Satellite remote sensing sea surface salinity prediction method, device, equipment and medium

Abstract

The application relates to a satellite remote sensing sea surface salinity prediction method, device, equipment and medium. Through preprocessing of obtained sea surface observation data, a training set is generated; historical salinity data is decomposed through an empirical orthogonal decomposition method to obtain time series information and spatial mode information; a spatial field is reconstructed based on the time series information and the spatial mode information, and then the reconstructed spatial field is supplemented into a channel dimension and combined with the training set to obtain a new training set; a satellite remote sensing sea surface salinity prediction model is constructed, the satellite remote sensing sea surface salinity prediction model is trained according to the new training set and a pre-constructed loss function, and a trained satellite remote sensing sea surface salinity prediction model is obtained; and the sea surface salinity is predicted through the trained satellite remote sensing sea surface salinity prediction model. The application can improve the prediction accuracy of the sea surface salinity and the interpretability of the prediction result, and provides prediction data support for development of marine resources and inversion of marine elements.

Description

Technical Field

[0001] This application relates to the field of marine engineering technology, and in particular to a satellite remote sensing method, device, equipment and medium for sea surface salinity forecasting. Background Technology

[0002] Sea surface salinity (SSS) plays a crucial role in studies of ocean cycles, climate change, and the global water cycle; therefore, the spatiotemporal variation patterns of SSS are of great research significance. In the past, limited observational methods resulted in relatively scarce research on ocean salinity, and forecasting studies were even rarer. In the 21st century, with the development of ocean salinity satellite remote sensing technology, such as the launches of the European Soil Moisture and Ocean Salinity Satellite (SMOS) in 2009, the US Aquarius / SAC-D satellite in 2011, and NASA's Soil Moisture Active Passive Sounding Satellite (SMAP) in 2015, real-time, large-scale salinity data have been obtained, greatly supporting research and forecasting of sea surface salinity.

[0003] However, the models currently used, such as the sea surface salinity statistical model that does not rely on physical mechanisms, the SMOS satellite sea surface salinity prediction model with support vector machine (GA-SVM) under genetic algorithm parameter optimization, and the sea surface salinity prediction model established by the BP neural network method, usually use purely data-driven statistical methods to predict sea surface salinity. They lack the constraints of physical laws and are therefore lacking in forecast accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide a satellite remote sensing sea surface salinity forecasting method, device, equipment, and medium that considers physical constraints and can improve the accuracy and interpretability of short-term sea surface salinity forecasts, thereby addressing the aforementioned technical problems.

[0005] A satellite remote sensing method for sea surface salinity prediction, the method comprising:

[0006] Acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window;

[0007] Historical salinity data is acquired, and the historical salinity data is decomposed using the empirical orthogonal decomposition method to obtain time series information and spatial modal information;

[0008] The spatial field is reconstructed based on the time series information and spatial modal information. Then, the reconstructed spatial field is added to the channel dimension and combined with the training set to obtain a new training set.

[0009] A satellite remote sensing sea surface salinity forecasting model is constructed. The model is trained using a new training set and a pre-built loss function to obtain a trained model. The model includes several backbone convolutional layers, with convolutional attention layers between adjacent backbone layers, and a residual mechanism is used for information transfer. The pre-built loss function is designed based on EOF physical constraints.

[0010] The trained satellite remote sensing sea surface salinity prediction model is used to predict sea surface salinity.

[0011] In one embodiment, sea surface observation data is acquired, the sea surface observation data is preprocessed, and a training set is generated by sliding a time window, including:

[0012] Acquire sea surface observation data, which includes: daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data;

[0013] The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data are processed using bilinear interpolation to obtain daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution.

[0014] The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are spliced ​​together along the time dimension, and the latitude and longitude range is selected; and

[0015] The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are stitched together along the channel dimension, and then missing values ​​and land grid points are marked.

[0016] Based on the selected latitude and longitude range, the daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are normalized to obtain normalized data.

[0017] Based on the labeled missing values ​​and land grid points, singular value processing is performed on the normalized data. Then, training data is selected according to the requirements, the input time step and prediction time step of the model are set, and the training set is generated by sliding the time window.

[0018] In one embodiment, historical salinity data is acquired, and the historical salinity data is decomposed using an empirical orthogonal decomposition method to obtain time series information and spatial modal information, including:

[0019] Based on the training set, historical salinity data is obtained, and the historical salinity data is decomposed using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0020] The expression for the empirical orthogonal decomposition method is:

[0021] X m×n =EOF m×m ×PCs m×n ;

[0022] In the formula, X represents the original data, EOF represents the spatial modal information obtained from the decomposition, PCs represents the time series information obtained from the decomposition, m represents the number of grid points in the spatial field, and n represents the observation time series length of each grid point.

