A Multimodal Fusion-Based Data Correlation Analysis Method for Deep-Sea Environmental Environments Outside the Yangtze Estuary

Abstract

This invention provides a multimodal fusion-based method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary, belonging to the field of deep-sea environmental analysis technology. This invention achieves accurate data registration by establishing a spatiotemporal standardization model for multi-source data and employing an adaptive spatiotemporal kriging interpolation algorithm. It utilizes a sparse coding marine signal separation algorithm to construct an overcomplete dictionary to separate mixed signals and extract pure features. Based on an attention mechanism, a feature extraction network is constructed to automatically learn deep feature representations of physical, chemical, and biological parameters. A matrix rank deficiency detection algorithm is used to automatically supplement feature compensation vectors to ensure feature integrity. A cross-scale attention mechanism based on wavelet transform is constructed to fuse multi-scale features. A Shapley value interpretability evaluation system is established to quantify the importance of correlated features. This invention solves the technical problem of inaccurate correlated feature extraction during the fusion of multi-source heterogeneous marine environmental data.

Description

Technical Field

[0001] This invention belongs to the field of offshore environmental analysis technology, and more specifically, relates to a method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary based on multimodal fusion. Background Technology

[0002] Correlation analysis of deep-sea environmental data is a crucial technological foundation for marine scientific research and marine environmental monitoring. Traditional marine environmental data analysis primarily employs statistical analysis methods based on single data sources or simple multi-data source overlay processing techniques. Mathematical statistical methods such as Pearson correlation coefficient, multiple linear regression, and principal component analysis are used to identify correlations between environmental parameters and are widely applied in various fields, including marine ecosystem assessment, marine disaster early warning, fisheries resource management, and marine pollution monitoring, providing vital technical support for marine scientific research and marine management decision-making. However, traditional multi-source data fusion techniques, when processing marine environmental data from different sources such as satellite remote sensing data, buoy observation data, and ship monitoring data, often use simple linear weighted averaging or statistical interpolation methods for feature extraction. These methods, based on linear and independence assumptions, cannot effectively capture the complex nonlinear correlations and dynamic coupling mechanisms between marine environmental parameters, resulting in a lack of accuracy and representativeness in the extracted correlation features, particularly exhibiting significant limitations when processing multimodal heterogeneous data. In current marine environmental monitoring applications, data acquired from different observation platforms have varying spatiotemporal resolutions, observation accuracies, and error characteristics. Traditional feature extraction methods struggle to accurately identify deep correlation patterns among physical, chemical, and biological parameters, failing to effectively uncover cross-modal collaborative variation patterns and causal driving relationships. This severely impacts the accuracy and reliability of marine environmental data correlation analysis. In other words, existing technologies suffer from inaccurate feature extraction during the fusion of multi-source heterogeneous marine environmental data. Summary of the Invention

[0003] In view of this, the present invention provides a method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary based on multimodal fusion, which can solve the technical problem of inaccurate extraction of correlation features in the process of fusion of multi-source heterogeneous marine environmental data in the prior art.

[0004] This invention is implemented as follows: This invention provides a multimodal fusion-based method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary, comprising: using an adaptive spatiotemporal kriging interpolation algorithm to perform spatiotemporal registration of multi-source data, forming a sparse matrix of environmental parameters; processing the sparse matrix of environmental parameters using a sparse coding-based marine signal separation algorithm, constructing an overcomplete dictionary to learn the sparse representation of marine signals, and using an orthogonal matching pursuit algorithm to solve for sparse coefficients, generating a signal separation coefficient vector; processing the multimodal data and extracting key feature vectors for each modality; establishing a multimodal feature sparse real-valued adjacency matrix; using a matrix rank deficiency detection algorithm to detect the multimodal feature sparse real-valued adjacency matrix, automatically supplementing feature compensation vectors to restore the full-rank characteristic of the matrix when a rank deficiency is detected in the multimodal feature sparse real-valued adjacency matrix; and employing a multimodal graph attention network-based method. The algorithm constructs an ocean parameter correlation map and establishes a parameter correlation sparse filling matrix. A Bayesian optimization-based hyperparameter automatic tuning algorithm optimizes model parameters, utilizes Gaussian processes to model the mapping relationship between hyperparameters and model performance, selects the next evaluation point through a data acquisition function, and generates an optimized hyperparameter configuration vector. Based on wavelet transform decomposition of features at different scales, a cross-scale attention mechanism is constructed, and a multi-scale fusion distortion index is calculated. The processing method is adjusted according to the multi-scale fusion distortion index. A Bayesian network is used for causal reasoning analysis to identify key driving factors, calculate the fusion weight scale index for each modality, and dynamically adjust the weight allocation of each modality's data based on the magnitude of the fusion weight scale index. An interpretive evaluation system based on Shapley values ​​is established, the fusion correlation attention index is calculated, and an interactive three-dimensional correlation heatmap is generated as the analysis result.

[0005] Specifically, the sparse matrix of environmental parameters is a data storage structure with a large number of zero-value elements formed by standardizing multi-source environmental monitoring data with different spatiotemporal resolutions. It is used to efficiently store and process the spatiotemporal distribution information of marine environmental parameters.

[0006] The signal separation coefficient vector is specifically a vector representing the weights of different signal components obtained after separating the ocean observation signal through a sparse coding algorithm. It is used to quantify the contribution of each signal component in the mixed signal.

[0007] The multimodal feature sparse real-valued adjacency matrix is ​​specifically a matrix that describes the connection relationship between different modal features. Non-zero elements represent the strength value of the correlation between features, which is used to quantify the degree of interaction between physical, chemical, and biological parameters.

[0008] Specifically, the feature compensation vector is a supplementary vector automatically generated when the multimodal feature sparse real-valued adjacency matrix has a rank deficiency, which restores the full-rank property of the matrix through linearly independent feature combinations.

[0009] Specifically, the parameter association sparse filling matrix is ​​a sparse representation matrix of the association relationships between marine parameters learned through a graph attention network. A filling algorithm is used to complete the missing association information to fully describe the network structure of marine environmental parameters.

[0010] Specifically, the temporal gradient change matrix is ​​a gradient information matrix that describes the trend of marine environmental parameters changing over time. It is formed by calculating the rate of change of parameters between adjacent time steps and is used to identify the dynamic change patterns of environmental parameters.

[0011] Specifically, the step of dynamically adjusting the signal processing frequency based on the gradient value of the time-series gradient transformation matrix is ​​as follows: when the gradient value of the time-series gradient transformation matrix G∈[0, 0.15], the signal processing frequency is reduced to 60% of the original frequency; when the gradient value of the time-series gradient transformation matrix G∈(0.15, 0.30], the current signal processing frequency is maintained; and when the gradient value of the time-series gradient transformation matrix G∈(0.30, 1.0], the signal processing frequency is increased to 150% of the original frequency.

[0012] The optimized hyperparameter configuration vector is specifically the optimal hyperparameter combination vector of the model obtained by searching through the Bayesian optimization algorithm. It includes key parameters such as learning rate, regularization coefficient, and number of network layers, and is used to improve model performance.

[0013] The step of adjusting the processing method according to the multi-scale fusion distortion index is as follows: when the multi-scale fusion distortion index F∈[0, 0.85), super-resolution reconstruction processing is added; when the multi-scale fusion distortion index F∈[0.85, 1.0], the current processing method is maintained.

[0014] The multi-source data includes satellite remote sensing data, buoy observation data, and ship monitoring data.

[0015] Specifically, the process of processing multimodal data and extracting key feature vectors for each modality involves: constructing a feature extraction network based on an attention mechanism to process physical, chemical, and biological parameter data and extract key feature vectors for each modality.

[0016] The step of establishing the parameter-related sparse filling matrix involves treating each ocean parameter as a graph node, calculating the initial edge weights through mutual information, and dynamically updating the node features using a multi-head attention mechanism to form the parameter-related sparse filling matrix.

[0017] Furthermore, it also includes the steps of introducing an LSTM timing module to capture the dynamic process of Yangtze River diluent expansion and seasonal algal blooms, calculating the timing gradient change matrix, and dynamically adjusting the signal processing frequency based on the gradient value of the timing gradient change matrix.

[0018] The multi-scale fusion distortion index is a quantitative indicator that assesses the degree of information loss during the fusion of data at different spatiotemporal scales. It is calculated by comparing the changes in information entropy of the data before and after fusion. The closer the value is to 1, the more complete the information is.

