A Deep Learning-Based Fusion Process for Marine Hydrological and Meteorological Data

Abstract

This invention relates to the field of data processing technology, specifically to a method for fusion processing of marine hydrological and meteorological data based on deep learning, comprising the following steps: S1, data acquisition: monitoring marine hydrological and meteorological data through a distributed marine monitoring network; S2, construction of a regional collaborative fusion network; S3, spatiotemporal consistency optimization: introducing a multi-scale spatiotemporal attention mechanism to capture both local short-term changes and global long-term trends simultaneously present in the data, dynamically adjusting the weights of different time scales; and optimizing the data fusion results by combining marine physical characteristics as constraints, including the periodic patterns of tides and ocean currents. This invention, by introducing a multi-scale spatiotemporal attention mechanism, achieves accurate capture of short-term dynamic changes and long-term trends based on globally fused data, and through a dynamic time window adjustment mechanism, can adapt to the changing characteristics at different time scales.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for fusion processing of marine hydrological and meteorological data based on deep learning fusion. Background Technology

[0002] With the increasing demand for global climate change and marine environmental monitoring, the collection and processing of hydrological and meteorological data plays a crucial role in marine scientific research. Accurate fusion and processing of marine hydrological and meteorological data is of great significance for extreme weather warnings, climate trend predictions, and marine resource development. However, the complexity, variability, and extreme nature of the marine environment present the following challenges to existing data processing methods:

[0003] Ocean data exhibits complex and unstable spatiotemporal variations. Extreme events (such as tropical cyclones and tsunamis) may occur in the short term, while periodic patterns (such as tidal changes and ocean current paths) are observed in the long term. Existing methods are clearly insufficient in handling the balance between short-term sudden dynamics and long-term trend changes.

[0004] Marine hydrological and meteorological data are strongly constrained by natural physical laws (such as tidal cycles and ocean current diffusion), but most existing deep learning models rely on data-driven approaches and fail to fully integrate domain knowledge, resulting in poor physical consistency of data results. Summary of the Invention

[0005] This invention provides a method for fusion processing of marine hydrological and meteorological data based on deep learning fusion.

[0006] The deep learning-based fusion method for marine hydrological and meteorological data includes the following steps:

[0007] S1, Data Acquisition: Monitoring marine hydrological and meteorological data through a distributed marine monitoring network, which includes multiple monitoring nodes. Each monitoring node is equipped with observation equipment to collect hydrological and meteorological data, including temperature, salinity, ocean current, wind speed, air pressure, and precipitation.

[0008] S2, Construction of Regional Collaborative Integration Network:

[0009] S21, Node Feature Embedding: Generate node features from the hydrological and meteorological data collected within the coverage area of ​​each monitoring node, including the temporal and spatial features of the nodes, and characterize the temporal and spatial correlation of the data through the edge connections between nodes.

[0010] S22, Edge weight definition: The weight of the edge is defined based on the spatial distance component and the data correlation component between the monitoring nodes;

[0011] S23, Graph Convolutional Network Modeling: Based on the global topology of the monitoring network, graph convolutional networks are used to jointly model all nodes. Through convolution operations, node features and edge connections are comprehensively learned to generate a globally fused representation of marine hydrological and meteorological data.

[0012] S3, Spatiotemporal Consistency Optimization: Introducing a multi-scale spatiotemporal attention mechanism to capture both local short-term changes and global long-term trends that coexist in the data, dynamically adjusting the weights of different time scales, and classifying and weighting the fused data through the spatiotemporal attention mechanism to ensure the consistency of spatiotemporal consistency between sudden and regular change data; combining ocean physical characteristics as constraints to optimize the data fusion results, including the periodicity of tides and ocean currents, to ensure that the consistency of the data conforms to the actual physical meaning.

[0013] Optionally, the monitoring nodes include deep-sea monitoring nodes, shallow-sea monitoring nodes, and near-shore monitoring nodes to ensure that the monitoring network covers the entire target sea area;

[0014] Each monitoring node is equipped with observation equipment, including:

[0015] Temperature sensors are used to monitor changes in seawater temperature.

