A red tide anomaly prediction method and system based on dual-channel spatiotemporal graph neural network

Abstract

The present invention relates to the field of marine red tide prediction technology, and in particular to a method and system for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network. The method comprises converting the preprocessed marine environment monitoring data into a graph structure, in which the nodes represent the marine monitoring stations; constructing an adjacency matrix based on the Pearson correlation coefficient, and the elements of the adjacency matrix represent the spatial correlation between the nodes; extracting spatial and temporal features from the graph structure using a spatiotemporal graph neural network; the present invention combines the deep feature extraction capabilities of the spatiotemporal graph neural network with remote sensing image data, and realizes accurate prediction of marine red tide anomalies through graph convolution and convolutional neural network (CNN) multi-layer operations.

Description

Technical Field

[0001] The present invention relates to the technical field of marine red tide prediction, and in particular to a method and system for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network. Background Art

[0002] Marine red tide is a natural phenomenon caused by the large-scale reproduction of certain phytoplankton, which usually leads to a series of environmental problems such as changes in seawater color, decreased dissolved oxygen, and eutrophication of water bodies. In severe cases, it may even harm marine ecosystems and fishery resources. The occurrence of red tide not only has a direct impact on the ecological environment, but may also pose a threat to human health and economic activities. Therefore, how to accurately and real-time predict the occurrence and development trend of marine red tide is an important issue in current marine environmental monitoring and management.

[0003] The formation of marine red tides is closely related to a variety of factors, including water temperature, salinity, nutrient concentration, current speed, wind direction and other environmental factors. These factors usually have spatiotemporal dependence, that is, they not only change over time, but are also affected by spatial location. Traditional detection methods often focus on the analysis of a single factor, lacking the comprehensive use of multi-source data and comprehensive modeling of spatiotemporal changes. The spatiotemporal graph neural network combines the powerful graph structure modeling capabilities of the graph neural network and the spatiotemporal characteristics of time series data, and can effectively process graph data with spatiotemporal dependence.

[0004] Traditional methods for detecting marine red tides focus on the analysis of single environmental factors, such as water temperature, salinity, nutrient concentration, etc., but there are complex dependencies between these factors in space and time. Traditional statistical methods and machine learning methods often find it difficult to fully consider the spatiotemporal changes in the occurrence of marine red tides. In addition, the prediction of marine red tides not only depends on the data from environmental monitoring stations, but also needs to consider image data such as remote sensing images and satellite images. However, many traditional prediction methods do not make full use of these heterogeneous data. In traditional methods, image data and environmental data are usually processed separately, lacking an effective fusion mechanism. This single data processing method may lead to information loss or inefficient use, making it difficult to obtain comprehensive and accurate prediction results. Currently, there are few methods that can accurately and efficiently identify marine red tides.

[0005] Most existing spatiotemporal data prediction methods directly input the original spatiotemporal data into the model for training and prediction. However, experimental studies have found that this method usually faces the problems of spatiotemporal data sparsity and noise interference, which significantly affects the prediction accuracy of the model. How to preprocess the marine environmental monitoring data is a key technical issue. Secondly, spatiotemporal graph neural networks usually rely on graph convolution operations on graph structures and process the time series data of nodes through time series modeling modules. For marine environmental monitoring data, how to reasonably model to capture the spatial and temporal dependencies between these points is a key technical issue. In addition, current spatiotemporal graph neural network models usually focus on extracting features from a single data source (such as environmental monitoring data or remote sensing image data), ignoring the collaborative use of multimodal data. In the task of marine red tide prediction, environmental monitoring data (such as water temperature, salinity, nutrient concentration, etc.) and remote sensing image data (algae density, etc.) have different characteristics, and the processing of a single data source cannot fully capture the formation and expansion process of marine red tides. Therefore, how to effectively integrate heterogeneous data sources in spatiotemporal graph neural networks and improve the prediction accuracy of the model while maintaining high efficiency is a key technical problem to be solved. Summary of the invention

[0006] In order to solve the above-mentioned problems, the present invention provides a red tide anomaly prediction method and system based on a dual-channel spatiotemporal graph neural network.

