Soil pollutant identification and route tracking method and system based on artificial intelligence

Abstract

The present invention provides a soil pollutant identification and route tracking method and system based on artificial intelligence, which relates to the field of pollution monitoring technology, including collecting soil samples in a target area, obtaining sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; inputting the sample spectral data and the sample chemical composition into a pre-trained pollutant identification model, the pollutant identification model is constructed based on a convolutional neural network, extracting sample component characteristics according to the sample spectral data and the sample chemical composition, and identifying the sample pollutant type and sample pollutant concentration by classifying the sample component characteristics; constructing a pollutant heat map based on the sample pollutant type and the sample pollutant concentration in combination with the sample location information; inputting the pollutant heat map into a pre-trained route tracking model, the route tracking model is constructed based on a graph neural network, and predicting the migration route and diffusion state of the pollutant by constructing temporal and spatial relationships.

Description

Technical Field

[0001] The present invention relates to the field of pollution monitoring technology, and in particular to a soil pollutant identification and route tracking method and system based on artificial intelligence. Background Art

[0002] There are many types of soil pollutants, including heavy metals, organic compounds, pesticides, etc. Pollutants not only destroy soil structure and reduce soil fertility, but also affect human health through the food chain. Therefore, accurately identifying and tracking the source and diffusion path of soil pollutants has become a key link in environmental protection and governance.

[0003] Traditional soil pollutant detection methods mainly rely on laboratory analysis, such as gas chromatography, liquid chromatography, mass spectrometry, etc. Although they can provide high-precision analysis results, they have problems such as complex sampling process, long detection cycle, and high cost, which makes it difficult to meet the needs of large-scale and rapid detection. In addition, tracking the diffusion path of pollutants also requires a large amount of on-site investigation and historical data, which is cumbersome and difficult to achieve real-time monitoring.

[0004] In recent years, the rapid development of artificial intelligence technology has provided new solutions for soil pollutant identification and tracking. The soil pollutant identification and route tracking method based on artificial intelligence can significantly improve the efficiency and accuracy of pollutant identification through big data analysis and model training. At the same time, through GIS technology and dynamic modeling, the diffusion path of pollutants can be tracked and their future diffusion trends can be predicted, providing a scientific basis for environmental governance. In summary, the present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a soil pollutant identification and route tracking method and system based on artificial intelligence, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a soil pollutant identification and route tracking method based on artificial intelligence, comprising:

[0008] Collect soil samples from the target area, obtain sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; input the sample spectral data and the sample chemical composition into a pre-trained pollutant identification model, the pollutant identification model is constructed based on a convolutional neural network, extracts sample component characteristics according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and sample pollutant concentration by classifying the sample component characteristics;

[0009] Based on the sample pollutant type and the sample pollutant concentration, combined with the sample location information, construct a pollutant heat map;

[0010] The pollutant heat map is input into a pre-trained route tracking model, which is built based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships.

[0011] In an optional embodiment,

[0012] According to the sample spectral data and sample chemical composition, the sample component characteristics are extracted, and the sample pollutant type and sample pollutant concentration are identified by classifying the sample component characteristics, including:

[0013] The sample spectrum data and the sample chemical composition data are combined into a two-dimensional matrix to determine the model input matrix;

[0014] Performing feature extraction on the input matrix, extracting local features and global features of the sample spectral data through the convolution layer and the pooling layer of the pollutant identification model to form a sample spectral feature, extracting the sample chemical composition feature from the sample chemical composition data through the fully connected layer, and fusing the sample spectral feature with the sample chemical composition feature to generate a comprehensive feature vector;

[0015] Two subnetworks are set at the output layer of the pollutant identification model, including a type classification subnetwork and a concentration regression subnetwork; wherein the type classification subnetwork uses a softmax activation function to map the comprehensive feature vector to the probability distribution of the pollutant type; and the concentration regression subnetwork uses a linear activation function to map the comprehensive feature vector to the pollutant concentration value;

[0016] A multi-task learning strategy is used to jointly optimize the type classification task and the concentration regression task, a multi-task loss function is constructed, and the weight parameters of the pollutant identification model are updated through the back-propagation algorithm until the preset convergence conditions are reached;

[0017] Based on the trained soil pollutant identification model, the sample pollutant type and sample pollutant concentration of the soil samples in the target area are determined.

[0018] In an optional embodiment,

[0019] Based on the sample pollutant type and the sample pollutant concentration, combined with the sample location information, constructing a pollutant heat map includes:

[0020] Based on the sample location information and the sample pollutant concentration, a spatial interpolation method is used to solve the Kriging weight by minimizing the estimated variance, and the unit pollutant concentration at any location in the target area is determined by combining the semivariogram and the spatial location relationship, and a grid concentration representation is constructed;

[0021] The grid concentration representation is converted into a raster image, and combined with color coding, the unit pollutant concentration is mapped to a continuous color scale to determine the spatial variation of the pollution degree, and the pollutant heat map is constructed by combining the sample location information and the sample pollutant concentration.

[0022] In an optional embodiment,

[0023] The semivariogram function has the following formula:

[0024] ;

[0025] Among them, γ(h) represents the semivariogram, h represents the distance vector of the spatial position, N(h) represents the total number of samples with distance vector h, n represents the sample ordinal number, and Z(s n ) represents the pollutant concentration value at the corresponding position of sample n, θ represents the rotation angle, R θ A rotation matrix representing the distance vector.

[0026] In an optional embodiment,

[0027] The route tracking model is built based on a graph neural network. By determining the temporal and spatial relationships, the migration routes and diffusion states of pollutants are predicted, including:

[0028] The pollutant heat maps at different time steps are used as nodes, and temporal edges are added between nodes adjacent in time, and spatial edges are added between nodes adjacent in space. Based on the migration intensity of pollutants in time and space, the temporal edge weights and spatial edge weights are determined respectively to construct a spatiotemporal relationship diagram.