[0019] Among them, the information content and reliability of each modality data are analyzed and comprehensively calculated to dynamically optimize the fusion weight of multimodal data.

[0020] The fusion association attention index is a quantitative indicator that measures the importance of key association patterns in the multimodal fusion result. It is calculated by analyzing the marginal contribution of each modality to the final association result through Shapley value analysis, and is used to identify and highlight the most important cross-modal association features.

[0021] Specifically, the adaptive spatiotemporal kriging interpolation algorithm is an interpolation method that combines prior knowledge of ocean dynamics and geostatistical theory. It automatically determines the optimal interpolation parameters by fitting a variogram, analyzes the spatiotemporal correlation structure of the observed data, constructs an anisotropic variogram model, optimizes the interpolation weights using cross-validation, and introduces trend surface analysis to eliminate the influence of large-scale gradients, thereby achieving high-precision spatiotemporal registration of multimodal data and imputation of missing values.

[0022] Specifically, the ocean signal separation algorithm based on sparse coding utilizes sparse representation theory to separate different components in ocean observation signals. It constructs an overcomplete dictionary to learn the sparse representation of ocean signals, uses the orthogonal matching pursuit algorithm to solve for sparse coefficients, and optimizes the representation quality through dictionary updates, thereby achieving effective separation and feature extraction of mixed ocean signals.

[0023] Furthermore, it also includes the step of establishing an interpretive evaluation system based on Shapley values, specifically outputting key association pattern recognition results when the fused association attention index A∈[0.92, 1.0].

[0024] This invention constructs a feature extraction network based on an attention mechanism and employs deep learning technology to automatically learn deep feature representations of physical, chemical, and biological parameters. Compared to traditional statistical methods, it can accurately capture complex nonlinear correlation feature patterns, effectively addressing the shortcomings of traditional linear methods in feature extraction accuracy. This invention utilizes a graph attention network to construct a marine parameter correlation map, modeling each marine parameter as a graph node. Through a multi-head attention mechanism, it dynamically learns the correlation weights between nodes, automatically discovering and accurately quantifying complex cross-modal correlation features. Combined with a sparse coding algorithm, it separates mixed observation signals, extracting pure environmental features and avoiding the impact of signal interference on correlation feature extraction, significantly improving the extraction accuracy and representation capability of correlation features in the process of multi-source heterogeneous data fusion. This invention introduces an LSTM time-series module to capture the dynamic changes in the marine environment, combines it with Bayesian networks for causal reasoning analysis, and establishes a Shapley value interpretability evaluation system. This achieves accurate identification and quantification of dynamic correlation features, enabling the precise extraction and effective representation of complex multimodal correlation features. In summary, this invention solves the technical problem of inaccurate correlation feature extraction in the process of multi-source heterogeneous marine environmental data fusion mentioned in the background art. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention.

[0026] Figure 2 This is a diagram showing the spatiotemporal registration results of multi-source data in the embodiment.

[0027] Figure 3 This is a visualization of the marine parameter correlation map in the embodiment.

[0028] Figure 4 This is an interactive three-dimensional correlation heatmap in the embodiment.

[0029] Figure 5 This is a graph showing the time-series prediction results of chlorophyll concentration in the examples.

[0030] Figure 6 This is a vertical distribution feature map of environmental parameters in the embodiment.

[0031] Figure 7 This is a weight distribution diagram of multimodal feature fusion in the embodiment. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0033] like Figure 1The diagram shows a flowchart of the multimodal fusion-based deep-sea environmental data correlation analysis method for the Yangtze River Estuary provided by this invention. This method includes the following steps:

[0034] S01. Establish a multi-source data spatiotemporal standardization model, and use an adaptive spatiotemporal kriging interpolation algorithm to perform spatiotemporal registration of satellite remote sensing data, buoy observation data, and ship monitoring data, construct a unified spatiotemporal reference grid, and form a sparse matrix of environmental parameters.

[0035] S02. The marine signal separation algorithm based on sparse coding processes the sparse matrix of the environmental parameters, constructs a sparse representation of the marine signal by learning an overcomplete dictionary, and uses the orthogonal matching pursuit algorithm to solve for the sparse coefficients to generate a signal separation coefficient vector.

[0036] S03. Construct a feature extraction network based on the attention mechanism, process physical parameters, chemical parameters, and biological parameters data respectively, extract key feature vectors of each modality, and establish a multimodal feature sparse real-valued adjacency matrix.

[0037] S04. The multimodal feature sparse real-valued adjacency matrix is ​​detected by a matrix rank deficiency detection algorithm. When a rank deficiency is detected in the multimodal feature sparse real-valued adjacency matrix, a feature compensation vector is automatically added to restore the full rank property of the matrix.

[0038] S05. A multimodal association mining algorithm based on graph attention network is used to construct an ocean parameter association graph. Each ocean parameter is used as a graph node. The initial edge weight is calculated through mutual information, and the node features are dynamically updated using a multi-head attention mechanism to form a parameter association sparse filling matrix.

[0039] S06. An LSTM timing module is introduced to capture dynamic processes such as the expansion of freshwater in the Yangtze River and seasonal algal blooms. The timing gradient change matrix is ​​calculated. When the gradient value of the timing gradient change matrix G∈[0, 0.15], the signal processing frequency is reduced to 60% of the original frequency. When the gradient value of the timing gradient change matrix G∈(0.15, 0.30], the current signal processing frequency is maintained. When the gradient value of the timing gradient change matrix G∈(0.30, 1.0], the signal processing frequency is increased to 150% of the original frequency.

[0040] S07. The model parameters are optimized using a Bayesian optimization-based hyperparameter automatic tuning algorithm. The mapping relationship between hyperparameters and model performance is modeled using a Gaussian process. The next evaluation point is selected through the acquisition function, and an optimized hyperparameter configuration vector is generated.

[0041] S08. Based on wavelet transform decomposition of features at different scales, construct a cross-scale attention mechanism, calculate the multi-scale fusion distortion index, add super-resolution reconstruction processing when the multi-scale fusion distortion index F∈[0, 0.85), and maintain the current processing method when the multi-scale fusion distortion index F∈[0.85, 1.0].

[0042] S09. Use Bayesian networks for causal reasoning analysis, identify key driving factors, calculate the fusion weight scale index of each modality, and dynamically adjust the weight allocation of each modality based on the magnitude of the fusion weight scale index.

[0043] S10. Establish an interpretability evaluation system based on Shapley values, calculate the fusion association attention index, generate an interactive three-dimensional association heatmap, and output the key association pattern recognition result when the fusion association attention index A∈[0.92, 1.0].

[0044] Among them, the sparse matrix of environmental parameters refers to a data storage structure with a large number of zero-value elements formed by standardizing multi-source environmental monitoring data with different spatiotemporal resolutions. It is used to efficiently store and process the spatiotemporal distribution information of marine environmental parameters.

[0045] The signal separation coefficient vector is a vector representing the weights of different signal components obtained after separating ocean observation signals through a sparse coding algorithm. It is used to quantify the contribution of each signal component in the mixed signal.

[0046] A multimodal feature sparse real-valued adjacency matrix is ​​a matrix that describes the connection relationships between different modal features. Non-zero elements represent the strength of the correlation between features and are used to quantify the degree of interaction between physical, chemical, and biological parameters.

[0047] Feature compensation vectors are supplementary vectors automatically generated when the rank deficit of a multimodal feature sparse real-valued adjacency matrix occurs. They restore the full-rank property of the matrix through linearly independent feature combinations.

[0048] The parameter association sparse filling matrix is ​​a sparse representation matrix of the relationships between marine parameters learned through graph attention networks. It uses a filling algorithm to complete the missing association information and is used to fully describe the network structure of marine environmental parameters.

[0049] The temporal gradient variation matrix is ​​a gradient information matrix that describes the trend of marine environmental parameters changing over time. It is formed by calculating the rate of change of parameters between adjacent time steps and is used to identify dynamic change patterns of environmental parameters.

[0050] The optimized hyperparameter configuration vector refers to the optimal combination vector of hyperparameters of the model obtained by searching through the Bayesian optimization algorithm. It includes key parameters such as learning rate, regularization coefficient, and number of network layers, and is used to improve model performance.

[0051] The multi-scale fusion distortion index is a quantitative indicator that assesses the degree of information loss during the fusion of data at different spatiotemporal scales. It is calculated by comparing the changes in information entropy of the data before and after fusion. The closer the value is to 1, the more complete the information is.

