Deep learning reconstruction method for sea surface current field based on multi-source sea-air spatio-temporal characteristics

Abstract

This invention provides a deep learning method for reconstructing sea surface current fields based on multi-source air-sea spatiotemporal features, relating to the field of marine data processing. Specifically, it includes: acquiring multi-year continuous sea surface current velocity data and multi-source environmental feature data for a target sea area; constructing a full-field training sample in a continuous time series manner, using multi-source environmental features from multiple consecutive historical moments as input to candidate deep learning models, and using the full-field data of the eastward and northward current velocity components at a future target time as prediction labels; constructing velocity moduli using the eastward and northward current components; training candidate deep learning models using a joint loss function; evaluating candidate deep learning models, selecting the optimal model, and using the optimal model to output the full-field current velocity results at the target time. The technical solution of this invention overcomes the problem in existing technologies of difficulty in obtaining long-term, large-scale, spatially continuous sea surface current field data.

Description

Technical Field

[0001] This invention relates to the field of marine data processing, specifically to a deep learning method for reconstructing sea surface flow fields based on multi-source spatiotemporal features of the ocean and atmosphere. Background Technology

[0002] Ocean surface currents are crucial physical quantities characterizing ocean dynamic processes. They directly influence air-sea interactions, heat and salinity transport, nutrient diffusion, and marine ecological processes. Furthermore, they play a vital role in applications such as mesoscale eddy identification, ocean front analysis, nearshore pollutant dispersion early warning, maritime search and rescue, fisheries production, and shipping route optimization. Therefore, acquiring temporally continuous, spatially complete, and physically reliable ocean surface current data is a fundamental requirement for marine scientific research and marine engineering applications.

[0003] While the importance of sea surface currents is widely recognized, the acquisition of high-quality sea surface current velocity data remains significantly limited. Traditional ocean current observations primarily rely on mooring, drifting buoys, underway observation, high-frequency ground-wave radar, and satellite remote sensing. These methods each have their advantages, but they also suffer from insufficient spatial coverage, limited temporal continuity, high observation costs, or limitations in inversion conditions. Compared to variables such as sea surface temperature and sea level height, direct observational data on sea surface current velocity are much scarcer, making it difficult to generate data products that combine long-term series, high spatial resolution, and large-scale continuous coverage. This data shortcoming significantly restricts the study of long-term ocean circulation evolution patterns and limits the further use of ocean current information in ocean forecasting and engineering applications.

[0004] In recent years, the rapid development of reanalysis data and satellite remote sensing products has provided new possibilities for reconstructing sea surface current fields. High-quality reanalysis data, such as GLORYS and ERA5, can provide relatively complete ocean and atmospheric state variables, including sea surface temperature, sea level height, and sea surface wind field. These variables have a clear physical relationship with sea surface current fields. From a dynamic perspective, sea surface flow is mainly composed of geostrophic currents and wind-generated circulation. The horizontal gradient of sea level height directly reflects changes in the upper ocean pressure field and is an important basis for the formation of geostrophic currents; sea surface temperature can characterize ocean fronts, thermal structure, and some dynamic processes, and is closely related to the spatial distribution of ocean currents. At the same time, the sea surface wind field affects the upper ocean circulation through momentum input and is an important factor driving wind-generated currents. Therefore, establishing a mapping relationship between multi-source ocean-atmosphere variables and sea surface current velocities has strong physical rationality.

[0005] At the methodological level, traditional ocean current inversion and reconstruction largely rely on empirical formulas, statistical regression, or dynamic constraints. These methods typically characterize linear or local relationships well, but their expressive power is often limited when faced with the prevalent strong nonlinearity, multi-scale coupling, and complex spatial structures in ocean systems. Deep learning methods have significant advantages in high-dimensional feature extraction and nonlinear mapping fitting, and have shown great potential in recent years for tasks such as ocean environmental parameter inversion, ocean element reconstruction, and numerical forecast correction. Compared to traditional methods, deep learning can automatically learn complex relationships between multiple variables from a large number of samples and more effectively uncover the comprehensive impact of different air-sea factors on sea surface current fields.