[0029] The pollutant heat map corresponding to each node is encoded by convolutional neural network, and a high-dimensional feature vector is extracted as the node feature; based on the node feature, the information of neighboring nodes is aggregated through multi-layer iterative update, the position of the node in the spatiotemporal relationship graph and the contextual information of pollutant migration are encoded, and the hidden state of the node is determined;

[0030] The node hidden state is input into the decoder network to determine the future pollutant heat map of the future time step. The future pollutant heat map is used as input, and the future pollutant heat map of multiple consecutive time steps is determined through recursive repetition to form a pollutant migration route. In combination with the spatial change trend, the pollutant diffusion speed and direction are determined.

[0031] In an optional embodiment,

[0032] Also includes:

[0033] Based on the preset pollutant types and the preset environmental impact factors, a map entity is constructed, based on the physical mechanism of migration and diffusion and the geographical impact factors, a map relationship is constructed, and based on the map entity and the map relationship, a pollutant diffusion map is constructed;

[0034] Embed the pollutant diffusion map into the spatiotemporal relationship graph to generate diffusion knowledge fusion nodes, and determine the hidden state of the knowledge fusion nodes through multi-level iterative updates;

[0035] The decoder network, combined with the pollutant diffusion map, constructs a diffusion knowledge decoder network, inputs the hidden state of the knowledge fusion node into the diffusion knowledge decoder network, and generates a pollutant migration route that conforms to the diffusion knowledge through recursive iteration.

[0036] In an optional embodiment, constructing a spatiotemporal relationship graph includes:

[0037] The temporal edge weight and the spatial edge weight are formulated as follows:

[0038] ;

[0039] Among them, ω l ij represents the spatial edge weight, l i represents the spatial position coordinates of node i, l j represents the spatial position coordinates of node j, σ l represents the spatial scale parameter, ω t ij represents the temporal edge weight, t i represents the time step corresponding to node i, t j represents the time step corresponding to node j, σ t represents the time scale parameter;

[0040] The formula of the high-dimensional feature vector is as follows:

[0041] ;

[0042] Among them, x t represents the high-dimensional feature vector at time step t, ReLU(·) represents the activation function, and W out represents the weight parameter of the fully connected layer, H t represents the pollutant heat map at time step t, MaxPool(·) represents the maximum pooling operation, and W conv represents the convolution kernel, and Conv(·) represents the convolution operation performed on the pollutant heat map using the convolution kernel.

[0043] According to a second aspect of the embodiments of the present invention,

[0044] Provided is an artificial intelligence-based soil pollutant identification and route tracking system, including:

[0045] The first unit is used to collect soil samples in the target area, obtain sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; input the sample spectral data and the sample chemical composition into a pre-trained pollutant identification model, the pollutant identification model is constructed based on a convolutional neural network, extracts sample component characteristics according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and sample pollutant concentration by classifying the sample component characteristics;

[0046] A second unit is used to construct a pollutant heat map based on the sample pollutant type and the sample pollutant concentration in combination with the sample location information;

[0047] The third unit is used to input the pollutant heat map into a pre-trained route tracking model, where the route tracking model is built based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships.

[0048] According to a third aspect of the embodiments of the present invention,

[0049] An electronic device is provided, comprising:

[0050] processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] A fourth aspect of the embodiments of the present invention is:

[0054] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0055] In an embodiment of the present invention, by collecting the location information, sample spectral data and sample chemical composition of soil samples, and using a pre-trained pollutant identification model to classify and identify pollutants, the types of pollutants present in the soil samples can be quickly and accurately determined; the concentration of pollutants in the soil samples can also be estimated, and by extracting features from the sample spectral data and the sample chemical composition and inputting them into the pollutant identification model, a prediction result on the pollutant concentration can be obtained; by calculating the semivariogram of the sample data, the relationship between the difference in pollutant concentration between sample points and the spatial distance is described, and a Gaussian model is preferably used for fitting, which can accurately reflect the spatial variation characteristics of the pollutant concentration; the grid concentration representation is converted into a pollutant heat map to intuitively display the spatial distribution of the pollutant concentration, and the color coding is used to represent the spatial distribution of the pollutant concentration. By using a coding method to map different concentration values ​​onto a continuous color scale, areas with higher and lower pollution levels can be clearly identified, providing an intuitive reference for environmental assessment and management. By constructing a spatiotemporal relationship graph, the migration and diffusion relationship of pollutants in time and space can be effectively described. The establishment of temporal edges and spatial edges can accurately reflect the migration intensity and correlation degree of pollutants between different time steps and spatial positions. The pollutant heat map corresponding to each node is encoded using a convolutional neural network, and high-dimensional feature vectors are extracted as node features, which effectively characterizes the location information and migration context information of pollutants in the spatiotemporal relationship graph. By recursively and repeatedly predicting future pollutant heat maps for multiple consecutive time steps, the migration routes of pollutants are constructed, and the spatial distribution evolution process of pollutants in future time steps is effectively analyzed. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flowchart of a soil pollutant identification and route tracking method based on artificial intelligence according to an embodiment of the present invention;

[0057] Figure 2 Schematic diagram of the structure of the soil pollutant identification and route tracking system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of a soil pollutant identification and route tracking method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] S101. Collect soil samples from the target area, obtain sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; input the sample spectral data and the sample chemical composition into a pre-trained pollutant identification model, the pollutant identification model is constructed based on a convolutional neural network, extracts sample component characteristics according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and sample pollutant concentration by classifying the sample component characteristics;

[0062] The sample spectral data specifically refers to the reflection or absorption data of the soil sample at different wavelengths. The reflection or absorption data is usually represented in the form of a spectral curve, where each data point represents the spectral response value of the soil sample at a specific wavelength. The sample spectral data provides information about the composition and properties of the soil, such as the color, texture, moisture content, etc. of the soil. Spectral data is usually recorded in digital form and can be collected by equipment such as a spectrometer;

[0063] The chemical composition of the sample specifically refers to the content or composition information of different chemical substances in the soil sample. The chemical composition may include but is not limited to heavy metal elements, organic matter, nutrient elements, etc. It is usually obtained through chemical analysis methods such as spectral analysis, mass spectrometry and other techniques. The chemical composition of the sample provides quantitative or qualitative information about various chemical substances in the soil;

[0064] The sample component features specifically refer to features extracted from sample spectral data and sample chemical composition, which are used to describe the properties and composition of soil samples. Features can include spectral features, such as the shape of the spectral curve, peak position, etc., and chemical composition features, such as the content of various chemical substances. By extracting features from sample spectral data and sample chemical composition, complex soil sample information is converted into numerical features that can be processed by the model, and the type and concentration of pollutants are identified.