[0016] Salinity sensors are used to measure the salt concentration in water in real time.

[0017] Current meters are used to monitor the speed and direction of ocean currents;

[0018] Anemometers are used to collect data on wind speed and direction changes over the sea.

[0019] A barometric pressure sensor is used to record changes in atmospheric pressure.

[0020] A precipitation meter is used to measure precipitation in a sea area.

[0021] Optionally, the generation of the time features includes: extracting periodic, trend, and sudden change features based on time-series hydrological and meteorological data collected within the coverage area of ​​the monitoring nodes; and modeling the time-series data through a long short-term memory network to generate time feature vectors for the nodes, describing the dynamic change patterns of the data.

[0022] The generation of the spatial features includes: based on the spatial distribution within the coverage area of ​​the monitoring nodes, using a convolutional neural network to extract the spatial change patterns of the data, and converting the spatial features, including local temperature gradients and ocean current diffusion paths, into spatial feature vectors to describe the local characteristics of the area where the nodes are located.

[0023] Temporal and spatial features are concatenated, and a unified high-dimensional node feature representation is generated through dimensionality reduction embedding to construct the node feature vector for each monitoring node.

[0024] Optionally, the spatial distance component is calculated using the spherical distance formula for node v. i and node v j Spatial distance;

[0025] The data correlation component is calculated using the correlation coefficient at node v. i and v j Time series correlation;

[0026] The edge weights, which combine spatial distance and data correlation, are defined as follows:

[0027] Where, d ij Represents node v i and v j Spatial distance, r ij It is node v i and v j Correlation coefficient of time series, w ij It is node v i and v j The edge weights between the distance and correlation are adjusted parameters, α and β, to balance the importance of distance and correlation.

[0028] Optionally, the graph convolutional network learns a higher-order representation of the node feature vectors:

[0029] Among them, H (l) It is the node feature matrix of the l-th layer, initialized as [F1, F2, ..., F n ], It is a weighted adjacency matrix, determined by the edge weights w. ij Build, It is the degree matrix of the weighted adjacency matrix. W (l) σ is the learnable weight matrix of the l-th layer, and σ is the activation function.

[0030] Optionally, the globally fused marine hydrometeorological data is represented as follows:

[0031] F global =Aggregate(H (L) ), where F global It is a globally integrated representation of ocean hydrological and meteorological data, H (L) It is the feature matrix of the nodes in the last layer of the graph convolutional network. Aggregate(·) is an aggregation operation, such as global average pooling or max pooling.

[0032] Optionally, S3 includes multi-scale time window division:

[0033] Based on the temporal characteristics of the monitoring data, the time series is divided into multiple time windows (short-term window and long-term window) to capture local short-term changes and global long-term trends, respectively.

[0034] Optionally, the multi-scale spatiotemporal attention mechanism includes the calculation of temporal attention weights and spatial attention weights;

[0035] The time attention weight α t The calculation is based on multi-scale time window partitioning, with each time window containing a data segment x. t ;

[0036] The spatial attention weight β ij The calculation includes calculations for each monitoring node v i Calculate its data correlation with surrounding nodes;

[0037] The classification and weighting adjustment of fused data through the spatiotemporal attention mechanism specifically includes:

[0038] Spatiotemporal feature fusion: incorporating temporal attention weight α t Spatial attention weight β ij Fusion to generate spatiotemporal joint weight γ ij,t ;

[0039] Based on joint weight γ ij,t Dynamic fluctuations, setting a classification threshold θ:

[0040] If γ ij,t >θ indicates a sudden change in the data;

[0041] If γ ij,t If the value is less than or equal to θ, it is considered as data with normal variation.

[0042] Different weights are assigned to the classification results, with higher weights given to data with sudden changes to ensure that their dynamic change characteristics are captured.

