Air-sea coupling data sampling method based on asynchronous cross iteration random sampling strategy

Abstract

The invention discloses an air-sea coupling data sampling method based on an asynchronous cross iteration random sampling strategy, and belongs to the technical field of marine meteorology. According to the invention, through an asynchronous cross iteration random sampling strategy, many defects in the existing air-sea coupling modeling method are overcome, and the method has significant advantages in the aspects of data sampling efficiency, model precision, computing resource utilization rate and the like. In order to improve the data sampling efficiency, a sliding time window and a random sampling mechanism are adopted, the hysteresis effect and randomness between atmosphere and ocean variables are simulated, and meanwhile, the sampling efficiency is remarkably improved by using an asynchronous cross iteration strategy. Through a multi-card asynchronous scheduling and non-blocking sampling mechanism, each computing node can complete a sampling task in parallel, and compared with a traditional serial sampling method, the sampling efficiency is improved by about 30%. The mechanism of dynamically adjusting the sampling weight concentrates resources to pay attention to a high-error area, redundant sampling of irrelevant or low-influence data is avoided, and waste of computing resources is further reduced.

Description

Technical Field

[0001] The present invention belongs to the field of marine meteorology technology, and in particular, relates to a sea-air coupling data sampling method based on an asynchronous cross-iteration random sampling strategy. Background Art

[0002] The coupling between the ocean and the atmosphere is an important part of the Earth system and is of great significance for climate change, extreme weather events, and high-precision forecasts of ocean state variables. Changes in ocean state variables (such as sea surface temperature, sea surface height, and ocean currents) do not occur in isolation, but are driven by complex ocean-atmosphere interactions. These interactions involve multi-scale, multi-variable dynamic feedback mechanisms, which are of great value to the study of Earth system dynamics, disaster forecasting and prevention, and the sustainable use of marine resources. Therefore, developing efficient forecasting methods that can fully capture the ocean-atmosphere coupling relationship is an important direction for promoting Earth system science and practical applications.

[0003] Existing methods for modeling ocean-atmosphere coupling mainly include numerical simulation models, data-driven methods, and data splicing and fusion technologies. Among them, numerical simulation models can simulate the dynamic process of the ocean-atmosphere system by solving the physical equations of the ocean-atmosphere system, but the computational cost is high and the ability to capture small-scale changes is limited; data-driven methods, such as spatiotemporal convolutional networks and long short-term memory networks (LSTM), perform well in dealing with single subsystems of the ocean or atmosphere, but often fail to fully capture the complex coupling relationship between the two; in addition, some studies use simple data splicing methods for modeling, but this method fails to reflect the causal relationship and time lag effect between variables, resulting in limited forecast accuracy.

[0004] Although existing methods have made some progress, there are still many problems. First, the time scale differences between ocean and atmospheric changes have not been fully considered, and existing methods are difficult to effectively simulate the time-lagged effects of the atmosphere on ocean variables. Second, the inconsistency in the spatial resolution of ocean and atmospheric data can easily lead to information loss or error accumulation during data fusion. In addition, with limited computing resources, existing coupling models are difficult to balance the needs of high efficiency and high-resolution modeling. More importantly, current methods are insufficient in modeling complex and nonlinear ocean-atmosphere coupling relationships, and can often only capture local or shallow correlations, limiting the improvement of forecast performance. Summary of the invention

[0005] The purpose of the present invention is to construct an air-sea coupled data sampling method based on an asynchronous cross-iterative random sampling strategy, which aims to improve the accuracy and computational efficiency of ocean state variable prediction by simulating the complex interaction between the atmosphere and the ocean, so as to make up for the shortcomings of the existing technology.