[0065] In this embodiment, by collecting the location information, sample spectral data and sample chemical composition of soil samples, and using a pre-trained pollutant identification model to classify and identify pollutants, the types of pollutants present in the soil samples can be quickly and accurately determined; in addition to the types of pollutants, the concentrations of pollutants in the soil samples can also be estimated, and by extracting features from the sample spectral data and the sample chemical composition, and inputting them into the pollutant identification model, a predicted result on the pollutant concentration can be obtained; the technical solution involves collecting the location information of the soil samples, so that the pollutant identification results are associated with the specific geographical location, which is of great significance for the location of pollution sources and environmental monitoring.

[0066] In an optional embodiment, extracting sample component features according to sample spectral data and sample chemical composition, and identifying the sample pollutant type and sample pollutant concentration by classifying the sample component features includes:

[0067] The sample spectrum data and the sample chemical composition data are combined into a two-dimensional matrix to determine the model input matrix;

[0068] Performing feature extraction on the input matrix, extracting local features and global features of the sample spectral data through the convolution layer and the pooling layer of the pollutant identification model to form a sample spectral feature, extracting the sample chemical composition feature from the sample chemical composition data through the fully connected layer, and fusing the sample spectral feature with the sample chemical composition feature to generate a comprehensive feature vector;

[0069] Two subnetworks are set at the output layer of the pollutant identification model, including a type classification subnetwork and a concentration regression subnetwork; wherein the type classification subnetwork uses a softmax activation function to map the comprehensive feature vector to the probability distribution of the pollutant type; and the concentration regression subnetwork uses a linear activation function to map the comprehensive feature vector to the pollutant concentration value;

[0070] A multi-task learning strategy is used to jointly optimize the type classification task and the concentration regression task, a multi-task loss function is constructed, and the weight parameters of the pollutant identification model are updated through the back-propagation algorithm until the preset convergence conditions are reached;

[0071] Based on the trained pollutant identification model, the sample pollutant type and sample pollutant concentration of the soil samples in the target area are determined.

[0072] The pollutant identification model is a comprehensive model for identifying and predicting the type and concentration of pollutants in soil samples. The pollutant identification model is constructed as follows: an input layer receives input data of soil samples, including sample location information, sample spectral data, and sample chemical composition data; a convolution layer is used to extract features from the input data. The convolution layer performs convolution operations on the input data through sliding convolution kernels to generate feature maps; a pooling layer performs dimensionality reduction and sampling on the feature maps generated by the convolution layer, which can reduce the size of the feature maps while retaining the most important feature information; a fully connected layer flattens the feature maps output by the pooling layer and connects them to one or more fully connected layers to map the extracted features to the dimensional space of the model output; an output layer consists of a type classification subnetwork and a concentration regression subnetwork, wherein the type classification subnetwork is responsible for type classification of the comprehensive feature vector, usually including one or more fully connected layers and a softmax output layer, for outputting the probability distribution of different pollutant types; a concentration regression subnetwork is responsible for concentration regression prediction of the comprehensive feature vector, usually including one or more fully connected layers and a linear output layer, for outputting the predicted value of the pollutant concentration.

[0073] Collect the spectral data and chemical composition data of soil samples. Spectral data are usually represented in the form of one-dimensional vectors, recording the reflectivity or absorbance at different wavelengths. Chemical composition data contains the content information of various chemical elements or compounds in the soil samples. The spectral data and chemical composition data of each soil sample are spliced ​​in a fixed order to form a two-dimensional matrix as the input matrix of the pollutant identification model. The number of rows in the input matrix corresponds to the number of samples, and the number of columns corresponds to the total dimension of the spectral data and chemical composition data.

[0074] Feature extraction is performed on the input matrix to obtain a higher-level feature representation. The convolutional neural network is used to extract features from the sample spectral data. By designing appropriate convolutional layers and pooling layers, local and global features in the spectral data can be effectively captured. The convolutional layer extracts local patterns and correlations in the spectral data through convolution operations; the pooling layer reduces the size of the feature map through downsampling operations while retaining the most significant features. After multiple layers of convolution and pooling operations, the sample spectral features are obtained. For the sample chemical composition data, a fully connected layer is used for feature extraction. The fully connected layer maps the chemical composition data to a low-dimensional space through matrix multiplication and nonlinear transformation to extract the sample chemical composition features; the extracted sample spectral features and sample chemical composition features are fused, preferably using element-level multiplication to generate a comprehensive feature vector, which contains the sample's spectral information and chemical composition information;

[0075] In the output layer of the pollutant identification model, two parallel sub-networks are set up, which are used for the pollutant type classification task and concentration regression task respectively. For the pollutant type classification sub-network, the softmax activation function is used as the output layer. The softmax activation function maps the comprehensive feature vector to a probability distribution, which indicates the probability that the sample belongs to each pollutant type. The type with the highest probability is the predicted pollutant type. For the pollutant concentration regression sub-network, the linear activation function is used as the output layer. The linear function maps the comprehensive feature vector to a real value, which indicates the predicted pollutant concentration. Through the dual-task sub-network output layer design, the pollutant identification model can simultaneously predict the pollutant type and concentration of soil samples, providing more comprehensive pollution assessment results.

[0076] In order to simultaneously optimize the pollutant type classification and concentration regression tasks, a multi-task learning strategy is adopted for model training, and a multi-task loss function including classification loss and regression loss is constructed. The classification loss adopts the cross entropy loss function to measure the difference between the predicted type probability distribution and the true type label; the regression loss adopts the mean square error loss function to measure the difference between the predicted concentration value and the true concentration value; the two loss functions are weighted and summed according to the predetermined weight coefficient to obtain the final multi-task loss function; during the training process, the back propagation algorithm is used to calculate the gradient of the loss function with respect to the model weight parameters, and the Adam optimization algorithm is used to update the model weight parameters to minimize the multi-task loss function, and multiple training cycles are iterated until the model performance reaches the preset convergence condition. The convergence condition is preferably that the training is iterated when the loss function drops to a preset threshold.