[0007] In the first aspect, the present invention provides a red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network, which adopts the following technical solution:

[0008] A red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network, comprising:

[0009] Obtain marine environment monitoring data and remote sensing image data;

[0010] Preprocess the marine environment monitoring data and remote sensing image data respectively;

[0011] The preprocessed marine environment monitoring data is converted into a graph structure, in which the nodes represent the marine monitoring stations; and an adjacency matrix is ​​constructed based on the Pearson correlation coefficient, and the elements of the adjacency matrix represent the spatial correlation between the nodes;

[0012] The spatial and temporal features of the graph structure are extracted using a spatiotemporal graph neural network. The Chebyshev polynomial is used to perform graph convolution operations to extract spatial features, and the gated recurrent unit GRU is used to extract temporal features.

[0013] Convolutional neural network (CNN) is used to extract spatial features and texture information of remote sensing image data, where multi-layer convolution operations are used to capture the key features of remote sensing images that reflect abnormal marine red tides.

[0014] The graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data are fused and input into the multi-layer perceptron MLP for predictive analysis to achieve accurate prediction of marine red tide anomalies.

[0015] Furthermore, the marine environment monitoring data and remote sensing image data are preprocessed, including eliminating outliers and applying a linear interpolation method to ensure the accuracy and integrity of the data, which is expressed as:

[0016]

[0017] in, x k filled Represents at a point in time t k The interpolation result at x k+1 and x k-1 Distribution represents time points t k+1 , t k-1 The sensor measurement value.

[0018] Furthermore, the preprocessing of the marine environment monitoring data and remote sensing image data also includes selecting a size of k × k The coordinates of the center pixel of the window are (0, 0). The window slides gradually on the image, and the pixel values ​​contained in the window at each position are weighted. The calculation formula of Gaussian weight is:

[0019]

[0020] in, G ( i , j ) is the Gaussian weight, ( i , j ) is the offset of the center pixel in the window, and σ is the standard deviation, which controls the width of the Gaussian distribution. I ( x , y ), using Gaussian weights G ( i , j ) is weighted averaged to obtain the enhanced image; it is expressed as:

[0021]

[0022] in, Ienhanced ( x , y ) is the pixel value of the processed image, I ( x + i , y + j ) is the value of each pixel in the window in the original image, k Represents the window size.

[0023] Furthermore, constructing the adjacency matrix based on the Pearson correlation coefficient includes using the Pearson correlation coefficient to calculate the spatial correlation between the monitoring stations and establish the adjacency matrix. A t Each element a in t ij The value of is:

[0024]

[0025] in, and Site v i and v j Data in the time period t The time series on and They respectively represent and The average value of .

[0026] Furthermore, the use of Chebyshev polynomials to perform graph convolution operations to extract spatial features includes using Chebyshev polynomials to approximate graph convolutions, and graph convolutions learn node representations by iteratively aggregating information about neighboring nodes in the graph. , the node at time T represents It is expressed as:

[0027]

[0028] in, θ is a learnable parameter, K is the order of the polynomial, yes k Chebyshev polynomial of order, X T Represents the original features at time T.

[0029] Furthermore, the extraction of temporal features by combining the gated recurrent unit GRU includes extracting spatial features As the input of GRU X t , GRU receives the input features of the current time stepX t and the hidden state at the previous moment H t-1 ( H 0=0), and determines how information flows through a gating mechanism, using the update gate z t Determine the degree of integration between the current hidden state and the previous hidden state:

[0030]

[0031] Where σ is the sigmoid activation function, W z is the weight matrix;

[0032] Reset Gate r t Control the current input and the previous hidden state H t-1 Relevance:

[0033]

[0034] in, W r is the weight matrix. By resetting the gate r t Acting on the hidden state of the previous moment H t-1 and the input of the current time step X t , calculate the candidate hidden state at the current moment :

[0035]

[0036] Among them, tanh is the hyperbolic tangent function, which is used to ensure that the range of candidate hidden states is between [-1, 1].

[0037] The final temporal feature is updated by the gate z t Weighted average of controls:

[0038] .

