Ocean current space-time sequence reconstruction and prediction method and system based on multi-source data fusion

Abstract

The invention belongs to the technical field of marine environment monitoring and prediction, and discloses an ocean current space-time sequence reconstruction and prediction method and system based on multi-source data fusion. The method comprises the following steps: firstly, acquiring and preprocessing multi-source data such as satellite remote sensing, field observation and a numerical model; through a multi-modal feature fusion network fusing a three-dimensional convolutional neural network (3D-CNN), a long short-term memory network (LSTM) and an attention mechanism, spatial-temporal features of multi-source data are extracted, and adaptive weighted fusion is carried out; and then the fusion features are input into a space-time diagram neural network (ST-GNN), the network explicitly models a non-Euclidean spatial relationship in an ocean region through a diagram structure, and is coupled with a space-time convolution operation, so that the space-time evolution law of ocean current is learned with high precision, and an ocean current sequence prediction result in a future time period is output. The method realizes high-precision and high-efficiency ocean current prediction, and is widely applied to the fields of ocean environment monitoring, shipping safety guarantee, climate research and the like.

Description

Technical Field

[0001] This invention belongs to the field of marine environmental monitoring and prediction technology, and in particular relates to a high-precision reconstruction and prediction method for the spatiotemporal sequence of ocean current velocity and direction that combines multi-source marine observation data, satellite remote sensing data and artificial intelligence models. Background Technology

[0002] Ocean currents are large-scale, directional water masses in the ocean that have a crucial impact on global climate regulation, shipping safety, fisheries resources, pollutant dispersion, and military activities. Therefore, achieving accurate prediction of the spatiotemporal variations of ocean currents has significant scientific value and practical implications.

[0003] Currently, ocean current prediction technologies are mainly divided into two categories: physical numerical simulation methods and data-driven methods.

[0004] The first category comprises numerical simulation methods based on physical mechanisms, such as the Regional Ocean Model System (ROMS) and the Hybrid Coordinate Ocean Model (HYCOM). These methods are based on the Navier-Stokes hydrodynamic equations and are solved numerically through discretized grids. However, these methods suffer from the following significant technical drawbacks in practical applications: First, they consume enormous computational resources; high-resolution simulations rely on high-performance computing clusters, making it difficult to meet the real-time requirements of tasks such as maritime emergency search and rescue. Second, the uncertainty of the physical parameterization scheme leads to error accumulation due to the model's approximate handling of complex processes such as turbulent mixing. Third, there is a bottleneck in data assimilation efficiency; when processing multi-source, heterogeneous, and asynchronous ocean observation data, it is difficult to achieve efficient fusion within a limited time window.

[0005] The second category comprises data-driven methods based on artificial intelligence that have emerged in recent years. These methods primarily utilize convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) to uncover patterns in historical data. This is the closest existing technology to this application. Although such methods reduce computational costs, they still suffer from the following unresolved technical shortcomings when dealing with complex ocean current spatiotemporal evolution problems: Constraints of data quality and resolution (lack of reconstruction mechanism): Existing AI prediction models are usually trained directly on satellite remote sensing data (which may be missing due to cloud cover) or low-resolution reanalysis data. Due to the lack of an effective "super-resolution reconstruction" step, the model can only learn in a low-dimensional and coarse data space, which often results in the loss of key high-frequency dynamic details such as mesoscale eddies and fronts, and makes it difficult to handle the spatiotemporal discontinuities of observation data.

[0006] The lack of deep fusion capabilities for multi-source heterogeneous data: Most existing models rely on only a single data source or simply stitch together channels of multi-source data. Due to the lack of a dedicated fusion network for the heterogeneous characteristics of point observations (such as Argo) and area observations (such as satellite data), it is impossible to effectively extract the nonlinear coupling features between different environmental elements (such as wind fields and flow fields), which limits the upper limit of prediction accuracy.

[0007] Modeling the non-Euclidean topology of the ocean fails: Traditional CNN models perform convolutions based on regular grids (Euclidean space), assuming translation invariance. However, ocean regions are physically constrained by irregular coastlines and islands, exhibiting typical non-Euclidean topological characteristics. When using standard convolutions to process the land-sea boundary, the introduced padding noise significantly increases the prediction error in nearshore areas.

[0008] In summary, overcoming the high computational cost of traditional numerical models while addressing the technical bottlenecks of existing AI methods, such as low data quality, insufficient heterogeneous fusion, and limited spatial modeling, are key technical issues that urgently need to be addressed in this field.

[0009] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) The "smoothing effect" and loss of high-frequency details in data-driven models stem from the lack of a high-quality reconstruction mechanism: Existing ocean current prediction methods often ignore the quality bottleneck of the input data itself. Due to the low resolution of widely used reanalysis data and the large amount of spatiotemporal gaps in satellite observation data, models trained directly on such data tend to learn low-frequency average states. The fundamental reason is that the existing technology system lacks the key preliminary step of "super-resolution reconstruction," making it impossible to restore high-resolution flow field details from low-quality data through generative adversarial networks or physical constraint mechanisms. This results in blurry prediction results in high-gradient regions such as vortex edges and abrupt flow field changes, failing to meet the requirements for refined prediction.

[0010] (2) Shallow stitching of multi-source heterogeneous data leads to low utilization of physical coupling features: Although some existing technologies attempt to introduce multi-source data, most adopt simple linear interpolation or channel stitching methods. The technical defect is that it fails to solve the essential heterogeneity in spatial topology between sparse, discrete vertical point observation data (such as Argo) and continuous, planar satellite remote sensing data. This crude processing method makes it impossible for the model to learn the complex nonlinear driving mechanism of wind stress and temperature-salinity structure on surface ocean currents, resulting in the potential value of massive multi-source observation data failing to be converted into improved prediction accuracy.

[0011] (3) Traditional convolutional architectures have poor adaptability in "non-Euclidean space", leading to distortion in boundary prediction: Mainstream CNN or ConvLSTM models are designed based on regular grids in Euclidean space. Their core defect is that they cannot explicitly model the complex geometric topology of the ocean composed of islands, ridges and irregular coastlines. When dealing with land-ocean boundaries, traditional convolutional kernels must be zero-filled, which introduces non-physical boundary noise, resulting in significant numerical distortion in the flow field prediction in nearshore waters or narrow straits, which seriously affects the applicability of the model in complex geographical environments. Summary of the Invention

[0012] To overcome the problems existing in related technologies, the present invention discloses an embodiment that provides a method and system capable of efficiently fusing multi-source heterogeneous data and accurately capturing the spatiotemporal evolution patterns of ocean currents using a deep learning model, thereby achieving high-precision, long-term ocean current prediction. This provides ocean current data-driven support for underwater search and rescue of lost equipment. The technical solution is as follows: This invention is implemented as follows: a method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion, comprising the following steps: S1. Data Acquisition and Preprocessing: Collect multi-source ocean current related data, including satellite remote sensing data, field observation data, and numerical model reanalysis data, and perform spatiotemporal alignment and missing value processing to form a unified spatiotemporal dataset. S2. Multimodal feature fusion: Construct a multimodal feature fusion network, extract spatial features of raster data through a three-dimensional convolutional neural network, extract dynamic evolution features of time series data through a recurrent neural network, and fuse the spatial features, dynamic evolution features and point observation data features to generate a fusion feature tensor containing multi-source complementary information; S3. Spatiotemporal resolution reconstruction: The fused feature tensor is input into a pre-trained generative ocean current field super-resolution reconstruction network. Through an encoder-decoder structure and adversarial training strategy, a high spatiotemporal resolution and complete historical ocean current field sequence is generated. S4. Spatiotemporal sequence prediction: Construct a spatiotemporal graph neural network model, define the grid points of the prediction area as graph nodes and construct edges based on marine geographic information; take the high spatiotemporal resolution historical ocean current field sequence as the core input and the fused feature tensor as the auxiliary feature input, perform spatiotemporal coupling modeling through the spatiotemporal graph neural network model, and output the ocean current spatiotemporal sequence prediction results for a specified future time period.