[0077] In this embodiment, by combining the sample spectral data and the sample chemical composition data into a two-dimensional matrix, and using a convolutional neural network and a fully connected layer to extract features from the input matrix, a comprehensive feature vector containing spectral information and chemical composition information is obtained, which can effectively capture the characteristic information of the soil sample and provide richer and more accurate input features for the pollutant type classification task and concentration regression task; two parallel sub-networks are set in the output layer of the pollutant identification model, which are used for the pollutant type classification task and concentration regression task respectively, and the pollutant type and concentration of the soil sample are predicted at the same time, providing users with more comprehensive pollution assessment results; by adopting a multi-task learning strategy for model training, a multi-task loss function containing classification loss and regression loss is constructed, and the pollutant type classification and concentration regression tasks are optimized at the same time, thereby improving the generalization ability and prediction performance of the model.

[0078] S102. Based on the sample pollutant type and the sample pollutant concentration, combined with the sample location information, construct a pollutant heat map;

[0079] In an optional embodiment, constructing a pollutant heat map based on the sample pollutant type and the sample pollutant concentration in combination with the sample location information includes:

[0080] Based on the sample location information and the sample pollutant concentration, a spatial interpolation method is used to solve the Kriging weight by minimizing the estimated variance, and the unit pollutant concentration at any location in the target area is determined by combining the semivariogram and the spatial location relationship, and a grid concentration representation is constructed;

[0081] The grid concentration representation is converted into a raster image, and combined with color coding, the unit pollutant concentration is mapped to a continuous color scale to determine the spatial variation of the pollution degree, and the pollutant heat map is constructed by combining the sample location information and the sample pollutant concentration.

[0082] The Kriging weight specifically refers to an interpolation method based on spatial variation, the core idea of ​​which is to determine the interpolation weight by estimating the semivariogram based on the spatial relationship between sample points and the correlation of variable values. The Kriging weight refers to the weight used to calculate the predicted value of the unknown location in the Kriging interpolation, which is determined by minimizing the variance between the predicted value and the known value. The Kriging weight determines the contribution of each sample point to the predicted point, and is usually calculated based on the distance and the semivariogram.

[0083] The semivariogram specifically refers to a function that describes the correlation between random field variables in space, reflects the degree of change of variable values ​​as the distance increases, and is used to describe how the degree of variation between samples changes as the distance increases. In spatial interpolation, the semivariogram is usually used to measure the spatial correlation between samples. By analyzing the semivariogram, the spatial correlation structure between samples can be understood, thereby guiding the selection of interpolation methods and parameter settings;

[0084] The color coding specifically refers to a method of representing data changes or different categories through colors. In the visualization of spatial changes in pollution levels, color coding is often used to map pollutants of different concentrations to different colors in order to intuitively display the pollution levels in different areas. Usually, high-concentration pollutants are represented by highly saturated colors, while low-concentration pollutants are represented by low-saturation colors. Color coding can clearly display the spatial distribution and change trend of pollution levels, providing a visual reference for environmental monitoring and pollution control.

[0085] Organizing the sample location information and pollutant concentration data into corresponding data pairs to form a sample data set; in order to estimate the pollutant concentration at any location in the target area, a spatial interpolation method is used, preferably Kriging interpolation, and interpolation weights are determined by minimizing the estimated variance based on the spatial autocorrelation of the sample data;

[0086] Calculate the semivariogram based on the sample data to describe the relationship between the difference in pollutant concentration between sample points and the spatial distance. Fit the semivariogram model by calculating the semivariogram values ​​of sample point pairs at different distances, preferably the Gaussian model. For any position in the target area, calculate the kriging weight by minimizing the estimated variance based on the spatial position relationship between the position and the sample point, solve the kriging equation group, and obtain the contribution weight of each sample point to the estimated position. Finally, perform a weighted summation of the kriging weight and the pollutant concentration of the sample point to obtain the pollutant concentration value at the estimated position. Repeat the process to estimate all positions in the target area, construct a grid concentration representation, divide the target area into a regular grid, and each grid cell corresponds to an estimated pollutant concentration value.

[0087] The grid concentration representation is converted into a pollutant heat map to intuitively display the spatial distribution of pollutant concentration. First, the grid concentration representation is converted into a raster image, where the raster image is a digital image composed of a pixel grid, each pixel corresponds to a grid unit, and the pixel value represents the pollutant concentration of the unit; the pollutant concentration value is mapped to a continuous color scale using a color coding method, preferably using a red-yellow-green gradient color scheme, where a higher concentration value corresponds to red, indicating a heavier degree of pollution; a lower concentration value corresponds to green, indicating a lighter degree of pollution; according to the sample location information and the sample pollutant concentration, the location and concentration value of the sample point are marked on the raster image, and a pollutant heat map is generated through color gradient and sample point marking. The pollutant heat map intuitively displays the spatial distribution of pollutant concentration in the target area, and can clearly identify areas with higher and lower pollution levels; based on the generated pollutant heat map, further result analysis and application can be carried out, and the spatial change trend of pollutant concentration can be identified by observing the color distribution and sample point location in the heat map.

[0088] In this embodiment, the Kriging interpolation method is used in combination with the spatial autocorrelation of the sample data to effectively estimate the pollutant concentration at any location in the target area. The interpolation weight is determined by minimizing the estimated variance to obtain an accurate estimate of the pollutant concentration, which provides an important reference for environmental monitoring and pollution control. The semivariogram of the sample data is calculated to describe the relationship between the difference in pollutant concentration and the spatial distance between the sample points. The Gaussian model is preferably used for fitting, which can accurately reflect the spatial variation characteristics of the pollutant concentration. The grid concentration representation is converted into a pollutant heat map to intuitively display the spatial distribution of the pollutant concentration. Different concentration values ​​are mapped to a continuous color scale through color coding, which can clearly identify areas with higher and lower pollution levels, providing an intuitive reference for environmental assessment and management. Based on the generated pollutant heat map, the spatial variation trend of the pollutant concentration is further analyzed. By observing the color distribution and sample point locations in the heat map, the spatial distribution characteristics of the pollutant concentration in the target area can be identified, providing support for further environmental monitoring and control.

