A sea surface wind field spatial downscaling method and system based on multi-task learning

Abstract

The application discloses a sea surface wind field spatial downscaling method based on multi-task learning, and comprises the following steps: obtaining a training data set and a verification data set of sea surface wind field data; constructing a generative adversarial network comprising a generator network and a discriminator network; constructing a down-sampling network; the down-sampling network realizes a regression task of the generator network task, and two tasks are combined into a pair of dual tasks and combined into a dual learning network; constructing an auxiliary task network; the auxiliary task network is connected with the generator network of the sea surface wind field spatial downscaling to form a sea surface wind field spatial downscaling network based on multi-task learning, and the training data set is used for training and optimization, and the verification data set is used for testing the trained network. Through the above method, high-precision and high-resolution sea surface wind field data are obtained.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for spatial downscaling of sea surface wind fields based on multi-task learning. Background Technology

[0002] Sea surface wind field is one of the main sources of ocean momentum and a key parameter in various marine and meteorological studies. Accurate and timely observation of sea surface wind field is of great significance for applications such as numerical weather prediction, marine environmental monitoring and transportation, maritime search and rescue, and wind energy assessment.

[0003] Sea surface wind fields can be obtained from in-situ observations by ships, buoys, and monitoring stations. These measurements are considered accurate, but their spatial coverage is limited, spatial sampling is sparse and unevenly distributed, and they are easily affected by severe weather conditions such as typhoons. Remote sensing observations, such as scatterometers, radiometers, altimeters, and spaceborne synthetic aperture radar, have the advantages of large scale, high frequency, and dynamic capabilities, and are widely used in meteorology and oceanography. In addition, some studies use sea surface wind field products from numerical models and reanalysis data. However, most existing sea surface wind field data products have limited spatial resolution, making it difficult to meet the needs of refined research and applications. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and system for spatial downscaling of sea surface wind fields based on multi-task learning, wherein the method includes the following steps:

[0005] S1. Obtain sea surface wind field data and preprocess the data, dividing the preprocessed data into training dataset and validation dataset.

[0006] S2. Construct a generative adversarial network, which consists of a generator network and a discriminator network. The generator network is used for downscaling of low-resolution sea surface wind field data.

[0007] S3. Construct a downsampling network to degrade the high-resolution data generated by the generator network in S2 to low-resolution data. The downsampling network realizes the regression task of the generator network's downscaling task. The generator network task and the downsampling network task are combined into a pair of dual tasks. The generator network in S2 and the downsampling network are combined into a dual learning network. The network for the spatial downscaling task of the sea surface wind field is constructed through S2 and S3.

[0008] S4. The spatial downscaling task of additional variables other than the sea surface wind field is used as an auxiliary task. An auxiliary task network is constructed. The auxiliary task network is connected to the generator network in S2 using a bridging network to form a multi-task learning spatial downscaling network for the sea surface wind field. The multi-task learning spatial downscaling network for the sea surface wind field is implemented using the PyTorch framework.

[0009] S5. Process the data of the extra variables according to the steps in S1 to obtain the training dataset and validation dataset of the extra variables. Summarize the training dataset and validation dataset of the two tasks to obtain the multi-task training dataset and multi-task validation dataset.

[0010] S6. The multi-task learning sea surface wind field spatial downscaling network is trained and optimized using a multi-task training dataset. The trained multi-task learning sea surface wind field spatial downscaling network is tested using a multi-task validation dataset. The downscaling results are validated using buoy data.

[0011] This invention also proposes a spatial downscaling system for sea surface wind fields based on multi-task learning, comprising:

[0012] The dataset acquisition module is used to acquire sea surface wind field data and preprocess the data, dividing the preprocessed data into training datasets and validation datasets.

[0013] A generative adversarial network (GAN) construction module is used to construct a GAN, which comprises a generator network and a discriminator network.

