prediction model training stage and the soil classification prediction model prediction stage. Retaining only the second inference layer in the soil classification prediction model prediction stage can reduce the inference time and ensure efficient single-node inference.

[0056] Inputting the implicit representation of node features of the first hidden layer and the implicit representation of node features of the second hidden layer into the first inference layer and the second inference layer, respectively, and obtaining the prediction results of the first inference layer and the second inference layer after linear transformation in the first inference layer and the second inference layer. The prediction results of the inference layer are the hidden layer features perceived by the graph topology structure. The prediction probability distribution is obtained by softmax normalization processing, and dual knowledge self-distillation is performed based on the node topology information and the known corresponding soil classification labels. The loss is calculated based on the results of the dual knowledge self-distillation.

[0057] The first hidden layer and the second hidden layer output the target node Implicit representation of node features , target node Neighbor nodes Implicit representation of node features and negative samples Implicit representation of node features , is the number of hidden layers; then two simulation points are constructed through the attention mechanism and , the target node Its neighboring nodes The node feature implicitly represents the weighted fitting of two simulation points and Implicit representation of node features and ; The target node Neighbor nodes Implicit representation of node features , negative samples Implicit representation of node features The two simulation points and Implicit representation of node features and Input into the first reasoning layer; the target node Implicit representation of node features and the target node Neighbor nodes Implicit representation of node features Input into the second reasoning layer; and perform dual knowledge self-distillation and calculate loss by combining the corresponding node topology information and the known corresponding soil classification label;

[0058] The dual knowledge self-distillation includes: feature-level self-distillation and label-level self-distillation;

[0059] The overall loss function for calculating the loss based on the results of the dual knowledge self-distillation is:

[0060] ;

[0061] Where, is the overall loss function of dual knowledge self-distillation; is the loss function for label-level self-distillation; is the feature-level self-distillation loss function; To balance the important hyperparameters of feature-level self-distillation and label-level self-distillation;

[0062] The loss function of the feature-level self-distillation is:

[0063] ;

[0064] Where, is the target node; Target node Neighbor nodes of For the target node Non-adjacent nodes, i.e. negative samples; is a simulation point constructed by the attention mechanism; is the number of nodes; is the number of hidden layers; Target node Neighbor nodes A collection of For the Target node of hidden layer output Implicit representation of node features; For the Negative samples of the hidden layer output Implicit representation of node features; Target node Its neighboring nodes The node features implicitly represent the simulation points obtained by weighted fitting Implicit representation of node features; is the target node after linear transformation in the second inference layer The prediction results; is the target node after SoftMax conversion in the second inference layer The predicted probability distribution of is the simulated point after linear transformation in the first inference layer The prediction results; is the negative sample after linear transformation in the first inference layer The prediction results; is the negative sample after SoftMax conversion in the first inference layer The predicted probability distribution of Negative samples the number of Target node With negative samples The edge between is the set of edges between nodes;

[0065] The loss function of the label-level self-distillation is:

[0066] ;

[0067] Where, is the cross loss entropy function; is a set of nodes with known corresponding soil classification labels; Target node Known soil classification labels, ; For the The target node of the hidden layer output Neighbor nodes Implicit representation of node features; is the target node after linear transformation in the first inference layer Neighbor nodes The prediction results; is the target node after SoftMax conversion in the first inference layer Neighbor nodes The predicted probability distribution of .

[0068] All parameters are updated through the back propagation algorithm to obtain the optimized soil classification prediction model.

[0069] Step three, soil classification prediction model prediction stage: the node features in the soil data of the soil to be classified are input into the optimized soil classification prediction model, and pass through the input layer, the first hidden layer and the second hidden layer in sequence to obtain the implicit representation of the node features of the optimized second hidden layer, and input it into the second inference layer. After linear transformation in the second inference layer, the prediction result of the second inference layer is obtained, and then the prediction probability distribution is obtained through SoftMax normalization processing; the output layer outputs the probability corresponding to each category.

[0070] Step 4: From the probabilities corresponding to each category output by the output layer, the category corresponding to the highest prediction probability is taken as the final result of soil category prediction.