[0013] Furthermore, the ocean current field super-resolution reconstruction network in the spatiotemporal resolution reconstruction step is a generative model based on deep learning. Its structure is a U-Net architecture that includes an encoder and a decoder, and it uses skip connections to pass the high-frequency features of the encoder to the decoder. The training process of the reconstructed network includes an adversarial training strategy, and its loss function includes at least pixel-level reconstruction loss and adversarial loss generated by the discriminator network.

[0014] Furthermore, the loss function of the reconstructed network includes gradient difference loss, which is used to constrain the consistency of the reconstructed ocean current field with the high-resolution real field in terms of spatial gradient.

[0015] Furthermore, the multimodal feature fusion step includes: A multimodal feature fusion network is constructed. A three-dimensional convolutional neural network is used to extract spatial features from raster data, and a recurrent neural network is used to extract dynamic evolution features from time-series data. An attention mechanism is used to adaptively weight and fuse the spatial features, dynamic evolution features, and embedded features from point observation data to generate a fusion feature tensor for auxiliary prediction.

[0016] Furthermore, the spatiotemporal sequence prediction step includes: The high spatiotemporal resolution historical ocean current field sequence is used as the core input; A spatiotemporal graph structure is constructed, high-resolution grid points in the predicted area are defined as graph nodes, and edges between nodes are constructed based on marine geographic information and physical relationships. Using the aforementioned spatiotemporal graph neural network model, the core and auxiliary features of the input are spatiotemporally coupled and modeled through alternating spatial graph convolution and temporal convolution operations, thereby completing the prediction.

[0017] Furthermore, the spatial graph convolution in the spatiotemporal graph neural network model employs a graph attention network to dynamically learn the weights of spatial dependencies between nodes; the temporal convolution employs a temporal convolutional network or a gated recurrent unit.

[0018] Furthermore, the method also includes a model training step, which is divided into two stages: In the first stage, the ocean current field super-resolution reconstruction network is trained separately; In the second stage, the pre-trained reconstruction network is connected to the spatiotemporal graph neural network to jointly fine-tune the overall ocean current spatiotemporal sequence reconstruction and prediction system in an end-to-end manner.

[0019] Another objective of this invention is to provide a system for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion, for implementing the aforementioned method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion, characterized in that the system comprises: The data acquisition and preprocessing module is used to acquire and preprocess multi-source ocean current related data; The spatiotemporal resolution reconstruction module has a built-in pre-trained ocean current field super-resolution reconstruction network, which is used to generate a complete historical ocean current field with high spatiotemporal resolution. The spatiotemporal prediction engine module has a built-in pre-trained spatiotemporal graph neural network model, which is used to predict future ocean currents based on the historical ocean current field and multi-source features. The results visualization and publishing module is used to visualize and publish the reconstructed historical ocean current fields and the predicted future ocean current fields.

[0020] Furthermore, the system also includes a multi-source data fusion module with a built-in multimodal feature fusion network, which is used to extract and fuse features from the preprocessed multi-source data, and provides the generated fused feature tensors to the spatiotemporal resolution reconstruction module and the spatiotemporal prediction engine module, respectively.

[0021] Furthermore, the results visualization and publishing module is used to display historical reconstructed flow fields and future predicted flow fields in parallel, and provides dynamic backtracking and forward-looking animation functions.

[0022] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention adopts a multi-layered, end-to-end deep learning framework. Its core idea is as follows: First, through a carefully designed feature fusion network, the advantages of multi-source heterogeneous data are deeply fused in a "complementary" manner; then, a spatiotemporal graph neural network that can explicitly express the complex relationships in ocean space is used to learn and predict the spatiotemporal dynamics of ocean currents from the fused high-quality features.

[0023] This invention first collects and preprocesses multi-source data, including satellite remote sensing, on-site observations, and numerical models. Then, it extracts spatiotemporal features from the multi-source data and performs adaptive weighted fusion by integrating a multimodal feature fusion network that combines a 3D convolutional neural network (3D-CNN), a long short-term memory network (LSTM), and an attention mechanism. Subsequently, the fused features are input into a spatiotemporal graph neural network (ST-GNN). This network explicitly models non-Euclidean spatial relationships in the ocean region through a graph structure and couples spatiotemporal convolution operations, thereby learning the spatiotemporal evolution patterns of ocean currents with high accuracy and outputting predictions of future ocean current sequences. This invention also provides a corresponding system implementation. This method effectively overcomes the limitations of high computational costs in traditional numerical simulations and insufficient feature utilization in existing artificial intelligence methods, achieving high-precision and high-efficiency ocean current prediction. It can be widely applied in fields such as marine environmental monitoring, shipping safety assurance, and climate research.

[0024] Secondly, by combining the above methods and systems, this invention constructs a complete, automated, and intelligent ocean current prediction solution from data to decision-making, which has the following beneficial effects: (1) This invention solves the prediction bias problem caused by low-quality data sources and significantly improves the spatiotemporal continuity and physical consistency of the data: Before prediction, this invention introduces an ocean current field super-resolution reconstruction network, using a generative adversarial network based on the U-Net architecture and a gradient difference loss function. This technique can reconstruct historical ocean current fields with high spatiotemporal resolution and complete physical structure (such as retaining the gradient at the edge of eddies) from satellite remote sensing and reanalysis data that have cloud cover, low resolution, or other issues. This not only overcomes the problems of flow field smoothing and loss of details caused by traditional interpolation methods, but also provides a high-quality, low-noise input benchmark for subsequent spatiotemporal sequence prediction models, ensuring the robustness of the prediction system from the data source.

[0025] (2) This invention overcomes the technical bottleneck of deep coupling of multi-source heterogeneous data and significantly improves the utilization rate of observation information: The invention constructs a multimodal feature fusion network, uses a three-dimensional convolutional neural network (3D-CNN) to extract the spatial features of raster data, uses a recurrent neural network (RNN) to extract the dynamic evolution features of time-series data such as wind fields, and introduces point observation embedding and attention mechanisms. This technique effectively solves the problem of heterogeneity between sparse point data (such as Argo) and continuous areal data in spatial topology, and realizes the adaptive weighted fusion of nonlinear driving mechanisms between different physical elements (current, temperature, wind), thereby enabling the prediction model to capture complex ocean dynamic coupling features that cannot be reflected by a single data source.