[0052] The fusion weight scale index is a standardized indicator that quantifies the contribution of each modality of data in the fusion process. It is calculated by analyzing the information content and reliability of each modality of data and is used to dynamically optimize the fusion weight of multimodal data.

[0053] The Fusion Association Attention Index is a quantitative indicator that measures the importance of key association patterns in multimodal fusion results. It is calculated by analyzing the marginal contribution of each modality to the final association result through Shapley value analysis, and is used to identify and highlight the most important cross-modal association features.

[0054] Adaptive spatiotemporal kriging interpolation is an interpolation method that combines prior knowledge of ocean dynamics with geostatistical theory. It automatically determines the optimal interpolation parameters by fitting a variogram. First, it analyzes the spatiotemporal correlation structure of the observation data and constructs an anisotropic variogram model. Then, it optimizes the interpolation weights using cross-validation and introduces trend surface analysis to eliminate the influence of large-scale gradients. Finally, it achieves high-precision spatiotemporal registration of multimodal data and imputation of missing values.

[0055] The ocean signal separation algorithm based on sparse coding refers to a method that uses sparse representation theory to separate different components in ocean observation signals. The algorithm constructs a complete dictionary to learn the sparse representation of ocean signals, uses the orthogonal matching pursuit algorithm to solve for the sparse coefficients, and optimizes the representation quality through dictionary updates, ultimately achieving effective separation and feature extraction of mixed ocean signals.

[0056] The matrix rank deficiency detection algorithm is a method to identify matrix rank defects through singular value decomposition. When there are singular values ​​in the matrix that are close to zero, it is judged as rank deficiency. The algorithm automatically calculates the effective rank of the matrix and restores the full rank property by adding linearly independent vectors.

[0057] The multimodal association mining algorithm based on graph attention network refers to a method that constructs an association graph of ocean parameters and uses an attention mechanism to automatically learn the interaction weights between parameters. The algorithm treats each ocean parameter as a graph node, calculates the initial edge weights through mutual information, dynamically updates the node features using a multi-head attention mechanism, and combines a graph convolutional network to propagate neighborhood information, ultimately identifying key cross-modal association patterns and driving factors.

[0058] The Bayesian optimization-based hyperparameter automatic tuning algorithm refers to a method that uses Bayesian optimization to automatically search for the optimal hyperparameter configuration of a marine data analysis model. The algorithm uses Gaussian processes to model the mapping relationship between hyperparameters and model performance, explores and utilizes the acquisition function balance, and selects the next evaluation point using expected improvement or confidence upper bound criteria, avoiding the computational overhead of grid search and efficiently finding the optimal hyperparameter combination.

[0059] Feature extraction networks refer to deep learning models that employ a combination of convolutional neural networks and recurrent neural networks. The structure of such a network includes an input layer that receives 512-dimensional multimodal marine environmental data; a first convolutional layer containing 64 3×3 convolutional kernels; a second convolutional layer containing 128 3×3 convolutional kernels; a third convolutional layer containing 256 3×3 convolutional kernels; each convolutional layer is followed by a ReLU activation function and a batch normalization layer, then a max pooling layer for dimensionality reduction; after a flattening layer, a first fully connected layer containing 512 neurons; a second fully connected layer containing 256 neurons; and an output layer containing 64 neurons that outputs a feature vector.

[0060] The process of establishing the training dataset for the feature extraction network first involves collecting multi-source observational data from the deep-sea area outside the Yangtze River Estuary. This includes sea surface temperature, chlorophyll concentration, and suspended matter concentration data acquired by satellite remote sensing; water temperature, salinity, dissolved oxygen, and pH data acquired by buoy observations; and nutrient salts, phytoplankton abundance, and sediment particle size data acquired by ship monitoring. A total of continuous observational data from 2018 to 2022 was collected. In the data preprocessing stage, quality control was performed on the raw data, removing outliers and missing values. Standardization methods were used to normalize data of different dimensions to the 0-1 range. Data from different sources were aligned according to a spatiotemporal grid to form a standardized multimodal dataset. The data annotation process involved manual annotation by oceanographic experts based on the correlation between physical, chemical, and biological parameters. Classification labels were established for physical, chemical, and biological parameters, ultimately forming a training dataset containing 10,000 annotated samples, divided into training, validation, and test sets in an 8:1:1 ratio.

[0061] The training steps of the feature extraction network first involve data preprocessing. The training dataset is divided into batches of 32, and the network weights are initialized using the Xavier initialization method with a learning rate of 0.001. The Adam optimizer is then used for parameter optimization. During the forward propagation phase, the input multimodal marine environment data is processed through each layer of the network to obtain the predicted output. In the loss calculation phase, the cross-entropy loss function is used to calculate the error between the predicted value and the true label. During the backpropagation phase, the gradient of each layer's parameters is calculated using the chain rule. In the parameter update phase, the network weights are updated based on the calculated gradients. The training process is set to a maximum of 1000 iterations. Training is terminated early when the validation set loss does not decrease for 10 consecutive iterations. Finally, the model weights with the best performance on the validation set are saved.

[0062] The LSTM temporal module refers to a time series processing model based on a Long Short-Term Memory (LSTM) network. The structure of the LSTM temporal module includes an input gate to control the input of new information, a forget gate to control the retention of historical information, an output gate to control the selection of output information, and cell states to store long-term memory information. The LSTM temporal module includes an input layer that receives a 72-hour time series of marine environmental parameters, a first LSTM layer with 128 hidden units, a second LSTM layer with 64 hidden units, a dropout layer after each LSTM layer to prevent overfitting, and a fully connected layer with 32 neurons. The output layer predicts the parameter change trend for the next 24 hours.

[0063] The training dataset creation process for the LSTM time-series module first collects historical marine environmental time-series data from 2017 to 2022 for the deep-sea area outside the Yangtze River Estuary, including hourly observation records of water temperature, salinity, current velocity, current direction, chlorophyll concentration, and dissolved oxygen concentration. During the data cleaning stage, outlier data due to instrument malfunctions and extreme weather conditions are removed, and linear interpolation is used to supplement any missing data, ensuring the continuity of the time-series data. The time-series segmentation process uses a sliding window method to segment the continuous time-series data into an input sequence and a target sequence. The input sequence is 72 hours long, the target sequence is 24 hours long, and the sliding step is 1 hour, ultimately forming a training dataset containing 50,000 time-series samples. Data standardization uses the Z-score method to standardize each parameter, eliminating the influence of differences in the dimensions of different parameters.

[0064] The training steps of the LSTM time series module first involve sequence segmentation, dividing the time series data into batches of 64. During network construction, the He initialization method is used to initialize the LSTM weights with a learning rate of 0.01, and the RMSprop optimizer is used for parameter optimization. In the gradient calculation phase, the gradient at each time step is calculated using the backpropagation algorithm. In the weight optimization phase, the LSTM network weights are updated based on the gradient information. The training process uses the mean squared error loss function to measure prediction accuracy, sets a maximum training epoch of 500 epochs, and employs an early stopping strategy to prevent overfitting; training stops when the validation set loss fails to improve for 15 consecutive epochs.

[0065] Graph Attention Networks (GANs) are deep learning models that combine graph convolutional networks (GANs) and attention mechanisms. The structure of a GAN includes graph convolutional layers for aggregating neighborhood node information, multi-head attention layers for learning the association weights between nodes, and an output layer for generating the final node representation. The GAN includes an input layer that receives the ocean parameter node feature matrix and adjacency matrix; a first graph convolutional layer with 8 attention heads, each containing 64-dimensional features; a second graph convolutional layer with 4 attention heads, each containing 128-dimensional features; multi-head attention layers that use a scaled dot product attention mechanism to calculate attention weights; and finally, the output layer is connected to generate a 256-dimensional node embedding representation.

[0066] The training dataset construction process for the graph attention network first involves building a marine parameter correlation graph. Parameters such as water temperature, salinity, dissolved oxygen, chlorophyll concentration, nutrient concentration, and phytoplankton abundance are used as graph nodes. The correlations between these parameters are determined based on oceanographic theory and statistical analysis results, serving as graph edges. The node feature construction process extracts the statistical, temporal, and spatial distribution features of each parameter as initial feature vectors for each node, with each node having a 128-dimensional feature dimension. Edge weight calculation quantifies the nonlinear correlation strength between parameters using mutual information, forming a weighted adjacency matrix. Graph data augmentation employs node dropout and edge dropout methods to increase data diversity, ultimately resulting in a training dataset containing 1000 graph samples, each containing 50 nodes and 200 edges.