[0043] Optionally, the optimization of the data fusion results by incorporating ocean physical characteristics as constraints specifically includes:

[0044] Periodicity embedding: Extracting the tidal and ocean current cycles of the monitoring area and constructing a constraint function: L physics =∑ t ‖F t -f physics (t)‖ 2 , of which F t For globally fused ocean hydrometeorological data over time t, f physics (t) is a physical law function that describes the periodic behavior of tides or ocean currents.

[0045] Optionally, an overall optimization objective can also be included: after joint weight adjustment, ocean physical property constraints are incorporated as a loss function into the overall optimization objective: L total =L attention +λ·L physics , where L attention It is the attention mechanism optimization loss, where λ is an adjustment parameter that controls the weight of the marine physical property constraints.

[0046] The beneficial effects of this invention are:

[0047] This invention, by introducing a multi-scale spatiotemporal attention mechanism, achieves accurate capture of short-term dynamic changes and long-term trends based on globally fused data. Through a dynamic time window adjustment mechanism, it can adapt to the changing characteristics of different time scales. Combined with the calculation of spatial correlation between nodes, it ensures the consistency optimization of the spatiotemporal characteristics of monitoring nodes in the global network. At the same time, the step of weighted adjustment of spatiotemporal features significantly improves the coordination between sudden changes and regular changes in data. The generated globally fused representation has higher spatiotemporal consistency and dynamic adaptability.

[0048] This invention further optimizes the global consistency of data by embedding constraints based on ocean physical characteristics (such as tidal cycles and ocean current paths). The constraints based on physical laws not only effectively suppress the influence of abnormal data, but also ensure that the optimized data conforms to the actual physical meaning of the ocean hydrological and meteorological system. This method, which combines deep learning and domain knowledge, is difficult to achieve in traditional deep learning models, and greatly improves the scientificity and credibility of the data results.

[0049] This invention constructs a regional collaborative fusion network based on graph convolutional networks to achieve global correlation modeling of data between distributed monitoring nodes. The combination of node feature embedding and dynamic definition of edge weights ensures the data integration capability in multi-source heterogeneous and complex environments. The resulting global fusion data representation, after spatiotemporal consistency optimization, can be directly used in various scenarios such as real-time monitoring, disaster early warning, and climate trend analysis, significantly improving the availability and diversity of data and providing strong technical support for accurate decision-making in the field of marine hydrology and meteorology. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention. Detailed Implementation

[0052] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0053] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0054] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0055] like Figure 1 As shown, the deep learning-based fusion processing method for marine hydrological and meteorological data includes the following steps:

[0056] S1, Data Acquisition: Monitoring marine hydrological and meteorological data through a distributed marine monitoring network. The distributed marine monitoring network includes multiple monitoring nodes, each equipped with observation equipment to collect hydrological and meteorological data, including temperature, salinity, ocean currents, wind speed, air pressure, and precipitation.

[0057] S2, Construction of Regional Collaborative Integration Network:

[0058] S21, Node Feature Embedding: Generate node features from the hydrological and meteorological data collected within the coverage area of ​​each monitoring node, including the temporal and spatial features of the nodes, and characterize the temporal and spatial correlation of the data through the edge connections between nodes.

[0059] S22, Edge weight definition: The weight of the edge is defined based on the spatial distance component and the data correlation component between the monitoring nodes;

[0060] S23, Graph Convolutional Network Modeling: Based on the global topology of the monitoring network, graph convolutional networks are used to jointly model all nodes. Through convolution operations, node features and edge connections are comprehensively learned to generate a globally fused representation of marine hydrological and meteorological data.

[0061] S3, Spatiotemporal Consistency Optimization: Introducing a multi-scale spatiotemporal attention mechanism to capture both local short-term changes and global long-term trends that coexist in the data, dynamically adjusting the weights of different time scales, and classifying and weighting the fused data through the spatiotemporal attention mechanism to ensure the consistency of spatiotemporal consistency between sudden and regular change data; combining ocean physical characteristics as constraints to optimize the data fusion results, including the periodicity of tides and ocean currents, to ensure that the consistency of the data conforms to the actual physical meaning.