[0026] (3) Overcoming the problem of ocean current boundary prediction distortion in non-Euclidean space, achieving refined forecasting of complex sea areas: This invention adopts a spatiotemporal graph neural network (ST-GNN) as the core prediction architecture, explicitly modeling the prediction area grid as graph nodes, and constructing edge connections in conjunction with marine geographic information. This technique abandons the regular convolution operation of traditional convolutional neural networks (CNN) based on the Euclidean space assumption, avoiding non-physical boundary noise introduced by zero-padding when dealing with irregular coastlines and island boundaries. This makes the model's flow field prediction accuracy significantly better than traditional methods in complex geographical environments such as nearshore areas, straits, and island peripheries, and can accurately capture the fluid motion patterns constrained by topography.

[0027] (4) End-to-end collaborative optimization integrating "reconstruction-prediction" is achieved, enhancing the overall generalization ability of the system: This invention proposes a two-stage joint fine-tuning training strategy, cascading the super-resolution reconstruction network with the spatiotemporal graph neural network. This end-to-end architecture design enables the reconstruction module not only to restore the historical field, but also to dynamically adjust the feature extraction strategy according to the gradient feedback of the prediction task, ensuring that the reconstructed flow field features are most conducive to subsequent spatiotemporal extrapolation. This mechanism effectively avoids the error accumulation caused by step-by-step processing, providing a complete solution for marine environmental monitoring from raw multi-source data input to future high-precision flow field output. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the ocean current spatiotemporal sequence reconstruction and prediction method based on multi-source data fusion provided in this embodiment of the invention; Figure 2 This is a diagram of the Generative Adversarial Network (GAN) framework provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the ocean current time prediction principle provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the ocean current spatial prediction principle provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the online ocean current prediction results provided in an embodiment of the present invention; Figure 6 This is a comparison chart of numerical and experimental results for the Bohai Sea area provided in this embodiment of the invention; Figure 7 This is a comparison chart of numerical data and experimental results for the East China Sea provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0030] Example 1, as Figure 1As shown in the embodiments of the present invention, the ocean current spatiotemporal sequence reconstruction and prediction method based on multi-source data fusion includes the following core processes: multi-source data acquisition and refined preprocessing (S1), spatiotemporal resolution reconstruction of ocean current data (S2), multimodal spatiotemporal feature fusion based on attention mechanism (S3), core prediction based on spatiotemporal graph neural network (S4), end-to-end model training and optimization (S5), and online prediction and visualization service (S6). The following provides a detailed description of each step: S1: Multi-source data acquisition and refined preprocessing.

[0031] This step aims to address the heterogeneity, inconsistency, and incompleteness of multi-source data, providing high-quality, standardized input for subsequent models.

[0032] Data Acquisition: Systematically collect three types of core data: Satellite remote sensing data includes, but is not limited to, sea surface height (SSH) or sea level anomaly (SLA), sea surface temperature (SST), sea surface wind field (including wind speed and direction), and seawater chlorophyll concentration. This data provides large-scale, near-synchronous sea surface information.

[0033] Field observation data includes, but is not limited to, temperature, salinity, and depth profile data provided by Argo buoys, drifting buoy trajectory data, shipborne ADCP (Acoustic Doppler Current Profiler) measurement data, and surface flow field data acquired by coastal high-frequency ground wave radar. This data provides valuable vertical profile information and near-real-time verification.

[0034] Numerical model reanalysis data: Obtain historical reanalysis data that has been assimilated from authoritative institutions (such as HYCOM, NCEP) as a supplement and benchmark for model training.

[0035] Table 1 Information from domestic flow field databases

[0036] Table 2 Information from foreign flow field databases

[0037] Refined preprocessing: Spatiotemporal alignment: This involves uniformly interpolating or resampling all data to a common spatial grid (such as a latitude and longitude grid) and a uniform temporal frequency (such as a daily average). This resolves the differences in spatial resolution and temporal sampling rate between different data sources.

[0038] Missing value handling: To address the missing data in satellite data caused by cloud cover and the sparse distribution of field data, advanced interpolation methods (such as the empirical orthogonal function DINEOF) or data imputation models based on generative adversarial networks (GANs) are used to reconstruct the data and ensure the continuity of the data space.

[0039] Standardization / Normalization: Standardize each physical variable (such as SST, wind speed, and flow velocity) to make its mean 0 and standard deviation 1, in order to accelerate model training convergence and improve performance.

[0040] Satellite remote sensing data (SSH, SST, wind field), field observation data (Argo, buoy, radar), and numerical model reanalysis data were collected. Preprocessing included spatiotemporal alignment, missing value labeling (rather than simple imputation), and standardization, ultimately forming a unified dataset for model input.

[0041] S2: Spatiotemporal resolution reconstruction of ocean current data: generating a high-resolution, complete reference field.

[0042] This step aims to address the problems of spatial sparsity, temporal discontinuity, and low overall resolution of the original observation data, providing a near-realistic, high-quality starting point for subsequent predictions. This invention constructs an ocean current field super-resolution reconstruction network, which is an encoder-decoder structure and incorporates an adversarial training strategy.

[0043] Input and output: Input: Low-resolution, multi-source fusion features with missing values ​​(from preprocessed data of S1, but with their original resolution preserved) and their corresponding spatial missing mask.

[0044] Output: A high-resolution, complete ocean current field (U and V components) with improved spatial resolution (from 1°×1° to 0.25°×0.25°), continuous temporal resolution, and no missing data.

[0045] Network structure and process: Feature Encoding: A hierarchical spatiotemporal feature extraction encoder is constructed, consisting of multiple feature extraction blocks concatenated. Each feature extraction block contains, sequentially, a 3D convolutional layer (3D-CNN), a non-linear activation layer, and a downsampling layer. Specifically, the 3D-CNN layer uses a size of... The (time × height × width) 3D convolution kernel slides simultaneously across the spatiotemporal dimensions of the input data to explicitly capture the dynamic evolution patterns between adjacent time steps and the local correlations within the spatial neighborhood of ocean current data. After the convolution operation, a batch normalization layer is connected to a ReLU or Leaky ReLU activation layer to accelerate network convergence, prevent gradient vanishing, and endow the network with the ability to process nonlinear ocean dynamic features. At the end, max pooling or strided convolution with a stride of 2 is used for downsampling, gradually halving the spatial resolution of the feature map. This reduces the computational load while multiplying the receptive field of subsequent network layers, enabling the extraction of deep abstract feature maps covering a larger spatial range (such as the overall topological structure of mesoscale eddies) from low-resolution input.