[0014] The dual learning network construction module is used to construct a dual learning network. First, a downsampling network is constructed. The downsampling network realizes the regression task of the generator network task. The generator network task and the downsampling network task are combined into a pair of dual tasks. The generator network in the generative adversarial network and the downsampling network are combined into a dual learning network.

[0015] A network construction module for the spatial downscaling task of sea surface wind field is used to construct the network for the spatial downscaling task of sea surface wind field through generative adversarial networks and dual learning networks.

[0016] The auxiliary task network construction module is used to construct an auxiliary task network by treating the spatial downscaling task of additional variables other than the sea surface wind field as an auxiliary task.

[0017] A multi-task learning sea surface wind field spatial downscaling network construction module is used to connect the auxiliary task network with the generator network in the generative adversarial network using a bridging network to construct a multi-task learning sea surface wind field spatial downscaling network.

[0018] The data aggregation module is used to divide the data of the additional variables into training datasets and validation datasets for the auxiliary tasks, and aggregate the training datasets and validation datasets of the two tasks to obtain multi-task training datasets and multi-task validation datasets.

[0019] The training and testing module is used to train and optimize the multi-task learning sea surface wind field spatial downscaling network using a multi-task training dataset, test the trained multi-task learning sea surface wind field spatial downscaling network using a multi-task validation dataset, and validate the downscaling results using buoy data.

[0020] The beneficial effects of the technical solution provided by this invention are:

[0021] This invention first constructs a spatial downscaling network for sea surface wind fields based on generative adversarial networks (GANs). This network consists of a generator and a discriminator. The generator performs spatial downscaling, generating high-resolution sea surface wind field data from low-resolution data. The discriminator evaluates the generated data to ensure it has more texture details. Then, using a dual learning strategy, an opposite downsampling task is constructed for the generator's downscaling task, forming a closed-loop dual learning circuit. This circuit is applied to generate higher-resolution data, thus generating high-resolution sea surface wind field data even in the absence of high-resolution training data. Finally, using a multi-task learning method, auxiliary meteorological data related to the sea surface wind field is used as an additional variable. Spatial downscaling of this additional variable is used as an auxiliary task, forming a multi-task learning process with the spatial downscaling task. A parameter sharing mechanism is designed to provide additional information for sea surface wind field downscaling, resulting in high-precision, high-resolution sea surface wind field data. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for spatial downscaling of sea surface wind fields based on multi-task learning according to an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of data edge filling in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of the generator network structure according to an embodiment of the present invention;

[0025] Figure 4 This is a diagram of the downscaled network structure of the sea surface wind field in the generative adversarial network of this invention.

[0026] Figure 5 This is a flowchart of the downscaling process of sea surface wind field in dual learning according to an embodiment of the present invention;

[0027] Figure 6 This is a diagram of the downscaling network structure of the sea surface wind field for multi-task learning in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0029] Relevant terms in this invention:

[0030] Downscaling: Downscaling is the process of transforming large-scale, low-resolution information into small-scale, high-resolution information.

[0031] Generative Adversarial Networks (GANs) are deep learning models that consist of two modules: a generative model and a discriminative model. Their interactive learning produces quite good outputs. In this invention, the generative model is used to downscale low-resolution sea surface wind field data, generating high-resolution wind field data from the original low-resolution data. The discriminator network's task is to distinguish between the original wind field data and the generated high-resolution wind field data.

[0032] A flowchart of a sea surface wind field spatial downscaling method based on multi-task learning in this embodiment is shown below. Figure 1 This includes the following steps:

[0033] S1. Obtain sea surface wind field data and preprocess the data, dividing the preprocessed data into training dataset and validation dataset.

[0034] This embodiment uses WindSat sea surface wind field downscaling as an example. The WindSat sea surface wind field product is downloaded with a spatial resolution of 0.25° and a time range of 2017-2019.

[0035] S11. Extract sea surface wind field data, and process wind speed and direction into u and v components, i.e., zonal wind and meridional wind, as shown in the following formula:

[0036] u = w sinθ

[0037] v = w cosθ

[0038] Where w is the wind speed and θ is the clockwise angle between the wind direction and due north. After processing the wind field data into zonal and meridional winds, the values ​​of u and v range from -50 to 50. After normalizing u and v, their values ​​are reduced to 0-1.