[0006] Therefore, there is a need for a deep learning method for reconstructing sea surface flow fields based on multi-source air-sea spatiotemporal characteristics, which can obtain long-term, large-scale, spatially continuous sea surface flow field data. Summary of the Invention

[0007] The main objective of this invention is to provide a deep learning reconstruction method for sea surface flow fields based on multi-source air-sea spatiotemporal characteristics, so as to solve the problem that it is difficult to obtain long-term, large-scale, spatially continuous sea surface flow field data in the prior art.

[0008] To achieve the above objectives, this invention provides a deep learning method for reconstructing sea surface current fields based on multi-source air-sea spatiotemporal characteristics, specifically including the following steps: S1 acquires multi-year continuous ocean surface current velocity data and multi-source environmental characteristic data of the target sea area, and cleans and standardizes the acquired data.

[0009] S2, based on preprocessed multi-source environmental feature data, constructs the entire training sample in a continuous time series manner, uses multi-source environmental features from multiple consecutive historical moments as input to candidate deep learning models, and uses the entire data of the eastward and northward ocean surface velocity components at the future target moment as prediction labels.

[0010] S3 uses the eastward and northward velocity components to construct the velocity modulus.

[0011] S4 uses a joint loss function to train the candidate deep learning model. The joint loss function includes: velocity component error term, spatial gradient error term, and velocity magnitude error term.

[0012] S5 uses a comprehensive scoring function to evaluate candidate deep learning models, selects the optimal model, and uses the optimal model to output the overall flow velocity result at the target time.

[0013] Furthermore, the ocean surface current velocity data in step S1 includes: eastward and northward ocean surface current velocity components; the multi-source environmental characteristic data includes: sea surface temperature, sea surface height, eastward wind speed, and northward wind speed.

[0014] Furthermore, the standardized processing expression in step S1 is: ; in, Represents the original feature values. This represents the mean of the features on the training set. This represents the standard deviation of the feature on the training set. This represents the standardized feature value.

[0015] Furthermore, step S2 specifically includes the following steps: S2.1, set the time... Multi-source environmental characteristics Represented as: ; in, The number of input feature channels, This represents the number of latitudinal grids. This represents the number of grid cells in the longitudinal direction.

[0016] S2.2, Let the length of the input sequence be... The predicted step size is Then the first Training samples at each time point Represented as: ; The corresponding supervision label is the future target time. Ocean surface velocity field : ; in, Indicates the future target time Eastward velocity component Indicates the future target time Northward velocity component.

[0017] S2.3, the candidate deep learning models include: PureCNN (pure convolutional model), CNNLSTM (a hybrid model of convolution and long short-term memory), CNNTransformer (a hybrid model of convolution and Transformer), and PureTransformer (a pure Transformer model).

[0018] Furthermore, the true velocity modulus at any grid point in step S3 Defined as: ; in, and These represent the true eastward velocity component and the true northward velocity component, respectively.

[0019] Candidate deep learning models predict flow velocity magnitude for: ; in, and These represent the eastward and northward velocity components predicted by the candidate deep learning model, respectively.

[0020] Furthermore, the joint loss function in step S4 Including velocity component error term, spatial gradient error term, and velocity modulus error term: ; in, This represents the error term of the flow velocity component. Represents the spatial gradient error term. This indicates the error term in the velocity modulus. These are the weighting coefficients.

[0021] The velocity component error term is defined as: ; in, Indicates the ocean area mask in the first The values ​​at each grid point and These represent the predicted first and second halves of the series. The eastward velocity component and the predicted northward velocity component at each grid point and They represent the real first The eastward velocity component and the true northward velocity component at each grid point.

[0022] The spatial gradient error term is defined as: ; in, For the meridional gradient error term, This is the latitudinal gradient error term.

[0023] The velocity modulus error term is defined as: ; in, For the predicted first Velocity modulus at each grid point For the true first The velocity modulus at each grid point.