[0046] Feature Fusion and Enhancement: An improved adaptive spatiotemporal dual attention gating mechanism is introduced at the bottleneck layer of the encoder. This aims to address the problem that traditional stitching and fusion methods cannot distinguish the differences in feature contributions from multiple data sources (such as sea surface temperature and wind fields), and to enhance focus on key marine dynamic regions. The specific processing logic of this mechanism is as follows: the input deep feature tensor... The process proceeds in parallel through channel attention and spatial attention branches. In the channel dimension, global average pooling and max pooling are used to aggregate spatial information, and a multilayer perceptron (MLP) is used to learn the channel weight vectors for different physical variables. The calculation formula is: This automatically suppresses data channels with significant noise interference. In the spatial dimension, spatial descriptors are generated through pooling operations along the channel axis, and large-kernel convolution is used to extract local correlations to generate a spatial saliency mask. The calculation formula is: This allows for precise location of the vortex edge and front position; finally, the weighted features are fused with the original features using a residual connection method. (in, It is the Sigmoid activation function. (For element-wise multiplication), the output information-enhancing feature tensor can guide the decoder to prioritize the recovery of high-frequency dynamic details, significantly improving the physical consistency of the flow field reconstruction.

[0047] High-resolution decoding and generation: A cascaded decoder consisting of multi-layer sub-pixel convolutional networks is constructed to progressively restore attention-enhanced deep abstract features into ocean current physical fields with high spatiotemporal resolution. The specific processing logic of the decoder is as follows: it adopts an efficient upsampling strategy of "3D convolution + subpixel rearrangement", which enlarges the spatial dimension of the feature map through step-by-step feature mapping, and adjusts the number of channels to match the output requirements. In this process, in order to effectively solve the problem of blurred flow field boundaries caused by the loss of spatial information in multiple downsampling operations in deep networks, a feature reuse strategy based on skip connections is implemented: the low-level high-frequency feature maps containing rich texture and edge information retained in each level of the encoder are directly passed to the corresponding level of the decoder through lateral connections, and are spliced ​​and fused with the current decoded features in the channel dimension. This multi-scale feature sequence that integrates deep semantics and shallow details is finally directly mapped into high spatiotemporal resolution eastward (U) and northward (V) velocity components after nonlinear activation and the 3D convolutional reconstruction layer at the end, thereby accurately reconstructing fine fluid dynamic structures such as vortex edge gradients and front positions while ensuring numerical accuracy.

[0048] Adversarial training to enhance physical realism: To effectively overcome the technical shortcomings of traditional mean square error loss functions, which easily lead to the loss of high-frequency details in the reconstructed flow field and excessive smoothing during optimization, this step introduces a generative adversarial network (GAN) framework (such as...) during the model training phase. Figure 2 As shown in the system flow diagram, an end-to-end joint adversarial training is conducted by constructing a discriminator network based on the PatchGAN structure and a reconstruction network acting as a generator. Specifically, the discriminator receives paired inputs of "generated high-resolution reconstruction field" and "real HYCOM high-resolution reanalysis field" and outputs a probability matrix representing the realism of local regions. During training, the generator and discriminator execute an alternating optimization zero-sum game strategy. That is, the discriminator aims to maximize its ability to distinguish between generated and real flow fields, while the generator strives to generate a realistic flow field that the discriminator can classify as a real sample. Through this adversarial gradient feedback mechanism, the generator is forced to deeply learn the high-order statistical distribution features and physical laws implicit in real ocean current fields (such as the edge sharpness of vortex structures, the continuity of flow field gradients, and energy cascade features). This essentially imposes a data distribution-based regularization constraint on the flow field, ultimately making the generated reconstruction field not only close to the true value in numerical accuracy but also highly consistent with the real physical ocean environment in terms of topological structure and texture details.

[0049] Training the reconstructed network: This network is pre-trained using historical high-resolution reanalysis data as supervised labels, and its loss function incorporates: Pixel-level reconstruction loss ( ): As a fundamental constraint term, it is calculated using the mean square error formula, i.e. ,in Represents a true high-resolution flow field. Represents the generated reconstructed flow field. The number of samples is denoted by . This term aims to minimize the Euclidean distance between the generated field and the real field in terms of pixel values, ensuring that the reconstruction result possesses basic numerical accuracy and convergence stability.

[0050] Combating loss ( As a term enhancing physical realism, it adopts the loss form of least-squares generative adversarial networks, i.e. ,in This represents the output probability of the discriminator. Its purpose is to drive the generator to approximate the higher-order statistical distribution of real ocean currents by minimizing the difference between the discriminator's output probability and the target true value, thereby effectively suppressing the smoothing effect of the generated results and improving the physical realism of the flow field texture.

[0051] Gradient difference loss ( ): As a specific improvement proposed in this invention for fluid dynamics characteristics, its formula is defined as follows: ,in and These represent the spatial gradient operators along the longitude and latitude directions, respectively. It is the L1 norm. This term imposes a constraint in the gradient domain rather than just the pixel domain, forcing the reconstructed flow field to maintain the same rate of change in space as the real flow field, thereby accurately preserving key high-frequency physical structures such as vortex edges and fronts.

[0052] S3: Multimodal spatiotemporal feature fusion based on attention mechanism.

[0053] This step is crucial for improving prediction accuracy. Its goal is to integrate features from different sensors and with different physical meanings into a unified feature representation that is information-rich and complementary. This invention constructs a multimodal feature fusion network, the structure of which is as follows: Parallel feature extraction channels: Spatial Feature Extraction Branch: For continuous spatiotemporal data with regular grid structures, such as sea surface temperature (SST) and sea surface height (SSH), a 3D convolutional neural network (3D-CNN) module is constructed to perform deep feature extraction. The specific implementation logic of this extraction process is as follows: using a preset size (e.g., The 3D convolution kernel performs stride-sliding convolution operations on the spatiotemporal tensor of the input data. Specifically, regarding the process of capturing spatial neighborhood correlations and short-term evolution patterns: the convolution kernel first uses its planar receptive field in the spatial dimension (height × width) to weight and aggregate grid point data within the local neighborhood, thereby identifying static spatial structures in the flow field (such as closed contour lines of mesoscale eddies and strong gradient bands of fronts); simultaneously, the convolution kernel utilizes its depth extension in the temporal dimension to perform joint convolution across multiple consecutive time steps of the input sequence. By calculating the differences in numerical changes at the same spatial location and neighborhood at adjacent times, the dynamic drift trend and intensity evolution rate of the ocean current are encoded into the output feature map, thereby achieving integrated extraction of the spatiotemporal coupling features of ocean dynamic processes. Specifically, the mathematical model of the 3D convolution operation can be expressed as: ; In the formula, Indicates the first Layer in spacetime Eigenvalues ​​at; For bias terms; To connect the first Layer The first feature map and the second The weight coefficients of the convolutional kernel; These represent the dimensions of the 3D convolutional kernel in the time, height, and width dimensions, respectively. This formula indicates that the model extracts features by simultaneously weighting and summing across the three spatiotemporal dimensions.