[0062] The monitoring nodes include deep-sea monitoring nodes, shallow-sea monitoring nodes, and near-shore monitoring nodes to ensure that the monitoring network covers the entire target sea area;

[0063] Each monitoring node is equipped with observation equipment, including:

[0064] Temperature sensors are used to monitor changes in seawater temperature.

[0065] Salinity sensors are used to measure the salt concentration in water in real time.

[0066] Current meters are used to monitor the speed and direction of ocean currents;

[0067] Anemometers are used to collect data on wind speed and direction changes over the sea.

[0068] A barometric pressure sensor is used to record changes in atmospheric pressure.

[0069] A precipitation meter is used to measure precipitation in a sea area.

[0070] The generation of time features includes: extracting periodic, trend and sudden change features from time-series hydrological and meteorological data collected within the coverage area of ​​monitoring nodes; and modeling the time-series data through a long short-term memory network to generate time feature vectors for nodes, describing the dynamic change patterns of the data.

[0071] The generation of spatial features includes: based on the spatial distribution within the coverage area of ​​the monitoring nodes, using a convolutional neural network to extract the spatial variation patterns of the data, and transforming spatial features, including local temperature gradients and ocean current diffusion paths, into spatial feature vectors to describe the local characteristics of the area where the nodes are located.

[0072] Temporal and spatial features are concatenated, and a unified high-dimensional node feature representation is generated through dimensionality reduction embedding to construct the node feature vector for each monitoring node.

[0073] Temporal Feature Generation: Modeling Node v using a Long Short-Term Memory (LSTM) Network i The dynamic characteristics of the time series are expressed by the following formula: in, It is node v i Input the hydrological and meteorological data collected at time t. It is the hidden state at time t, representing the time characteristic. Represents the hidden state and memory state of the previous time step, and LSTM(·) represents the activation function of the Long Short-Term Memory network;

[0074] Spatial Feature Generation: Extracting Node v using a Convolutional Neural Network (CNN) i Spatial characteristics:

[0075] in, Represents node v i The spatial data matrix within the coverage area includes local temperature gradients and ocean current diffusion paths, S i Represents node v i The spatial feature representation, where CNN(·) represents the convolutional neural network operation used to extract spatial features;

[0076] Node feature vector: Incorporating time features and spatial features S i The features are concatenated and then reduced in dimensionality to generate a unified node feature vector. Among them, F i It is node v i The final eigenvectors, where W is the feature dimensionality reduction weight matrix. This indicates a feature splicing operation.

[0077] The spatial distance components are calculated using the spherical distance formula for node v. i and node v j Spatial distance:

[0078] d ij =R·arccos(sinφ) i ·sinφ j +cosφ i ·cosφ j ·cos(λ i -λ j )),in,

[0079] d ij Represents node v i and v j Spatial distance, R is the Earth's radius, φ i ,φ j For node vi and v j latitude, λ i ,λ j For node v i and v j Longitude;

[0080] The correlation component of the data is calculated using the correlation coefficient at node v. i and v j Time series correlation:

[0081] Where, r ij It is node v i and v j Correlation coefficient of time series It is node v i and v j At time t, It is node v i and v j Time series average;

[0082] Edge weights combine spatial distance and data relevance, and are defined as follows:

[0083] Where, d ij Represents node v i and v j Spatial distance, r ij It is node v i and v j Correlation coefficient of time series, w ij It is node v i and v j The edge weights between the distance and correlation are adjusted parameters, α and β, to balance the importance of distance and correlation.

[0084] Graph convolutional networks learn higher-order representations of node feature vectors:

[0085] Among them, H (l) It is the node feature matrix of the l-th layer, initialized as [F1, F2, ..., F n ], It is a weighted adjacency matrix, determined by the edge weights w. ij Build, It is the degree matrix of the weighted adjacency matrix. W (l) σ is the learnable weight matrix of the l-th layer, and σ is the activation function.