[0067] The training steps of the graph attention network first involve graph construction, converting the associated ocean parameters into a graph structure representation. In the feature initialization phase, the Glorot initialization method is used to initialize the network parameters, setting the learning rate to 0.005, and the AdamW optimizer is used for parameter optimization. In the attention calculation phase, a multi-head attention mechanism is used to calculate the association weights between nodes. In the information aggregation phase, neighboring node information is aggregated according to weights to update the node representation. In the model optimization phase, a contrastive learning loss function is used to train the network, learning effective node representations by maximizing positive sample similarity and minimizing negative sample similarity. The training process sets a maximum of 300 iterations, employing a learning rate decay strategy, decreasing the learning rate by a factor of 0.9 every 50 iterations. The optimal model parameters are saved when the training loss converges.

[0068] The specific implementation methods of the above steps are described in detail below.

[0069] The specific implementation of step S01 involves establishing a multi-source data spatiotemporal standardization model. First, satellite remote sensing data undergoes quality assessment and preprocessing to remove cloud cover and atmospheric noise. Radiometric calibration and geometric correction are used to ensure data accuracy. Next, buoy observation data is sorted according to time series and missing value detection is performed. Linear interpolation and mean replacement methods are used to handle data breakpoints. Then, ship monitoring data undergoes GPS trajectory verification and observation time synchronization to ensure spatial location accuracy. The adaptive spatiotemporal kriging interpolation algorithm automatically fits the optimal interpolation parameters by constructing an anisotropic variogram model. Cross-validation is used to determine the optimal spatial correlation parameters, setting the search radius to 5km and the maximum number of nearest neighbors to 15. The algorithm calculates spatial autocorrelation through the variogram and introduces trend surface analysis to eliminate the influence of large-scale gradients such as ocean circulation. Finally, a 1km × 1km unified spatiotemporal reference grid is constructed, interpolating multi-source data of different resolutions onto the standard grid points to form a model including temperature, salinity, and other parameters. Sparse matrix of environmental parameters such as concentration.

[0070] The specific implementation of step S02 is a marine signal separation algorithm based on sparse coding to process the sparse matrix of environmental parameters. The algorithm constructs an overcomplete dictionary to learn the sparse representation of the marine signal, optimizes the representation quality through dictionary updates, sets the number of dictionary atoms to 512, and the sparsity constraint coefficient to 0.1. The orthogonal matching pursuit algorithm selects the atom most relevant to the residual in each iteration, and ensures the global optimum through Schmitt orthogonalization. The maximum number of iterations is set to 50, and the residual threshold is [value missing]. The algorithm first initializes the residual vector to the original signal, calculates the inner product coefficients of each atom in the dictionary with the residual, and selects the atom with the largest absolute value of the inner product to add to the support set. Then, it orthogonals the atoms in the support set, solves for the sparse coefficients using the least squares method, updates the residual vector, and determines the termination condition. Finally, it generates a signal separation coefficient vector, effectively separating ocean signals of different frequencies, such as tides, internal waves, and turbulence.

[0071] The specific implementation of step S03 is to construct a feature extraction network based on an attention mechanism. The network adopts a multi-channel parallel architecture to process three types of data: physical parameters, chemical parameters, and biological parameters. The physical parameter channel processes data such as temperature, salinity, and flow rate, while the chemical parameter channel processes... Data such as concentrations of phosphorus, dissolved oxygen, and nutrients are processed, while biological parameter channels handle data such as chlorophyll concentration and phytoplankton abundance. Each channel employs a self-attention mechanism to calculate feature weights, with 8 attention heads, a hidden layer dimension of 256, and a dropout rate of 0.2. The algorithm calculates attention weights through query, key, and value matrix transformations, and uses a softmax function for normalization to obtain key feature vectors for each modality. Finally, the feature vectors from the three channels are concatenated and fused, mapped to a 128-dimensional unified feature space through a fully connected layer, and a multimodal feature sparse real-valued adjacency matrix is ​​established. The matrix dimension is 128×128, and the sparsity is approximately 15%.

[0072] The specific implementation of step S04 is to use a matrix rank deficiency detection algorithm. The algorithm calculates the singular values ​​of the multimodal feature sparse real-valued adjacency matrix through singular value decomposition, and sets the rank detection threshold to [value missing]. When a singular value below a threshold exists, the matrix is ​​identified as rank-deficient. First, singular value decomposition (SVD) is performed on the matrix to obtain the left singular vector, the singular value diagonal matrix, and the right singular vector. Then, the number of singular values ​​greater than the threshold is counted to determine the effective rank, and the rank deficit is calculated as the difference between the theoretical full rank and the effective rank. When a rank deficit is detected, the algorithm automatically generates a feature compensation vector. A linearly independent compensation vector is constructed using principal component analysis (PCA) to ensure that the compensation vector is orthogonal to the existing feature vectors. The compensation vector is added to the original matrix, and the SVD is recalculated to verify the recovery of the full-rank property, ensuring that the matrix possesses good numerical stability and invertibility.

[0073] The specific implementation of step S05 involves using a multimodal association mining algorithm based on a graph attention network. The algorithm treats marine parameters as graph nodes, calculates the association weights between nodes through an attention mechanism, and achieves weighted aggregation of neighbors. The number of graph nodes is set to 50, with each node representing a class of marine environmental parameters. The algorithm first calculates initial edge weights using mutual information, setting a mutual information threshold of 0.3, and retaining only strongly correlated parameter pairs. The graph attention network uses a multi-head attention mechanism to dynamically update node features. Each attention head independently calculates association weights, with 4 attention heads and 64-dimensional features for each head. The algorithm calculates attention coefficients using the LeakyReLU activation function, masks non-existent edge connections using a masking mechanism, and employs residual connections and layer normalization to improve training stability. Network training uses a contrastive learning loss function, learning effective node representations through positive and negative sample comparisons, ultimately forming a parameter association sparse padding matrix. An interpolation algorithm is then used to complete the missing association information.

[0074] The specific implementation of step S06 involves introducing an LSTM time series module to capture dynamic processes. The module employs a two-layer LSTM structure, with the first layer containing 128 hidden units and the second layer containing 64 hidden units. The input sequence length is set to 72 hours, and the prediction sequence length is 24 hours. The algorithm first performs standardized preprocessing on time series data such as the Yangtze River diversion and seasonal algal blooms, and then constructs training samples using a sliding window method. The LSTM controls information flow through input gates, forget gates, and output gates, storing long-term memory information in cell states. The algorithm calculates the temporal gradient change matrix, measuring the gradient strength using the L2 norm of the parameter change rate between adjacent time steps. When the gradient value G∈[0, 0.15], the system identifies it as a stationary period, reducing the signal processing frequency to 60% of the original frequency to save computational resources. When the gradient value G∈(0.15, 0.30], the current processing frequency is maintained to ensure normal monitoring. When the gradient value G∈(0.30, 1.0], it is identified as a period of drastic change, increasing the processing frequency to 150% of the original frequency to ensure high-precision capture of dynamic processes.

[0075] The specific implementation of step S07 is to use a Bayesian optimization-based hyperparameter automatic tuning algorithm. The algorithm uses a Gaussian process as a surrogate model to establish a mapping relationship between hyperparameters and model performance. The Gaussian process kernel function is set as a radial basis function, the length scale parameter is 1.0, and the noise variance is 0.01. Optimization objectives include key hyperparameters such as learning rate, batch size, number of network layers, and regularization coefficient. The algorithm uses an expected improvement sampling function to select the next evaluation point, setting the exploration-utilization tradeoff parameter to 0.01 and the maximum number of evaluations to 100. First, initial points are randomly sampled in the hyperparameter space for model training to obtain performance feedback and update the posterior distribution of the Gaussian process. Then, the expected improvement value of each candidate point is calculated using the sampling function, and the point with the largest expected improvement is selected as the next evaluation point. The algorithm iteratively optimizes until convergence or the maximum number of evaluations is reached, generating an optimized hyperparameter configuration vector containing parameters such as the optimal learning rate of 0.001, batch size of 32, network depth of 5 layers, and L2 regularization coefficient of 0.0001.