[0039] Furthermore, the key features reflecting the abnormal marine red tide in the remote sensing image are captured by multi-layer convolution operations, including using the convolution kernel to gradually slide from the upper left corner on the image to cover the entire image. Each time the convolution kernel slides, the convolution kernel covers a k×k area in the image; for each position, each element of the convolution kernel is multiplied by the corresponding image area element, and the product results are added; the weighted sum result is used as the pixel value of the feature map at the position; the convolution operation is expressed as:

[0040]

[0041] in, f is the input image (in matrix form), g is the convolution kernel (filter), ( x , y ) is the pixel position in the output feature map, m and n is the index of the convolution kernel, indicating the local area of ​​the input image considered in the convolution operation.

[0042] In the second aspect, a red tide anomaly prediction system based on a dual-channel spatiotemporal graph neural network comprises:

[0043] A data acquisition module is configured to acquire marine environment monitoring data and remote sensing image data;

[0044] The preprocessing module is configured to preprocess the marine environment monitoring data and the remote sensing image data respectively;

[0045] A graph construction module is configured to convert the preprocessed marine environment monitoring data into a graph structure, wherein nodes represent marine monitoring stations; and construct an adjacency matrix based on a Pearson correlation coefficient, wherein elements of the adjacency matrix represent spatial correlations between nodes;

[0046] The feature extraction module is configured to extract spatial and temporal features of the graph structure using a spatiotemporal graph neural network; wherein the spatial features are extracted by using a graph convolution operation with Chebyshev polynomials, and the temporal features are extracted by combining a gated recurrent unit (GRU);

[0047] The image extraction module is configured to extract spatial features and texture information of remote sensing image data using a convolutional neural network (CNN), wherein key features reflecting abnormal marine red tides in remote sensing images are captured through multi-layer convolution operations;

[0048] The prediction module is configured to fuse the graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data, and input them into the multi-layer perceptron MLP for prediction analysis to achieve accurate prediction of marine red tide anomalies.

[0049] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network.

[0050] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded by the processor and executing the method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network.

[0051] In summary, the present invention has the following beneficial technical effects:

[0052] (1) This paper proposes a marine red tide anomaly prediction system based on a dual-channel spatiotemporal graph neural network, which combines the deep feature extraction capabilities of the spatiotemporal graph neural network and remote sensing image data, and achieves accurate prediction of marine red tide anomalies through graph convolution and convolutional neural network (CNN) multi-layer operations. This method effectively integrates spatiotemporal data and remote sensing image data from ocean monitoring stations, and can accurately capture the key features reflecting marine red tide anomalies.

[0053] (2) This invention innovatively integrates the spatiotemporal graph neural network with the dual channels of remote sensing image data, enabling the system to comprehensively consider the dynamic characteristics of space and time and extract spatial texture information in the image at the same time. This dual-channel design enhances the model's ability to predict complex marine red tide anomalies, especially in the process of processing marine environmental monitoring data, greatly improving the accuracy and efficiency of anomaly detection.

[0054] (3) By adopting advanced technologies such as Chebyshev polynomials and gated recurrent units (GRU), this system solves the limitations of traditional methods in extracting spatiotemporal features and can more accurately capture the spatiotemporal variation patterns related to red tide anomalies in the marine environment. At the same time, based on the deep fusion of image features and graph features, the system has higher robustness and adaptability when facing complex marine red tide prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of a red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network according to Example 1 of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0057] Example 1

[0058] Reference Figure 1 , a red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network in this embodiment includes:

[0059] Obtain marine environment monitoring data and remote sensing image data;

[0060] Preprocess the marine environment monitoring data and remote sensing image data respectively;

[0061] The preprocessed marine environment monitoring data is converted into a graph structure, in which the nodes represent the marine monitoring stations; and an adjacency matrix is ​​constructed based on the Pearson correlation coefficient, and the elements of the adjacency matrix represent the spatial correlation between the nodes;

[0062] The spatial and temporal features of the graph structure are extracted using a spatiotemporal graph neural network. The Chebyshev polynomial is used to perform graph convolution operations to extract spatial features, and the gated recurrent unit GRU is used to extract temporal features.

[0063] Convolutional neural network (CNN) is used to extract spatial features and texture information of remote sensing image data, where multi-layer convolution operations are used to capture the key features of remote sensing images that reflect abnormal marine red tides.

[0064] The graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data are fused and input into the multi-layer perceptron MLP for predictive analysis to achieve accurate prediction of marine red tide anomalies.