[0024] Furthermore, the comprehensive scoring function in step S5 for: ; ; ; ; in, This represents the root mean square error of the eastward velocity component. This represents the root mean square error of the northward velocity component. This represents the root mean square error of the velocity modulus. This represents the overall coefficient of determination. and These represent the average true eastward flow velocity and the average true northward flow velocity, respectively. This represents the coefficient of determination for the eastward flow velocity component. This represents the coefficient of determination for the northward velocity component.

[0025] The present invention has the following beneficial effects:

[0026] (1) Realize the whole-field reconstruction of the ocean surface velocity field This invention directly uses the entire two-dimensional grid velocity field as the output object to achieve joint reconstruction of the eastward and northward velocity components at the target time. Compared with single-point prediction or local block inference methods, the output results of this invention are more complete and continuous in space, and can better meet the application needs of regional flow field analysis and ocean dynamic process research.

[0027] (2) Multi-source feature fusion improves the model's expressive power. This invention integrates ocean state variables and atmospheric driving variables, including surface ocean temperature (thetao), sea surface height (zos), 10-meter eastward wind speed (u10), and 10-meter northward wind speed (v10), and incorporates flow field evolution information through continuous time series whole-field sample construction. Simultaneously, an ocean region mask is integrated throughout the sample constraint, loss calculation, and testing and evaluation processes, thereby improving the accuracy and physical plausibility of the overall reconstruction of the ocean surface velocity field.

[0028] (3) Continuous time-series input enhances the characterization of ocean current field evolution. This invention uses continuous multi-time feature sequences as input, which is better able to characterize the temporal continuity, lag and evolution trend of the ocean surface velocity field than the scheme that only uses single-time input. It helps to improve the reconstruction effect of the velocity field at future target time and enhances the model's ability to characterize the spatiotemporal evolution law of the flow field. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 The flowchart of a deep learning reconstruction method for sea surface flow field based on multi-source air-sea spatiotemporal characteristics is shown.

[0030] Figure 2 It shows the true distribution of ocean surface current modulus at target times in the target sea area from 90° to 130° east longitude and 0° to 45° north latitude.

[0031] Figure 3 The diagram shows the reconstructed distribution of ocean surface current modulus lengths at target times, based on the optimal model, for target sea areas ranging from 90° to 130° east longitude and 0° to 45° north latitude.

[0032] Figure 4 The test metrics for different candidate models are shown. Compare the bar charts.

[0033] Figure 5 The test metrics for different candidate models are shown. Compare the bar charts.

[0034] Figure 6 The test metrics for different candidate models are shown. Compare the bar charts.

[0035] Figure 7 The test metrics for different candidate models are shown. Compare the bar charts. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 The sea surface flow field deep learning reconstruction method based on multi-source air-sea spatiotemporal features, shown below, specifically includes the following steps: S1 acquires multi-year continuous ocean surface current velocity data and multi-source environmental characteristic data of the target sea area, and cleans and standardizes the acquired data.

[0038] S2, based on the preprocessed multi-source environmental feature data, i.e., multi-source environmental feature data, constructs the whole field training sample in the form of continuous time series, uses the multi-source environmental features of multiple consecutive historical moments as the input of the candidate deep learning model, and uses the whole field data of the eastward and northward current velocity components of the ocean surface at the future target time as the prediction label.

[0039] S3 uses the eastward and northward velocity components to construct the velocity modulus.

[0040] S4 uses a joint loss function to train the candidate deep learning model. The joint loss function includes: velocity component error term, spatial gradient error term, and velocity magnitude error term.

[0041] S5 uses a comprehensive scoring function to evaluate candidate deep learning models, selects the optimal model, and uses the optimal model to output the overall flow velocity result at the target time.

[0042] First, in the data preparation stage, this invention acquires GLORYS ocean reanalysis data and ERA5 atmospheric reanalysis data, and completes time alignment, outlier detection, ocean region mask extraction, and missing data processing, outputting cleaned annual data. Second, in the feature engineering stage, input features composed of thetao, zos, u10, and v10 are constructed and divided and standardized according to training, validation, and test sets to further generate time-series whole-field samples. Then, in the model training stage, candidate models such as PureCNN, CNNLSTM, CNNTransformer, and PureTransformer are constructed and trained using ocean region mask constraints and joint loss functions. Subsequently, in the model evaluation stage, based on error indices and a comprehensive scoring function within the effective ocean region, each candidate model is uniformly tested and the optimal model is selected. Finally, in the reconstruction and demonstration stage, the selected best model is used to output the whole-field velocity results at the target time, and real velocity modulus maps and predicted velocity modulus maps are generated to demonstrate the reconstruction effect of the method of this invention.