[0023] In one embodiment, the spatial field is reconstructed based on the time series information and spatial modal information, and then the reconstructed spatial field is added to the channel dimension and combined with the training set to obtain a new training set, including:

[0024] Several major modes related to the physical laws of sea surface salinity were selected, and then the time series information and spatial mode information of each mode were reconstructed back into the spatial field corresponding to each mode.

[0025] The reconstructed spatial field is added to the channel dimension and combined with the training set to obtain five-dimensional data, which is then used as a new training set.

[0026] In one embodiment, the preset loss function is expressed as:

[0027] EOF_Loss=SSS_Loss+ω×PCs_Loss;

[0028] In the formula, ω is the weight of the PCs loss value; SSS_Loss is the MSE loss of the dimensionless sea surface salinity; and PCs_Loss is the PCs loss of the dimensionless sea surface salinity.

[0029] In one embodiment, the expression for the dimensionless sea surface salinity MSE loss SSS_Loss is:

[0030]

[0031] In the formula, x(SSS) represents the true value of sea surface salinity; The value represents the predicted sea surface salinity; m represents the number of grid points in the spatial field; batch represents the number of samples set during model training; SSS train Sea surface salinity data representing the entire training set.

[0032] In one embodiment, the expression for the PCs loss (PCs_Loss) of dimensionless sea surface salinity is:

[0033]

[0034] In the formula, x(PCs) represents PCs obtained by projecting the true values; PCs represents the PCs obtained by projecting the predicted values; m represents the number of grid points in the spatial field; batch represents the number of samples set during the model training process; PCs train PCs represent the entire training set.

[0035] A satellite remote sensing sea surface salinity forecasting device, the device comprising:

[0036] The training set acquisition module is used to acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window.

[0037] The data decomposition module is used to acquire historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0038] The training set reorganization module is used to reconstruct the spatial field based on the time series information and spatial modality information, and then supplement the reconstructed spatial field into the channel dimension and combine it with the training set to obtain a new training set;

[0039] The model training module is used to construct a satellite remote sensing sea surface salinity forecasting model. It trains the model using a new training set and a pre-built loss function to obtain a trained model. The model includes several backbone convolutional layers, with convolutional attention layers between adjacent layers, and uses a residual mechanism for information transfer. The pre-built loss function is designed based on EOF physical constraints.

[0040] The satellite remote sensing sea surface salinity forecasting module is used to forecast sea surface salinity using the trained satellite remote sensing sea surface salinity forecasting model.

[0041] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods described above.

[0042] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0043] The aforementioned satellite remote sensing sea surface salinity forecasting method, device, equipment, and medium first acquire sea surface observation data, preprocess the data, and generate a training set through a sliding time window; acquire historical salinity data, decompose the historical salinity data using an empirical orthogonal decomposition method to obtain time series information and spatial modal information; reconstruct the spatial field based on the time series information and spatial modal information, then supplement the channel dimension with the reconstructed spatial field and combine it with the training set to obtain a new training set; construct a satellite remote sensing sea surface salinity forecasting model, train the model using the new training set and a pre-constructed loss function, and obtain a trained satellite remote sensing sea surface salinity forecasting model; the satellite remote sensing sea surface salinity forecasting model includes several backbone convolutional layers, with convolutional attention layers set between adjacent backbone convolutional layers, and a residual mechanism is used for information transfer; the pre-constructed loss function is designed based on EOF physical constraints; finally, the trained satellite remote sensing sea surface salinity forecasting model is used to forecast sea surface salinity.

[0044] This invention identifies the time-series and spatial modal information of data through empirical orthogonal decomposition, and reorganizes the training set based on the time-series and spatial modal information to ensure that the new training set is more reasonable and reliable, thereby improving the usability and interpretability of the model in practical applications. The constructed satellite remote sensing sea surface salinity forecasting model employs convolutional attention and residual mechanisms to fully extract the statistical features of sea surface salinity. Deep transmission of feature information ensures efficient information transfer. Furthermore, the design of the loss function considers the impact of the EOF physical constraint on sea surface salinity, effectively improving the forecast accuracy and interpretability of the forecast results, and providing forecast data support for marine resource development and the inversion of marine elements. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a satellite remote sensing sea surface salinity prediction method in one embodiment;

