Land property classification method based on double knowledge self-distillation multilayer perceptron

Abstract

The invention discloses a land property classification method based on a double-knowledge self-distillation multilayer perceptron, which comprises the following steps of: 1, taking soil properties obtained by sensors as node characteristics, and taking paths for connecting adjacent sensors as edges to obtain node topological information; 2, constructing a training sample set, wherein the training sample set comprises node features, node topological information and corresponding soil classification labels; constructing a soil classification prediction model; inputting node features in the training sample set into the soil classification prediction model, carrying out double knowledge self-distillation and calculating loss in combination with node topology information, and updating all parameters through a back propagation algorithm to obtain an optimized soil classification prediction model; 3, inputting node features of soil to be classified into the optimized soil classification prediction model, and outputting a probability corresponding to each category by an output layer; and 4, taking the category corresponding to the highest prediction probability as a final result of soil category prediction.

Description

Technical Field

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

[0002] Soil classification is a key task in multiple fields such as agriculture, environmental science, and geology, aiming to systematically classify soils according to their physical, chemical, and biological properties. Effective soil classification can help formulate reasonable land use policies, increase crop yields, manage water resources, and predict ecological changes. With the development of remote sensing technology and geographic information system (GIS), using graph data for soil classification has become increasingly common. Graph data can effectively represent the spatial relationships and similarities between soil samples. Traditional methods usually rely on manual feature extraction and empirical models, which are inefficient and vulnerable to human biases.

[0003] In recent years, the introduction of deep learning technology has brought new opportunities for soil classification. By automatically learning soil features, neural networks can improve classification accuracy. Graph data has become increasingly important in many fields, including social network analysis, bioinformatics, and the Internet of Things. Soil information, as graph information with a natural spatial structure and belonging to the category of homomorphic graphs with the same node properties (soil) and edge properties (adjacency), is thus best processed using a neural network framework for graph data. The traditional graph convolutional neural network (GCN) has been widely used because it can effectively capture the relationships between nodes, but its slow training speed and large model size limit its application on large-scale datasets. Other methods do not consider the soil spatial topology information, reducing the model's accuracy.

[0004] In the prior art, the multi-modal neural network classification method sets detectors in the soil to monitor and obtain various modal data of the soil in real time for training 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, thus affecting the classification accuracy. Moreover, the multi-modal framework also has disadvantages such as a large amount of data noise and easy overfitting.

[0005] Another technique, the graph convolutional neural network (GCN), although capable of processing the soil's spatial topology information, relies on feature aggregation of the graph structure and local neighborhood information, which causes computational bottlenecks when processing large-scale graph data. Although there are some improvement schemes such as GraphSAGE and ChebNet that 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 objective of the present invention is to provide a land property classification method based on a dual knowledge self-distillation multi-layer perceptron, which can improve the training speed, reduce the model complexity, integrate soil topological information, reduce the influence of overfitting and noise, and accurately classify soil properties of different scales and types.

[0007] The technical solution provided by the present invention is as follows: A land property classification method based on a dual knowledge self-distillation multi-layer perceptron, comprising the following steps: Step 1: Taking the sensors arranged in the soil as nodes, taking the soil properties obtained through the sensors as node features, and taking the paths connecting adjacent sensors as edges, to obtain the soil map topological structure and node topological information; Step 2: Constructing a training sample set, which includes: node features, node topological information, and corresponding soil classification labels; Constructing a soil classification prediction model based on a pure multi-layer perceptron network, which includes: an input layer, two hidden layers, two inference layers, and an output layer; inputting the node features in the training sample set into the soil classification prediction model, after being processed by the two hidden layers, respectively inputting them into the two inference layers, performing dual knowledge self-distillation in combination with the node topological information and the corresponding soil classification labels and calculating the loss, and updating all parameters through the backpropagation algorithm to obtain an optimized soil classification prediction model; Step 3: Inputting 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, taking the category corresponding to the highest predicted probability as the final result of the soil category prediction.