[0039] S12, Missing data point marking: Mark the missing data points and record them as a mark array, where 0 represents missing points and 1 represents valid points.

[0040] S13. Boundary value imputation: Missing data is filled using boundary values, with horizontal imputation prioritized. A data imputation diagram is shown below. Figure 2 As shown.

[0041] S14. Divide the data processed in the above steps into 80% as the training dataset and 20% as the validation dataset.

[0042] S2. Construct a generative adversarial network, which consists of a generator network and a discriminator network.

[0043] The generator network adopts an encoder-decoder structure, consisting of a downsampling module (Down) and an upsampling module (Up).

[0044] Figure 4 This is a diagram of the downscaled network structure of the sea surface wind field using a generative adversarial network according to an embodiment of the present invention. (Refer to...) Figure 4 The generator network in the model consists of a bicubic interpolation algorithm module, three downsampling modules, three upsampling pattern modules, and two convolutional layers. The input to the generator network is first interpolated by the bicubic interpolation algorithm module, and then sequentially passed through the first convolutional layer, the first downsampling module, the second downsampling module, the third downsampling module, the first upsampling pattern module, the second upsampling pattern module, the third upsampling pattern module, and the second convolutional layer before being output.

[0045] The convolutional layer in this embodiment includes a 3×3 convolution with a stride of 2 and a 3×3 convolution with a stride of 1.

[0046] The generator network employs skip connections, which integrates low-level feature connections into higher-level networks. This ensures full utilization of low-level information and avoids gradient vanishing and degradation issues. The input before the nth Down module is connected to the nth-to-last Up module. The generator network structure after adding skip connections is as follows: Figure 3 As shown, the input after the first convolutional layer and before the first downsampling module of the generator network is connected to the input after the third upsampling module and before the second convolutional layer; the input before the second downsampling module and after the first downsampling module is connected to the input after the second upsampling module and before the third upsampling module; and the input before the third downsampling module and after the second downsampling module is connected to the input after the first upsampling module and before the second upsampling module.

[0047] Channel attention can treat the features of each channel differently, multiplying each channel by a weight, so that the network can focus more on features with greater information content.

[0048] A generator network model was implemented using the deep learning framework PyTorch to downscale low-resolution sea surface wind field data and generate high-resolution sea surface wind field data.

[0049] The discriminator model consists of convolutional layers, LeakyReLU activation function, batch normalization layer, adaptive average pooling layer, and sigmoid function. The convolutional layers, LeakyReLU activation function, and batch normalization layer form four feature extraction layers with progressively increasing channel number and decreasing feature size. The final convolutional layer reduces the number of channels to 1, fusing all features into a single channel.

[0050] Figure 4 This is a diagram of the downscaled network structure of the sea surface wind field using a generative adversarial network according to an embodiment of the present invention. (Refer to...) Figure 4 The discriminator network in the model consists of nine convolutional layers, eight LeakyReLU activation functions, six batch normalization layers, one adaptive average pooling layer, and one sigmoid function.

[0051] The first feature extraction layer is formed by sequentially connecting the first convolutional layer, the first LeakyReLU activation function, the second convolutional layer, the first batch normalization layer, and the second LeakyReLU activation function of the discriminator network.

[0052] The third convolutional layer, the second batch normalization layer, the third LeakyReLU activation function, the fourth convolutional layer, the third batch normalization layer, and the fourth LeakyReLU activation function of the discriminator network are connected in sequence to form the second feature extraction layer;

[0053] The fifth convolutional layer, the fourth batch normalization layer, the fifth LeakyReLU activation function, the sixth convolutional layer, the fifth batch normalization layer, and the sixth LeakyReLU activation function of the discriminator network are connected in sequence to form the third feature extraction layer;

[0054] The seventh convolutional layer, the seventh LeakyReLU activation function, the eighth convolutional layer, the sixth batch normalization layer, and the eighth LeakyReLU activation function of the discriminator network are connected in sequence to form the fourth feature extraction layer;

[0055] The input to the discriminator network passes sequentially through the first feature extraction layer, the second feature extraction layer, the third feature extraction layer, the fourth feature extraction layer, the ninth convolutional layer, the adaptive average pooling layer, and the sigmoid function before being output.