[0089] In an optional embodiment, the semivariogram function has the following formula:

[0090] ;

[0091] Among them, γ(h) represents the semivariogram, h represents the distance vector of the spatial position, N(h) represents the total number of samples with distance vector h, n represents the sample ordinal number, and Z(s n ) represents the pollutant concentration value at the corresponding position of sample n, θ represents the rotation angle, R θ A rotation matrix representing the distance vector.

[0092] The formula is a semivariogram, which is used to measure the degree of variation in pollutant concentration between spatial locations. In the formula, the spatial correlation of pollutants is evaluated by comparing the difference in pollutant concentration between a given location and sample points near that location. Specifically:

[0093] The γ(h) in the formula represents the semivariogram of the pollutant, that is, the degree of variation on the distance vector h between given spatial locations;

[0094] The 1 / (2N(h)) part in the formula represents the normalization factor, where N(h) represents the total number of samples within the range of distance vector h;

[0095] The summation part in the formula represents the summation of all samples within the range of the distance vector h;

[0096] In the formula [Z(s n )-Z(s n +R θh)] represents the square of the difference in pollutant concentration between each sample point and the sample point after it is rotated by an angle θ in the direction of the distance vector h. The square of this difference is used to measure the degree of variation of pollutants in space;

[0097] The formula estimates the spatial correlation of pollutants by calculating the average of the squared differences between the pollutant concentrations of all samples within a given distance vector h.

[0098] According to the formula, the semivariogram can well describe the spatial correlation of pollutants. By comparing the differences in pollutant concentrations between different spatial locations, the spatial distribution law and change trend of pollutants can be revealed, providing an important reference for the spatial distribution model of environmental pollution. The rotation angle θ and the rotation matrix R in the formula are θ , as well as the normalization factor in the semivariogram, can be optimized and fitted through actual observation data to better adapt to the diffusion characteristics of pollutants under different environments and improve the accuracy and reliability of the model; the semivariogram value calculated by the formula can be used to analyze the diffusion trend of pollutants in space. By observing the changes in the semivariogram at different distance vectors, the diffusion rate and direction of pollutants can be determined, providing an important reference for environmental monitoring and pollution control.

[0099] S103. Input the pollutant heat map into a pre-trained route tracking model, which is built based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships.

[0100] In an optional embodiment, the route tracking model is constructed based on a graph neural network, and by determining the temporal relationship and spatial relationship, predicting the migration route and diffusion state of pollutants includes:

[0101] The pollutant heat maps at different time steps are used as nodes, and temporal edges are added between nodes adjacent in time, and spatial edges are added between nodes adjacent in space. Based on the migration intensity of pollutants in time and space, the temporal edge weights and spatial edge weights are determined respectively to construct a spatiotemporal relationship diagram.

[0102] The pollutant heat map corresponding to each node is encoded by convolutional neural network, and a high-dimensional feature vector is extracted as the node feature; based on the node feature, the information of neighboring nodes is aggregated through multi-layer iterative update, the position of the node in the spatiotemporal relationship graph and the contextual information of pollutant migration are encoded, and the hidden state of the node is determined;

[0103] The node hidden state is input into the decoder network to determine the future pollutant heat map of the future time step. The future pollutant heat map is used as input, and the future pollutant heat map of multiple consecutive time steps is determined through recursive repetition to form a pollutant migration route. In combination with the spatial change trend, the pollutant diffusion speed and direction are determined.

[0104] The time series edge specifically refers to the edge between nodes in the constructed time-space relationship graph that connects adjacent time steps. The time series edge is used to represent the migration relationship and evolution trend of pollutants between different time steps. The weight of the time series edge can reflect the migration intensity of pollutants in time, that is, the degree of influence of the distribution of pollutants in a certain time step on the next time step;

[0105] The spatial edge specifically refers to the edge connecting adjacent spatial position nodes in the constructed space-time relationship graph. The spatial edge is used to represent the migration relationship and propagation path of pollutants in space. The weight of the spatial edge can reflect the migration intensity of pollutants in space, that is, the diffusion speed and propagation direction of pollutants between different spatial positions;

[0106] The node hidden state specifically refers to a pollutant heat map corresponding to each node in the spatiotemporal relationship graph, and the node hidden state refers to the node feature vector obtained after encoding by the convolutional neural network. The node feature vector represents the high-dimensional feature information of the pollutant heat map corresponding to the node, including local features and global features. The node hidden state can capture the location of the pollutant in the spatiotemporal relationship graph and the contextual information of the pollutant migration, and encode and characterize the migration route and diffusion trend of the pollutant;

[0107] The pollutant heat maps at different time steps are used as nodes in the spatiotemporal relationship diagram. Each node represents the spatial distribution of pollutants at a certain time step. Time series edges are added between nodes adjacent in time. The time series edges connect the pollutant heat maps at adjacent time steps to represent the migration and change relationship of pollutants in time. Space edges are added between nodes adjacent in space. The space edges connect the pollutant heat maps adjacent in space to represent the migration and diffusion relationship of pollutants in space.

[0108] For time-series edges, the edge weight is determined according to the migration intensity of pollutants in time. The migration intensity can be measured by calculating the rate of change of pollutant concentrations in adjacent time steps. The larger the rate of change, the stronger the migration of pollutants in time, and the higher the corresponding time-series edge weight; for spatial edges, the edge weight is determined according to the migration intensity of pollutants in space. The migration intensity can be measured by calculating the gradient of pollutant concentrations in adjacent spatial positions. The larger the gradient, the stronger the migration of pollutants in space, and the higher the corresponding space edge weight; then a spatiotemporal relationship diagram containing time-series edges and spatial edges is constructed to determine the migration and diffusion relationship of pollutants in time and space;

[0109] For each node in the spatiotemporal relationship graph, a convolutional neural network is used to encode the corresponding pollutant heat map. The pollutant heat map is used as the input of the convolutional neural network. Through multi-layer convolution and pooling operations, the high-dimensional feature vector of the heat map is extracted, and the local and global pattern information of the spatial distribution of pollutants is extracted. The extracted high-dimensional feature vector is used as the feature representation of the node for subsequent graph neural network processing;