[0008] Preferably, the dual knowledge self-distillation includes: feature-level self-distillation and label-level self-distillation.

[0009] Preferably, the overall loss function of the dual knowledge self-distillation is:

[0010] ; In the formula, is the overall loss function of the dual knowledge self-distillation; is the loss function of the label-level self-distillation; is the feature-level self-distillation loss function of; is an important hyperparameter for weighing the feature-level self-distillation and the label-level self-distillation.

[0011] Preferably, the loss function of the feature-level self-distillation is: ; In the formula, is the target node; is the target node 's neighbor node; is the node that is not adjacent to the target node , that is, the negative sample; is the first simulated point constructed by the attention mechanism; is the number of nodes; is the number of hidden layers; is the target node 's neighbor nodes 's set; is at the layer of the hidden layer output of the target node 's implicit representation of node features; is at the layer of the hidden layer output of the negative sample 's implicit representation of node features; is the target node and its neighbor nodes 's first simulated point weighted and fitted from the implicit representation of node features 's implicit representation of node features; is the prediction result of the target node after linear transformation in the second inference layer; is the predicted probability distribution of the target node after SoftMax conversion in the second inference layer; is the prediction result of the simulated point after linear transformation in the first inference layer; is the predicted probability distribution of the negative sample after SoftMax conversion in the first inference layer; is the number of negative samples ; is the target node and the negative sample 's edge; is the set of edges between nodes.

[0012] Preferably, the loss function of the label-level self-distillation is: ; In the formula, is the cross-entropy loss function; is the set of nodes with known corresponding soil classification labels; is the target node 's known soil classification label; is at the layer of the hidden layer output of the target node 's neighbor nodes Implicit representation of node features; For the target node after SoftMax conversion in the first inference layer of the neighbor nodes predicted probability distribution.

[0013] Preferably, a fully connected connection method of neurons is adopted between each of the hidden layers.

[0014] Preferably, each of the hidden layers includes: batch normalization, activation function, dropout, and linear transformation.

[0015] Preferably, the representation formula of the input layer is: ; wherein, is the implicit representation of the node features output by the input layer; is the node feature of the soil data.

[0016] Preferably, the representation formula of the hidden layer is: ; wherein, is the implicit representation of the node features output by the th hidden layer; is the implicit representation of the node features input to the th hidden layer; is the parameter matrix of a specific layer; is the dropout function; is the batch normalization; is the activation function Leaky Relu.

[0017] Preferably, an edge sampling strategy is adopted to optimize the loss calculation of double knowledge self-distillation, and the optimized loss functions of feature-level self-distillation and label-level self-distillation are: ; wherein, is the batch size, is the mini-batch edge set; is all the sampling points in the mini-batch edge set, ; is the set of target nodes ; is the target node and its neighbor nodes edge; is the negative sample node obtained by optimizing the negative sample selection through the edge sampling strategy; is the probability distribution, and the uniform distribution is adopted by default, ; The predicted result of the neighbor nodes of the target node after linear transformation in the second inference layer; The predicted result; The predicted probability distribution of the neighbor nodes of the target node after SoftMax transformation in the second inference layer; The predicted probability distribution; The second simulated point constructed by the attention mechanism; The target node and its neighbor nodes The second simulated point weighted and fitted by the implicit representation of the node features The predicted result of the second simulated point after linear transformation in the first inference layer.

[0018] The beneficial effects of the present invention are: The land property classification method based on double 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 overfitting and noise effects, and accurately classify soil properties of different scales and types. Brief Description of the Drawings

[0019] Figure 1 It is a training flow chart of the prediction model of the land property classification method based on double knowledge self-distillation multi-layer perceptron described in the present invention.

[0020] Figure 2 It is a prediction flow chart of the prediction model of the land property classification method based on double knowledge self-distillation multi-layer perceptron described in the present invention.

[0021] Figure 3 It is a schematic diagram of feature-level self-distillation described in the present invention.