[0071] like Figure 3 As shown in the figure, in order to make the distribution of eigenvalues of adjacent nodes converge, the traditional Mixup data enhancement scheme is modified. , find all neighbor nodes of the target node , two simulation points are constructed by the attention mechanism and ; Inspect the target node through the attention mechanism and its neighboring nodes The difference between two points, obtaining a weight that increases as the difference increases , through the weight Weighted target node Its neighboring nodes The node features are implicitly represented and two simulation points are fitted. 、 The node feature implicitly represents the target node Its neighboring nodes The first simulation point generated after interpolation , the second simulation point The node feature is expressed as:

[0072] ;

[0073] Where, is the first simulation point constructed by the attention mechanism; is the second simulation point constructed by the attention mechanism; Target node Its neighboring nodes The node feature implicitly represents the first simulation point of weighted fitting Implicit representation of node features; Target node Its neighboring nodes The node feature implicitly represents the second simulation point obtained by weighted fitting Implicit representation of node features; is the interpolation coefficient;

[0074] The interpolation coefficient The definition formula is:

[0075] ;

[0076] Where, is the feature transformation matrix, which maps the input features to a new space; is the shared attention weight, used to calculate the similarity between nodes; Target node With neighboring nodes Feature splicing; is the attention mechanism function;

[0077] Due to the characteristics of the attention mechanism, when the two features converge, the weight As the value of tends to 0.5, the loss function will become smaller, which is equivalent to implicitly reducing the difference in features by reducing the difference in label distribution.

[0078] The feature-level self-distillation is as follows: extracting feature information in the neighborhood to the target node; With the first simulation point Feature prediction differences, neighbor nodes With the second simulation point The feature prediction difference is used as the positive term in the loss function, and the training expectation is to reduce the difference, and then select the node that does not match the target node. and neighbor nodes Adjacent points are used as negative samples , the target node With negative samples , neighbor nodes With negative samples The feature prediction difference is used as the negative term of the loss function, and the training expectation is to increase the difference; thus, the loss function of feature-level self-distillation is constructed, which reduces the feature difference between adjacent points and increases the feature difference between non-adjacent points, thereby realizing feature-level self-distillation; when the edge sampling strategy is not performed, the loss function needs to globally traverse all nodes and edges, and the calculation of neighbor nodes is implicit in the global traversal, so the loss function of the feature-level self-distillation is:

[0079] ;

[0080] The feature-level self-distillation encourages adjacent positive nodes to be closer together and pushes adjacent negative nodes away.

[0081] like Figure 4 As shown, the label-level self-distillation is: extracting the label information of the target node into the neighborhood; considering the target node The known corresponding soil classification labels and target nodes The difference in the predicted probability distribution of the target node The known corresponding soil classification labels and neighbor nodes The differences in the predicted probability distributions are calculated by the cross-loss entropy function Calculated, and are all positive, that is, in the normal prediction loss (target node The known corresponding soil classification labels and target nodes Based on the difference in the predicted probability distribution of Neighbor nodes The corresponding loss of the label level self-distillation is constructed to narrow the prediction difference of the labels between the neighborhoods, and the cross loss entropy function is used to The model learns how to keep the label prediction of the target node consistent with the label prediction of the neighboring nodes, thereby improving the effect of label propagation; the loss function of the label-level self-distillation is:

[0082] .

[0083] The overall loss function is the weighted sum of the label-level self-distillation loss and the feature-level self-distillation loss, with a hyperparameter To measure the weight between the two, the calculation formula is:

[0084] ;

[0085] During model training, by dynamically adjusting hyperparameters The value of , enhances the model's sensitivity to feature differences, and balances the impact of label-level self-distillation and feature-level self-distillation on model learning to achieve the best performance. For example, when When it is larger, the model focuses more on the optimization effect of feature-level self-distillation on the model; when When is smaller, the model focuses more on the direct impact of label-level self-distillation on classification accuracy.