[0065] Specifically,

[0066] The step S1 specifically includes:

[0067] S1.1: Marine environmental monitoring data includes sensor data of key physical quantities such as temperature, salinity, dissolved oxygen, etc. In data processing, we eliminate outliers and use linear interpolation technology to ensure the accuracy and integrity of the data:

[0068]

[0069] in, x k filled Represents at a point in time t k The interpolation result at x k+1 and x k-1 Distribution represents time points t k+1 , t k-1 The sensor measurement value.

[0070] S1.2: In order to eliminate the deviation caused by the dimension difference of different features, we standardized the data to make it follow the standard normal distribution with mean 0 and variance 1. In addition, in order to improve the efficiency of model training, we further normalized these standardized data to the interval [-1, 1].

[0071] S1.3: Select a size of k ×k The coordinates of the center pixel of the window are (0, 0), and the window slides gradually on the image. The pixel values ​​contained in the window at each position are weighted, and the calculation formula of Gaussian weight is:

[0072]

[0073] in, G ( i , j ) is the Gaussian weight, ( i , j ) is the offset of the center pixel in the window, and σ is the standard deviation, which controls the width of the Gaussian distribution. I ( x , y ), using Gaussian weights G ( i , j ) to perform weighted averaging to obtain the enhanced image. The specific operation is:

[0074]

[0075] in, I enhanced ( x , y ) is the pixel value of the processed image, I ( x + i , y + j ) is the value of each pixel in the window in the original image, k Represents the window size.

[0076] The step S2 specifically includes:

[0077] S2.1: Treat each monitoring station as a node in a graph v i , the node set is V={ v 1, v 2,…, v M}, M Represents the number of monitoring stations, each node v i With corresponding attribute characteristics X t i , represents the environmental monitoring data of each monitoring station at a certain time point.

[0078] S2.2: Use the Pearson correlation coefficient to calculate the spatial correlation between monitoring stations and establish an adjacency matrix. At Each element a in t ij The value of is:

[0079]

[0080] in, and Site v i and v j Data in the time period t The time series on and They respectively represent and The average value of .

[0081] The step S3 specifically includes:

[0082] In the spatial and temporal feature extraction stage of the spatiotemporal graph neural network, Chebyshev polynomials are used for graph convolution operations to extract spatial features, and gated recurrent units (GRUs) are used to extract temporal features.

[0083] S3.1: Chebyshev polynomials are used to approximate graph convolutions. Graph convolutions learn node representations by iteratively aggregating information from neighboring nodes in the graph. The node at time T represents It can be calculated by the following formula:

[0084]

[0085] in, θ is a learnable parameter, K is the order of the polynomial, yes k Chebyshev polynomial of order, X T Represents the original features at time T.

[0086] S3.2: Extract the spatial features As the input of GRU X t , GRU receives the input features of the current time step X t and the hidden state at the previous moment H t-1 ( H 0=0), and determines how information flows through a gating mechanism. Update gate z t Determine the degree of integration between the current hidden state and the previous hidden state:

[0087]

[0088] Where σ is the sigmoid activation function, W z is the weight matrix.

[0089] Reset Gate r t Control the current input and the previous hidden state H t-1 Relevance:

[0090]

[0091] in, W r is the weight matrix. By resetting the gate r t Acting on the hidden state of the previous moment H t-1 and the input of the current time step X t , calculate the candidate hidden state at the current moment :

[0092]

[0093] Among them, tanh is the hyperbolic tangent function, which is used to ensure that the range of candidate hidden states is between [-1, 1].

[0094] The final temporal feature is updated by the gate z t Weighted average of controls:

[0095]

[0096] After m iterations, the graph features are obtained f graph .

[0097] The step S4 specifically includes:

[0098] For remote sensing image data, convolutional neural network (CNN) is used to extract image feature information.

[0099] S4.1 The convolution kernel starts from the upper left corner of the image and slides gradually to cover the entire image. Each time it slides, the convolution kernel covers a k×k area in the image.

[0100] S4.2 For each position, each element of the convolution kernel is multiplied by the corresponding image area element, and the product results are added. The result of this weighted summation will be used as the pixel value of the feature map at that position. The formula for the convolution operation is:

[0101]

[0102] in, f is the input image (in matrix form), g is the convolution kernel (filter), ( x , y ) is the pixel position in the output feature map, m and n is the index of the convolution kernel, indicating the local area of ​​the input image considered in the convolution operation.