[0006] To achieve the above object, the present invention is implemented by the following specific technical solutions: A method for sampling air-sea coupled data based on an asynchronous cross-iterative random sampling strategy comprises the following steps: S1: Collect ocean variable data and atmospheric variable data, and perform data preprocessing; S2: Using a sliding time window mechanism to simulate the lag effect of the atmosphere on ocean variables and enhancing data diversity through random sampling methods; S3: The cross-iteration sampling strategy is carried out asynchronously, dynamically optimizing the sampling process under the multi-card parallel computing framework and achieving seamless connection between sampling and training; S4: Constructing a data processing model including: data sampling module, encoder and decoder; both encoder and decoder adopt Swin Transformer V2 structure; S5: The data processing model is trained based on the sampled data, and the forecasting performance and the sea-air coupling modeling capability are verified.

[0007] Furthermore, in S1, by unifying the spatiotemporal resolution of ocean and atmospheric variables, a consistent input basis is provided for subsequent sliding window sampling and model training.

[0008] Furthermore, the S1 includes: S1-1: The ocean variable data adopts GLORYS12 reanalysis data with a horizontal resolution of 1 / 12° and contains 50 vertical layers; the atmospheric variable data adopts ERA5 reanalysis data with a horizontal resolution of 0.25° and contains 37 vertical isobaric layers; S1-2: Interpolation: Use bilinear interpolation to interpolate ERA5 atmospheric variables to the resolution of GLORYS12 (1 / 12°). The interpolation formula is: ; in, represents the interpolated atmospheric variables, represents the original atmospheric variables, is the resolution of GLORYS12 ocean variables, is an interpolation function; the present invention adopts the nearest neighbor interpolation method; S1-3: Time alignment: Align the two data sets to a daily time step to ensure consistency in the temporal dimension. Through this step, multi-source coupled data with uniform resolution and time step are generated, providing standardized input for the subsequent sliding window sampling mechanism.

[0009] Further, the S2 includes: S2-1: Set a time impact window for the daily ocean state variable forecast to limit the sampling range of atmospheric variables, which is defined as follows: ; in, Indicates the current time time window, and They are the maximum and minimum lag times (set in the present invention = 7, = 3) days); the rationale for setting the time window is to characterize that the current ocean state variables may be affected by the atmospheric conditions of the past 3 to 7 days by setting the maximum and minimum lag times.

[0010] S2-2: Random sampling: in the time window Randomly select a time point , the corresponding atmospheric variables are sampled, and the sampling method is as follows: ; in, Indicates the selected time point The corresponding atmospheric variables, Indicates the time range of sampling; reflects the randomness and complexity of the atmosphere's impact on ocean variables through random sampling.

[0011] S2-3: Training sample combination: Combine the randomly sampled atmospheric variables with the ocean variables at the current time point to form a complete training sample: ; in, Represents the input features used for training, including the current time The ocean state variables and time Atmospheric variables Through this step, training samples that take into account hysteresis effects and randomness are generated, providing rich and diverse data support for subsequent modeling.

[0012] Further, the S3 includes: S3-1: Sampling task allocation: Assign random sampling tasks to multiple computing nodes to ensure that each node can complete the sampling operation independently to avoid resource conflicts; S3-2: Cross-iterative update: According to the forecast error of the data processing model, the sampling weights of the atmospheric and oceanic variables are dynamically adjusted, and the update status is performed using the following formula: ; in, and denote the sampling probabilities of atmospheric and oceanic variables, respectively, and are the forecast errors of atmospheric and oceanic variables respectively; the sampling probability in the sampling scheme is inversely proportional to the forecast error. The larger the error, the higher the sampling probability, thus increasing the attention to the high error area; S3-3: Asynchronous execution: A non-blocking scheduling mechanism is used to enable sampling tasks and model training to be performed simultaneously, further improving computing efficiency. The asynchronous cross-iteration strategy is used to achieve dynamic optimization of the sampling process, thereby enhancing the ability to model the air-sea coupling relationship while ensuring efficiency.

[0013] Furthermore, the data processing model in S4 specifically includes: (1) Data sampling module The data sampling module is responsible for processing the input raw data and generating training samples based on a specific sampling strategy (e.g., based on an asynchronous cross-iteration random sampling strategy). The function of this module is to preprocess and sample the air-sea coupling data to ensure the quality and diversity of the input data. The specific steps include: Data preprocessing: The raw data (such as atmospheric variables and ocean variables) are interpolated and normalized to ensure data consistency and availability.