[0022] Figure 4 It is a schematic diagram of label-level self-distillation described in the present invention. Detailed Embodiments

[0023] The following further describes the present invention in detail with reference to the drawings, so that those skilled in the art can implement it according to the description in the specification.

[0024] As Figure 1-2 shown, the present invention provides a land property classification method based on double knowledge self-distillation multi-layer perceptron, and the specific steps are as follows: Step 1: Arrange sensors in the soil. Take the sensors arranged in the soil as nodes, and take the soil properties obtained by the sensors as the node features of each node. The soil properties include soil temperature, soil humidity, etc. obtained by the sensors; take the paths connecting adjacent sensors as edges to form the soil map topology structure, thereby obtaining node topology information; the node features of each node and the node topology information together constitute the soil data.

[0025] In this embodiment, let represent the soil map, where is the set of nodes; is the set of edges between nodes, that is, the soil map topology structure; let be the node feature set; where is the number of nodes; each node has a -dimensional feature vector associated with it.

[0026] Step 2: Soil classification prediction model training stage: Construct a training sample set, which includes: nodes with known soil classification labels, the node features, node topology information of the nodes, and the known corresponding soil classification labels. In this embodiment, following the common semi-supervised node setting, a part of the nodes have corresponding and known soil classification labels , which are used as the training sample set data.

[0027] Construct a soil classification prediction model based on a pure multi-layer perceptron network, 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 ; input the node features in the training sample set into the soil classification prediction model, and after being processed by the input layer, the first hidden layer, and the second hidden layer in sequence, 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; Among them, the representation formula of the input layer is: ; In the formula, is the implicit representation of the node features output by the input layer; is the node feature of the soil data; 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 the output is the implicit feature representation of each node, which is a high-dimensional encoding of the input data and is usually used for subsequent feature optimization and inference processes; A fully connected connection method of neurons is adopted between each of the hidden layers; each of the hidden layers includes: batch normalization, activation function, dropout, and linear transformation; The representation formula of the hidden layer is: ; In the formula, is the implicit feature representation of the node features output by the -th hidden layer; is the implicit feature representation of the node features input to the -th hidden layer; is the parameter matrix of a specific layer; is the dropout function; is the batch normalization; is the activation function Leaky Relu.

[0028] Each of the inference layers includes: linear transformation and SoftMax normalization. Among them, the first inference layer only participates in the training stage of the soil classification prediction model, and the second inference layer participates in both the training stage and the prediction stage of the soil classification prediction model. Retaining only the second inference layer in the prediction stage of the soil classification prediction model can reduce the inference time and ensure efficient single-node inference.