[0046] Figure 2 This is a schematic diagram of the backbone convolutional layer structure of a satellite remote sensing sea surface salinity prediction model in one embodiment;

[0047] Figure 3 This is a schematic diagram of the channel attention layer structure in one embodiment;

[0048] Figure 4 This is a schematic diagram of the spatial attention layer structure in one embodiment;

[0049] Figure 5 This is a schematic diagram of the overall network structure of a satellite remote sensing sea surface salinity prediction model in one embodiment;

[0050] Figure 6This is a structural block diagram of a satellite remote sensing sea surface salinity forecasting device in one embodiment;

[0051] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Understandably, the inventors noticed that traditional sea surface salinity prediction generally employs statistical methods, and machine learning algorithms are essentially a type of statistical method. Therefore, some machine learning-based prediction models have emerged in the existing technology. However, these technologies lack the constraints of physical laws and are somewhat lacking in forecast accuracy. The inventors discovered that deep learning technology possesses powerful end-to-end learning capabilities. Based on this, the inventors provide a new approach and method for satellite remote sensing sea surface salinity prediction. Firstly, by processing the training set, a more reasonable and reliable new training set is obtained, thereby improving the usability and interpretability of the model in subsequent model training. Furthermore, in the model design, convolutional attention and residual mechanisms were adopted in the satellite remote sensing sea surface salinity forecast model, and combined with the EOF physical information of sea surface salinity, to propose a new satellite remote sensing sea surface salinity forecast model, which can also be called the EOF_SNet satellite remote sensing sea surface salinity forecast model. The EOF_SNet satellite remote sensing sea surface salinity forecast model can effectively improve the forecast accuracy and interpretability of sea surface salinity, and provide forecast data support for the development of marine resources and the inversion of marine elements.

[0054] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0055] In one embodiment, such as Figure 1 As shown, a satellite remote sensing method for sea surface salinity prediction is provided, including the following steps:

[0056] Step 202: Obtain sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window.

[0057] The selection of training sets plays a crucial role in machine learning. Choosing a reasonable and reliable training set can provide effective learning data and information for model training, which is very important for training high-quality models.

[0058] This embodiment first processes the acquired buoy observation data, unifying different elements in the data to the same spatial resolution. Then, it stitches the various elements in the buoy observation data along the time dimension, selecting a reasonable range of longitude and latitude. It also stitches the various elements along the channel dimension, marking missing values ​​and land grid points for each element. Normalization is then performed, and finally, outliers are distinguished from normal values ​​to generate a training set.

[0059] Step 204: Obtain historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0060] It is understandable that Empirical Orthogonal Decomposition (EOF analysis) can effectively understand the main patterns of change in the ocean, and rationally select ground-state PCs and ground-state EOFs based on the variance contribution rate of each mode. This can improve the predictive ability of future changes and help models predict possible future trends.

[0061] Step 206: Reconstruct the spatial field based on time series information and spatial modal information, then supplement the channel dimension with the reconstructed spatial field and combine it with the training set to obtain a new training set.

[0062] It is understandable that the Empirical Orthogonal Factorization (EOF analysis) in step 204 typically reflects the main physical change patterns in the data. By recombining the training set and introducing these physical constraints, the spatial modal information of EOF is added to the input training data. This makes the subsequent model learning more consistent with actual physical laws, improves the model's rationality and interpretability, enhances its generalization performance, and increases the diversity of training data, thus improving the model's robustness. This makes the model more adaptable to different spatial distributions and change patterns.

[0063] Step 208: Construct a satellite remote sensing sea surface salinity forecasting model. Train the satellite remote sensing sea surface salinity forecasting model based on the new training set and the pre-constructed loss function to obtain a trained satellite remote sensing sea surface salinity forecasting model. The satellite remote sensing sea surface salinity forecasting model includes several backbone convolutional layers, with convolutional attention layers set between each adjacent backbone convolutional layer, and a residual mechanism is used for information transfer. The pre-constructed loss function is designed based on EOF physical constraints.

[0064] As can be understood, this embodiment, inspired by the powerful end-to-end learning capabilities of deep learning, constructs a satellite remote sensing sea surface salinity forecasting model. By adding a convolutional attention layer to the constructed backbone convolutional layer, valuable features can be extracted, and information is transferred through a residual mechanism to ensure efficient information transfer. In constructing the model's loss function, the introduction of EOF physical constraints is considered. Based on the MSE loss function, the time-series information and spatial modal information of EOF are incorporated to design the model's loss function.