[0014] Sampling strategy: According to the model requirements, asynchronous cross-iteration random sampling is implemented to ensure that time series data that have an impact on future states are extracted from different time windows. This strategy helps capture dynamic changes on long time scales, especially when dealing with the complexity of ocean-atmosphere coupling.

[0015] Output: Pass the sampled and processed data to the encoder.

[0016] (2) Encoder The encoder is responsible for extracting high-level feature representations from the input data. Swin Transformer V2 has extremely high expressiveness when processing images and sequence data. It uses a window attention mechanism and a hierarchical feature extraction strategy to efficiently capture local and global features.

[0017] Patch Embedding: First, the input two-dimensional data (such as time series data of sea-air coupling) is divided into non-overlapping image patches, and the size of each image patch is adjusted according to the resolution of the input data. These image patches are then mapped into high-dimensional embedding vectors through linear transformation.

[0018] Window Attention: In Swin Transformer V2, the window self-attention mechanism (WindowAttention) is its core feature. Each Transformer layer captures local context information by calculating the relationship between data in a local window. In order to avoid the disconnection effect between windows, Swin V2 adopts the Shifted Window mechanism, which allows information to be passed across windows, thereby better modeling global dependencies.

[0019] Multi-layer structure: The encoder adopts a layered Transformer structure. As the number of layers increases, the spatial resolution gradually decreases, while the depth of the feature map increases. By aggregating information layer by layer, the encoder can effectively extract features of different scales.

[0020] Output: The output of the encoder is a high-dimensional feature representation, which contains deep information of the input data, including both local features and global information across windows.

[0021] (3) Decoder The decoder is responsible for mapping the features extracted by the encoder to the final output space, such as predicting ocean state variables and generating forecast results of sea-air coupling.

[0022] Backward construction: The decoder works similarly to the encoder, but its goal is to recover the specific output from the high-dimensional features generated by the encoder. The decoder gradually refines the output of each layer through back-propagated information to generate high-quality prediction results.

[0023] Cross-layer feature fusion: At each layer of the decoder, low-level features and high-level features from the encoder are combined through a specific fusion mechanism to ensure the accuracy of the output. The decoder not only relies on features from the encoder, but can also focus on different parts of the input data through a self-attention mechanism for effective feature recovery.

[0024] Generate output: The final output of the decoder is the result of the prediction and regression processing, such as the predicted value of variables such as sea surface temperature and ocean current speed.

[0025] Further, the S5 includes: S5-1: Training sample construction, using sampled data to generate input features and target output: ; in, is the input feature set, is the target output (i.e., the ocean state variable at the next time step); S5-2: Model training, by optimizing the loss function to minimize the forecast error to achieve the forecast of ocean state variables, the optimization formula is as follows: ; in, is the loss function, is the true value, is the model prediction value, is the total number of samples. By optimizing the loss function, we focus on minimizing the prediction error to improve the model's prediction ability for the target variable.

[0026] S5-2: Result verification: Use the test data set to evaluate the model's forecast accuracy and its ability to characterize the ocean-atmosphere coupling relationship.

[0027] Compared with the prior art, the present invention has the following beneficial effects: The present invention overcomes many deficiencies in existing sea-air coupling modeling methods through an asynchronous cross-iterative random sampling strategy, and shows significant advantages in data sampling efficiency, model accuracy, and computing resource utilization, which are specifically reflected as follows: 1. Improving data sampling efficiency. The present invention adopts a sliding time window and a random sampling mechanism to simulate the lag effect and randomness between atmospheric and oceanic variables, and at the same time significantly improves the sampling efficiency by using an asynchronous cross-iterative strategy. Through multi-card asynchronous scheduling and non-blocking sampling mechanisms, each computing node can complete the sampling task in parallel, and the sampling efficiency is improved by about 30% compared with the traditional serial sampling method. The mechanism of dynamically adjusting the sampling weights concentrates resources on high error areas, avoids redundant sampling of irrelevant or low-impact data, and further reduces the waste of computing resources.