[0110] Use graph neural network to encode the spatiotemporal relationship graph, learn the hidden state representation of nodes, initialize the hidden state of each node as the corresponding high-dimensional feature vector, and gradually aggregate the information of neighboring nodes through iterative updates of multi-layer graph neural network to update the hidden state of the node. In each layer of iteration, for each node, according to its own characteristics and the hidden state of neighboring nodes, the aggregation information is calculated through the aggregation function, and the aggregation information is combined with the hidden state of the node itself. Through nonlinear transformation, preferably ReLU activation function, the hidden state of the node is updated; repeat multiple iterative updates so that the hidden state of the node can encode the location information of the node in the spatiotemporal relationship graph and the context information of pollutant migration;

[0111] The hidden state of each node is input into the decoder network to predict the pollutant heat map of the future time step. The decoder network takes the hidden state of the node as input, generates the pollutant heat map of the future time step through multi-layer deconvolution and upsampling operations, and uses the predicted future pollutant heat map as the input of the next time step. Through recursive repetition, the future pollutant heat map of multiple consecutive time steps is predicted. The predicted series of future pollutant heat maps are connected in chronological order to form the migration route of pollutants, and the spatial distribution evolution process of pollutants in future time steps is determined;

[0112] Based on the predicted pollutant migration route, the spatial variation trend of pollutants is analyzed, and the diffusion speed and direction of pollutants are estimated. For each time step in the migration route, the spatial gradient vector of pollutant concentration is calculated. The gradient vector indicates the direction in which the pollutant concentration changes fastest in space. The gradient vectors of adjacent time steps are compared, and the temporal change rate of pollutant concentration is calculated. The larger the change rate, the faster the pollutant diffuses in the corresponding direction. The spatial and temporal variation trends of pollutant concentrations are comprehensively analyzed to estimate the diffusion speed and direction of pollutants. The diffusion speed indicates the migration distance of pollutants per unit time, and the diffusion direction indicates the main trend of pollutant migration.

[0113] In this embodiment, by constructing a spatiotemporal relationship graph, the migration and diffusion relationship of pollutants in time and space is effectively described. The establishment of temporal edges and spatial edges can accurately reflect the migration intensity and correlation degree of pollutants between different time steps and spatial positions; the convolutional neural network is used to encode the pollutant heat map corresponding to each node, and a high-dimensional feature vector is extracted as the node feature, which effectively represents the location information and migration context information of the pollutant in the spatiotemporal relationship graph; the spatiotemporal relationship graph is encoded using a graph neural network, and the hidden state representation of the node is learned. Through multi-layer iterative updates and aggregation of neighbor node information, the evolution law and diffusion trend of pollutants in the spatiotemporal relationship graph can be effectively captured; the hidden state of the node is input into the decoder network, and the pollutant heat map of the future time step can be predicted. By recursively and repeatedly predicting the future pollutant heat map of multiple consecutive time steps, the migration route of pollutants is constructed, and the spatial distribution evolution process of pollutants in the future time step is effectively analyzed; based on the predicted pollutant migration route, the spatial change trend of pollutants can be analyzed, the diffusion speed and direction of pollutants can be estimated, and the diffusion of pollutants in space and time can be accurately evaluated by calculating the gradient vector and the rate of change, providing an important reference for pollutant control and management.

[0114] In an optional embodiment, it also includes:

[0115] Based on the preset pollutant types and the preset environmental impact factors, a map entity is constructed, based on the physical mechanism of migration and diffusion and the geographical impact factors, a map relationship is constructed, and based on the map entity and the map relationship, a pollutant diffusion map is constructed;

[0116] Embed the pollutant diffusion map into the spatiotemporal relationship graph to generate diffusion knowledge fusion nodes, and determine the hidden state of the knowledge fusion nodes through multi-level iterative updates;

[0117] The decoder network, combined with the pollutant diffusion map, constructs a diffusion knowledge decoder network, inputs the hidden state of the knowledge fusion node into the diffusion knowledge decoder network, and generates a pollutant migration route that conforms to the diffusion knowledge through recursive iteration.

[0118] Based on the pre-set pollutant types, map entities are constructed. Each pollutant type is an entity node in the map, representing different types of pollutants. Based on the pre-set environmental impact factors, map entities are constructed. Environmental impact factors may include temperature, humidity, wind speed, terrain, soil type, etc. Environmental impact factors are entity nodes in the map, representing environmental conditions that affect the diffusion of pollutants; based on the physical mechanism of migration and diffusion, map relationships are constructed. The physical mechanism of migration and diffusion describes the migration and diffusion laws of pollutants in different media, such as convection, diffusion, adsorption, etc. The physical mechanism is used as the relationship edge in the map to connect related pollutant entities and environmental impact factor entities; based on geographical impact factors, map relationships are constructed. Geographical impact factors include spatial location, topography, hydrological conditions, etc. Geographical impact factors have an important influence on the migration and diffusion of pollutants; Geographical impact factors are used as relationship edges in the map to connect related pollutant entities and environmental impact factor entities; a pollutant diffusion map is constructed, which includes entity nodes such as pollutant types and environmental impact factors, as well as relationship edges such as migration and diffusion physical mechanisms and geographical impact factors, providing key factors and influencing mechanisms in the process of pollutant diffusion;

[0119] The constructed pollutant diffusion map is embedded into the spatiotemporal relationship graph to generate a diffusion knowledge fusion node. For each node in the spatiotemporal relationship graph, it is fused with the corresponding entity node in the pollutant diffusion map. The fusion process can be achieved through element-level weighted averaging; the fused node is called a diffusion knowledge fusion node, which contains both the pollutant concentration information in the spatiotemporal relationship graph and the prior knowledge information in the pollutant diffusion map;

[0120] Through multi-level iterative updates, the hidden state of the knowledge fusion node is determined; in each layer of iteration, for each knowledge fusion node, the neighbor nodes in the spatiotemporal relationship graph and the connected entity nodes in the pollutant diffusion map are combined, and the characteristics of the diffusion knowledge fusion node are aggregated with the hidden state of the neighbor nodes and the characteristics of the connected entity nodes, and the aggregate information is obtained through the convolution calculation of the aggregation function; the aggregate information is combined with the hidden state of the diffusion knowledge fusion node itself, and the hidden state of the diffusion knowledge fusion node is updated through nonlinear transformation; the iterative update is repeated multiple times, so that the hidden state of the knowledge fusion node can simultaneously encode the information in the spatiotemporal relationship graph and the pollutant diffusion map;