[0076] The specific implementation of step S08 is based on wavelet transform to decompose features at different scales. Wavelet transform decomposes the signal into wavelet coefficients of different frequencies, and multi-scale refinement is achieved through scaling and translation operations. Daubechies wavelet is used for four-level decomposition to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The algorithm first performs wavelet transform on the multimodal data to separate feature components at different spatiotemporal scales. Then, a cross-scale attention mechanism is constructed to fuse feature information at different scales through adaptive weights. The scale weight learning rate is set to 0.01, and the momentum parameter is set to 0.9. The algorithm calculates the multi-scale fusion distortion index and quantifies the degree of information loss by comparing the information entropy changes of the data before and after fusion. When the distortion index F∈[0, 0.85), the system judges that the information loss is large and starts super-resolution reconstruction processing. Interpolation and deconvolution methods are used to recover high-frequency detail information. The reconstruction factor is set to 2 times, and the filter size is 5×5. When the distortion index F∈[0.85, 1.0], the information integrity is relatively high, and the current processing method is maintained to ensure computational efficiency.

[0077] The specific implementation of step S09 involves using Bayesian networks for causal reasoning analysis. A Bayesian network is a directed graph model describing the interactions between random variables, representing causal semantics through conditional probability distributions. The network has 20 nodes, representing major marine environmental factors. The algorithm first constructs a directed acyclic graph to represent the causal relationships between variables. A structure learning algorithm is used to discover potential causal structures in the data, setting a causal strength threshold of 0.4 to eliminate weakly correlated connections. Then, the maximum likelihood estimation method is used to learn conditional probability parameters, and the EM algorithm is used to handle latent variables and missing data. The algorithm identifies key driving factors such as temperature, salinity, and nutrients, analyzing their impact paths and intensity on the ecosystem. Finally, the fusion weight scale index for each modality is calculated, quantifying the contribution of each modality's data through information gain and mutual information. The weight update frequency is set to once every 10 epochs to ensure adaptive adjustment of weight allocation. When the weight of physical modalities is high, the weight of temperature and salinity data is increased; when biological modalities are active, the weight of biological parameters such as chlorophyll is increased.

[0078] The specific implementation of step S10 involves establishing an interpretability evaluation system based on Shapley values. Shapley values ​​quantify the specific contribution of each feature to the model output using game theory, calculating the average marginal contribution of a feature across all possible combinations. The feature subset sampling number is set to 1000 times, and the convergence threshold is 0.001. The algorithm first constructs feature consortia, calculating the marginal contribution of each feature in different consortia, and estimating Shapley values ​​through Monte Carlo sampling to avoid the combinatorial explosion problem. Then, it calculates the fusion association attention index, comprehensively considering feature importance, interaction strength, and prediction confidence, setting the attention weights to a combination of importance 0.4, interaction 0.3, and confidence 0.3. The algorithm generates an interactive 3D association heatmap, using color mapping to represent association strength: red indicates positive correlation, blue indicates negative correlation, and color depth reflects the degree of association. The heatmap supports interactive operations; users can view detailed values ​​by hovering the mouse and adjust the viewpoint by zooming and rotating. When the fusion association attention index A∈[0.92, 1.0], the system judges the association pattern to be highly reliable and outputs the key association pattern recognition results, including the dominant driving factor, response variable and mechanism description.

[0079] The detailed structure of the feature extraction network includes an input layer that receives 512-dimensional multimodal data. The first convolutional layer uses 64 3×3 convolutional kernels to extract shallow features with a stride of 1 and padding of 1, resulting in an output feature map size of 512×64. The second convolutional layer uses 128 3×3 convolutional kernels to further extract features, combined with ReLU activation and batch normalization layers, and then a 2×2 max pooling layer to reduce the feature map size to 256×128. The third convolutional layer uses 256 3×3 convolutional kernels to extract deep features, again combined with activation and normalization layers, and then pooled again to 128×256. The network converts the two-dimensional feature map into a one-dimensional vector through a flattening layer. The first fully connected layer contains 512 neurons, using ReLU activation and dropout regularization. The second fully connected layer contains 256 neurons, and the final output layer contains 64 neurons to generate the feature vector. The training dataset contains continuous observation data from 2018 to 2022, with a total of 10,000 labeled samples, which are divided into training, validation and test sets in an 8:1:1 ratio.

[0080] The LSTM time series module employs an encoder-decoder architecture. The encoder consists of two LSTM layers: the first layer with 128 hidden units processes the 72-hour input sequence, and the second layer with 64 hidden units performs feature abstraction. The decoder also uses a two-layer structure, progressively generating 24-hour prediction sequences. The network uses the Adam optimizer with a learning rate of 0.01, employing gradient clipping to prevent gradient explosion with a clipping threshold of 1.0. The training dataset contains 50,000 time series samples, constructed using a sliding window method with a window step size of 1 hour. Data preprocessing uses Z-score normalization to eliminate the influence of dimensions. The training process uses the mean squared error loss function, with a maximum training epoch of 500 epochs, and an early stopping strategy to prevent overfitting.

[0081] The graph attention network architecture consists of two graph convolutional layers. The first layer uses eight attention heads, each with 64-dimensional features, employing the LeakyReLU activation function with a negative slope parameter of 0.2. The second layer uses four attention heads, each with 128-dimensional features, ultimately outputting 256-dimensional node embeddings. The network utilizes residual connections and layer normalization to improve training stability, employing the AdamW optimizer with a learning rate of 0.005 and weight decay of 0.01. The training dataset contains 1000 graph samples, each with 50 nodes and 200 edges, and employs a contrastive learning training strategy, learning effective representations through comparison of positive and negative samples.

[0082] It should be noted that the key technical ideas of this invention include an adaptive spatiotemporal kriging interpolation algorithm, a multimodal graph attention fusion mechanism, Shapley value-based interpretability analysis, and a dynamic frequency adjustment mechanism. The adaptive spatiotemporal kriging interpolation algorithm, compared to traditional interpolation methods, automatically adjusts interpolation parameters based on the anisotropic characteristics of ocean circulation. By modeling spatial correlation structures through variograms, it effectively handles the complex spatiotemporal distribution patterns of ocean data, significantly improving the spatiotemporal registration accuracy and consistency of multi-source heterogeneous data. The multimodal graph attention fusion mechanism models ocean parameters as graph nodes and automatically learns the nonlinear relationships between parameters using an attention mechanism. Compared to traditional linear fusion methods, it better captures the complex interaction patterns of the marine environmental system, achieving collaborative analysis of physical, chemical, and biological parameters. The Shapley value-based interpretability analysis framework quantifies the contribution of each modality to the final result from a game theory perspective, providing a reliable decision-making basis compared to black-box models. Interactive visualization enhances the model's transparency and user understanding. The dynamic frequency adjustment mechanism adaptively adjusts the processing frequency based on the severity of changes in the marine environment, optimizing computational resource allocation while ensuring monitoring accuracy and improving the system's real-time response capability compared to fixed-frequency processing. The synergistic effect of these four key technological approaches enables high-precision fusion analysis of marine environmental data. Adaptive interpolation ensures data quality, graph attention mechanisms uncover deep correlation patterns, interpretability analysis provides decision support, and dynamic adjustment guarantees system efficiency. This forms a complete technological chain from data preprocessing to correlation mining and result interpretation, demonstrating greater adaptability, accuracy, and practicality compared to existing technologies in handling complex marine environmental systems.

[0083] It should be noted that in marine environmental monitoring, satellite remote sensing data has the characteristics of large spatial coverage but relatively low temporal resolution, buoy observation data has the characteristics of high temporal resolution but limited spatial coverage, and ship monitoring data has the characteristics of high precision but irregular spatiotemporal sampling. This spatiotemporal heterogeneity makes it difficult for traditional data fusion methods to effectively handle the spatiotemporal registration problem between different data sources. This invention constructs an adaptive spatiotemporal kriging interpolation algorithm, combines it with prior knowledge of ocean dynamics to establish an anisotropic variogram model, optimizes interpolation parameters using cross-validation, and introduces trend surface analysis to eliminate systematic bias. This can accurately model the spatiotemporal correlation structure of different data sources, achieve high-precision multimodal data spatiotemporal registration and missing value imputation, and effectively solve the impact of spatiotemporal heterogeneity on data fusion accuracy. Marine environmental observation signals often contain mixed information of multiple physical, chemical, and biological processes, and are also affected by multiple factors such as ocean dynamics, seasonal changes, and human interference. Traditional signal processing methods, based on the linear separation assumption, are difficult to effectively separate the effective components in these complex mixed signals, resulting in insufficient extraction of effective information. This invention employs a sparse coding-based ocean signal separation algorithm. It utilizes overcomplete dictionary learning technology to construct a sparse representation model of ocean signals, solves for the optimal sparse coefficients through the orthogonal matching pursuit algorithm, and continuously optimizes the signal representation quality using a dictionary update strategy. This effectively separates different components in mixed ocean observation signals, accurately extracts effective information from each modality of data, and significantly improves the accuracy and robustness of signal processing.