[0028] In summary, the present invention has achieved a dual improvement in data sampling efficiency and model prediction accuracy through technological innovation, and is particularly suitable for high-resolution ocean-atmosphere coupling modeling tasks, while reducing the cost of using computing resources, which has important application value and practical significance. The present invention can provide technological breakthroughs in complex coupling relationship modeling, high-resolution data fusion, and computing efficiency optimization, and provide strong support for high-precision ocean state variable prediction and Earth system science research. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a design idea diagram of the sea-air coupling data of the asynchronous cross-iterative random sampling strategy of the present invention.

[0030] Figure 2 It is the basic architecture design diagram of the data processing model of the present invention.

[0031] Figure 3The figures are result comparison diagrams of the present invention and the existing methods; (a) is the RMSE comparison of the present invention and the existing methods in the east-west component of the ocean current; (b) is the RMSE comparison of the present invention and the existing methods in the north-south component of the ocean current; (c) is the RMSE comparison of the present invention and the existing methods in the sea temperature forecast; (d) is the RMSE comparison of the present invention and the existing methods in the salinity forecast. DETAILED DESCRIPTION

[0032] The technical solution of the present invention is further described below in conjunction with embodiments and drawings.

[0033] In view of the shortcomings of the existing sea-air coupling modeling methods, the present invention aims to solve the following technical problems: 1. Solve the time scale difference problem: The existing methods do not fully consider the time scale differences between ocean and atmospheric changes, and cannot accurately simulate the time lag effect of the atmosphere on ocean variables. The present invention realizes dynamic modeling of the impact of atmospheric changes on ocean state variables by designing a sliding time window and a random sampling mechanism, thereby improving the physical consistency and accuracy of the forecast. 2. Realize cross-resolution data fusion: In view of the information loss and error accumulation problems caused by the inconsistency of spatial resolution between ocean and atmospheric data, the present invention proposes a cross-iterative sampling strategy, which realizes the effective fusion of data with different resolutions through flexible data selection and dynamic optimization mechanisms. 3. Improve computing efficiency and adaptability: Considering the high demand for computing resources for high-resolution sea-air coupling models, the present invention performs sampling operations in an asynchronous manner under a multi-level multi-card parallel computing framework, which significantly improves computing efficiency while maintaining the accuracy of coupled modeling, and is suitable for practical application scenarios with limited resources.

[0034] This paper proposes an "asynchronous cross-iterative random sampling strategy" for sea-air coupled data sampling (the design idea is as follows Figure 1 As shown in the figure), by simulating the time lag effect of the atmosphere on the ocean and combining the sliding time window with the random sampling mechanism, it can effectively capture the complex dynamic characteristics of the ocean-atmosphere coupled system, which is particularly suitable for high-resolution modeling needs under resource-constrained conditions.

[0035] In the following examples, the prediction of air-sea coupled data is carried out, combined with the newly proposed asynchronous cross-iterative random sampling strategy, and compared with the existing numerical model (PSY4 model of Mercator Ocean International). The data set includes: 1. GLORYS12 dataset GLORYS12 is a global ocean reanalysis dataset designed under the framework of the European Copernicus Marine Environment Monitoring Service (CMEMS), with a spatial resolution of 1 / 12° and 50 vertical levels, covering data from January 1993 to December 2020. As the main training dataset, GLORYS12 contains high-quality ocean state variable data, such as sea surface temperature (SST), salinity, ocean currents, etc. These data have high temporal continuity and spatial resolution, and are suitable for high-resolution ocean state prediction tasks.

[0036] 2. ERA5 atmospheric dataset: In order to combine atmospheric and ocean data for coupled modeling, we used the ERA5 dataset, which is provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). It has a spatial resolution of 0.25° and includes atmospheric variable data such as wind speed, air temperature, and humidity at 37 vertical layers.