[0054] Temporal Feature Extraction Branch: For dynamic environmental data with strong temporal driving characteristics, such as sea surface wind fields, a bidirectional long short-term memory network (Bi-LSTM) or gated recurrent unit (GRU) model is constructed to perform deep encoding of temporal dynamic features. The specific implementation logic of this capture process is as follows: the wind speed vector sequence of continuous time steps is mapped into a high-dimensional hidden state vector, and the historical state information is adaptively filtered and updated using the unique gating mechanism inside the network (such as the forget gate and input gate of LSTM, or the update gate and reset gate of GRU). Specifically, the forward hidden layer is responsible for accumulating the input effect of wind stress along the time axis, simulating the continuous dragging and momentum transfer process of wind on the surface seawater, thereby accurately capturing the inertial lag effect of wind-generated current (i.e., the current flow velocity is affected by the cumulative effect of the wind field over a period of time); the backward hidden layer scans the temporal data in reverse, using the evolution of the flow field at future moments to infer the weight of the current wind field to enhance the robustness of the features. Finally, the hidden state vectors generated from the forward and backward directions are concatenated or weighted and summed along the channel dimension to fully construct a temporal feature representation that can characterize the long-term, nonlinear, and continuous driving effect of wind stress on ocean currents. Taking the gated cyclic unit (GRU) as an example, its core time step update logic is defined by the following state equation: ; In the formula, To update the gate, control the degree to which historical state information is retained; To reset the gate and control the degree of forgetting of historical information; Output the hidden state at the current moment; Input the wind field at the current moment; This is the Sigmoid activation function. This set of formulas demonstrates how the network filters long-term wind stress history information through a gating mechanism.

[0055] Point observation feature embedding branch: For point observation data such as Argo buoys, which exhibit sparse and unstructured spatial distribution, a multilayer perceptron (MLP) or a fully connected neural network (Dense Network) is constructed to perform feature space projection of discrete physical information. The specific implementation logic of this mapping process is as follows: First, the multivariate physical attributes of each observation point (including but not limited to temperature) are embedded into the feature space. ,salinity ,pressure Combined into an initial input vector Subsequently, through the learnable weight matrix between network layers... With bias vector Perform linear transformation And via a nonlinear activation function Processing methods such as ReLU map the data from a low-dimensional physical space to a high-dimensional latent feature space, generating feature vectors containing rich semantic information. Building upon this, to align discrete features with a regular grid, an adaptive weighted aggregation algorithm based on spatial distance is used for feature assimilation when multiple observation points fall within the same grid cell. The calculation principle is described as follows: Let the coordinates of the target grid center be... The grid contains The observation point, the first The feature vector of each point is Its Euclidean distance from the center of the grid is Then the final embedding feature of this grid node. The following is calculated using the inverse distance weighted formula: ; Among them, the weighting coefficient ( To prevent the use of tiny constants with a denominator of zero, this mechanism can automatically assign information weights based on the spatial proximity of the observation locations, ensuring that discrete observation data can be accurately embedded into the background field grid in a manner consistent with the first law of geography, effectively solving the technical problem of spatial structure mismatch in multi-source data.

[0056] Adaptive Feature Fusion: Performs dimensional alignment and depth concatenation of multi-source heterogeneous features. Spatiotemporal dimensional standardization is performed on the feature tensors output by each branch. For the one-dimensional feature vectors output by the temporal branch, spatial broadcasting is used to expand them to a size matching the raster feature map. Subsequently, a tensor concatenation operator is used to linearly stack spatial texture features, temporal dynamic features, and point observation embedding features along the channel dimension, constructing a joint feature tensor containing comprehensive physical information. To achieve accurate fusion of multi-source features, this invention proposes an improved channel-space dual attention fusion algorithm. This algorithm differs from traditional methods that only use single pooling by introducing multi-scale pooling aggregation and a large-kernel convolution strategy to enhance feature representation. Specifically, in the channel dimension, it simultaneously utilizes global average pooling (… ) and global max pooling ( This is used to capture background information and extreme textures of features, and its calculation formula is as follows: In the formula The Sigmoid activation function is used, and MLP is a multilayer perceptron with shared weights. This formula enables the model to adaptively identify and amplify high-contribution physical variable channels (such as SSH). In the spatial dimension, to focus on key dynamic regions, a spatial saliency mask is constructed based on large-kernel convolution, and its calculation formula is as follows: In the formula The kernel size represents the convolution kernel size. The convolutional operation is used to extract a wide range of spatial contextual relationships; finally, the original multi-source features are refined and reconstructed through dual attention weights, and the fusion output formula is as follows: In the formula Broadcast multiplication of the representative tensor generates an attention-weighted, unified fusion feature tensor. .

[0057] At the channel dimension: feature recalibration based on global context information is performed. The specific process of automatic learning and assignment of these weights is as follows: First, the input joint feature tensor... Global average pooling is performed to aggregate the values ​​of each channel across all spatial locations into a global statistical scalar, generating a compressed channel descriptor vector. This vector is then input into a multilayer perceptron (MLP) that incorporates dimensionality reduction and expansion operations. Nonlinear transformations between layers are used to explicitly model the interdependencies between channels with different physical variables (such as sea surface temperature and wind field channels). Next, the output of the MLP is mapped to normalized weight coefficients between 0 and 1 using a sigmoid activation function. Finally, these weight coefficients are multiplied channel-by-channel with the original feature tensor, thereby adaptively amplifying the channel responses containing key dynamic information (such as SSH features in geostrophic regions) while suppressing the expression of channels with significant noise interference or irrelevant features.

[0058] In the spatial dimension: Perform feature enhancement based on spatial saliency. The specific implementation logic of this dynamic focusing is as follows: First, in order to capture the importance differences of different spatial locations in the flow field, the input joint feature tensor is... Average pooling and max pooling operations are performed along the channel axis to generate two two-dimensional spatial descriptor maps, thereby compressing high-dimensional features into a planar graph containing spatial statistics. Subsequently, these two descriptors are concatenated along the channel dimension and fed into a large-kernel convolutional layer (e.g., ...). The convolution kernel extracts contextual features (such as gradient continuity at the edge of a vortex) within the spatial neighborhood using its large planar receptive field. Then, a spatial saliency mask with a value between 0 and 1 is generated by the Sigmoid activation function. Finally, the mask is multiplied element-wise with the original feature tensor to suppress background ocean noise (such as the steady current region of the open ocean) at the spatial pixel level and highlight key dynamic structures that have a decisive impact on the prediction target (such as the frontal region or vortex center of the Kuroshio extension).

[0059] Finally, an attention-weighted, unified fusion feature tensor is output.

[0060] S4: Core prediction based on spatiotemporal graph neural network (ST-GNN).

[0061] This step is the core innovation of the present invention, aiming to explicitly model the complex non-Euclidean spatial dependencies in ocean current fields caused by islands, ridges, and irregular boundaries using graph theory. Specifically, a depth prediction network is constructed, consisting of a dynamic graph structure generation module and multi-layer spatiotemporal graph convolutional blocks.

[0062] Dynamic Graph Structure Construction: To discretize the irregular ocean space into a graph structure In this invention, each high-resolution grid point within the prediction region is defined as a graph node. The total number of nodes is To introduce physical constraints into the graph, a weighted adjacency matrix is ​​constructed. The weight of its edges The calculation formula is as follows: ; In the formula, Indicates the great circle distance between nodes; Represents a node of A set of nearest neighbors (in this embodiment) ); This means that if the connection between two nodes crosses land or islands, the connection is forcibly disconnected (i.e., the weight is reset to 0), thus eliminating non-real cross-land correlations through this physical constraint; It is a tiny constant.

[0063] Spatiotemporal graph convolution operation: This invention constructs a deep network consisting of multiple spatiotemporal graph convolutional blocks (ST-Conv Blocks) stacked together.