[0043] Specifically, the ocean surface current velocity data in step S1 includes: eastward and northward ocean surface current velocity components; the multi-source environmental characteristic data includes: sea surface temperature (thetao), sea surface height (zos), eastward wind speed, and northward wind speed. In this embodiment, the eastward wind speed is 10 m eastward wind speed u10, and the northward wind speed is 10 m northward wind speed v10.

[0044] In this embodiment, the target sea area spans from 90°E to 130°E and from 0°N to 45°N. Ocean surface current velocity data and ocean state variables are derived from the GLORYS ocean reanalysis data product, while meteorological driving variables are derived from the ERA5 atmospheric reanalysis data product. Multi-source environmental characteristic data includes ocean state variables and meteorological driving variables. Ocean state variables include sea surface height and sea surface temperature, while meteorological driving variables include sea surface wind field; the eastward and northward components of the ocean surface current velocity are the model's prediction targets.

[0045] The raw data for each year were cleaned and processed separately, including time sorting, dimension identification, outlier checking, missing test checking, document integrity checking, and marine area mask extraction.

[0046] In this invention, the ocean region mask is not directly derived from an externally fixed land-sea template, but is automatically generated based on the long-term effectiveness of the target current velocity components. For any grid point, the proportion of it simultaneously possessing an effective eastward current velocity component uo and an effective northward current velocity component vo over the time dimension is statistically analyzed. When this proportion is not lower than a preset threshold, the grid point is determined to be a valid ocean region; otherwise, it is determined to be an invalid region. The ocean region mask generated in this way is used for effective region constraints in subsequent sample construction, loss calculation, and test evaluation.

[0047] Specifically, the standardized processing expression in step S1 is: ; in, Represents the original feature values. This represents the mean of the features on the training set. This represents the standard deviation of the feature on the training set. This represents the standardized feature value.

[0048] By employing the above methods, the impact of differences in the units and ranges of different variables on the stability of model training can be reduced. Standardized parameters are preferably obtained statistically from the training set and used consistently on both the validation and test sets to ensure a consistent processing flow. In this specific implementation, the training set spans from 2001 to 2016, the validation set from 2017 to 2018, and the test set from 2019 to 2020. The standardized parameters are obtained solely from the training set and remain unchanged during the validation and test set phases.

[0049] Specifically, step S2 includes the following steps: S2.1, In order to simultaneously learn the temporal evolution and spatial structure characteristics of the ocean current field, this invention adopts a continuous time series whole-field sample construction method. Let time... Multi-source environmental characteristics Represented as: ; in, The number of input feature channels, This represents the number of latitudinal grids. This represents the number of grid cells in the longitudinal direction.

[0050] S2.2, Let the length of the input sequence be... The predicted step size is Then the first Training samples at each time point Represented as: .

[0051] The corresponding supervision label is the future target time. Ocean surface velocity field : ; in, Indicates the future target time Eastward velocity component Indicates the future target time Northward velocity component.

[0052] For each training sample, a marine region mask M corresponding to the output grid is generated simultaneously. Valid values ​​are used for marine grid positions, while invalid values ​​are used for land regions. This mask is used as a valid region constraint for both training loss and evaluation metrics.

[0053] S2.3, the candidate deep learning models include: PureCNN (pure convolutional model), CNNLSTM (a hybrid model of convolution and long short-term memory), CNNTransformer (a hybrid model of convolution and Transformer), and PureTransformer (a pure Transformer model). These candidate deep learning models are used to extract spatial structure features, temporal evolution features, and long-range dependencies from the entire dataset.