[0037] (III) Numerical model (PSY4) data set: PSY4 is a numerical ocean forecast model released by Mercator Océan International of France. It provides forecasts of global ocean state variables, including temperature, salinity, current velocity, etc. PSY4 uses traditional numerical weather forecasting methods to numerically solve the equations of ocean-atmosphere coupling to predict ocean states. This model is widely used in short-term forecasting tasks of global ocean-atmosphere interactions.

[0038] Example 1 The sea surface temperature (SST) forecast method based on GLORYS12 and ERA5 data includes the following steps: 1. Data preprocessing and resolution matching 1.1. Input Data: Ocean data: GLORYS12 reanalysis data, including variables such as sea surface temperature (SST) and sea surface salinity (SSS), with a horizontal resolution of 1 / 12° and a time range from 1993 to 2020.

[0039] Atmospheric data: ERA5 reanalysis data, including variables such as wind speed (U10, V10), air pressure, temperature, etc., with a horizontal resolution of 0.25° and a time range from 1993 to 2020.

[0040] 1.2. Interpolation processing: Using bilinear interpolation, the ERA5 data is interpolated to a resolution of 1 / 12° to match the spatial resolution of the GLORYS12 data. The specific interpolation formula is as follows: ; in, represents the interpolated value, are the values ​​of the four adjacent grid points, is the interpolation weight, calculated based on the distance.

[0041] 1.3. Time Alignment: The two datasets were aligned based on a daily time step, ensuring temporal consistency of atmospheric and oceanic variables.

[0042] 2. Sliding time window definition and random sampling 2.1. Time window definition: Assume that the current time is Day, for forecast The SST of the day is defined as the time window of influence of atmospheric variables. Days, generate window range: ; 2.2. Random Sampling: In the window Randomly select a time point , extract the corresponding atmospheric variables, the formula is as follows: ; For example, for = 50 days, randomly sampled = 47, then the atmospheric variables on the 47th day are selected as one of the input features to participate in the SST forecast.

[0043] 2.3. Training sample construction: The atmospheric variables obtained by sampling are compared with the The combination of ocean variables of the day forms the training sample: ; in, is the input feature, is the target output, i.e. Days of SST.

[0044] 3. Asynchronous cross-iteration sampling strategy 3.1. Multi-card parallel sampling: Through the multi-card parallel computing framework, data sampling tasks at different time steps are distributed to multiple computing nodes and executed asynchronously: Node 1: Processing =1 to =100 days of data; Node 2: Processing =101 to =200 days of data; The same goes for other nodes.

[0045] 3.2. Sampling weight adjustment: Dynamically adjust the sampling weight according to the prediction error of the model to increase the sampling probability in the high error area: ; in, For the The forecast error of ocean state variables for the day, is the sampling probability.

[0046] 3.3. Sampling and training collaboration: Asynchronous sampling and model training are carried out simultaneously. After each sampling task is completed, the result is immediately sent to the training queue.

[0047] 4. Model training and validation 4.1. Training model (model architecture design such as Figure 2 shown): Based on the sampled data, train a deep learning model (such as STNet) and optimize the loss function: ; in, is the mean square error loss, is the true value of SST, is the forecast value of SST.

[0048] 4.2. Verify performance: The prediction accuracy of the model is evaluated using test data and compared with traditional methods: Mean Square Error (MSE): Reduced by about 15%.

[0049] Extreme conditions forecast: error reduced by about 18%.

[0050] Through the above steps, an efficient sea-air coupling data sampling method was successfully constructed. Each step is closely connected. Data preprocessing lays the foundation for sliding window sampling, random sampling enhances data diversity, asynchronous cross-iteration strategy optimizes sampling efficiency, and model training ultimately achieves high-precision forecasts of complex sea-air coupling relationships.