[0064] Spatial graph convolution operation: Within each ST-Conv Block, spatial feature aggregation is first performed using a graph attention network (GAT) to dynamically learn the anisotropic spatial dependencies between nodes. Its core computational process includes three steps: feature transformation, attention coefficient calculation, and weighted aggregation. The specific mathematical model is as follows: (1) : ; (2) : ; (3) : ; In the formula, For the first Layer nodes Input features, The weight matrix is ​​a learnable matrix. For attention weight vectors, This represents a vector concatenation operation. For nodes The neighborhood set, For non-linear activation functions (such as ELU), parallel computation is achieved through a multi-head attention mechanism. The above processes are combined and spliced ​​together to capture spatial dynamic patterns at different scales.

[0065] Temporal convolution operation: After spatial aggregation, dilated causal convolution is performed on the temporal series features of each node to capture multi-scale temporal evolution patterns. For a one-dimensional input sequence... and convolution kernel At any moment The output of a dilated convolution is defined as: ; In the formula, The dilation factor increases exponentially with the number of network layers. ); The kernel size; This represents the dilated convolution operator. The formula demonstrates that the model can effectively capture multi-scale temporal dependencies, ranging from semi-diurnal tides to seasonal oscillations, through an exponentially expanded temporal receptive field, without increasing the number of parameters.

[0066] Residual Connections and Output: To ensure effective gradient propagation, each convolutional block employs a residual connection structure, with the output formula as follows: Finally, the deep spatiotemporal features extracted by the stacked ST-Conv Blocks are linearly decoded through a point-based fully connected layer: ; Thus directly generating the future A time-step ocean current velocity vector prediction sequence, covering zonal current velocities. and meridional velocity Two physical components.

[0067] S5: End-to-end model training and optimization.

[0068] Loss function: A composite loss function is adopted, which integrates data-driven error and physical mechanism constraints. The specific mathematical expression and physical meaning of this function are as follows: Let... The overall optimization objective for network training is defined as the weighted sum of the prediction accuracy term and the physical constraint term, i.e. ,in, and These are hyperparameters used to balance numerical accuracy and physical consistency; where the numerical accuracy loss term... The goal is to minimize the Euclidean distance between the predicted and actual flow fields, and its calculation formula is as follows: In the formula This represents the total number of spatial nodes. To predict the time step, and Representing the first Time of the first The actual and predicted latitude and longitude velocity components at each node; simultaneously, the fluid dynamics constraint loss term. The aim is to introduce the physical conservation laws of incompressible fluids as soft constraints into neural networks. A divergence zero-constraint penalty term is constructed to address the mass conservation property; its formula is as follows: This term, by penalizing non-physical source-sink divergence terms in the predicted flow field, forces the flow field to satisfy the continuity equation (…). This guides the model to generate prediction results that conform to the laws of ocean hydrodynamics, significantly improving the physical consistency of the prediction field.

[0069] Training process: Data from a historical period is used as the training set, and subsequent periods are used as the validation and test sets. The AdamW optimizer is used for end-to-end joint training, optimizing all parameters from feature fusion to the ST-GNN in one go. Early stopping is used during training to prevent overfitting.

[0070] The prediction principles for ocean currents in time and space are shown in the diagrams below. Figure 3 , Figure 4 As shown.

[0071] S6: Online prediction and visualization services.

[0072] The trained model is deployed as an online service. The system periodically and automatically acquires the latest multi-source data, performs the same preprocessing steps, and then inputs it into the model to generate ocean current prediction fields for the next few hours to days in real time. The prediction results are interactively visualized through a WebGIS platform in various forms such as dynamic particle streamline maps, velocity color-filled maps, and vector arrow maps, and a data download API is provided to serve various end users.

[0073] The following section provides a detailed explanation of the above method steps using a specific application example from the East China Sea region (reconstruction and prediction of daily ocean currents in the East China Sea region based on multi-source data fusion).

[0074] 1. Data preparation and preprocessing; 1.1 Data Source Selection and Acquisition: Satellite remote sensing data: ERA5-day mean sea surface 10-meter wind field (U10, V10) data were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF); global ocean grid data, including sea surface height (SSH) and sea surface temperature (SST), were obtained from the Copernicus Marine Environment Monitoring Service (CMEMS), with a spatial resolution of 0.25°×0.25°.

[0075] Field observation data: Argo buoy profile data, including temperature, salinity, and pressure, were obtained from the global Argo data center for the East China Sea and adjacent areas; surface flow observation data from the China Ocean Network were obtained from the coastal high-frequency ground wave radar.

[0076] Numerical model reanalysis data: Reanalysis data (GLBu0.08) at a global 1 / 12° resolution were obtained from the Hybrid Coordinate Ocean Model (HYCOM) as ground truth and validation data for training the reconstructed network and prediction model.

[0077] 1.2 Refined Preprocessing Process: Spatiotemporal domain definition: The study area is defined as (117°E-135°E, 20°N-40°N). The time span is from January 1, 2015 to December 31, 2020, with 2015-2018 serving as the training set, 2019 as the validation set, and 2020 as the test set. All data are uniformly scaled to a daily scale.

[0078] Spatial resampling and alignment: All raster data (SST, SSH, wind field, HYCOM velocity) were resampled to a uniform 0.25°×0.25° latitude and longitude grid using bilinear interpolation, totaling 72 (longitude) × 80 (latitude) grid points.

[0079] Missing value handling: For satellite data, missing values ​​are filled using the Empirical Orthogonal Function Interpolation (DINEOF) method.

[0080] For sparse point data such as Argo buoys, spatial interpolation is not performed; instead, their original position information is preserved and processed separately in subsequent feature fusion steps.

[0081] Data standardization: Calculate the mean and standard deviation for each physical variable (e.g., SST, U10, V10, SSH, U_curr, V_curr) across the entire training set, and perform Z-score standardization on all data: X_norm=(X-μ) / σ.