[0065] Step 210: Forecast sea surface salinity using the trained satellite remote sensing sea surface salinity forecasting model.

[0066] The aforementioned satellite remote sensing sea surface salinity forecasting method first acquires sea surface observation data, preprocesses the data, and generates a training set through a sliding time window. It then acquires historical salinity data and decomposes it using an empirical orthogonal decomposition method to obtain time-series and spatial modal information. Based on the time-series and spatial modal information, it reconstructs the spatial field, adds the reconstructed spatial field to the channel dimension, and combines it with the training set to obtain a new training set. Next, it constructs a satellite remote sensing sea surface salinity forecasting model, training it using the new training set and a pre-built loss function. This model includes several backbone convolutional layers, with convolutional attention layers between adjacent backbone layers, and uses a residual mechanism for information transfer. The pre-built loss function is designed based on EOF physical constraints. Finally, it uses the trained satellite remote sensing sea surface salinity forecasting model to forecast sea surface salinity.

[0067] This invention identifies the time-series and spatial modal information of data through empirical orthogonal decomposition, and reorganizes the training set based on the time-series and spatial modal information to ensure that the new training set is more reasonable and reliable, thereby improving the usability and interpretability of the model in practical applications. The constructed satellite remote sensing sea surface salinity forecasting model employs convolutional attention and residual mechanisms to fully extract the statistical features of sea surface salinity. Deep transmission of feature information ensures efficient information transfer. Furthermore, the design of the loss function considers the impact of the EOF physical constraint on sea surface salinity, effectively improving the forecast accuracy and interpretability of the forecast results, and providing forecast data support for marine resource development and the inversion of marine elements.

[0068] In one embodiment, sea surface observation data is acquired, preprocessed, and a training set is generated by sliding a time window, including:

[0069] Acquire sea surface observation data, including daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data.

[0070] Bilinear interpolation was used to process daily sea surface salinity, sea surface temperature, wind speed, and sea surface height data to obtain daily sea surface salinity, sea surface temperature, wind speed, and sea surface height data with the same spatial resolution.

[0071] Daily sea surface salinity, temperature, wind speed, and height data with the same spatial resolution are stitched together along the time dimension, selecting a latitude and longitude range; and

[0072] Daily sea surface salinity, temperature, wind speed, and height data with the same spatial resolution are stitched together along the channel dimension, and then missing values ​​and land grid points are marked.

[0073] Based on the selected latitude and longitude range, daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are normalized to obtain normalized data.

[0074] Based on the labeled missing values ​​and land grid points, singular value processing is performed on the normalized data. Then, training data is selected according to the requirements, the input time step and prediction time step of the model are set, and the training set is generated by sliding the time window.

[0075] Specifically, obtain daily average sea surface salinity data with a spatial resolution of r1°×r1°, daily average sea surface temperature data with a spatial resolution of r2°×r2°, daily average sea surface wind speed (meridian and zonal) data with a spatial resolution of r3°×r3°, and daily average spatial resolution data from the official data website. The sea surface height data mentioned above cover a consistent spatial range, all in longitude. arrive Dimension arrive The time span is n years.

[0076] To unify different elements in the data to the same spatial resolution, the Bilinear interpolation method was used to unify the resolution of daily sea surface salinity, daily sea surface temperature, daily wind speed, and daily sea surface height data to r°×r°. In the Bilinear interpolation method, each grid of the target resolution is derived from the values ​​of the four grids surrounding the target in the original resolution, that is, linear interpolation is performed on both the x-axis and y-axis.

[0077] First, perform two linear interpolations along the x-axis to obtain the values ​​of T1 and T2, as shown in the following formula:

[0078]

[0079]

[0080] In the formula, P 11 P 21 P 12 P 22 Represents the values ​​of any four grid points in the original resolution feature, where P 11 P 21 P 12 P 22 The coordinates are (x1, y1), (x1, y2), (x2, y1), and (x2, y2).

[0081] Then, linear interpolation is performed along the y-axis to obtain the final value Q, as shown in the following formula:

[0082]

[0083] It is understandable that bilinear interpolation results are smoother and satisfy the linear transition law, and is often used for interpolation processing of meteorological and oceanographic element products.

[0084] The daily sea surface salinity, temperature, wind speed, and height data at the same spatial resolution were sequentially stitched together along the time dimension, and the latitude and longitude range was selected according to the actual study area; and

[0085] Daily sea surface salinity, temperature, wind speed, and height data with the same spatial resolution are concatenated along the channel dimension, and then missing values ​​and land grid points are labeled. For example, missing values ​​and land grid points among the four data factors are set to NaN.