[0054] The aforementioned candidate deep learning models uniformly accept multi-time and multi-variable whole-field features as input, uniformly output dual-channel ocean surface current field results at the target time, and conduct performance comparison and optimal model selection under a unified dataset and unified evaluation system.

[0055] Specifically, to enhance the model's ability to represent the velocity intensity distribution, this invention further constructs the velocity modulus during training and evaluation. The true velocity modulus at any grid point in step S3... Defined as: ; in, and These represent the true eastward velocity component and the true northward velocity component, respectively.

[0056] Candidate deep learning models predict flow velocity magnitude for: ; in, and These represent the eastward and northward velocity components predicted by the candidate deep learning model, respectively.

[0057] Specifically, to balance numerical accuracy, spatial structure preservation, and flow velocity intensity recovery, this invention employs a joint loss function for model training. Training set samples are input into each candidate deep learning model, and model parameters are optimized based on the joint loss function. Model performance changes are monitored on the validation set. For each candidate model, the weights of the best-performing models during the validation phase, along with corresponding metadata, are saved, including input feature names, number of input channels, sequence length, prediction step size, spatial size, and spatial downsampling factor.

[0058] Joint loss function in step S4 Including velocity component error term, spatial gradient error term, and velocity modulus error term: ; in, This represents the error term of the flow velocity component. Represents the spatial gradient error term. This indicates the error term in the velocity modulus. These are the weighting coefficients.

[0059] The velocity component error term is defined as: ; in, Indicates the ocean area mask in the first The values ​​at each grid point and These represent the predicted first and second halves of the series. The eastward velocity component and the predicted northward velocity component at each grid point and They represent the real first The eastward velocity component and the true northward velocity component at each grid point.

[0060] The spatial gradient error term is defined as: ; in, For the meridional gradient error term, The latitudinal gradient error term is used to constrain the gradient consistency between the predicted flow field and the actual flow field in two spatial directions.

[0061] In this invention, the meridional gradient can be obtained by the difference between adjacent meridional grids, the zonal gradient can be obtained by the difference between adjacent zonal grids, and the gradient error is only statistically analyzed when both adjacent grid points are located in the effective sea area.

[0062] The velocity modulus error term is defined as: ; in, For the predicted first Velocity modulus at each grid point For the true first The velocity modulus at each grid point.

[0063] In this invention, all loss terms in the joint loss function are calculated only within the effective sea area corresponding to the ocean region mask, so as to avoid invalid regions interfering with model training. The value is 0.2. The value is 0.15.

[0064] Specifically, the comprehensive scoring function in step S5 for: ; ; ; ; in, This represents the root mean square error of the eastward velocity component. This represents the root mean square error of the northward velocity component. This represents the root mean square error of the velocity modulus. This represents the overall coefficient of determination. and These represent the average true eastward flow velocity and the average true northward flow velocity, respectively. This represents the coefficient of determination for the eastward flow velocity component. This represents the coefficient of determination for the northward velocity component.

[0065] During the testing phase, candidate deep learning models were evaluated using a unified evaluation metric. The evaluation metrics included the root mean square error (RMSE) of the eastward velocity component, the RMSE of the northward velocity component, the RMSE of the velocity modulus, the coefficient of determination for the eastward velocity component, the coefficient of determination for the northward velocity component, and the overall coefficient of determination.

[0066] The lower the overall score, the better the overall performance of the model. Based on this scoring function, the best model is selected from all candidate models, and its model name, model weight path, test metric results, input feature configuration, test year information, and standardized parameter path are saved as a best model summary file.

[0067] The principles of this invention are explained below: First, we acquire multi-year continuous ocean surface current velocity data and related multi-source environmental characteristic data for the target sea area. We then perform time sorting, spatial alignment, dimensional unification, anomaly checking, missing data checking, and file integrity checking on the raw data, and extract the target variables and input variables.

[0068] Based on the long-term validity of the target ocean current components uo and vo in the time dimension, an ocean region mask is automatically generated. Grids with long-term stable effective current velocity values ​​are identified as effective ocean regions, and this mask is used for subsequent sample construction, loss statistics, and test evaluation.