[0084] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of inaccurate feature extraction during the fusion of multi-source heterogeneous marine environmental data. Its fundamental principle lies in constructing a multimodal feature extraction and correlation analysis technology system based on deep learning and graph neural networks. First, the adaptive spatiotemporal kriging interpolation algorithm, through the fusion of variogram modeling and prior knowledge of ocean dynamics, can accurately model the spatiotemporal correlation structure of different data sources, achieving high-quality data preprocessing and providing a unified standard data foundation for subsequent feature extraction, avoiding interference from data quality issues in the extraction of correlation features. Second, the marine signal separation algorithm based on sparse coding utilizes overcomplete dictionary learning technology to decompose mixed observation signals into different independent components, extracting pure marine environmental features and eliminating the impact of signal aliasing on the identification of correlation features. Crucially, the feature extraction network based on the attention mechanism adopts a combined architecture of convolutional neural networks and attention mechanisms, which can automatically learn the deep nonlinear feature representation of marine environmental parameters. Through multi-layer feature abstraction and feature selection, it accurately captures the complex correlation patterns between physical, chemical, and biological parameters. Graph attention networks further model ocean parameter relationships as a graph structure. Through a multi-head attention mechanism, they dynamically calculate the association weights between nodes, adaptively discovering and quantifying cross-modal association features. Compared to traditional correlation analysis methods, they exhibit stronger nonlinear modeling capabilities and higher feature representation accuracy. The introduction of an LSTM time-series module enables the system to capture the temporal dynamics of ocean environmental parameters and identify dynamic association patterns. Furthermore, the causal reasoning capabilities of Bayesian networks enhance the interpretability and accuracy of association features. The entire technical solution, through multi-level feature learning, multi-scale information fusion, and multi-dimensional association modeling, forms a complete feature extraction chain from raw data to high-level semantic features, significantly improving the accuracy and robustness of association feature extraction from multi-source heterogeneous ocean environmental data.

[0085] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0086] The specific implementation of step S01 is to establish a multi-source data spatiotemporal standardization model. The mathematical expression of the adaptive spatiotemporal kriging interpolation algorithm is:

[0087] ;

[0088] In the formula, This is the estimated value at the point to be interpolated; For the first Measured values ​​at each observation point; For the first Weighting coefficients for each observation point; This represents the number of neighboring observation points, with a value of 15. The weighting coefficients are obtained by solving the Kriging equations:

[0089] ;

[0090] In the formula, For observation point and The values ​​of the variation function between them; Let be the Lagrange multiplier. The anisotropic variogram model is expressed as:

[0091] ;

[0092] In the formula, For the nugget effect, the value is 0.1; This is the base value, which is 0.9. For variable range, the value is 5000m; This represents the lag distance. The sparse matrix of environmental parameters is represented as:

[0093] ;

[0094] In the formula, For the first The first spatiotemporal grid point One environmental parameter value; The number of spatiotemporal grid points. This refers to the number of types of environmental parameters.

[0095] The specific implementation of step S02 is a marine signal separation algorithm based on sparse coding. The sparse representation model is expressed as:

[0096] ;

[0097] In the formula, The observed signal vector; For an overcomplete dictionary matrix, the dimension is ,in For signal dimension, The number of atoms in the dictionary; It is a sparse coefficient vector; Let be the noise vector. The residual update formula for the orthogonal matching pursuit algorithm is:

[0098] ;

[0099] In the formula, For the first The residual vector of the next iteration; For the first The second selected dictionary atom; This is an inner product operation. The atom selection criterion is:

[0100] ;

[0101] In the formula, For the first The atom index selected in the next iteration. The signal separation coefficient vector is represented as:

[0102] ;

[0103] In the formula, For the first The contribution coefficient of each dictionary atom.

[0104] The specific implementation of step S03 is based on constructing a feature extraction network using an attention mechanism. The formula for calculating the self-attention weights is:

[0105] ;

[0106] In the formula, For query matrix; The key matrix; It is a value matrix; Let be the dimension of the key vector. The multi-head attention mechanism is expressed as:

[0107] ;

[0108] In the formula, ; For the number of attention heads; This is the output weight matrix. The sparse real-valued adjacency matrix of multimodal features is represented as:

[0109] ;

[0110] In the formula, For the first The first feature and the second The strength of the association between features; The dimension of the feature vector.

[0111] The specific implementation of step S04 is to use a matrix rank deficit detection algorithm. Singular value decomposition is expressed as:

[0112] ;

[0113] In the formula, It is a left singular vector matrix; It is a singular value diagonal matrix; Let be a right singular vector matrix. The formula for calculating the effective rank is:

[0114] ;

[0115] In the formula, For the first One singular value; The threshold for rank detection; The indicator function is used. The feature compensation vector is constructed through principal component analysis:

[0116] ;

[0117] In the formula, For compensation vector; The matrix is ​​a basis vector matrix in the null space; It is a vector of random coefficients.

[0118] The specific implementation of step S05 is to use a multimodal association mining algorithm based on graph attention networks. The mutual information calculation formula is:

[0119] ;

[0120] In the formula, For parameters and Mutual information between them; For joint probability density; and Let be the marginal probability density. The attention coefficient of the graph attention mechanism is calculated as follows:

[0121] ;

[0122] In the formula, For nodes To the node Attention coefficient; This is the attention parameter vector; It is a linear transformation matrix; For nodes eigenvectors; This is for the splicing operation. The normalized attention weights are:

[0123] ;

[0124] In the formula, For nodes The neighbor set. The parameter-associative sparse padding matrix is ​​represented as:

[0125] ;

[0126] In the formula, For parameters and The strength of the correlation between them; This represents the number of nodes for ocean parameters.

[0127] The specific implementation of step S06 involves introducing an LSTM timing module to capture the dynamic process. The mathematical expression of the LSTM gating mechanism is:

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] In the formula, , , These are the forget gate, input gate, and output gate, respectively. In cellular state; It is in a hidden state; and These are the weights and bias parameters; It is the sigmoid function; This is an element-wise multiplication. The temporal gradient change matrix is ​​expressed as:

[0135] ;

[0136] In the formula, For the first The parameters at time... The gradient value; The number of parameters; Where is the time step. The gradient value is calculated using the following formula:

[0137] ;

[0138] In the formula, For a moment The parameter vector; It is an L2 norm.

[0139] The specific implementation of step S07 is to use a hyperparameter automatic tuning algorithm based on Bayesian optimization. The Gaussian process model is expressed as:

[0140] ;

[0141] In the formula, The objective function is... It is a mean function; Let be the covariance function. The radial basis function kernel function is defined as:

[0142] ;

[0143] In the formula, The variance of the signal; Let be the length scale parameter. The desired improved acquisition function expression is:

[0144] ;

[0145] In the formula, and To predict the mean and variance; This is the current optimal value; ; and Let be the cumulative distribution function and probability density function of the standard normal distribution. The optimized hyperparameter configuration vector is represented as:

[0146] ;

[0147] In the formula, For the first The optimal values ​​of each hyperparameter; This represents the total number of hyperparameters.

[0148] The specific implementation of step S08 is based on wavelet transform decomposition of features at different scales. The wavelet transform is expressed as:

[0149] ;

[0150] In the formula, These are wavelet coefficients; For scale parameters; These are translation parameters; This is the complex conjugate of the mother wavelet function. The formula for calculating the multi-scale fusion distortion index is:

[0151] ;

[0152] In the formula, For data Information entropy; For the merged data; This is the original data. Information entropy is defined as:

[0153] ;

[0154] In the formula, For the first The probability of each data value occurring.

[0155] The specific implementation of step S09 involves using Bayesian networks for causal reasoning analysis. The joint probability distribution of a Bayesian network is expressed as:

[0156] ;

[0157] In the formula, For the first One random variable; for The set of parent nodes; This represents the total number of network nodes. The formula for calculating the fusion weight metric index is:

[0158] ;

[0159] In the formula, For the first The weights of each modality; For information gain; Mutual information value; This represents the total number of modes.