[0086] The globally fused marine hydrometeorological data is represented as follows:

[0087] F global=Aggregate(H (L) ), where F global It is a globally integrated representation of ocean hydrological and meteorological data, H (L) It is the feature matrix of the nodes in the last layer of the graph convolutional network. Aggregate(·) is an aggregation operation, such as global average pooling or max pooling.

[0088] The global representation generated by the graph convolutional network integrates the data features and interrelationships of all monitoring nodes into a unified high-dimensional vector, reflecting the spatiotemporal dynamic changes of the entire monitoring network. The global fusion representation preserves the spatiotemporal correlation information between monitoring nodes, ensuring that subsequent steps can handle sudden changes and long-term trends based on the context of the overall network.

[0089] In spatiotemporal consistency optimization, the global fusion representation provides a complete spatiotemporal feature vector. Based on these features, spatiotemporal consistency optimization can further classify and adjust weights to ensure the coordination between sudden and regular changes. The high-dimensional feature vectors in the global fusion representation can be combined with ocean physical laws (such as tidal and ocean current cycles) to provide specific data support for the constraints in spatiotemporal consistency optimization.

[0090] In the process of spatiotemporal consistency optimization, the input data of the multi-scale spatiotemporal attention mechanism is a global fusion representation generated by a graph convolutional network. These global features represent the spatial relationships and temporal dynamics of the monitoring network as a whole, ensuring that the optimization steps operate directly based on global context information.

[0091] The global fusion features already include the temporal characteristics (short-term and long-term trends) of each node. The multi-scale spatiotemporal attention mechanism further refines these temporal characteristics, dynamically adjusting the weights at different time scales to make the optimization results more closely reflect global dynamic changes. In the global fusion representation, the spatial relationships between nodes (correlation defined by edge weights) are passed to the spatiotemporal attention mechanism as the basis for calculating spatial attention weights. These global spatial relationships ensure that the distribution of influence among different nodes reflects the real ocean hydrological and meteorological distribution patterns during the optimization process. S3 includes multi-scale time window partitioning:

[0092] Based on the temporal characteristics of the monitoring data, the time series is divided into multiple time windows (short-term window and long-term window) to capture local short-term changes and global long-term trends, respectively. The length of the time window is dynamically adjusted and determined by the volatility of the data within the window. The higher the volatility, the shorter the window, and the lower the volatility, the longer the window.

[0093] Multi-scale spatiotemporal attention mechanisms include the calculation of temporal attention weights and spatial attention weights;

[0094] Time attention weight α tThe calculation is based on multi-scale time window partitioning, with each time window containing a data segment x. t The time attention weights are calculated as follows: Where, α t W represents the attention weight at time t. q W k It is a time-attention query and key mapping matrix, d k It is the dimension of the key vector;

[0095] Spatial attention weight β ij The calculation includes calculations for each monitoring node v i Calculate its data correlation with surrounding nodes: Where, β ij Represents node v i and v j Spatial attention weights between F i ,F j It is the node feature vector, W q W k It is a spatial attention query and key mapping matrix;

[0096] The classification and weighting adjustment of fused data through a spatiotemporal attention mechanism specifically includes:

[0097] Spatiotemporal feature fusion: incorporating temporal attention weight α t Spatial attention weight β ij Fusion to generate spatiotemporal joint weight γ ij,t γ ij,t =α t ·β ij , where γ ij,t This represents node v within the time t and spatial range. i With v j Joint weights, classification of sudden changes and regular changes:

[0098] Based on joint weight γ ij,t Dynamic fluctuations, setting a classification threshold θ:

[0099] If γ ij,t >θ indicates a sudden change in the data;

[0100] If γ ij,t If the value is less than or equal to θ, it is considered as data with normal variation.

[0101] Different weights are assigned to the classification results, with higher weights given to data with sudden changes to ensure that their dynamic change characteristics are captured.