[0086] Based on the selected latitude and longitude range, daily sea surface salinity, temperature, wind speed, and height data with the same spatial resolution are normalized to obtain normalized data. The normalization formula is expressed as follows:

[0087] x new =(x old -min) / (max-min)*50;

[0088] Where, x old x represents the data before normalization. new This represents the normalized data.

[0089] Based on the marked missing values ​​and land grid points, singular value processing is performed on the normalized data. For example, the NaN value is set to -50 to distinguish it from normal values.

[0090] Then, select training data according to requirements, set the input time step and prediction time step of the model, and generate the training set by sliding the time window. For example, you can select historical data of 10 years or more as the training set and reserve data of 1 year or more as the test set, and then set the input time step and prediction time step of the model, and obtain the training set and test set respectively by sliding the time window.

[0091] Finally, the processed data is stored as an npz file.

[0092] In one embodiment, historical salinity data is acquired, and the historical salinity data is decomposed using an empirical orthogonal decomposition method to obtain time series information and spatial modal information, including:

[0093] Based on the training set, historical salinity data is obtained, and the historical salinity data is decomposed using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0094] The expression for the empirical orthogonal decomposition method is:

[0095] X m×n =EOF m×m ×PCs m×n ;

[0096] In the formula, X represents the original data, EOF represents the spatial modal information obtained from the decomposition, PCs represents the time series information obtained from the decomposition, m represents the number of grid points in the spatial field, and n represents the observation time series length of each grid point.

[0097] Specifically, the EOF analysis formula is used to decompose selected historical sea surface salinity (SSS) data of 10 years or more into time series (PCs) and spatial modes (EOFs).

[0098] It is understandable that EOF analysis will obtain the variance contribution rate of each mode. Based on the variance contribution rate of each mode, the number of modes whose total variance contribution rate is above 85% or 90% is selected as the ground state PCs and ground state EOFs according to actual needs.

[0099] In one embodiment, the spatial field is reconstructed based on time series information and spatial modal information. Then, the reconstructed spatial field is added to the channel dimension and combined with the training set to obtain a new training set, including:

[0100] Several major modes related to the physical laws of sea surface salinity were selected, and then the time series information and spatial mode information of each mode were reconstructed back into the spatial field corresponding to each mode.

[0101] The reconstructed spatial field is added to the channel dimension and combined with the training set to obtain five-dimensional data, which is then used as a new training set.

[0102] Specifically, based on actual needs, three main modes related to the physical laws of sea surface salinity are selected, and the EOF spatial and temporal modes of each mode are reconstructed back to the spatial field corresponding to each mode.

[0103] The reconstructed spatial field is added to the channel dimension, forming a new dataset with the original training set. The constructed dataset is now five-dimensional, consisting of [number of samples, time series length, number of channels, height, and width]. This five-dimensional dataset is used as the new training set.

[0104] It is worth noting that the selection of the main modes is determined according to actual needs. The three modes proposed in this embodiment can be used, or more or fewer main modes related to the physical laws of sea surface salinity can be selected.

[0105] In one embodiment, the loss function of the satellite remote sensing sea surface salinity prediction model is designed based on the MSE loss function, incorporating time series and spatial modal information of EOF. The preset loss function is expressed as follows:

[0106] EOF_Loss=SSS_Loss+ω×PCs_Loss;

[0107] In the formula, ω is the weight of the PCs loss value; SSS_Loss is the MSE loss of the dimensionless sea surface salinity; and PCs_Loss is the PCs loss of the dimensionless sea surface salinity.

[0108] Specifically, using MSE as the basic loss function, the grid-by-grid loss value of SSS is calculated. The formula for calculating MSE is as follows:

[0109]

[0110] In the formula, x is the actual value. This is the predicted value, where m represents the number of grid points in the spatial field.

[0111] Project the predicted and actual values ​​of the SSS of each batch onto the ground state EOFs to obtain the PCs of the predicted and actual values. Calculate the PCs loss value using the following formula:

[0112]

[0113] In the formula, x(PCs) represents PCs obtained by projecting the true values; PCs represent the projected values ​​of the forecast; batch represents the number of samples set during the model training process.