[0056] An adaptive average pooling layer is used to average the features of this channel, resulting in a unique value to accommodate inputs of any size. A Simgoid function is then used to obtain a true / false value between 0 and 1. The discriminator uses real raw sea surface wind field data (y) and sea surface wind field data generated by the generator. As input. When the input is y, the discriminator is expected to output 1, and when the input is... The discriminator is expected to output 0.

[0057] By combining the generator and discriminator, a sea surface wind field downscaling network based on a generative adversarial network is obtained, and its structure diagram is shown below. Figure 4 As shown, adversarial learning is used to give the generated high-resolution sea surface wind field data more texture details.

[0058] S3. Construct a downsampling network and implement it using the PyTorch deep learning framework to degrade the downscaling results to low-resolution data. The downsampling network implements the regression task of the generator network's downscaling task, forming a dual task between the generator network task and the downsampling network task. The generator network and the downsampling network in S2 are combined into a dual learning network. The network for the spatial downscaling task of the sea surface wind field is constructed through S2 and S3.

[0059] The downsampling network contains multiple downsampling modules identical to those in the generator. It is used to degrade the high-resolution sea surface wind field data generated by the generator into low-resolution sea surface wind field data, thus constituting a regression task for the downscaling process.

[0060] A dual learning network is trained using the training data to obtain a supervised dual learning network. This trained dual learning network is then applied to the task of generating higher resolution data, producing constrained high-resolution sea surface wind field data. The flowchart is as follows. Figure 5 As shown, Figure 5 This is a flowchart of the downscaling process of sea surface wind field in dual learning according to an embodiment of the present invention.

[0061] S4. The spatial downscaling task of additional variables other than the sea surface wind field is used as an auxiliary task. An auxiliary task network is constructed. The auxiliary task network is connected to the generator network in S2 using a bridging network to form a multi-task learning spatial downscaling network for the sea surface wind field. The multi-task learning spatial downscaling network for the sea surface wind field is implemented using the PyTorch framework.

[0062] The auxiliary task includes the same generator network as the one used in the sea surface wind field spatial downscaling task, and its network structure diagram is shown below. Figure 3 As shown.

[0063] In this embodiment, the correlation between sea surface wind and variables such as sea surface temperature and humidity is utilized, and sea surface temperature and water vapor are treated as additional variables.

[0064] Due to the addition of extra variables, the network model now has 4 input channels, namely:

[0065] channel_1 = (u + 50) / 100

[0066] channel_2 = (v + 50) / 100

[0067] channel_3 = (sst + 10) / 50

[0068] channel_4 = vapor / 75

[0069] Where u and v are zonal and meridional winds, sst is sea surface temperature, and vapor is water vapor. The above formula normalizes the four variables and then uses them as the input to the network.

[0070] Design a soft sharing method to share features between auxiliary tasks and the main task of downscaling sea surface wind fields through a bridging network.

[0071] The bridging network comprises four bridge modules, each consisting of 16 stacked residual channel attention modules (RCABs) used to transform and filter shared features, enabling the network to adaptively share features and parameters across different tasks. The bridging network is then used to connect the auxiliary task network to the generator network used for spatial downscaling of sea surface wind fields, forming a multi-task learning-based spatial downscaling network for sea surface wind fields. The network structure diagram is shown below. Figure 6 As shown.