[0086] As a preferred method, in the actual training process of the model, if the entire soil map data is directly used as input, the number of negative samples (i.e., non-adjacent points) will be too large, resulting in too high computational complexity and greatly affecting the training efficiency. Therefore, an optimization scheme for the edge sampling strategy is proposed, which optimizes the selection of negative samples. The edge sampling strategy no longer uses the complete graph as input, but selects samples from the entire edge set of the graph. Select several edges as minibatch and distribute them according to the defined probability , accurately select negative samples from the minibatch. In a complete training process, the model adopts a batch-by-batch training method, that is, for any target node , no need to check all neighbor nodes For consideration, we only need to focus on the neighboring nodes in the selected edge set, while negative samples are randomly selected based on a probability distribution. This approach effectively reduces computational complexity and improves overall training efficiency while ensuring model learning effectiveness, enabling the model to more efficiently handle soil classification tasks on large-scale graph data.

[0087] After optimizing the negative sample selection using the edge sampling strategy, the neighbor nodes The computation is explicitly restricted to the sampled edges, and local computation is achieved through batch processing, which not only preserves the interaction of neighbor information but also avoids the overhead of global traversal. This strategy achieves a balance between efficiency and effect by replacing the global with the local. Therefore, the loss function of the feature-level self-distillation is:

[0088] ;

[0089] Where, is the batch size, is a small batch of edge sets; are all sampling points in the mini-batch edge set, ; Target node A collection of Target node Its neighboring nodes edge; The negative sample nodes are obtained after optimizing the negative sample selection through the edge sampling strategy; is a probability distribution, and the default is uniform distribution. ; is the target node after linear transformation in the second inference layer Neighbor nodes The prediction results; is the target node after SoftMax conversion in the second inference layer Neighbor nodes The predicted probability distribution of is the second simulated point after linear transformation in the first inference layer The prediction results;

[0090] The loss function of the label-level self-distillation is expressed as follows after optimizing the negative sample selection using the edge sampling strategy:

[0091] ;

[0092] The overall loss function of the dual knowledge self-distillation remains unchanged.

[0093] The land property classification method based on dual knowledge self-distillation multi-layer perceptron provided by the present invention can improve the training speed, reduce the model complexity, integrate soil topology information, reduce the influence of overfitting and noise, and accurately classify soil properties of different scales and types; the use of pure multi-layer perceptron avoids the computational bottleneck of GCN in the feature aggregation process, significantly improves the training speed, and makes the application on large-scale data sets more efficient; compared with GCN, the pure multi-layer perceptron has a simpler structure and fewer parameters, simplifies the model structure, and reduces the complexity of the model, thereby making the model easier to deploy and run in resource-limited environments; based on the homomorphic graph smoothing hypothesis, through the dual knowledge self-distillation method, the neighbor information is distilled to the target node by combining feature-level self-distillation and label-level self-distillation, and the topological information of the soil is fully integrated to improve the classification accuracy, overcome the defect that the existing method fails to consider the spatial structure of the soil, and overcomes the defect that the existing method fails to consider the spatial structure of the soil. The attention mechanism dynamically calculates the feature differences and label differences between nodes, optimizes the model learning process, better considers the spatial topological information of the soil, and makes full use of neighborhood information in feature learning and label prediction to improve classification accuracy and effectiveness; combines the feature-level and label-level self-distillation strategies to make the model more robust in the face of data noise, and reduces the risk of overfitting through effective positive and negative sample selection; introduces an edge sampling strategy to optimize negative sample selection, effectively reduces computational complexity, and improves the efficiency of the model when processing large-scale data, while retaining the effectiveness and accuracy of the model in the learning process, enhancing the model's robustness to data noise, and reducing the possibility of overfitting, thereby improving the model's generalization ability; achieves high-performance soil classification while keeping the model lightweight, ensuring that it can operate effectively in resource-limited environments; has good scalability and is adaptable to soil classification tasks of different scales and types.