[0029] The implicit feature representations of the node features of the first hidden layer and the second hidden layer are respectively input into the first inference layer and the second inference layer. After the linear transformation in the first inference layer and the second inference layer, the prediction results of the first inference layer and the prediction results of the second inference layer are obtained. The prediction results of the inference layer are the hidden layer features perceived by the graph topology structure. Then, after SoftMax normalization, the prediction probability distribution is obtained, and double knowledge self-distillation is performed in combination with the node topology information and the known corresponding soil classification labels. The loss is calculated according to the results of the double knowledge self-distillation; Among them, the first hidden layer and the second hidden layer output the implicit feature representations of the node features of the target node , the implicit feature representations of the node features of the neighbor nodes of the target node and negative samples Implicit representation of node features , is the number of hidden layers; then two simulated points are constructed through the attention mechanism and , and the target node and its neighbor nodes 's implicit representations of node features are weighted and fitted to obtain the implicit representations of node features of two simulated points and ; the implicit representations of node features of the neighbor nodes and of the target node , the implicit representations of node features of the negative samples , the implicit representations of node features of the two simulated points and and are input into the first inference layer; the implicit representation of node features of the target node and the implicit representation of node features of the neighbor nodes of the target node are input into the second inference layer; and double knowledge self-distillation and loss calculation are combined with the corresponding node topology information and the known corresponding soil classification labels ; The double knowledge self-distillation includes: feature-level self-distillation and label-level self-distillation The overall loss function for calculating the loss according to the results of the double knowledge self-distillation is: ; In the formula, is the overall loss function of the double knowledge self-distillation is the loss function of the label-level self-distillation is the feature-level self-distillation loss function is the important hyperparameter for weighing the feature-level self-distillation and the label-level self-distillation The loss function of the feature-level self-distillation is: ; In the formula, is the target node is the neighbor node of the target node ; is the node that is not adjacent to the target node , that is, the negative sample is the simulated point constructed by the attention mechanism is the number of nodes​​​​ is the number of hidden layers; is the target node of the neighbor nodes set; is at the target node output by the hidden layer latent representation of node features; is at the negative sample output by the hidden layer of the latent representation of node features; is the target node and its neighbor nodes simulation point weighted and fitted from the latent representation of node features latent representation of node features; is the prediction result of the target node after linear transformation in the second inference layer is the prediction probability distribution of the target node after SoftMax transformation in the second inference layer is the prediction result of the simulation point after linear transformation in the first inference layer is the prediction result of the negative sample after linear transformation in the first inference layer is the prediction probability distribution of the negative sample after SoftMax transformation in the first inference layer is the number of negative samples ; is the target node and the negative sample edge between; is the set of edges between nodes; The loss function of the label-level self-distillation is: ; In the formula, is the cross-loss entropy function; is the set of nodes with known corresponding soil classification labels; is the target node known soil classification label, ; is at the target node output by the hidden layer of the neighbor nodes latent representation of node features; is the prediction result of the neighbor nodes of the target node after linear transformation in the first inference layer The predicted probability distribution of the neighbor nodes of the target node after SoftMax conversion in the first inference layer.

[0030] Update all parameters through the backpropagation algorithm to obtain an optimized soil classification prediction model.

[0031] Step 3, Prediction stage of the soil classification prediction model: Input the node features in the soil data of the soil to be classified into the optimized soil classification prediction model, and successively pass through the input layer, the first hidden layer, and the second hidden layer to obtain the implicit representation of the node features of the optimized second hidden layer, and input it into the second inference layer. After the linear transformation in the second inference layer, the prediction result of the second inference layer is obtained, and then through SoftMax normalization processing to obtain the prediction probability distribution; the output layer outputs the probability corresponding to each category.

[0032] Step 4, From the probabilities corresponding to each category output by the output layer, take the category corresponding to the highest predicted probability as the final result of the soil category prediction.

[0033] As Figure 3 shown, in order to make the distribution of adjacent node feature values converge, the traditional Mixup data augmentation scheme is modified. For the target node , find all neighbor nodes of the target node , and construct two simulated points and by the attention mechanism; Examine the difference between the target node and its neighbor node through the attention mechanism to obtain a weight that increases as the difference increases. Weight the implicit representation of the node features of the target node and its neighbor node by the weight to fit the implicit representation of the node features of the two simulated points , . The implicit representation of the node features of the first simulated point and its neighbor node generated by interpolation, and the second simulated point is: ; ; In the formula, is the first simulated point constructed by the attention mechanism; is the second simulated point constructed by the attention mechanism; is the target node and its neighbor node The first simulated point weighted and fitted by the implicit representation of the node features The implicit representation of the node features; Is the target node And its neighbor nodes The second simulated point weighted and fitted by the implicit representation of the node features The implicit representation of the node features; Is the interpolation coefficient; The interpolation coefficient The definition formula is: ; In the formula, Is the feature transformation matrix, which maps the input features to a new space; Is the shared attention weight, which is used to calculate the similarity between nodes; Is the target node And neighbor nodes The feature concatenation; Is the attention mechanism function; Due to the characteristics of the attention mechanism, when the two-point features converge, the weight Tends to 0.5, and the loss function will become smaller, which is equivalent to implicitly reducing the feature difference by reducing the label distribution difference.