[0160] The specific implementation of step S10 is to establish an interpretability evaluation system based on the Shapley value. The formula for calculating the Shapley value is:

[0161] ;

[0162] In the formula, Features Shapley value; For features not included A subset of features; For the set of all features; For the alliance The feature function value. The fusion association attention index is expressed as:

[0163] ;

[0164] In the formula, Importance index; The interaction strength; To predict confidence levels.

[0165] It needs to be explained that the sparse matrix of environmental parameters is obtained through the Kriging interpolation equation. Achieving spatiotemporal standardization, weighting coefficients By minimizing the estimated variance, the unbiasedness and optimality of the interpolation results are guaranteed.

[0166] Sparse coding algorithms use the objective function To achieve signal decomposition, where Norm constraints ensure the sparsity of solutions; dictionary update formula. Optimize representation quality;

[0167] Feature weight allocation formula Weight normalization and importance score are achieved using the softmax function. Obtained through linear transformation;

[0168] Matrix rank test pass conditions To determine valid singular values, when Time-triggered compensation mechanism, compensation vector Linear independence from existing features is guaranteed through null space projection;

[0169] Mutual information function Quantifying the strength of nonlinear correlation, when Preserve parameters for connectivity, graph adjacency matrix Define the node connection relationships;

[0170] LSTM achieves long-term memory through gating mechanisms and gradient adaptive strategies. The processing frequency is dynamically adjusted according to the severity of the change. Frequency reduction by 60% saves resources;

[0171] Bayesian optimization uses Gaussian process priors Modeling hyperparameter space Expected to improve the criteria Balance exploration and utilization;

[0172] wavelet transform Multi-resolution analysis is achieved through scale scaling; distortion index Information retention is quantified by the information entropy ratio;

[0173] Bayesian networks utilize the chain rule Decomposition of joint distribution, causal strength Quantifying causal relationships between variables, weight update formula Achieve dynamic weight adjustment;

[0174] Shapley value is determined by marginal contribution. The importance of weighted average quantization features, and the interaction Shapley value. Fusion index measures feature synergy A comprehensive assessment of the credibility of the association patterns.

[0175] To better understand and implement this invention, Example 2 of a specific application scenario is provided below: A technical team needs to conduct environmental data correlation analysis on the deep-sea area outside the Yangtze River Estuary, focusing on analyzing the correlation driving factors of summer red tide outbreaks in this area. The study area involves deep-sea areas with water depths of 20–200 m. The technical team decided to adopt an environmental data correlation analysis method based on multimodal fusion to deeply mine multi-source observation data of this area from June to August 2023.

[0176] First, a multi-source data spatiotemporal standardization model was established. The technical team collected satellite remote sensing data for the region, including parameters such as sea surface temperature, chlorophyll concentration, and suspended matter concentration, with a spatial resolution of 1km×1km and a temporal resolution of daily averages. Simultaneously, water temperature, salinity, dissolved oxygen, and pH data from 12 buoy observation stations were acquired, with a temporal resolution of hourly levels. In addition, monitoring data on nutrients, phytoplankton abundance, and sediment particle size from eight research vessels were collected. An adaptive spatiotemporal kriging interpolation algorithm was used to register these data at different spatiotemporal resolutions, constructing a unified spatiotemporal reference grid of 500×400 with a grid spacing of 0.01°×0.01°. After processing, a sparse matrix of environmental parameters was formed, with a matrix dimension of 200000×15, where non-zero elements account for only 23.8%.

[0177] Next, a sparse coding-based ocean signal separation algorithm was used to process the sparse matrix of environmental parameters. The technical team constructed an overcomplete dictionary containing 512 atoms to learn the sparse representation of ocean signals. The orthogonal matching pursuit algorithm was used to solve for the sparse coefficients, with a sparsity threshold of 0.15, successfully separating different components such as tidal signals, seasonal variation signals, and anthropogenic disturbance signals. The generated signal separation coefficient vector had a length of 512, with an average of 76 effective coefficients, and the energy retention rate of the separated signal reached 96.7%.

[0178] In the feature extraction stage, the technical team constructed a feature extraction network based on an attention mechanism to process physical, chemical, and biological parameter data. Physical parameters included eight parameters such as water temperature, salinity, flow velocity, and flow direction; chemical parameters included seven parameters such as dissolved oxygen, pH, and nutrient concentration; and biological parameters included five parameters such as chlorophyll concentration and phytoplankton abundance. The feature extraction network extracted a 64-dimensional key feature vector for each type of parameter, ultimately establishing a 20×64 multimodal feature sparse real-valued adjacency matrix. Figure 2 As shown, the adjacency matrix demonstrates the strength of connectivity relationships between different modal features.

[0179] In the matrix rank deficit detection stage, the technical team used the singular value decomposition algorithm to detect the rank property of sparse real-valued adjacency matrices with multimodal features. The detection revealed a rank deficit in the matrix, with an effective rank of 18, lacking two linearly independent vectors. The system automatically generated feature compensation vectors of length 64, and ensured the linear independence of the vectors through a Gram-Schmidt orthogonalization process, successfully restoring the full-rank property of the matrix.

[0180] A multimodal association mining algorithm based on graph attention networks was employed to construct an ocean parameter association map. The team used 20 ocean parameters as graph nodes, calculated initial edge weights through mutual information, and identified 157 significant association edges. An 8-head attention mechanism was used to dynamically update node features, with each attention head containing 32-dimensional features. After three layers of graph convolution processing, a 20×20 parameter association sparse-filled matrix was formed, with non-zero elements accounting for 39.25% of the matrix. Figure 3 As shown, this correlation map clearly demonstrates the network structure of interactions among various ocean parameters.

[0181] To capture dynamic processes such as the expansion of freshwater runoff in the Yangtze River and seasonal algal blooms, the technical team introduced an LSTM time-series module for time-series analysis. A two-layer LSTM network was constructed, with the first layer containing 128 hidden units and the second layer containing 64 hidden units. The input sequence length was set to 72 hours to predict parameter change trends over the next 24 hours. A time-series gradient change matrix with dimensions of 20×24 was calculated to record the time-series gradient changes of each parameter. During the analysis, the gradient value G of chlorophyll concentration was 0.42, exceeding the 0.30 threshold, and the system automatically increased the signal processing frequency of this parameter to 150% of the original frequency. The gradient value G of water temperature change was 0.23, and the current processing frequency was maintained. The gradient value G of salinity change was 0.08, and the system reduced its processing frequency to 60% of the original frequency.

[0182] During the hyperparameter optimization phase, the technical team employed an automatic tuning algorithm based on Bayesian optimization. A Gaussian process was used to model the mapping relationship between 12 key hyperparameters and model performance, including learning rate, regularization coefficient, number of network layers, and number of attention heads. Evaluation points were selected by improving the expected acquisition function, and the optimal hyperparameter configuration was found after 85 iterations. As shown in Table 1, the optimized hyperparameters significantly improved model performance.

[0183] Table 1 Comparison of Key Hyperparameter Optimization Results

[0184]

[0185] Based on wavelet transform decomposition of features at different scales, the technical team constructed a cross-scale attention mechanism. Daubechies wavelets were used for 5-level decomposition to obtain feature representations at different frequency scales. The multi-scale fusion distortion index was calculated, revealing that the distortion index F was 0.78 when processing chlorophyll concentration data, which is less than the threshold of 0.85. Therefore, the system initiated super-resolution reconstruction processing, increasing the spatial resolution from 1 km to 500 m. For water temperature data, the distortion index F was 0.91, and the current processing method was maintained.

[0186] In the causal reasoning analysis phase, the technical team employed Bayesian networks to identify key driving factors. A Bayesian network structure with 20 nodes and 43 directed edges was constructed, and a structure learning algorithm was used to determine the causal relationships between parameters. Analysis revealed that nutrient concentration was the primary driving factor for changes in chlorophyll concentration, with a causal strength of 0.73. The scaling index of the fusion weights for each modality was calculated: 0.42 for physical parameters, 0.35 for chemical parameters, and 0.23 for biological parameters. The system dynamically adjusts the contribution of each modality based on these weights, and the multimodal feature fusion weight distribution is as follows: Figure 7 As shown.