[0072] In a further embodiment, the second and third upsampling modules of the auxiliary task network are connected to the first and second downsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the first bridge; the first and second downsampling modules of the auxiliary task network are connected to the second and third upsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the second bridge; the first and second upsampling modules of the auxiliary task network are connected to the second and third downsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the third bridge; and the second and third downsampling modules of the auxiliary task network are connected to the first and second upsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the fourth bridge.

[0073] The total content loss of the generator networks for both tasks is

[0074]

[0075] Where x is the input to the sea surface wind field downscaling network based on multi-task learning, and y is the original high-resolution data. wFor raw, high-resolution sea surface wind field data, y s For the original high-resolution auxiliary data, y is derived from y w and y s Composition; G w This represents the network for generating a downscaling task of sea surface wind fields. G represents the information shared by the auxiliary tasks, s represents the downscaling result of the additional variables, and G represents the information shared by the auxiliary tasks. s This represents an auxiliary downscaling task generator network. This represents the information shared by the main task. λ3 is the weight of the auxiliary task loss, used to adjust the speed and balance of the two tasks during network optimization. The larger the weight, the faster the network optimization speed, and the more it favors the corresponding branch during optimization.

[0076] The total loss function of the sea surface wind field downscaling network based on multi-task learning is designed as follows:

[0077]

[0078]

[0079]

[0080]

[0081] The loss function consists of two parts, x w This is low-resolution sea surface wind field data. Among them, S... p (x) is an indicator function; when the network input is low-resolution data, S p (x) is 1; when the network input is high-resolution data, S p (x) is 0.

[0082] To counteract the loss, the discriminator assesses the authenticity of the generated high-resolution sea surface wind y'. The loss for the regression task is the difference between the higher-resolution result generated by the generative model and the input (downsampled to the same resolution as the input), which is also measured by pixel-level L2 distance. When the network input is high-resolution data, the loss in the first half of the network is 0, and the loss in the second half is...

[0083] The loss of the discriminator is used to measure the discriminator's ability to distinguish raw sea surface wind field data y. w The error loss of the generated data y' is expected to be relative to the true original data y. wOutput 1 for the generator network's data y', and 0 for the data y', using the cross-entropy loss function. The coefficients in the formula are used to balance the weights of its components, where λ1 is set to 1 by default and λ2 is set to 0.1 by default.

[0084] S5. Process the data of the extra variables according to the steps in S1 to obtain the training dataset and validation dataset of the extra variables. Combine the training dataset and validation dataset of the two tasks to obtain the multi-task training dataset and multi-task validation dataset.

[0085] S6. The multi-task training dataset is used to train and optimize the multi-task learning sea surface wind field spatial downscaling network until the model converges. The trained multi-task validation dataset is used to test the network. NDBC and RAMA buoy data are downloaded, and buoy data in the study area are compiled. The downscaling results are validated using the buoy data. The COARE 4.0 algorithm is used to calibrate the NDBC and RAMA buoy-measured wind fields to a standard neutral equivalent wind at a 10-meter reference height. Data from the auxiliary tasks is input into the auxiliary task network, while data from the main task is input into the network constructed in S2 and S3. The two tasks exchange parameters through a bridging network.

[0086] The embodiment also includes a sea surface wind field spatial downscaling system based on multi-task learning, used to implement a sea surface wind field spatial downscaling method based on multi-task learning, including:

[0087] The dataset acquisition module is used to acquire sea surface wind field data and preprocess the data, dividing the preprocessed data into training datasets and validation datasets.

[0088] A generative adversarial network (GAN) construction module is used to construct a GAN, which comprises a generator network and a discriminator network.

[0089] The dual learning network construction module is used to construct a dual learning network. First, a downsampling network is constructed. The downsampling network realizes the regression task of the generator network task. The generator network task and the downsampling network task are combined into a pair of dual tasks. The generator network in the generative adversarial network and the downsampling network are combined into a dual learning network.

[0090] A network construction module for the spatial downscaling task of sea surface wind field is used to construct the network for the spatial downscaling task of sea surface wind field through generative adversarial networks and dual learning networks.