[0103] S4.3 The result of the convolution operation is a feature map matrix. A fully connected layer is used to map the feature map matrix to a one-dimensional feature vector f image .

[0104] The step S5 specifically includes:

[0105] The graph features and image features are concatenated into a fusion vector by column, and the fusion vector is input into a multi-layer perceptron to generate the prediction results of marine red tide anomalies.

[0106] S5.1 Map features f graph and image features f image Concatenate columns into a long vector f fused , forming a unified feature representation. The concatenation formula is as follows:

[0107]

[0108] S5.2 The concatenated features are input into the multi-layer perceptron (MLP) for processing. The trained MLP model is used to perform prediction analysis using the fused features to generate prediction results for abnormal marine red tides. The calculation of each layer can be expressed as:

[0109]

[0110] in, ReLU represents the activation function, W l It is l The weight matrix of the layer, b l It is l The weight of the layer, h l-1 is the output of the previous layer. Initially h 0 = f fused The output layer uses the Sigmoid function as the activation function for binary classification prediction.

[0111] Example 2

[0112] This embodiment provides a red tide anomaly prediction system based on a dual-channel spatiotemporal graph neural network, including:

[0113] A data acquisition module is configured to acquire marine environment monitoring data and remote sensing image data;

[0114] The preprocessing module is configured to preprocess the marine environment monitoring data and the remote sensing image data respectively;

[0115] A graph construction module is configured to convert the preprocessed marine environment monitoring data into a graph structure, wherein nodes represent marine monitoring stations; and construct an adjacency matrix based on a Pearson correlation coefficient, wherein elements of the adjacency matrix represent spatial correlations between nodes;

[0116] The feature extraction module is configured to extract spatial and temporal features of the graph structure using a spatiotemporal graph neural network; wherein the spatial features are extracted by using a graph convolution operation with Chebyshev polynomials, and the temporal features are extracted by combining a gated recurrent unit (GRU);

[0117] The image extraction module is configured to extract spatial features and texture information of remote sensing image data using a convolutional neural network (CNN), wherein key features reflecting abnormal marine red tides in remote sensing images are captured through multi-layer convolution operations;

[0118] The prediction module is configured to fuse the graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data, and input them into the multi-layer perceptron MLP for prediction analysis to achieve accurate prediction of marine red tide anomalies.

[0119] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for a method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network.

[0120] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to describe a red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network.

[0121] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A red tide anomaly prediction method based on a dual-channel spatiotemporal graph neural network, characterized in that: include: Obtain marine environment monitoring data and remote sensing image data; Preprocess the marine environment monitoring data and remote sensing image data respectively; The preprocessed marine environment monitoring data is converted into a graph structure, in which the nodes represent the marine monitoring stations; and an adjacency matrix is ​​constructed based on the Pearson correlation coefficient, and the elements of the adjacency matrix represent the spatial correlation between the nodes; The spatial and temporal features of the graph structure are extracted using a spatiotemporal graph neural network. The Chebyshev polynomial is used to perform graph convolution operations to extract spatial features, and the gated recurrent unit GRU is used to extract temporal features. Convolutional neural network (CNN) is used to extract spatial features and texture information of remote sensing image data, where multi-layer convolution operations are used to capture the key features of remote sensing images that reflect abnormal marine red tides. The graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data are fused and input into the multi-layer perceptron MLP for predictive analysis to achieve accurate prediction of marine red tide anomalies.

2. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 1 is characterized in that: The preprocessing of the marine environment monitoring data and remote sensing image data includes eliminating outliers and applying a linear interpolation method to ensure the accuracy and integrity of the data, which is expressed as: in, x k filled Represents at a point in time t k The interpolation result at x k+1 and x k-1 Distribution represents time points t k+1 , t k-1 The sensor measurement value.

3. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 2 is characterized in that: The preprocessing of the marine environment monitoring data and remote sensing image data also includes selecting a size of k × k The coordinates of the center pixel of the window are (0, 0). The window slides gradually on the image, and the pixel values ​​contained in the window at each position are weighted. The calculation formula of Gaussian weight is: in, G ( i , j ) is the Gaussian weight, ( i , j ) is the offset of the center pixel in the window, σ is the standard deviation, which controls the width of the Gaussian distribution. For each pixel in the window I ( x , y ), using Gaussian weights G ( i , j ) is weighted averaged to obtain the enhanced image; it is expressed as: in, I enhanced ( x , y ) is the pixel value of the processed image, I ( x + i , y + j ) is the value of each pixel in the window in the original image, k Represents the window size.

4. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 3 is characterized in that: The adjacency matrix is ​​constructed based on the Pearson correlation coefficient, including using the Pearson correlation coefficient to calculate the spatial correlation between the monitoring stations, and establishing the adjacency matrix. A t Each element a in t ij The value of is: in, and Site v i and v j Data in the time period t The time series on and They respectively represent and The average value of .

5. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 4 is characterized in that: The method uses Chebyshev polynomials to perform graph convolution operations to extract spatial features, including using Chebyshev polynomials to approximate graph convolutions. Graph convolutions learn node representations by iteratively aggregating information about neighboring nodes in a graph. , the node at time T represents It is expressed as: in, θ is a learnable parameter, K is the order of the polynomial, yes k Chebyshev polynomial of order, X T Represents the original features at time T.

6. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 5 is characterized in that: The combined gated recurrent unit GRU extracts temporal features, including using the extracted spatial features as the input of the GRU X t , GRU receives the input features of the current time step X t and the hidden state at the previous moment H t-1 ( H 0=0), and determines how information flows through a gating mechanism, using the update gate z t Determine the degree of integration between the current hidden state and the previous hidden state: Where σ is the sigmoid activation function, W z is the weight matrix; Reset Gate r t Control the current input and the previous hidden state H t-1 Relevance: in, W r is the weight matrix, by resetting the gate r t Acting on the hidden state of the previous moment H t-1 and the input of the current time step X t , calculate the candidate hidden state at the current moment : Among them, tanh is the hyperbolic tangent function, which is used to ensure that the range of candidate hidden states is between [-1, 1]; The final temporal feature is updated by the gate z t Weighted average of controls: , after m iterations, we get the graph features f graph .

7. The method for predicting red tide anomalies based on a dual-channel spatiotemporal graph neural network according to claim 6 is characterized in that: The key features reflecting abnormal marine red tide in remote sensing images are captured by multi-layer convolution operations, including using a convolution kernel to gradually slide from the upper left corner of the image to cover the entire image, and each time the convolution kernel slides, the convolution kernel covers a k×k area in the image; for each position, each element of the convolution kernel is multiplied by the corresponding image area element, and the product results are added; The result of the weighted sum is used as the pixel value of the feature map at that position; the convolution operation is expressed as: in, f is the input image in matrix form, g is the convolution kernel of the filter, ( x , y ) is the pixel position in the output feature map, m and n is the index of the convolution kernel, indicating the local area of ​​the input image considered in the convolution operation.

8. A red tide anomaly prediction system based on a dual-channel spatiotemporal graph neural network, characterized in that: include: A data acquisition module is configured to acquire marine environment monitoring data and remote sensing image data; The preprocessing module is configured to preprocess the marine environment monitoring data and the remote sensing image data respectively; A graph construction module is configured to convert the preprocessed marine environment monitoring data into a graph structure, wherein nodes represent marine monitoring stations; and construct an adjacency matrix based on a Pearson correlation coefficient, wherein elements of the adjacency matrix represent spatial correlations between nodes; The feature extraction module is configured to extract spatial and temporal features of the graph structure using a spatiotemporal graph neural network; wherein the spatial features are extracted by using a graph convolution operation with Chebyshev polynomials, and the temporal features are extracted by combining a gated recurrent unit (GRU); The image extraction module is configured to extract spatial features and texture information of remote sensing image data using a convolutional neural network (CNN), wherein key features reflecting abnormal marine red tides in remote sensing images are captured through multi-layer convolution operations; The prediction module is configured to fuse the graph features extracted from the spatiotemporal graph neural network and the image features extracted from the remote sensing image data, and input them into the multi-layer perceptron MLP for prediction analysis to achieve accurate prediction of marine red tide anomalies.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images