[0114] The MSE loss value of SSS and the loss value of PCs are dedimensionalized, and their variances are removed. The formula for calculating the variance is as follows:

[0115]

[0116] In the formula, Let n be the average value and n be the number of sample points. Then, the dimensionless MSE loss and PCs loss of SSS are respectively:

[0117]

[0118]

[0119] In the formula, x(SSS) represents the true value of sea surface salinity; The value represents the predicted sea surface salinity; m represents the number of grid points in the spatial field; SSS train Sea surface salinity data representing the entire training set; PCs train PCs represent the entire training set.

[0120] In one embodiment, a satellite remote sensing sea surface salinity prediction model is built, and multiple satellite remote sensing sea surface salinity prediction models (hereinafter referred to as EOF_SNet models) are constructed and trained according to different EOF weights in the loss function. The optimal model is found through comparative experiments.

[0121] Specifically, you can set up a deep learning environment, for example, by installing the Python-based deep learning framework PyTorch on a server.

[0122] Construct the backbone convolutional layers of the EOF_SNet model, such as Figure 2 As shown, the backbone convolutional layer consists of three ConvLSTM layers. The first layer contains 512 5×5 convolutional kernels with a stride of 1, zero-padding, and two layers of edge padding. The second layer contains 512 3×3 convolutional kernels with a stride of 1, zero-padding, and one layer of edge padding. The third layer contains 256 3×3 convolutional kernels with a stride of 1, zero-padding, and one layer of edge padding. Within each ConvLSTM layer, the cell state and hidden state are obtained through the Sigmoid activation function and the Tanh activation function, respectively.

[0123] The convolutional attention layer (CBAM layer) of the EOF_SNet model is divided into channel attention layer (CAM) and spatial attention layer (SAM), as follows: Figure 3 and Figure 4 As shown, CAM processes the input 3D feature map (channels × height × width) through global max pooling and global average pooling, respectively, to obtain two channel-dimensional feature maps (channels × 1 × 1). These are then fed into a first fully connected layer, where the number of neurons is one-quarter of the number of channels, and the activation function is ReLU. The second fully connected layer, where the number of neurons is the same as the number of channels, also uses ReLU. This two-layer network is shared. The output feature values ​​of both layers are then summed and activated using a Sigmoid activation function to generate the final channel attention weights. Finally, the channel attention weights are multiplied by the input 3D feature map to obtain a feature map with superimposed channel attention. SAM first performs channel-based global max pooling and global average pooling on the input 3D feature map (channels × height × width) to obtain a feature map with two spatial dimensions (1 × height × width). Then, these two feature maps are concatenated based on the channel dimension, doubling the number of channels to 2 × height × width. Next, a convolutional network is passed through this layer with a kernel size of 7 × 7, one kernel, and a stride of 1, outputting a single channel (1 × height × width). Then, a sigmoid activation function is applied to obtain spatial attention weights. Finally, the spatial attention weights are multiplied by the input 3D feature map to obtain a feature map with superimposed spatial attention.

[0124] The final EOF_SNet model is constructed by embedding CBAM layers between the backbone convolutional layers and employing two Res residual structures to ensure efficient information transfer, such as... Figure 5 As shown, a CBAM layer is embedded between the first and second backbone convolutional layers. The output of the first backbone layer is added to the output of the CBAM and then input into the second backbone layer, forming the first type of Res residual structure. The result of the second backbone layer is input into the second CBAM layer, and the result of the first CBAM layer is added to the result of the second CBAM layer and then input into the third backbone layer, resulting in the second type of Res residual structure. The result of the third backbone layer is input into a convolutional layer with a kernel size of 1, a stride of 1, and the number of kernels equal to the number of channels in the salinity data, thus obtaining the final output of the model.

[0125] The weight values ​​ω in the EOF loss function are preset based on prior knowledge. For example, four different weight values ​​can be preset to train EOF_SNet models with different weight values. The number of iterations is initialized to epochs, where epochs are positive integers; the initial learning rate is lr, usually less than or equal to 0.001; the initial batch size is batch, usually greater than or equal to 16; and the Adam iterative optimizer is used. Finally, the optimal EOF_SNet model is obtained by evaluating the root mean square error (RMSE), mean absolute error (MAE), and spatial structure similarity index (SSIM).

[0126] Save the parameters of the optimal EOF_SNet model for later use.

[0127] Finally, the optimal EOF_SNet model parameters are loaded, and the reserved sea surface salinity test set is input to obtain the short-term forecast results of sea surface salinity.