[0051] Example 2: Forecast of ocean dynamic variables In this embodiment, the same technical solution is used to predict ocean dynamic variables (such as flow velocity and flow direction).

[0052] 1. Data preprocessing: perform resolution matching and time alignment on flow velocity and flow direction data.

[0053] 2. Time Windows and Random Sampling: Definition Atmospheric variables are randomly sampled in a time window of 2 days and combined into input samples.

[0054] 3. Asynchronous sampling and training: Same as the first embodiment, a multi-card parallel and weight adjustment mechanism is adopted.

[0055] Experiments show that the present invention reduces the mean square error of ocean dynamic variable prediction by about 12% and improves the calculation efficiency by about 20%. Figure 3 As shown: Under the same hardware conditions, the time required for the sampling process of the present invention is shortened by about 25%-40% compared with the traditional method.

[0056] The present invention can improve the prediction accuracy of the model. Since the present invention comprehensively considers the complexity of the sea-air interaction, the generated training samples can better reflect the real sea-air coupling relationship, which significantly improves the prediction performance of the model. The sliding time window mechanism effectively captures the lagged effect of the atmosphere on ocean variables, making the prediction results closer to the actual situation. The diversity of random sampling enhances the generalization ability of the model and avoids the overfitting problem caused by a single data distribution. Experimental results show that in forecasting tasks such as sea surface temperature (SST), the root mean square error (RMSE) of the method of the present invention is reduced by 15%-20% compared with the traditional method.

[0057] The present invention can save computing resources and support high-resolution modeling. The present invention is particularly optimized for the resource limitation problem of high-resolution modeling, and has the following characteristics: i) Through data resolution matching and interpolation methods, the spatiotemporal resolution of different data sources is unified, the inconsistency of model input is reduced, and the foundation for high-resolution modeling is laid. ii) The asynchronous cross-iteration strategy makes full use of computing resources, reducing computing overhead while ensuring model performance. Compared with traditional methods, on 1 / 12° high-resolution global reanalysis data, the computing resource requirements of the present invention are reduced by about 20%-30%, but high-precision forecast results consistent with the original resolution can still be achieved.

[0058] The present invention realizes the simplicity and adaptability of operation. The method of the present invention has a clear structure and is easy to be integrated into the existing sea-air coupling model. The data sampling and model training process is completely modularized and applicable to a variety of reanalysis data (such as GLORYS12 and ERA5) and different forecasting tasks (such as sea surface temperature, ocean current, sea surface height, etc.).

[0059] Through the above embodiments, the present invention has demonstrated its superior performance in ocean-atmosphere coupling modeling and provided an effective solution for high-resolution forecasting tasks under resource-constrained conditions.

[0060] Finally, although this specification is described according to implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for sampling air-sea coupled data based on an asynchronous cross-iterative random sampling strategy, characterized in that: The following steps are involved: S1: Collect ocean variable data and atmospheric variable data, and perform data preprocessing; S2: Using a sliding time window mechanism to simulate the lag effect of the atmosphere on ocean variables and enhancing data diversity through random sampling methods; S3: The cross-iterative sampling strategy is performed asynchronously, and the sampling process is dynamically optimized under the multi-card parallel computing framework; S4: Constructing a data processing model including: data sampling module, encoder and decoder; both encoder and decoder adopt Swin Transformer V2 structure; S5: The data processing model is trained based on the sampled data, and the forecasting performance and the sea-air coupling modeling capability are verified.

2. The sea-air coupling data sampling method according to claim 1, characterized in that: The S1 includes: S1-1: The ocean variable data adopts GLORYS12 reanalysis data with a horizontal resolution of 1 / 12° and 50 vertical layers; the atmospheric variable data adopts ERA5 reanalysis data with a horizontal resolution of 0.25° and 37 vertical isobaric layers, unifying the temporal and spatial resolutions of ocean and atmospheric variables; S1-2: Interpolation processing: Use bilinear interpolation to interpolate the atmospheric variables of ERA5 to the resolution of GLORYS12 (1 / 12°). The interpolation formula is: ； in, represents the interpolated atmospheric variables, represents the original atmospheric variables, is the resolution of the GLORYS12 ocean variables, is the interpolation function; S1-3: Time alignment: Align the two data to daily time steps to ensure consistency in the time dimension.