[0082] 2. Spatiotemporal resolution reconstruction of ocean current data; 2.1 Reconstructing the network model structure (PyTorch pseudocode illustration): Python import torch import torch.nn as nn # Adaptive Spatiotemporal Dual Attention Module class DualAttentionBlock(nn.Module): def __init__(self, channels): super().__init__() # Define the MLP (Multilayer Perceptron) required for channel attention self.mlp = nn.Sequential( nn.Linear(channels, channels / / 16), # Dimensionality reduction nn.ReLU(), nn.Linear(channels / / 16, channels) # Upscaling ) # Define the convolutional layers required for spatial attention self.conv = nn.Conv3d(2, 1, kernel_size=7, padding=3) self.sigmoid = nn.Sigmoid() def forward(self, x): # Input x dimension: [Batch, Channel, Time, Height, Width] # Channel Attention (Mc) # Logic: Global Average Pooling + Global Max Pooling -> MLP -> Summation -> Sigmoid avg_out = self.mlp(global_avg_pool(x)) max_out = self.mlp(global_max_pool(x)) mc = self.sigmoid(avg_out + max_out) x = x * mc # Channel weighting using broadcast mechanism # Spatial Attention (Ms) # Logic: Average pooling + max pooling along the channel axis -> concatenation -> convolution -> Sigmoid avg_map = torch.mean(x, dim=1, keepdim=True) max_map = torch.max(x, dim=1, keepdim=True) scale_map = torch.cat([avg_map, max_map], dim=1) # Concatenate ms = self.sigmoid(self.conv(scale_map)) #Adaptive Fusion Output return x * ms + x # Spatially weighted and introduce residual join class OceanSRNet(nn.Module): # Ocean Super-Resolution Network def __init__(self, in_channels, out_channels=2, scale_factor=4): # Assuming the target resolution is increased by 4 times super().__init__() # Encoder (Downsampling) self.enc1 = nn.Sequential(nn.Conv3d(in_channels, 64, 3, padding=1),nn.ReLU()) self.pool1 = nn.MaxPool3d(2) # Downsample by 2 times self.enc2=nn.Sequential(nn.Conv3d(64,128,3,padding=1), nn.ReLU()) self.pool2 = nn.MaxPool3d(2) # Downsample by 2 times, for a total of 4 times # Bottleneck layer self.bottleneck=nn.Sequential(nn.Conv3d(128,256,3,padding=1), nn.ReLU(),DualAttentionBlock(256)) # Decoder (Upsampling) # Efficient Upsampling Using Subpixel Convolution self.up1 = nn.Sequential( nn.Conv3d(256, 512, 3, padding=1), nn.PixelShuffle(2), # Upsample by 2 times, output channels become 512 / (2^3)=64 nn.ReLU() ) self.up2 = nn.Sequential( nn.Conv3d(192, 256, 3, padding=1), # Number of channels after concatenation with enc2 (skip connection). nn.PixelShuffle(2), # Upsample by 2 times, for a total of 4 times nn.ReLU() ) # After skip connection fusion, output the final flow rate. self.final_conv = nn.Conv3d(96, out_channels, 3, padding=1) # U, V components def forward(self, x_lr, mask): # x_lr: Low-resolution input [B, C, T, H_lr, W_lr] # mask: Missing value mask x = torch.cat([x_lr, mask], dim=1) # Input the mask as an extra channel e1 = self.enc1(x) e2 = self.enc2(self.pool1(e1)) b = self.bottleneck(self.pool2(e2)) d1 = self.up1(b) # Skip connection 1: Concatenate with e2 after upsampling d1 = torch.cat([d1, nn.functional.interpolate(e2, scale_factor=2)],dim=1) d2 = self.up2(d1) # Skip connection 2: Concatenate with e1 after upsampling d2 = torch.cat([d2, nn.functional.interpolate(e1, scale_factor=4)],dim=1) output = self.final_conv(d2) # [B, 2, T, H_hr, W_hr] return output # Discriminator network class Discriminator(nn.Module): # ...The PatchGAN structure is used to output a feature matrix instead of a single value to improve training stability.

[0083] 2.2 Reconstructing Network Training Details: Input and output: Input: Low-resolution (0.25°) multi-source data stack (SST, SSH, U10, V10, and corresponding missing masks), with a total of 5 channels in_channels. Time dimension T=7 (past week).

[0084] Output: High-resolution (0.0625°) U and V velocity fields, out_channels=2.

[0085] Training objective: The network learns to map from low-resolution, missing multi-source inputs to high-resolution HYCOM reanalysis flow fields.

[0086] Loss function: Total_Loss = λ1*MSE_Loss + λ2*Adv_Loss + λ3*Grad_Loss MSE_Loss: Mean squared error loss, λ1=1.0.

[0087] Adv_Loss: Adversarial loss (least square GAN loss), λ2=0.01.

[0088] Grad_Loss: Gradient difference loss, calculates the difference between the generated result and the true value in the longitudinal and latitudinal directions, λ3=0.5.

[0089] Optimizer: The Adam optimizer is used, with a learning rate of 1e-4 for both the generator and discriminator, and a batch size of 8.

[0090] 3. Spatiotemporal prediction based on reconstructed fields and multi-source features; 3.1 Multimodal feature fusion; Raster data from SST and SSH are input into a lightweight 3D-CNN; wind field time series data are input into a bidirectional GRU; Argo point data are embedded through a fully connected layer and assigned to the grid according to their position. Finally, spatial broadcasting technology is used to align the feature dimensions of each branch, and the stitched features are adaptively weighted and fused using the aforementioned dual channel and spatial attention mechanism to output a unified feature tensor.

[0091] 3.2 Spatiotemporal Graph Neural Network (ST-GNN) Prediction Model: Graph structure construction: Nodes: The reconstructed high-resolution grid (0.0625°) consists of 288 (longitude) × 320 (latitude) nodes.

[0092] Edges: A combination of K-nearest neighbors (K=8) and physical constraints is used. The great circle distance between nodes is calculated, and edges that cross land are removed (using ETOPO1 terrain data for determination).

[0093] Model architecture: A 2-layer Graph Attention Network (GAT) is used as the spatial convolutional layer, with each attention head having a dimension of 16, for a total of 8 heads.

[0094] Temporal convolutional networks (TCNs) were used as temporal convolutional layers with dilation coefficients set to [1,2,4,8] to capture multi-scale temporal dependencies.

[0095] Three spatiotemporal convolutional blocks (ST-Conv Blocks) are stacked. Within each block, GAT convolution is performed first, followed by TCN convolution.

[0096] Input and output: Input: Core input: High-resolution (0.0625°) historical ocean current sequences (U, V) over the past 30 days generated by the reconstructed network.

[0097] Auxiliary input: Multi-source fusion features generated over the past 30 days by a multimodal feature fusion network.

[0098] Output: High-resolution ocean current fields (U, V) for the next 7 days, daily.

[0099] 3.3 Model Training and Optimization; Two-stage training: Phase 1: Fix the pre-trained reconstructed network and train the ST-GNN prediction model using only data from 2015-2018. The loss function is the MSE of the predicted U / V versus the ground truth value of HYCOM.

[0100] Phase Two (Joint Fine-tuning): Using the 2019 validation set data, the entire system (from multi-source data input to prediction output) is fine-tuned end-to-end with a small learning rate to optimize the coordination between the reconstruction and prediction stages.

[0101] 4. System deployment and operation; 4.1 System Architecture: This system adopts a microservice architecture and is deployed on a cloud-based Kubernetes cluster. The data acquisition and preprocessing module, as a scheduled Cron job, automatically pulls the latest data from APIs of data sources such as CMEMS and ECMWF at 00:00 UTC every day and completes the preprocessing pipeline.

[0102] The spatiotemporal resolution reconstruction module and the spatiotemporal prediction engine module are each encapsulated as independent GPU inference services (Docker containers). The preprocessed data first triggers the reconstruction service to generate a high-resolution reconstruction field for the current day and the past 29 days; subsequently, this sequence triggers the prediction service to generate a forecast for the next 7 days.

[0103] The results visualization and publishing module uses a front-end visualization interface developed based on the Vue.js and Deck.gl frameworks, and a RESTful API provided by Python Flask on the back-end. This module reads the reconstruction and prediction results from the database and displays them in a side-by-side comparison format. Left side: Displays the reconstructed flow animation from the past 7 days, and allows for retrospective viewing.