[0034] The feature-level self-distillation is as follows: refining the feature information in the neighborhood to the target node; the feature prediction difference between the target node And the first simulated point , the feature prediction difference between the neighbor node And the second simulated point Are used as the positive terms in the loss function, and the training expects to reduce the difference. Then, points that are not adjacent to the target node And neighbor nodes Are selected as negative samples , and the feature prediction differences between the target node And the negative sample , the neighbor node And the negative sample Are used as the negative terms in the loss function, and the training expects to increase the difference; thus, the loss function of the feature-level self-distillation is constructed to reduce the feature difference between adjacent points and increase the feature difference between non-adjacent points, realizing the 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. Therefore, the loss function of the feature-level self-distillation is: ; The feature-level self-distillation encourages adjacent positive nodes to be closer and pushes away adjacent negative nodes.

[0035] As shown Figure 4 below, the label-level self-distillation is to refine the label information of the target node into its neighborhood; considering the difference between the known corresponding soil classification label of the target node and the predicted probability distribution of the target node , as well as the difference between the known corresponding soil classification label of the target node and the predicted probability distribution of its neighbor node . These differences are all calculated by the cross entropy loss function and are all positive. That is, on the basis of the normal prediction loss (the difference between the known corresponding soil classification label of the target node and the predicted probability distribution of the target node ), the corresponding loss of the neighbor node of the target node is additionally considered to construct the loss function of label-level self-distillation, so as to reduce the prediction difference of labels between neighborhoods. Through the cross entropy loss function , the model learns how to make the label prediction of the target node consistent with that of the neighbor node, thereby improving the effect of label propagation; the loss function of the label-level self-distillation is: .

[0036] The overall loss function is the weighted sum of the label-level self-distillation loss and the feature-level self-distillation loss, and a hyperparameter is used to measure the weight between the two. The calculation formula is: ;During the model training process, by dynamically adjusting the value of the hyperparameter , the sensitivity of the model to feature differences is enhanced, and the influence of label-level self-distillation and feature-level self-distillation on model learning is balanced to achieve the best performance. For example, when is large, the model focuses more on the optimization effect of feature-level self-distillation on the model; when is small, the model focuses more on the direct impact of label-level self-distillation on classification accuracy.

[0037] As a preference, during the actual model training process, if the entire soil map data is directly used as the input, it will result in too many negative samples (i.e., non-adjacent points), leading to too high computational complexity and greatly affecting the training efficiency. Therefore, an optimized solution of the edge sampling strategy is proposed to optimize the selection of negative samples through the edge sampling strategy. The edge sampling strategy no longer uses the complete graph as the input, but selects several edges from the entire edge set of the graph as the minibatch and selects according to the defined probability distribution ​, select negative samples accurately from the minibatch. During a complete training process, the model adopts a batch-by-batch training method, that is, for any target node , there is no need to consider all neighbor nodes anymore. It only needs to focus on the neighbor nodes in the selected edge set, and the negative samples are randomly drawn according to the probability distribution. In this way, while ensuring the learning effect of the model, the computational complexity is effectively reduced, the overall training efficiency is improved, and the model can more efficiently handle the soil classification task of large-scale graph data.

[0038] After optimizing the negative sample selection using the described edge sampling strategy, the calculation of neighbor nodes is explicitly restricted to the sampled edges. Local calculations are achieved through batch processing, which not only retains the neighbor information interaction 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: ; In the formula, is the batch size, is the mini-batch edge set; are all the sampling points in the mini-batch edge set, ; is the set of target nodes ; is the target node and its edge with its neighbor node ; are the negative sample nodes obtained after optimizing the negative sample selection using the edge sampling strategy; is the probability distribution, and the uniform distribution is adopted by default, ; is the prediction result of the neighbor node of the target node after linear transformation in the second inference layer; is the predicted probability distribution of the neighbor node of the target node after SoftMax conversion in the second inference layer; is the prediction result of the second simulation point after linear transformation in the first inference layer; The loss function of the label-level self-distillation is expressed as follows after optimizing the negative sample selection using the described edge sampling strategy: ; The overall loss function of the dual knowledge self-distillation remains unchanged.