[0069] The preprocessed multi-source environmental feature data is standardized. The standardized parameters are obtained only from the training set statistics and remain unchanged in the validation and test set stages to avoid information leakage from non-training data into the training process.

[0070] Based on the preprocessed multi-source environmental feature data, a whole-field training sample is constructed in a continuous time series manner. The multivariate whole-field features of multiple consecutive historical moments are used as model inputs, and the whole-field data of the eastward and northward velocity components of the ocean surface at the future target time are used as prediction labels.

[0071] Candidate deep learning models are constructed, including pure convolutional models, hybrid convolutional and long short-term memory models, hybrid convolutional and Transformer models, and pure Transformer models. Each candidate model uniformly receives multi-time- and multi-variable whole-field feature inputs and uniformly outputs the dual-channel ocean surface current field at the target time.

[0072] During model training, a marine region mask constraint is introduced to calculate the loss only within the effective sea area. At the same time, a joint loss function is constructed by combining the velocity component error, spatial gradient error, and velocity modulus error to improve the numerical accuracy and spatial structure preservation capability of the velocity field reconstruction.

[0073] Different candidate models were trained, validated, and tested separately. The performance of the models was uniformly evaluated and compared using indicators such as the root mean square error of the eastward velocity component, the root mean square error of the northward velocity component, the root mean square error of the velocity modulus, and the comprehensive coefficient of determination.

[0074] Based on the test results, a comprehensive scoring function is constructed to select the model with the best comprehensive performance from multiple candidate models, which is then used as the best model for the final reconstruction of the ocean surface velocity field.

[0075] Finally, using the best model obtained through screening, whole-field inference is performed on the input features at a specified time in the test year to obtain the reconstructed ocean surface velocity field at the target time. A comparison chart of the real velocity field and the reconstructed velocity field is generated to demonstrate the whole-field reconstruction effect of the method of the present invention.

[0076] Figure 2 and Figure 3 The results show that the method of the present invention can recover the position of the main axis, the strong current region and some mesoscale structure of the ocean surface velocity field in a good way, and has good reconstruction accuracy and spatial structure consistency in most sea areas.

[0077] In this embodiment, the quantitative evaluation results of the four candidate models within the effective ocean region of the test set are shown in Table 1. The CNNLSTM model in... , , , The model outperforms the other three models in key indicators, indicating that it can more effectively learn the mapping relationship between multi-source air-sea spatiotemporal characteristics and ocean surface velocity field, and is therefore the optimal model in the embodiments of the present invention.

[0078] Table 1. Performance comparison of different deep learning models on the test set for whole ocean current reconstruction

[0079] like Figure 2 and Figure 3 As shown, Figure 2 and Figure 3 The true distribution of ocean surface current velocity modulus at target times (90°E to 130°E, 0°N to 45°N) and the reconstructed distribution based on the optimal model are presented respectively. A comparison of the two shows that the method of this invention can effectively recover the position of the main current axis, the range of strong current areas, and the main spatial structure characteristics of the target sea area.

[0080] like Figures 4 to 7 As shown, a bar chart is used to compare the root mean square errors (RMS) of the eastward and northward current velocity components, the RMS error of the current magnitude, and the overall coefficient of determination of four candidate models—PureCNN, CNNLSTM, CNNTransformer, and PureTransformer—within the effective ocean region of the test set. Figures 4 to 7 As can be seen, the CNNLSTM model performs best overall in terms of error and goodness of fit, and is therefore selected as the optimal model in this invention.

[0081] This invention utilizes reanalysis data to establish a nonlinear mapping relationship between multiple ocean-atmosphere variables such as sea surface temperature, sea surface height, and sea surface wind field and sea surface velocity components. Based on this, it uses readily available multi-source ocean-atmosphere data to reconstruct the sea surface current field of the target sea area. By jointly introducing multiple driving factors characterizing geostrophic flow and wind-generated current, a deep learning reconstruction model oriented towards sea surface velocity is constructed, thereby achieving the reconstruction of a spatially continuous sea surface current field under the condition of scarce direct observation of sea surface velocity.