3. The sea-air coupling data sampling method according to claim 1, characterized in that: The S2 includes: S2-1: Set a time impact window for the daily ocean state variable forecast to limit the sampling range of atmospheric variables, which is defined as follows: ； in, Indicates the current time time window, and are the maximum and minimum lag times, respectively; S2-2: Random sampling: in the time window Randomly select a time point , the corresponding atmospheric variables are sampled, and the sampling method is as follows: ； in, Indicates the selected time point The corresponding atmospheric variables, Indicates the time range of sampling; S2-3: Training sample combination: Combine the randomly sampled atmospheric variables with the ocean variables at the current time point to form a complete training sample: ； in, Represents the input features used for training, including the current time The ocean state variables and time Atmospheric variables .

4. The sea-air coupling data sampling method according to claim 1, characterized in that: The S3 includes: S3-1: Sampling task allocation: Assign random sampling tasks to multiple computing nodes to ensure that each node can complete the sampling operation independently; S3-2: Cross-iterative update: According to the forecast error of the data processing model, the sampling weights of the atmospheric and oceanic variables are dynamically adjusted, and the update status is performed using the following formula: ； in, and denote the sampling probabilities of atmospheric and oceanic variables, respectively, and are the forecast errors of atmospheric and oceanic variables, respectively; S3-3: Asynchronous execution: A non-blocking scheduling mechanism is used to enable sampling tasks and model training to proceed simultaneously; an asynchronous cross-iteration strategy is used to achieve dynamic optimization of the sampling process.

5. The sea-air coupling data sampling method according to claim 1, characterized in that: The data processing model in S4 specifically includes: (1) Data sampling module The data sampling module is responsible for processing the input raw data and generating training samples according to the asynchronous cross-iteration random sampling strategy; preprocessing the raw data; implementing asynchronous cross-iteration random sampling to ensure that time series data that have an impact on future states are extracted from different time windows; and passing the sampled and processed data to the encoder; (2) The encoder is responsible for extracting high-level feature representations from the input data. It uses the window attention mechanism and hierarchical feature extraction strategy to efficiently capture local and global features. Patch Embedding: The input two-dimensional data is divided into non-overlapping image blocks. The size of each image block is adjusted according to the resolution of the input data. These image blocks are then mapped to high-dimensional embedding vectors through linear transformation. Window Attention: Each Transformer layer captures local context information by calculating the relationship between data within a local window. Multi-level structure: The encoder adopts a hierarchical Transformer structure. As the number of layers increases, the spatial resolution gradually decreases, while the depth of the feature map increases. The output of the encoder is a high-dimensional feature representation, which contains deep information of the input data, including both local features and global information across windows. (3) The decoder maps the features extracted by the encoder to the final output space; the decoder gradually refines the output of each layer through back-propagation information to generate high-quality prediction results; at each layer of the decoder, the low-level features and high-level features from the encoder are combined through a specific fusion mechanism; the decoder not only relies on the features from the encoder, but can also focus on different parts of the input data through the self-attention mechanism to perform effective feature restoration; the final output of the decoder is the result of prediction and regression processing.

6. The sea-air coupling data sampling method according to claim 1, characterized in that: The S5 includes: S5-1: Training sample construction, using sampled data to generate input features and target output: ； Among them, X is the input feature set and Y is the target output; S5-2: Model training, by optimizing the loss function to minimize the forecast error to achieve the forecast of ocean state variables, the optimization formula is as follows: ； Where L is the loss function, Y i is the true value, is the model prediction value, N is the total number of samples; S5-3: Result verification: Use the test data set to evaluate the model's forecast accuracy and its ability to characterize the ocean-atmosphere coupling relationship.

Images