[0128] The EOF_SNet model constructed in this invention utilizes convolutional attention and Res residual mechanisms to fully extract the statistical features of sea surface salinity and deeply propagates this feature information, ensuring efficient information transfer. Compared to purely data-driven deep learning forecasting models, this invention introduces the EOF physical information of sea surface salinity to construct a deep learning forecasting model incorporating physical constraints. Compared to existing research on deep learning forecasting of sea surface salinity, this invention improves the accuracy and interpretability of short-term sea surface salinity forecasts, providing forecast data support for the development and utilization of marine resources.

[0129] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0130] In one embodiment, such as Figure 6 As shown, a satellite remote sensing sea surface salinity forecasting device is provided, comprising: a training set acquisition module 402, a data decomposition module 404, a training set recombination module 406, a model training module 408, and a satellite remote sensing sea surface salinity forecasting module 410, wherein:

[0131] The training set acquisition module 402 is used to acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window.

[0132] The data decomposition module 404 is used to acquire historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0133] The training set reorganization module 406 is used to reconstruct the spatial field based on time series information and spatial modal information, and then supplement the channel dimension with the reconstructed spatial field and combine it with the training set to obtain a new training set.

[0134] The model training module 408 is used to construct a satellite remote sensing sea surface salinity forecasting model. It trains the satellite remote sensing sea surface salinity forecasting model based on a new training set and a pre-constructed loss function to obtain a trained satellite remote sensing sea surface salinity forecasting model. The satellite remote sensing sea surface salinity forecasting model includes several backbone convolutional layers, with convolutional attention layers set between each adjacent backbone convolutional layer, and a residual mechanism is used for information transfer. The pre-constructed loss function is designed based on EOF physical constraints.

[0135] The satellite remote sensing sea surface salinity forecasting module 410 is used to forecast sea surface salinity using a trained satellite remote sensing sea surface salinity forecasting model.

[0136] Specific limitations regarding the satellite remote sensing sea surface salinity forecasting device can be found in the limitations of the satellite remote sensing sea surface salinity forecasting method described above, and will not be repeated here. Each module in the aforementioned satellite remote sensing sea surface salinity forecasting device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0137] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores satellite remote sensing sea surface salinity forecast data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a satellite remote sensing sea surface salinity forecasting method.

[0138] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps:

[0140] Step 202: Obtain sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window.

[0141] Step 204: Obtain historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0142] Step 206: Reconstruct the spatial field based on time series information and spatial modal information, then supplement the channel dimension with the reconstructed spatial field and combine it with the training set to obtain a new training set.

[0143] Step 208: Construct a satellite remote sensing sea surface salinity forecasting model. Train the satellite remote sensing sea surface salinity forecasting model based on the new training set and the pre-constructed loss function to obtain a trained satellite remote sensing sea surface salinity forecasting model. The satellite remote sensing sea surface salinity forecasting model includes several backbone convolutional layers, with convolutional attention layers set between each adjacent backbone convolutional layer, and a residual mechanism is used for information transfer. The pre-constructed loss function is designed based on EOF physical constraints.

[0144] Step 210: Forecast sea surface salinity using the trained satellite remote sensing sea surface salinity forecasting model.

[0145] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0146] Step 202: Obtain sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window.

[0147] Step 204: Obtain historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information.

[0148] Step 206: Reconstruct the spatial field based on time series information and spatial modal information, then supplement the channel dimension with the reconstructed spatial field and combine it with the training set to obtain a new training set.

[0149] Step 208: Construct a satellite remote sensing sea surface salinity forecasting model. Train the satellite remote sensing sea surface salinity forecasting model based on the new training set and the pre-constructed loss function to obtain a trained satellite remote sensing sea surface salinity forecasting model. The satellite remote sensing sea surface salinity forecasting model includes several backbone convolutional layers, with convolutional attention layers set between each adjacent backbone convolutional layer, and a residual mechanism is used for information transfer. The pre-constructed loss function is designed based on EOF physical constraints.

[0150] Step 210: Forecast sea surface salinity using the trained satellite remote sensing sea surface salinity forecasting model.

[0151] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A satellite remote sensing method for predicting sea surface salinity, characterized in that, The method includes: Acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window; Historical salinity data is acquired, and the historical salinity data is decomposed using the empirical orthogonal decomposition method to obtain time series information and spatial modal information; The spatial field is reconstructed based on the time series information and spatial modal information. Then, the reconstructed spatial field is added to the channel dimension and combined with the training set to obtain a new training set. A satellite remote sensing sea surface salinity forecasting model is constructed. The model is trained using a new training set and a pre-built loss function to obtain a trained model. The model includes several backbone convolutional layers, with convolutional attention layers between adjacent backbone layers, and a residual mechanism is used for information transfer. The pre-built loss function is designed based on EOF physical constraints. The trained satellite remote sensing sea surface salinity prediction model is used to predict sea surface salinity.