[0091] The auxiliary task network construction module is used to construct an auxiliary task network by treating the spatial downscaling task of additional variables other than the sea surface wind field as an auxiliary task.

[0092] A multi-task learning sea surface wind field spatial downscaling network construction module is used to connect the auxiliary task network with the generator network in the generative adversarial network using a bridging network to construct a multi-task learning sea surface wind field spatial downscaling network.

[0093] The data aggregation module is used to divide the data of the additional variables into training datasets and validation datasets for the auxiliary tasks, and aggregate the training datasets and validation datasets of the two tasks to obtain multi-task training datasets and multi-task validation datasets.

[0094] The training and testing module is used to train and optimize the multi-task learning sea surface wind field spatial downscaling network using a multi-task training dataset, test the trained multi-task learning sea surface wind field spatial downscaling network using a multi-task validation dataset, and validate the downscaling results using buoy data.

[0095] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A spatial downscaling method for sea surface wind fields based on multi-task learning, characterized in that, Includes the following steps: S1. Obtain sea surface wind field data and preprocess the data, dividing the preprocessed data into training dataset and validation dataset. S2. Construct a generative adversarial network, which consists of a generator network and a discriminator network. The generator network is used for downscaling of low-resolution sea surface wind field data. S3. Construct a downsampling network to degrade the high-resolution data generated by the generator network in S2 to low-resolution data. The downsampling network realizes the regression task of the generator network's downscaling task. The generator network task and the downsampling network task are combined into a pair of dual tasks. The generator network in S2 and the downsampling network are combined into a dual learning network. The network for the spatial downscaling task of the sea surface wind field is constructed through S2 and S3. S4. The spatial downscaling task of additional variables other than the sea surface wind field is used as an auxiliary task. An auxiliary task network is constructed. The auxiliary task network contains the same network as the generator network in the sea surface wind field spatial downscaling task network. The auxiliary task network is connected to the generator network in S2 using a bridging network to form a multi-task learning sea surface wind field spatial downscaling network. The multi-task learning sea surface wind field spatial downscaling network is implemented using the PyTorch framework. S5. Process the data of the extra variables according to the steps in S1 to obtain the training dataset and validation dataset of the extra variables. Summarize the training dataset and validation dataset of the two tasks to obtain the multi-task training dataset and multi-task validation dataset. S6. The multi-task learning sea surface wind field spatial downscaling network is trained and optimized using a multi-task training dataset. The trained multi-task learning sea surface wind field spatial downscaling network is tested using a multi-task validation dataset. The downscaling results are validated using buoy data. In step S2, the generator network adopts an encoder-decoder structure, including a bicubic interpolation algorithm module, three downsampling modules, three upsampling pattern modules, and two convolutional layers. The input of the generator network passes through the bicubic interpolation algorithm module, the first convolutional layer, the first downsampling module, the second downsampling module, the third downsampling module, the first upsampling pattern module, the second upsampling pattern module, the third upsampling pattern module, and the second convolutional layer in sequence before being output. The generator network uses skip connections, connecting the input after the first convolutional layer and before the first downsampling module to the input after the third upsampling module and before the second convolutional layer; connecting the input before the second downsampling module and after the first downsampling module to the input after the second upsampling module and before the third upsampling module; and connecting the input before the third downsampling module and after the second downsampling module to the input after the first upsampling module and before the second upsampling module. A bridging network is constructed to share features between the auxiliary task and the sea surface wind field spatial downscaling task. The bridging network module contains four bridge modules, each composed of 16 stacked residual channel attention modules (RCABs). This bridging network connects the auxiliary task network with the generator network for sea surface wind field spatial downscaling, forming a multi-task learning-based sea surface wind field downscaling network. The second and third upsampling modules of the auxiliary task network are connected to the first and second downsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the first bridge module. The first downsampling module of the auxiliary task network... The second upsampling module and the second downsampling module are connected to the second upsampling module and the third upsampling module of the generator network of the sea surface wind field spatial downscaling task network via the second bridge module; the first upsampling module and the second upsampling module of the auxiliary task network are connected to the second downsampling module and the third downsampling module of the generator network of the sea surface wind field spatial downscaling task network via the third bridge module; the second downsampling module and the third downsampling module of the auxiliary task network are connected to the first upsampling module and the second upsampling module of the generator network of the sea surface wind field spatial downscaling task network via the fourth bridge module.