[0187] Finally, an interpretability evaluation system based on Shapley values ​​was established. The technical team calculated the marginal contribution of each feature to the final association results and found that the temperature-nutrient association pattern had the highest Shapley value of 0.28. The salinity-chlorophyll association pattern had a Shapley value of 0.22, and the dissolved oxygen-phytoplankton association pattern had a Shapley value of 0.19. The calculated fusion association attention index A was 0.94, exceeding the 0.92 threshold, indicating that the system successfully output the key association pattern identification results. As shown in Table 2, the system identified five main cross-modal association patterns.

[0188] Table 2 Key Cross-Modal Association Patterns Identified

[0189]

[0190] The generated interactive 3D correlation heatmap visually displays the distribution of correlation strength between various parameters, such as... Figure 4 As shown in the image. Through this visualization interface, the technical team can clearly observe the temperature and... The concentrations showed a significant negative correlation, with a correlation coefficient of -0.84. In surface seawater, there was a clear lag response between salinity changes and chlorophyll-a concentration, with a lag time of approximately 18 hours.

[0191] Through analysis using the time-series prediction module, the technical team discovered a clear red tide warning signal in the area in late July. For example... Figure 5 As shown, the predicted chlorophyll concentration is 3.2 mg / L from the normal level. It rose sharply to 15.8 mg / L. The error compared to the actual observed value is only 8.7%. Meanwhile... Concentration and water temperature showed a synergistic upward trend, providing favorable conditions for red tide outbreaks. Throughout the analysis, the system demonstrated excellent computational efficiency. Multi-source data fusion processing took 45 minutes, feature extraction and association mining took 32 minutes, and time-series analysis took 28 minutes, for a total processing time of 105 minutes. Compared to traditional statistical analysis methods, this method is approximately 4.2 times more efficient when processing the same amount of data. Figure 6 As shown, the technical team also analyzed the vertical distribution characteristics of environmental parameters at different depths and found that there are obvious parameter gradient changes in the strata region (water depth 20-40m), which provides an important basis for understanding the mixing process of Yangtze River desalinated water and offshore seawater.

[0192] This invention represents a significant technological advancement over traditional marine environmental data analysis methods. Traditional methods primarily rely on single data sources and statistical correlation analysis, making them ill-suited for handling multimodal, multi-scale, and nonlinear marine environmental data. This invention achieves effective separation of mixed signals through sparse coding technology, overcoming the limitations of traditional methods in handling signal overlap. The use of graph attention networks to construct parameter correlation maps enables automatic learning of complex nonlinear relationships, avoiding the limitations of traditional linear analysis methods. The introduction of spatiotemporal adaptive mechanisms and multi-scale fusion strategies significantly improves the ability to capture dynamic ocean processes. Deep learning-based feature extraction networks automatically discover hidden patterns in the data, avoiding the subjectivity and limitations of manual feature engineering. Bayesian-optimized hyperparameter tuning algorithms achieve automatic optimization of model parameters, improving the reliability and stability of the analysis results. The establishment of an interpretability evaluation system makes the analysis results more transparent and credible, providing a scientific basis for marine environmental management decisions.

[0193] It should be noted that the variables involved in this invention are explained in detail in Tables 3, 4, and 5.

[0194] Table 3. Variable Explanation Table (Part 1)

[0195]

[0196] Table 4. Variable Explanation Table (Part Two)

[0197]

[0198] Table 5. Variable Explanation Table (Part 3)

[0199]

[0200] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary based on multimodal fusion, characterized in that, Includes the following steps: S01. Establish a multi-source data spatiotemporal standardization model, and use an adaptive spatiotemporal kriging interpolation algorithm to perform spatiotemporal registration of satellite remote sensing data, buoy observation data, and ship monitoring data to construct a unified spatiotemporal reference grid and form a sparse matrix of environmental parameters. Among them, satellite remote sensing data includes sea surface temperature, chlorophyll concentration, and suspended matter concentration; buoy observation data includes water temperature, salinity, dissolved oxygen, and pH value of buoy observation stations; and ship monitoring data includes nutrients, phytoplankton abundance, and sediment particle size. S02. The marine signal separation algorithm based on sparse coding processes the sparse matrix of the environmental parameters, constructs an overcomplete dictionary to learn the sparse representation of the marine signal, uses the orthogonal matching pursuit algorithm to solve the sparse coefficients, generates a signal separation coefficient vector, and separates the marine signals of different frequency components. S03. Construct a feature extraction network based on the attention mechanism to process physical parameters, chemical parameters, and biological parameters data respectively, extract key feature vectors of each modality, and establish a multimodal feature sparse real-valued adjacency matrix; among which, physical parameters include water temperature, salinity, flow velocity, and flow direction, chemical parameters include dissolved oxygen, pH value, and nutrient concentration, and biological parameters include chlorophyll concentration and phytoplankton abundance. S04. The multimodal feature sparse real-valued adjacency matrix is ​​detected by a matrix rank deficiency detection algorithm. When a rank deficiency is detected in the multimodal feature sparse real-valued adjacency matrix, a feature compensation vector is automatically added to restore the full rank property of the matrix. S05. A multimodal association mining algorithm based on graph attention network is used to construct an ocean parameter association graph. Each ocean parameter is used as a graph node. The initial edge weight is calculated through mutual information. The node features are dynamically updated using a multi-head attention mechanism to form a parameter association sparse filling matrix. S06. An LSTM timing module is introduced to capture the dynamic process of the expansion of the Yangtze River's freshwater runoff and seasonal algal blooms. The timing data of each parameter is input into the LSTM timing module to calculate the timing gradient change matrix and record the timing gradient of each parameter. When the gradient value of the timing gradient change matrix G∈[0, 0.15], the signal processing frequency is reduced to 60% of the original frequency. When the gradient value of the timing gradient change matrix G∈(0.15, 0.30], the current signal processing frequency is maintained. When the gradient value of the timing gradient change matrix G∈(0.30, 1.0], the signal processing frequency is increased to 150% of the original frequency. S07. The model parameters are optimized using a Bayesian optimization-based hyperparameter automatic tuning algorithm. The mapping relationship between hyperparameters and model performance is modeled using a Gaussian process. The next evaluation point is selected through the acquisition function, and an optimized hyperparameter configuration vector is generated. S08. Based on wavelet transform decomposition of features at different scales, construct a cross-scale attention mechanism, calculate the multi-scale fusion distortion index, add super-resolution reconstruction processing when the multi-scale fusion distortion index F∈[0, 0.85), and maintain the current processing method when the multi-scale fusion distortion index F∈[0.85, 1.0]. S09. Use Bayesian networks for causal reasoning analysis, identify key driving factors, calculate the fusion weight scale index of each modality, and dynamically adjust the weight allocation of each modality data according to the magnitude of the fusion weight scale index. S10. Establish an interpretability evaluation system based on Shapley values, calculate the fusion association attention index, generate an interactive three-dimensional association heatmap, and output the key association pattern recognition result when the fusion association attention index A∈[0.92, 1.0].

2. The method according to claim 1, characterized in that, The sparse matrix of environmental parameters is specifically a data storage structure with a large number of zero-value elements formed by standardizing multi-source environmental monitoring data with different spatiotemporal resolutions.

3. The method according to claim 2, characterized in that, The signal separation coefficient vector is specifically a vector representing the weights of different signal components obtained after separating ocean observation signals using a sparse coding algorithm.

4. The method according to claim 3, characterized in that, The multimodal feature sparse real-valued adjacency matrix is ​​specifically a matrix that describes the connection relationship between different modal features, where non-zero elements represent the strength value of the correlation between features.

5. The method according to claim 4, characterized in that, The feature compensation vector is specifically a supplementary vector automatically generated when the rank deficiency occurs in the multimodal feature sparse real-valued adjacency matrix.

6. The method according to claim 5, characterized in that, The parameter association sparse filling matrix is ​​specifically a sparse representation matrix of the association between ocean parameters learned through a graph attention network, which uses a filling algorithm to complete the missing association information.

7. The method according to claim 6, characterized in that, The time-series gradient change matrix is ​​specifically a gradient information matrix describing the trend of marine environmental parameters changing over time, which is formed by calculating the rate of change of parameters between adjacent time steps.

8. The method according to claim 7, characterized in that, The optimized hyperparameter configuration vector is specifically the optimal hyperparameter combination vector of the model obtained by searching through the Bayesian optimization algorithm.

Images