[0121] Based on the pollutant diffusion map, a diffusion knowledge decoder network is constructed. The diffusion knowledge decoder network introduces the prior knowledge in the pollutant diffusion map on the basis of the original decoder network. For each diffusion knowledge fusion node, its hidden state is fused with the features of the corresponding entity node in the pollutant diffusion map as the input of the diffusion knowledge decoder network. The diffusion knowledge decoder network decodes the hidden state of the knowledge fusion node into the pollutant heat map of the future time step through multi-layer deconvolution and upsampling operations. In the decoding process, the diffusion knowledge decoder network uses the prior knowledge in the pollutant diffusion map to constrain and guide the migration and diffusion of pollutants. For example, according to the physical mechanism of migration and diffusion and the geographical influencing factors, the parameters and structure of the decoder network are adjusted to make the generated pollutant heat map conform to the diffusion knowledge.

[0122] The hidden state of the diffusion knowledge fusion node is input into the diffusion knowledge decoder network, and the pollutant migration route that conforms to the diffusion knowledge is generated through recursive iteration. In each iteration, the diffusion knowledge decoder network predicts the pollutant heat map of the next time step according to the hidden state of the knowledge fusion node of the current time step, and uses the predicted pollutant heat map as the input of the next time step to update the hidden state of the knowledge fusion node, and continue to predict the pollutant heat map of the subsequent time step. Through recursive iteration, a series of pollutant heat maps of consecutive time steps are generated to form the pollutant migration route. The generated pollutant migration route not only considers the changes in pollutant concentration in the spatiotemporal relationship diagram, but also integrates the prior knowledge in the pollutant diffusion map, so that the migration route is more in line with the actual migration and diffusion laws.

[0123] In this embodiment, by constructing a pollutant diffusion map, prior knowledge such as environmental influencing factors, physical mechanisms of migration and diffusion, and geographical influencing factors are introduced into the model, which can comprehensively consider the impact of environmental factors on pollutant diffusion and make the model closer to the actual situation; the relationship between different entity nodes such as pollutant type, environmental influencing factors, and physical mechanisms of migration and diffusion are modeled by using graph relationships, which can more accurately describe the migration process of pollutants in time and space; the constructed pollutant diffusion map is embedded in the spatiotemporal relationship graph to generate a diffusion knowledge fusion node, which can simultaneously fuse the information in the spatiotemporal relationship graph and the pollutant diffusion map, so that the model has stronger knowledge representation and reasoning capabilities; in the diffusion knowledge decoder network, the prior knowledge in the pollutant diffusion map is introduced to constrain and guide the migration and diffusion of pollutants, which can better ensure that the generated pollutant heat map conforms to the actual migration and diffusion laws, and improve the prediction accuracy and interpretability of the model.

[0124] In an optional embodiment, constructing a spatiotemporal relationship graph includes:

[0125] The temporal edge weight and the spatial edge weight are formulated as follows:

[0126] ;

[0127] Among them, ω l ij represents the spatial edge weight, l i represents the spatial position coordinates of node i, l j represents the spatial position coordinates of node j, σ l represents the spatial scale parameter, ω t ij represents the temporal edge weight, t i represents the time step corresponding to node i, t j represents the time step corresponding to node j, σ t represents the time scale parameter;

[0128] The formula of the high-dimensional feature vector is as follows:

[0129] ;

[0130] Among them, x t represents the high-dimensional feature vector at time step t, ReLU(·) represents the activation function, and W out represents the weight parameter of the fully connected layer, H t represents the pollutant heat map at time step t, MaxPool(·) represents the maximum pooling operation, and W conv represents the convolution kernel, and Conv(·) represents the convolution operation performed on the pollutant heat map using the convolution kernel.

[0131] According to the formula, the calculation of spatial edge weights takes into account the spatial position relationship between nodes, and uses the Gaussian kernel function to weight the spatial distance, which can better reflect the spatial propagation law of pollutants; the calculation of temporal edge weights takes into account the temporal relationship between nodes, and uses the Gaussian kernel function to weight the difference between time steps, which can better capture the temporal change trend of pollutants; the calculation of high-dimensional feature vectors uses convolutional neural networks to extract features from pollutant heat maps, which can extract rich spatial and temporal information from pollutant heat maps; the maximum pooling operation can downsample the feature map and retain the most significant features, which helps to reduce feature dimensions and computational complexity; the combination of fully connected layers and ReLU activation functions can perform nonlinear transformations on the extracted features, enhance the representation ability of the features, and thus better describe the spatial and temporal characteristics of pollutants.

[0132] Figure 2 FIG. 1 is a schematic diagram of the structure of a soil pollutant identification and route tracking system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0133] The first unit is used to collect soil samples in the target area, obtain sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; input the sample spectral data and the sample chemical composition into a pre-trained pollutant identification model, the pollutant identification model is constructed based on a convolutional neural network, extracts sample component characteristics according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and sample pollutant concentration by classifying the sample component characteristics;

[0134] A second unit is used to construct a pollutant heat map based on the sample pollutant type and the sample pollutant concentration in combination with the sample location information;

[0135] The third unit is used to input the pollutant heat map into a pre-trained route tracking model, where the route tracking model is built based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships.