2. The method for spatial downscaling of sea surface wind fields based on multi-task learning according to claim 1, characterized in that, Step S1 is as follows: S11. Extract sea surface wind field data, and process wind speed and direction into... u , v The components, and the formula, are as follows: in, w For wind speed, ɵ The angle between the wind direction and due north, clockwise, and... u , v Perform normalization processing; S12. Mark the points with missing data; S13. Fill in missing data using boundary values; S14. Divide the data processed in the above steps into a training dataset and a validation dataset.

3. The method for spatial downscaling of sea surface wind fields based on multi-task learning according to claim 1, characterized in that, The discriminator network consists of nine convolutional layers, eight LeakyReLU activation functions, six batch normalization layers, one adaptive average pooling layer, and one sigmoid function; The first feature extraction layer is formed by sequentially connecting the first convolutional layer, the first LeakyReLU activation function, the second convolutional layer, the first batch normalization layer, and the second LeakyReLU activation function of the discriminator network. The third convolutional layer, the second batch normalization layer, the third LeakyReLU activation function, the fourth convolutional layer, the third batch normalization layer, and the fourth LeakyReLU activation function of the discriminator network are connected in sequence to form the second feature extraction layer; The fifth convolutional layer, the fourth batch normalization layer, the fifth LeakyReLU activation function, the sixth convolutional layer, the fifth batch normalization layer, and the sixth LeakyReLU activation function of the discriminator network are connected in sequence to form the third feature extraction layer; The seventh convolutional layer, the seventh LeakyReLU activation function, the eighth convolutional layer, the sixth batch normalization layer, and the eighth LeakyReLU activation function of the discriminator network are connected in sequence to form the fourth feature extraction layer; The input to the discriminator network passes sequentially through the first feature extraction layer, the second feature extraction layer, the third feature extraction layer, the fourth feature extraction layer, the ninth convolutional layer, the adaptive average pooling layer, and the sigmoid function before being output.

4. The method for spatial downscaling of sea surface wind fields based on multi-task learning according to claim 1, characterized in that, The downsampling network contains three downsampling modules identical to those in the generator.

5. The method for spatial downscaling of sea surface wind fields based on multi-task learning according to claim 1, characterized in that, The total content loss of the generator networks for the sea surface wind field spatial downscaling task and auxiliary tasks is: : in, x As input to a multi-task learning-based network for downscaling sea surface wind fields, y For the original high-resolution data, This is the original high-resolution sea surface wind field data. For the original high-resolution auxiliary data, This represents the network for generating a downscaling task of sea surface wind fields. This represents information shared by auxiliary tasks, and s represents the downscaling result of additional variables. This represents the auxiliary task generator network. This indicates information shared regarding the spatial downscaling of sea surface wind fields. To aid in the weighting of task loss, Represents the L2 norm; The total loss function of the multi-task learning sea surface wind field downscaling network is designed as follows: This is low-resolution sea surface wind field data, among which, This is an indicator function that indicates when the network input is low-resolution data. The value is 1; when the network input is high-resolution data, =0, () represents the natural logarithm; To counteract the loss, the discriminator represents the high-resolution sea surface wind generated by the generator network. Authenticity assessment The input is low-resolution wind field data. For regression tasks, the loss represents the difference between the higher-resolution result generated by the generator network and the data downsampled to the same resolution as the generator network input, measured by pixel-level L2 distance. When the network input is high-resolution data, the loss of the first half of the network is 0, and the loss of the second half is... ; The loss of the discriminator is used to measure the discriminator's ability to distinguish raw sea surface wind field data. and generating data The error loss, which is expected to affect the true original data. Output 1, for the data of the generator network. The output is 0, using the cross-entropy loss function. The coefficients in the formula are used to balance the weights of its components, where... and As weight.