[0094] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A land property classification method based on dual knowledge self-distillation multilayer perceptron, characterized in that: The steps include: Step 1: The sensors placed in the soil are regarded as nodes, the soil properties obtained by the sensors are regarded as node features, and the paths connecting adjacent sensors are regarded as edges, thereby obtaining the soil graph topology structure and node topology information; Step 2: Use edge sampling strategy to construct a training sample set, which includes: node features, node topology information and corresponding soil classification labels; A soil classification prediction model based on a pure multi-layer perceptron network is constructed, comprising: an input layer, two hidden layers, two inference layers, and an output layer; node features in the training sample set are input into the soil classification prediction model, processed by the two hidden layers, and then respectively input into the two inference layers; dual knowledge self-distillation is performed and loss is calculated by combining the node topology information and the corresponding soil classification labels; all parameters are updated through a back-propagation algorithm to obtain an optimized soil classification prediction model; Step 3: Input the node features of the soil data to be classified into the optimized soil classification prediction model, and the output layer outputs the probability corresponding to each category; Step 4: From the probabilities corresponding to each category output by the output layer, the category corresponding to the highest prediction probability is taken as the final result of soil category prediction.

2. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 1 is characterized in that: The dual knowledge self-distillation includes feature-level self-distillation and label-level self-distillation.

3. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 2 is characterized in that: The overall loss function of the dual knowledge self-distillation is: ； Where, is the overall loss function of dual knowledge self-distillation; is the loss function for label-level self-distillation; is the feature-level self-distillation loss function; An important hyperparameter for balancing feature-level self-distillation and label-level self-distillation.

4. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 3 is characterized in that: The loss function of the feature-level self-distillation is: ； Where, is the target node; Target node Neighbor nodes of For the target node Non-adjacent nodes, i.e. negative samples; is the first simulation point constructed by the attention mechanism; is the number of nodes; is the number of hidden layers; Target node Neighbor nodes A collection of For the The target node of the hidden layer output Implicit representation of node features; For the Negative samples of the hidden layer output Implicit representation of node features; Target node Its neighboring nodes The node feature implicitly represents the first simulation point of weighted fitting Implicit representation of node features; is the target node after linear transformation in the second inference layer The prediction results; is the target node after SoftMax conversion in the second inference layer The predicted probability distribution of is the simulated point after linear transformation in the first inference layer The prediction results; is the negative sample after SoftMax conversion in the first inference layer The predicted probability distribution of Negative samples the number of Target node With negative samples The edge between is the set of edges between nodes.

5. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 4 is characterized in that: The loss function of the label-level self-distillation is: ； Where, is the cross loss entropy function; is a set of nodes with known corresponding soil classification labels; Target node Known soil classification labels; For the The target node of the hidden layer output Neighbor nodes Implicit representation of node features; is the target node after SoftMax conversion in the first inference layer Neighbor nodes The predicted probability distribution of .

6. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 1 is characterized in that: The hidden layers are connected in a fully connected manner.

7. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 6 is characterized in that: Each of the hidden layers includes batch normalization, activation function, random dropout, and linear transformation.

8. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 7 is characterized in that: The representation formula of the input layer is: ； Where, Implicit representation of node features output by the input layer; is the node feature of soil data.

9. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 8 is characterized in that: The hidden layer is represented by: ； Where, For the The node features of the hidden layer output are implicitly represented; For the The node features of the hidden layer input are implicitly represented; is the parameter matrix of a specific layer; is the random dropout function; is batch normalization; is the activation function Leaky Relu.

10. The land property classification method based on dual knowledge self-distillation multilayer perceptron according to claim 5 is characterized in that: The edge sampling strategy is used to optimize the loss calculation of dual knowledge self-distillation. The optimized loss functions of feature-level self-distillation and label-level self-distillation are: ； Where, is the batch size, is a small batch of edge sets; are all sampling points in the mini-batch edge set, ; Target node A collection of Target node Its neighboring nodes edge; The negative sample nodes are obtained after optimizing the negative sample selection through the edge sampling strategy; is a probability distribution, and the default is uniform distribution. ; is the target node after linear transformation in the second inference layer Neighbor nodes The prediction results; is the target node after SoftMax conversion in the second inference layer Neighbor nodes The predicted probability distribution of is the second simulation point constructed by the attention mechanism; Target node Its neighboring nodes The node feature implicitly represents the second simulation point obtained by weighted fitting Implicit representation of node features is the second simulated point after linear transformation in the first inference layer prediction results.

Images