[0104] Right side: Shows a 7-day forecast flow field animation, with particle streamline diagrams clearly showing the Kuroshio Current's path, vortex movement, and evolution.

[0105] 4.2 Performance and Effect Verification: Ocean current online prediction results are as follows Figure 5 As shown, the numerical and experimental results for the Bohai Sea and East China Sea are compared as follows: Figure 6 , Figure 7 As shown. Quantitative evaluation on the 2020 test set indicates: Reconstruction results: Compared with the HYCOM reanalysis data, the root mean square error (RMSE) of the reconstructed flow field by this method is reduced by more than 60% at a resolution of 0.0625° compared with the simple bilinear interpolation result, and it can clearly reconstruct the sub-mesoscale vortex structure present in HYCOM.

[0106] Prediction Performance: In the prediction of the next 7 days, its RMSE is reduced by about 22% compared to the model that only uses raw resolution data for prediction, and by about 35% compared to the traditional CNN-LSTM model. Especially in the Kuroshio main axis region, the prediction ability of the main axis oscillation is significantly improved.

[0107] Example 2: This embodiment of the invention provides a system for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion to implement the above method. This system adopts a modular design and includes: Data Acquisition and Preprocessing Module: As the system's data entry point, it is responsible for API integration with external databases (such as CMEMS and NOAA) to automate the crawling, parsing, and storage of multi-source data. It includes the aforementioned refined preprocessing pipeline.

[0108] Multi-source data fusion module: It encapsulates a pre-trained multimodal feature fusion network, receives pre-processed standardized data, performs feature extraction and attention fusion in parallel, and outputs a high-quality fused feature tensor.

[0109] Spatiotemporal prediction engine module: As the "brain" of the system, it encapsulates the trained ST-GNN model. This module receives fused feature tensors, performs efficient forward inference computation, and outputs high-precision ocean current spatiotemporal sequence prediction results. This module can be deployed on a GPU server to ensure prediction speed.

[0110] The results visualization and publishing module serves as the user interface and is developed based on a B / S architecture. It utilizes front-end visualization libraries (such as Cesium and Deck.gl) to render the predicted data into intuitive 2D / 3D ocean current field animations. Simultaneously, it provides a RESTful API, allowing third-party systems such as shipping companies and environmental protection departments to directly access structured predicted data.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion, characterized in that, The method includes the following steps: S1. Data Acquisition and Preprocessing: Collect multi-source ocean current related data, including satellite remote sensing data, field observation data, and numerical model reanalysis data, and perform spatiotemporal alignment and missing value processing to form a unified spatiotemporal dataset. S2. Multimodal feature fusion: Construct a multimodal feature fusion network, extract spatial features of raster data through a three-dimensional convolutional neural network, extract dynamic evolution features of time series data through a recurrent neural network, and fuse the spatial features, dynamic evolution features and point observation data features to generate a fusion feature tensor containing multi-source complementary information; S3. Spatiotemporal resolution reconstruction: The fused feature tensor is input into a pre-trained generative ocean current field super-resolution reconstruction network. Through an encoder-decoder structure and adversarial training strategy, a high spatiotemporal resolution and complete historical ocean current field sequence is generated. S4. Spatiotemporal sequence prediction: Construct a spatiotemporal graph neural network model, define the grid points of the prediction area as graph nodes and construct edges based on marine geographic information; take the high spatiotemporal resolution historical ocean current field sequence as the core input and the fused feature tensor as the auxiliary feature input, perform spatiotemporal coupling modeling through the spatiotemporal graph neural network model, and output the ocean current spatiotemporal sequence prediction results for a specified future time period.

2. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 1, characterized in that, The ocean current field super-resolution reconstruction network in the spatiotemporal resolution reconstruction step is a generative model based on deep learning. Its structure is a U-Net architecture that includes an encoder and a decoder, and it uses skip connections to pass the high-frequency features of the encoder to the decoder. The training process of the reconstructed network includes an adversarial training strategy, and its loss function includes at least pixel-level reconstruction loss and adversarial loss generated by the discriminator network.

3. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 2, characterized in that, The loss function of the reconstructed network includes gradient difference loss, which is used to constrain the consistency of the reconstructed ocean current field with the high-resolution real field in terms of spatial gradient.

4. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 1, characterized in that, The multimodal feature fusion step includes: A multimodal feature fusion network is constructed. A three-dimensional convolutional neural network is used to extract spatial features from raster data, and a recurrent neural network is used to extract dynamic evolution features from time-series data. An attention mechanism is used to adaptively weight and fuse the spatial features, dynamic evolution features, and embedded features from point observation data to generate a fusion feature tensor for auxiliary prediction.

5. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 4, characterized in that, The spatiotemporal sequence prediction step includes: The high spatiotemporal resolution historical ocean current field sequence is used as the core input; A spatiotemporal graph structure is constructed, high-resolution grid points in the predicted area are defined as graph nodes, and edges between nodes are constructed based on marine geographic information and physical relationships. Using the aforementioned spatiotemporal graph neural network model, the core and auxiliary features of the input are spatiotemporally coupled and modeled through alternating spatial graph convolution and temporal convolution operations, thereby completing the prediction.

6. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 5, characterized in that, The spatial graph convolution in the spatiotemporal graph neural network model employs a graph attention network to dynamically learn the weights of spatial dependencies between nodes; the temporal convolution employs a temporal convolution network or a gated recurrent unit.

7. The method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion according to claim 1, characterized in that, The method also includes a model training step, which is divided into two stages: In the first stage, the ocean current field super-resolution reconstruction network is trained separately; In the second stage, the pre-trained reconstruction network is connected to the spatiotemporal graph neural network to jointly fine-tune the overall ocean current spatiotemporal sequence reconstruction and prediction system in an end-to-end manner.

8. A system for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion, used to implement the method for reconstructing and predicting ocean current spatiotemporal sequences based on multi-source data fusion as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire and preprocess multi-source ocean current related data; The spatiotemporal resolution reconstruction module has a built-in pre-trained ocean current field super-resolution reconstruction network, which is used to generate a complete historical ocean current field with high spatiotemporal resolution. The spatiotemporal prediction engine module has a built-in pre-trained spatiotemporal graph neural network model, which is used to predict future ocean currents based on the historical ocean current field and multi-source features. The results visualization and publishing module is used to visualize and publish the reconstructed historical ocean current fields and the predicted future ocean current fields.

9. The ocean current spatiotemporal sequence reconstruction and prediction system based on multi-source data fusion according to claim 8, characterized in that, The system also includes a multi-source data fusion module with a built-in multimodal feature fusion network, which is used to extract and fuse features from preprocessed multi-source data, and provides the generated fusion feature tensors to the spatiotemporal resolution reconstruction module and the spatiotemporal prediction engine module, respectively.

10. The ocean current spatiotemporal sequence reconstruction and prediction system based on multi-source data fusion according to claim 8, characterized in that, The results visualization and publishing module is used to display historical reconstructed flow fields and future predicted flow fields in parallel, and provides dynamic backtracking and forward-looking animation functions.

Images