[0102] The classification threshold θ should be dynamically adjusted based on the statistical analysis of historical data. The threshold, based on the standard deviation, is calculated using the following formula: θ = μ + k·σ, where μ is the joint weight γ.ij,t The mean of γ, where σ is the joint weight γ ij,t The standard deviation is given, and k is an adjustment factor (taken as 1.52, but can also be adjusted according to the sensitivity of the actual data).

[0103] Assigning classification weights:

[0104] Data with sudden changes: The weighting is relatively high, and w is set. 突发 =0.7 to 0.9, to ensure that sudden changes have a more significant impact on optimization;

[0105] Regularly changing data: The weighting is relatively low, so w is set... 常规 =0.3 to 0.5, used to maintain the background characteristics of regular data.

[0106] Optimizing the data fusion results by incorporating ocean physical characteristics as constraints specifically includes:

[0107] Periodicity embedding: Extracting the tidal and ocean current cycles of the monitoring area and constructing a constraint function: L physics =∑ t ‖F t -f physics (t)‖ 2 , of which F t This represents globally fused ocean hydrometeorological data over time t, belonging to F. global In spatiotemporal consistency optimization, to combine the dynamic characteristics of time with physical constraints, it is necessary to slice F into time slices. global Decompose the data along the time dimension, extract the global data slice corresponding to time t, and represent it as F. t f physics (t) is a physical law function that describes the periodic behavior of tides or ocean currents.

[0108] It also includes the overall optimization objective: after joint weight adjustment, ocean physical property constraints are added as a loss function to the overall optimization objective: L total =L attention +λ·L physics , where L attention It is the attention mechanism optimization loss, where λ is an adjustment parameter that controls the weight of the marine physical property constraints.

[0109] Based on the optimized spatiotemporal feature weights and classification adjustment results, a spatiotemporally consistent marine hydrological and meteorological data representation is generated for subsequent application scenarios.

[0110] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0111] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for fusion processing of marine hydrological and meteorological data based on deep learning, characterized in that, Includes the following steps: S1, Data Acquisition: Monitoring marine hydrological and meteorological data through a distributed marine monitoring network, which includes multiple monitoring nodes. Each monitoring node is equipped with observation equipment to collect hydrological and meteorological data, including temperature, salinity, ocean current, wind speed, air pressure, and precipitation. S2, Construction of Regional Collaborative Integration Network: S21, Node Feature Embedding: Generate node features from the hydrological and meteorological data collected within the coverage area of ​​each monitoring node, including the temporal and spatial features of the nodes, and characterize the temporal and spatial correlation of the data through the edge connections between nodes. The generation of the time features includes: extracting periodic, trend, and sudden change features from time-series hydrological and meteorological data collected within the coverage area of ​​the monitoring nodes; and modeling the time-series data using a long short-term memory network to generate time feature vectors for the nodes. The generation of the spatial features includes: based on the spatial distribution within the coverage area of ​​the monitoring nodes, using a convolutional neural network to extract the spatial change patterns of the data, and converting the spatial features, including local temperature gradients and ocean current diffusion paths, into spatial feature vectors to describe the local characteristics of the area where the nodes are located. Temporal and spatial features are concatenated, and a unified high-dimensional node feature representation is generated through dimensionality reduction embedding to construct the node feature vector for each monitoring node. S22, Edge weight definition: The weight of the edge is defined based on the spatial distance component and the data correlation component between the monitoring nodes; S23, Graph Convolutional Network Modeling: Based on the global topology of the monitoring network, graph convolutional networks are used to jointly model all nodes. Through convolution operations, node features and edge connections are comprehensively learned to generate a globally fused representation of marine hydrological and meteorological data. S3, Spatiotemporal Consistency Optimization: Introducing a multi-scale spatiotemporal attention mechanism to capture the simultaneous local short-term changes and global long-term trends in the data, dynamically adjusting the weights of different time scales, and classifying and weighting the fused data through the spatiotemporal attention mechanism to ensure the consistency of spatiotemporal consistency between sudden change data and regular change data; combining ocean physical characteristics as constraints to optimize the data fusion results, including the periodic patterns of tides and ocean currents; S3 also includes multi-scale time window partitioning: Based on the temporal characteristics of the monitoring data, the time series is divided into multiple time windows, which are used to capture local short-term changes and global long-term trends, respectively. The multi-scale spatiotemporal attention mechanism includes the calculation of temporal attention weights and spatial attention weights; The time attention weight α t The calculation is based on multi-scale time window partitioning, with each time window containing a data segment x. t ; The spatial attention weight β ij The calculation includes calculations for each monitoring node v i Calculate its data correlation with surrounding nodes; The classification and weighting adjustment of fused data through the spatiotemporal attention mechanism specifically includes: Spatiotemporal feature fusion: incorporating temporal attention weight α t Spatial attention weight β ij Fusion to generate spatiotemporal joint weight γ ij,t Based on joint weight γ ij,t Dynamic fluctuations, setting a classification threshold θ: If γ ij,t >θ indicates a sudden change in the data; If γ ij,t If the value is less than or equal to θ, it is considered a normally changing data point. Different weights are assigned to the classification results.

2. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 1, characterized in that, The monitoring nodes include deep-sea monitoring nodes, shallow-sea monitoring nodes, and near-shore monitoring nodes; Each monitoring node is equipped with observation equipment, including: Temperature sensors are used to monitor changes in seawater temperature. Salinity sensors are used to measure the salt concentration in water in real time. Current meters are used to monitor the speed and direction of ocean currents; Anemometers are used to collect data on wind speed and direction changes over the sea. A barometric pressure sensor is used to record changes in atmospheric pressure. A precipitation meter is used to measure precipitation in a sea area.

3. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 1, characterized in that, The spatial distance components are calculated using the spherical distance formula for node v. i and node v j Spatial distance; The data correlation component is calculated using the correlation coefficient at node v. i and v j Time series correlation; The edge weights, which combine spatial distance and data correlation, are defined as follows: Where, d ij Represents node v i and v j Spatial distance, r ij It is node v i and v j Correlation coefficient of time series, w ij It is node v i and v j The edge weights between the distance and correlation are adjusted parameters, α and β, to balance the importance of distance and correlation.

4. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 3, characterized in that, The graph convolutional network learns a high-order representation of the node feature vectors: Among them, H (l) It is the node feature matrix of the l-th layer, initialized as [F1, F2, ..., F n ], It is a weighted adjacency matrix, determined by the edge weights w. ij Build, It is the degree matrix of the weighted adjacency matrix. W (l) σ is the learnable weight matrix of the l-th layer, and σ is the activation function.

5. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 4, characterized in that, The globally fused marine hydrometeorological data is represented as follows: F global =Aggregate(H (L) ), where F global It is a globally integrated representation of ocean hydrological and meteorological data, H (L) It is the feature matrix of the nodes in the last layer of the graph convolutional network, and Aggregate(·) is the aggregation operation.

6. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 1, characterized in that, The optimization of the data fusion results by incorporating ocean physical characteristics as constraints specifically includes: Periodicity embedding: Extracting the tidal and ocean current cycles of the monitoring area and constructing a constraint function: L physics =Σ t ‖F t -f physics (t)‖ 2 , of which F t For globally fused ocean hydrometeorological data over time t, f physics (t) is a physical law function that describes the periodic behavior of tides or ocean currents.

7. The method for fusion processing of marine hydrological and meteorological data based on deep learning fusion according to claim 6, characterized in that, It also includes the overall optimization objective: after joint weight adjustment, ocean physical property constraints are added as a loss function to the overall optimization objective: L total =L attention +λ·L physics , where L attention It is the attention mechanism optimization loss, where λ is an adjustment parameter that controls the weight of the marine physical property constraints.

Images