[0082] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for reconstructing sea surface current fields based on multi-source air-sea spatiotemporal characteristics, characterized in that, Specifically, the steps include the following: S1, acquire multi-year continuous ocean surface current velocity data and multi-source environmental characteristic data of the target sea area, and clean and standardize the acquired data; S2, based on the preprocessed multi-source environmental feature data, constructs the whole field training sample in the manner of continuous time series, uses the multi-source environmental features of multiple consecutive historical moments as the input of the candidate deep learning model, and uses the whole field data of the eastward and northward velocity components of the ocean surface at the future target time as the prediction label. S3, using the eastward and northward velocity components to construct the velocity modulus; S4. The candidate deep learning model is trained using a joint loss function, which includes: velocity component error term, spatial gradient error term and velocity magnitude error term. S5 uses a comprehensive scoring function to evaluate candidate deep learning models, selects the optimal model, and uses the optimal model to output the overall flow velocity result at the target time. Joint loss function in step S4 Including velocity component error term, spatial gradient error term, and velocity modulus error term: ； in, This represents the error term of the flow velocity component. Represents the spatial gradient error term. This indicates the error term in the velocity modulus. These are the weighting coefficients; The velocity component error term is defined as: ； in, Indicates the ocean area mask in the first The values ​​at each grid point and These represent the predicted first... The eastward velocity component and the predicted northward velocity component at each grid point and They represent the real first The eastward velocity component and the true northward velocity component at each grid point; The spatial gradient error term is defined as: ； in, For the meridional gradient error term, This is the latitudinal gradient error term; The velocity modulus error term is defined as: ； in, For the predicted first Velocity modulus at each grid point For the true first The velocity modulus at each grid point.

2. The deep learning reconstruction method for sea surface current field based on multi-source air-sea spatiotemporal characteristics according to claim 1, characterized in that, The ocean surface current velocity data in step S1 includes: eastward and northward ocean surface current velocity components; multi-source environmental characteristic data includes: sea surface temperature, sea surface height, eastward wind speed, and northward wind speed.

3. The deep learning reconstruction method for sea surface current field based on multi-source air-sea spatiotemporal characteristics according to claim 1, characterized in that, The standardized processing expression in step S1 is: ； in, Represents the original feature values. This represents the mean of the features on the training set. This represents the standard deviation of the feature on the training set. This represents the standardized feature value.

4. The deep learning reconstruction method for sea surface current field based on multi-source air-sea spatiotemporal characteristics according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1, set the time... Multi-source environmental characteristics Represented as: ； in, The number of input feature channels, This represents the number of latitudinal grids. This represents the number of grid cells in the meridional direction. S2.2, Let the length of the input sequence be... The predicted step size is Then the first Training samples at each time point Represented as: ； The corresponding supervision label is the future target time. Ocean surface velocity field : ； in, Indicates the future target time Eastward velocity component, Indicates the future target time Northward velocity component; S2.3, the candidate deep learning models include: PureCNN (pure convolutional model), CNNLSTM (a hybrid model of convolution and long short-term memory), CNNTransformer (a hybrid model of convolution and Transformer), and PureTransformer (a pure Transformer model).

5. The deep learning reconstruction method for sea surface current field based on multi-source air-sea spatiotemporal characteristics according to claim 1, characterized in that, The true velocity modulus at any grid point in step S3 Defined as: ； in, and These represent the true eastward velocity component and the true northward velocity component, respectively. Candidate deep learning models predict flow velocity magnitude for: ； in, and These represent the eastward and northward velocity components predicted by the candidate deep learning model, respectively.

6. The deep learning reconstruction method for sea surface current field based on multi-source air-sea spatiotemporal characteristics according to claim 1, characterized in that, The comprehensive scoring function in step S5 for: ； ； ； ； in, This represents the root mean square error of the eastward velocity component. This represents the root mean square error of the northward velocity component. This represents the root mean square error of the velocity modulus. This represents the overall coefficient of determination. and These represent the average true eastward flow velocity and the average true northward flow velocity, respectively. This represents the coefficient of determination for the eastward flow velocity component. This represents the coefficient of determination for the northward velocity component.

Images