[0039] The land property classification method based on a dual knowledge self-distillation multi-layer perceptron provided by the present invention can improve the training speed, reduce the model complexity, integrate soil topological information, reduce the influence of overfitting and noise, and accurately classify soil properties of different scales and types; by using a pure multi-layer perceptron, it avoids the computational bottleneck in the feature aggregation process of the GCN, significantly improves the training speed, and makes the application on large-scale datasets more efficient; compared with the GCN, the structure of the pure multi-layer perceptron is simpler and has fewer parameters, simplifies the model structure, and reduces the complexity of the model, thus making the model easier to deploy and run in an environment with limited resources; based on the homomorphic graph smoothing assumption, through the dual knowledge self-distillation method, by combining feature-level self-distillation and label-level self-distillation, distill the neighbor information to the target node, fully integrate the topological information of the soil to improve the classification accuracy, overcome the defect that the existing methods fail to consider the soil spatial structure, dynamically calculate the feature differences and label differences between nodes through the attention mechanism, optimize the model learning process, can better consider the soil spatial topological information, make full use of the neighborhood information in feature learning and label prediction, and improve the classification accuracy and effectiveness; by combining the strategies of feature-level and label-level self-distillation, the model is more robust in the face of data noise, and reduces the overfitting risk through effective positive and negative sample selection; introduce the edge sampling strategy to optimize the negative sample selection, effectively reduce the computational complexity, improve the efficiency of the model in processing large-scale data, while retaining the effectiveness and accuracy of the model in the learning process, enhance the robustness of the model to data noise, and reduce the possibility of overfitting, thereby improving the generalization ability of the model; achieve high-performance soil classification while keeping the model lightweight, ensuring its effective operation in an environment with limited resources; have good scalability and adapt to soil classification tasks of different scales and types.

[0040] Although the embodiments of the present invention have been disclosed as above, they are not limited to only the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. A land property classification method based on dual knowledge self-distillation multi-layer perceptron, characterized in that: The steps include: Step 1: Take the sensors arranged in the soil as nodes, the soil properties obtained by the sensors as node features, and the paths connecting adjacent sensors as edges to obtain the soil graph topology and node topology information; Step 2: 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, which includes: an input layer, two hidden layers, two inference layers and an output layer; the node features in the training sample set are input into the soil classification prediction model, and after being processed by the two hidden layers, they are respectively input into the two inference layers, and the node topology information and the corresponding soil classification labels are combined to perform double knowledge self-distillation and calculate the loss, and all parameters are updated by the 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, take the category corresponding to the highest prediction probability 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: ； In the formula, 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; Important hyperparameters 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: ； In the formula, is the target node; The 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; The 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 hidden layer output Implicit representation of node features; The target node Its neighbor nodes The node feature implicitly represents the first simulation point 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 SoftMax conversion in the first inference layer The predicted probability distribution of Negative samples the number of The 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: ； In the formula, is the cross loss entropy function; is a set of nodes with known corresponding soil classification labels; The 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 multi-layer 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 multi-layer perceptron according to claim 7 is characterized in that: The representation formula of the input layer is: ； In the formula, It is the 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 multi-layer perceptron according to claim 8 is characterized in that: The hidden layer is represented by: ； In the formula, 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 drop function; is batch normalization; is the activation function Leaky Relu.

10. The land property classification method based on dual knowledge self-distillation multi-layer 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 loss functions of the optimized feature-level self-distillation and label-level self-distillation are: ； In the formula, is the batch size, is a small batch edge set; is all the sampling points in the mini-batch edge set, ; The target node A collection of; The target node Its neighbor nodes The edge of The negative sample nodes are obtained after optimizing the negative sample selection through the edge sampling strategy; is a probability distribution, uniform distribution is used by default, ; 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; The target node Its neighbor 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 reasoning layer prediction results.

Images