6. A spatial downscaling system for sea surface wind fields based on multi-task learning, characterized in that, include: The dataset acquisition module is used to acquire sea surface wind field data and preprocess the data, dividing the preprocessed data into training datasets and validation datasets. A generative adversarial network (GAN) construction module is used to construct a GAN, which consists of a generator network and a discriminator network. The generator network is used for downscaling low-resolution sea surface wind field data. The dual learning network construction module is used to construct a dual learning network. First, a downsampling network is constructed to degrade the high-resolution data generated by the generator network to low-resolution data. The downsampling network realizes the regression task of the generator network task. The generator network task and the downsampling network task are combined into a pair of dual tasks. The generator network in the generative adversarial network and the downsampling network are combined into a dual learning network. A network construction module for the spatial downscaling task of sea surface wind field is used to construct the network for the spatial downscaling task of sea surface wind field through generative adversarial networks and dual learning networks. The auxiliary task network construction module is used to construct an auxiliary task network by taking the spatial downscaling task of additional variables other than the sea surface wind field as an auxiliary task. The auxiliary task network contains the same network as the generator network in the spatial downscaling task network of the sea surface wind field. A multi-task learning sea surface wind field spatial downscaling network construction module is used to connect the auxiliary task network with the generator network in the generative adversarial network using a bridging network to construct a multi-task learning sea surface wind field spatial downscaling network, and the multi-task learning sea surface wind field spatial downscaling network is implemented using the PyTorch framework. The data aggregation module is used to divide the data of the additional variables into training datasets and validation datasets for the auxiliary tasks, and aggregate the training datasets and validation datasets of the two tasks to obtain multi-task training datasets and multi-task validation datasets. The training and testing module is used to train and optimize the multi-task learning sea surface wind field spatial downscaling network using a multi-task training dataset, test the trained multi-task learning sea surface wind field spatial downscaling network using a multi-task validation dataset, and validate the downscaling results using buoy data. The generator network adopts an encoder-decoder structure, including a bicubic interpolation algorithm module, three downsampling modules, three upsampled pattern modules, and two convolutional layers. The input of the generator network passes through the bicubic interpolation algorithm module, the first convolutional layer, the first downsampling module, the second downsampling module, the third downsampling module, the first upsampled pattern module, the second upsampled pattern module, the third upsampled pattern module, and the second convolutional layer in sequence before being output. The generator network uses skip connections, connecting the input after the first convolutional layer and before the first downsampling module to the input after the third upsampling module and before the second convolutional layer; connecting the input before the second downsampling module and after the first downsampling module to the input after the second upsampling module and before the third upsampling module; and connecting the input before the third downsampling module and after the second downsampling module to the input after the first upsampling module and before the second upsampling module. A bridging network is constructed to share features between the auxiliary task and the sea surface wind field spatial downscaling task. The bridging network module contains four bridge modules, each composed of 16 stacked residual channel attention modules (RCABs). This bridging network connects the auxiliary task network with the generator network for sea surface wind field spatial downscaling, forming a multi-task learning-based sea surface wind field downscaling network. The second and third upsampling modules of the auxiliary task network are connected to the first and second downsampling modules of the generator network of the sea surface wind field spatial downscaling task network via the first bridge module. The first downsampling module of the auxiliary task network... The second upsampling module and the second downsampling module are connected to the second upsampling module and the third upsampling module of the generator network of the sea surface wind field spatial downscaling task network via the second bridge module; the first upsampling module and the second upsampling module of the auxiliary task network are connected to the second downsampling module and the third downsampling module of the generator network of the sea surface wind field spatial downscaling task network via the third bridge module; the second downsampling module and the third downsampling module of the auxiliary task network are connected to the first upsampling module and the second upsampling module of the generator network of the sea surface wind field spatial downscaling task network via the fourth bridge module.

Images