[0136] According to a third aspect of the embodiments of the present invention,

[0137] An electronic device is provided, comprising:

[0138] processor;

[0139] a memory for storing processor-executable instructions;

[0140] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0141] A fourth aspect of the embodiments of the present invention is:

[0142] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0143] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A soil pollutant identification and route tracking method based on artificial intelligence, characterized in that: include: Collect soil samples from the target area, and obtain sample location information, sample spectral data, and sample chemical composition corresponding to the soil samples; The sample spectral data and the sample chemical composition are input into a pre-trained pollutant identification model, wherein the pollutant identification model is constructed based on a convolutional neural network, extracts sample component features according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and the sample pollutant concentration by classifying the sample component features; Based on the sample pollutant type and the sample pollutant concentration, combined with the sample location information, construct a pollutant heat map; Inputting the pollutant heat map into a pre-trained route tracking model, wherein the route tracking model is constructed based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships; The route tracking model is built based on a graph neural network. By determining the temporal and spatial relationships, the migration routes and diffusion states of pollutants are predicted, including: The pollutant heat maps at different time steps are used as nodes, and temporal edges are added between nodes adjacent in time, and spatial edges are added between nodes adjacent in space. Based on the migration intensity of pollutants in time and space, the temporal edge weights and spatial edge weights are determined respectively to construct a spatiotemporal relationship diagram. The pollutant heat map corresponding to each node is encoded by convolutional neural network, and a high-dimensional feature vector is extracted as the node feature; based on the node feature, the information of neighboring nodes is aggregated through multi-layer iterative update, the position of the node in the spatiotemporal relationship graph and the contextual information of pollutant migration are encoded, and the hidden state of the node is determined; The node hidden state is input into the decoder network to determine the future pollutant heat map of the future time step. The future pollutant heat map is used as input, and the future pollutant heat map of multiple consecutive time steps is determined through recursive repetition to form a pollutant migration route. In combination with the spatial change trend, the pollutant diffusion speed and direction are determined.

2. The method according to claim 1, characterized in that According to the sample spectral data and sample chemical composition, the sample component characteristics are extracted, and the sample pollutant type and sample pollutant concentration are identified by classifying the sample component characteristics, including: The sample spectrum data and the sample chemical composition data are combined into a two-dimensional matrix to determine the model input matrix; Performing feature extraction on the input matrix, extracting local features and global features of the sample spectral data through the convolution layer and the pooling layer of the pollutant identification model to form a sample spectral feature, extracting the sample chemical composition feature from the sample chemical composition data through the fully connected layer, and fusing the sample spectral feature with the sample chemical composition feature to generate a comprehensive feature vector; Two subnetworks are set at the output layer of the pollutant identification model, including a type classification subnetwork and a concentration regression subnetwork; wherein the type classification subnetwork uses a softmax activation function to map the comprehensive feature vector to the probability distribution of the pollutant type; and the concentration regression subnetwork uses a linear activation function to map the comprehensive feature vector to the pollutant concentration value; A multi-task learning strategy is used to jointly optimize the type classification task and the concentration regression task, a multi-task loss function is constructed, and the weight parameters of the pollutant identification model are updated through the back-propagation algorithm until the preset convergence conditions are reached; Based on the trained soil pollutant identification model, the sample pollutant type and sample pollutant concentration of the soil samples in the target area are determined.

3. The method according to claim 1, characterized in that Based on the sample pollutant type and the sample pollutant concentration, combined with the sample location information, constructing a pollutant heat map includes: Based on the sample location information and the sample pollutant concentration, a spatial interpolation method is used to solve the Kriging weight by minimizing the estimated variance, and the unit pollutant concentration at any location in the target area is determined by combining the semivariogram and the spatial location relationship, and a grid concentration representation is constructed; The grid concentration representation is converted into a raster image, and combined with color coding, the unit pollutant concentration is mapped to a continuous color scale to determine the spatial variation of the pollution degree, and the pollutant heat map is constructed by combining the sample location information and the sample pollutant concentration.

4. The method according to claim 3, characterized in that: The semivariogram function has the following formula: Among them, γ(h) represents the semivariogram, h represents the distance vector of the spatial position, N(h) represents the total number of samples with distance vector h, n represents the sample ordinal number, and Z(s n ) represents the pollutant concentration value at the corresponding position of sample n, θ represents the rotation angle, R θ A rotation matrix representing the distance vector.

5. The method according to claim 1, characterized in that Also includes: Based on the preset pollutant types and the preset environmental impact factors, a map entity is constructed, based on the physical mechanism of migration and diffusion and the geographical impact factors, a map relationship is constructed, and based on the map entity and the map relationship, a pollutant diffusion map is constructed; Embed the pollutant diffusion map into the spatiotemporal relationship graph to generate diffusion knowledge fusion nodes, and determine the hidden state of the knowledge fusion nodes through multi-level iterative updates; The decoder network, combined with the pollutant diffusion map, constructs a diffusion knowledge decoder network, inputs the hidden state of the knowledge fusion node into the diffusion knowledge decoder network, and generates a pollutant migration route that conforms to the diffusion knowledge through recursive iteration.

6. The method according to claim 1, characterized in that Constructing a spatiotemporal relationship graph includes: The temporal edge weight and the spatial edge weight are formulated as follows: Among them, ω l ij represents the spatial edge weight, l i represents the spatial position coordinates of node i, l j represents the spatial position coordinates of node j, σ l represents the spatial scale parameter, ω t ij represents the temporal edge weight, t i represents the time step corresponding to node i, t j represents the time step corresponding to node j, σ t represents the time scale parameter; The formula of the high-dimensional feature vector is as follows: x t =ReLU(W out *MaxPool(Conv(H t ,IN conv ))); Among them, x t represents the high-dimensional feature vector at time step t, ReLU(·) represents the activation function, and W out represents the weight parameter of the fully connected layer, H t represents the pollutant heat map at time step t, MaxPool(·) represents the maximum pooling operation, and W conv represents the convolution kernel, and Conv(·) represents the convolution operation performed on the pollutant heat map using the convolution kernel.

7. A soil pollutant identification and route tracking system based on artificial intelligence, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect soil samples in the target area and obtain sample location information, sample spectral data and sample chemical composition corresponding to the soil samples; The sample spectral data and the sample chemical composition are input into a pre-trained pollutant identification model, wherein the pollutant identification model is constructed based on a convolutional neural network, extracts sample component features according to the sample spectral data and the sample chemical composition, and identifies the sample pollutant type and the sample pollutant concentration by classifying the sample component features; A second unit is used to construct a pollutant heat map based on the sample pollutant type and the sample pollutant concentration in combination with the sample location information; The third unit is used to input the pollutant heat map into a pre-trained route tracking model, where the route tracking model is built based on a graph neural network and predicts the migration route and diffusion state of pollutants by constructing temporal and spatial relationships.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images