2. The satellite remote sensing sea surface salinity prediction method according to claim 1, characterized in that, Acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding a time window, including: Acquire sea surface observation data, which includes: daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data; The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data are processed using bilinear interpolation to obtain daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution. The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are spliced ​​together along the time dimension, and the latitude and longitude range is selected; and The daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are stitched together along the channel dimension, and then missing values ​​and land grid points are marked. Based on the selected latitude and longitude range, the daily sea surface salinity data, daily sea surface temperature data, daily wind speed data, and daily sea surface height data with the same spatial resolution are normalized to obtain normalized data. Based on the labeled missing values ​​and land grid points, singular value processing is performed on the normalized data. Then, training data is selected according to the requirements, the input time step and prediction time step of the model are set, and the training set is generated by sliding the time window.

3. The satellite remote sensing sea surface salinity prediction method according to claim 2, characterized in that, Historical salinity data is acquired and decomposed using an empirical orthogonal decomposition method to obtain time series information and spatial modal information, including: Based on the training set, historical salinity data is obtained, and the historical salinity data is decomposed using the empirical orthogonal decomposition method to obtain time series information and spatial modal information. The expression for the empirical orthogonal decomposition method is: X m×n =EOF m×m ×PCs m×n ; In the formula, X represents the original data, EOF represents the spatial modal information obtained from the decomposition, PCs represents the time series information obtained from the decomposition, m represents the number of grid points in the spatial field, and n represents the observation time series length of each grid point.

4. The satellite remote sensing sea surface salinity prediction method according to claim 3, characterized in that, The spatial field is reconstructed based on the time series information and spatial modal information. Then, the reconstructed spatial field is added to the channel dimension and combined with the training set to obtain a new training set, including: Several major modes related to the physical laws of sea surface salinity were selected, and then the time series information and spatial mode information of each mode were reconstructed back into the spatial field corresponding to each mode. The reconstructed spatial field is added to the channel dimension and combined with the training set to obtain five-dimensional data, which is then used as a new training set.

5. The satellite remote sensing sea surface salinity forecasting method according to any one of claims 1 to 4, characterized in that, The preset loss function is expressed as: EOF_Loss=SSS_Loss+ω×PCs_Loss; In the formula, ω is the weight of the PCs loss value; SSS_Loss represents the MSE loss of sea surface salinity after dimensionless measurement; PCs_Loss represents the PCs loss of sea surface salinity after dimensionless measurement.

6. The satellite remote sensing sea surface salinity prediction method according to claim 5, characterized in that, The expression for the dimensionless sea surface salinity MSE loss SSS_Loss is: In the formula, x(SSS) represents the true value of sea surface salinity; The value represents the predicted sea surface salinity; m represents the number of grid points in the spatial field; batch represents the number of samples set during model training; SSS train Sea surface salinity data representing the entire training set.

7. The satellite remote sensing sea surface salinity prediction method according to claim 5, characterized in that, The expression for PCs loss (PCs_Loss) of dimensionless sea surface salinity is: In the formula, x(PCs) represents PCs obtained by projecting the true values; PCs represents the PCs obtained by projecting the predicted values; m represents the number of grid points in the spatial field; batch represents the number of samples set during the model training process; PCs train PCs represent the entire training set.

8. A satellite remote sensing sea surface salinity forecasting device, characterized in that, The device includes: The training set acquisition module is used to acquire sea surface observation data, preprocess the sea surface observation data, and generate a training set by sliding through a time window. The data decomposition module is used to acquire historical salinity data and decompose the historical salinity data using the empirical orthogonal decomposition method to obtain time series information and spatial modal information. The training set reorganization module is used to reconstruct the spatial field based on the time series information and spatial modality information, and then supplement the reconstructed spatial field into the channel dimension and combine it with the training set to obtain a new training set; The model training module is used to construct a satellite remote sensing sea surface salinity forecasting model. It trains the model using a new training set and a pre-built loss function to obtain a trained model. The model includes several backbone convolutional layers, with convolutional attention layers between adjacent layers, and uses a residual mechanism for information transfer. The pre-built loss function is designed based on EOF physical constraints. The satellite remote sensing sea surface salinity forecasting module is used to forecast sea surface salinity using the trained satellite remote sensing sea surface salinity forecasting model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images