Bias correction method for domestic satellite altimeter fusion products based on deep learning

Abstract

The present invention belongs to the field of satellite altimeter data deviation correction, and specifically relates to a method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning. The present invention includes the following steps: S1, obtaining satellite remote sensing data as input and label data, and integrating the obtained input and label data sets into a certain organizational structure and dividing them into subsets for model training, verification and testing; S2, building a deep learning network A2A; S3, using the data set obtained in step S1 to alternately train and verify the deep learning network A2A obtained in step S2; S4, using the trained deep learning network A2A to correct the domestic satellite altimeter fusion product, that is, to obtain domestic satellite altimeter data with higher accuracy. The present invention builds a model based on a deep learning framework and uses public data sets for model training and prediction. It is more convenient to operate and has stronger generalization ability than traditional methods, and provides a new and effective way to improve the accuracy of domestic satellite altimeter fusion products.

Description

Technical Field

[0001] The present invention belongs to the field of satellite altimeter data deviation correction, and specifically relates to a domestic satellite altimeter fusion product deviation correction method based on deep learning. Background Art

[0002] Satellite altimeters are active, spaceborne microwave radars that can accurately measure global dynamic topography, providing a powerful new tool for in-depth understanding of ocean topography. To overcome discrepancies between altimeter observations from different satellites, gridded, multi-satellite altimeter fusion products are often generated based on appropriate fusion algorithms. For example, the Chinese invention patent "A System and Method for Adaptive Fusion of Vertical Bias Based on Multi-Source Satellite Altimeter Data" (Application No. 202410884624.8) employs a classification and preprocessing method for multi-source altimeter data. In 2023, my country's National Satellite Ocean Application Center released the ALT MUL (A fusion product of sea surface satellite altimeters). However, based on independent measured sea temperature data and drifting buoy data, an analysis and evaluation found that while the domestically produced ALT MUL (A fusion product of sea surface satellite altimeters) has achieved certain accuracy and precision standards compared to mainstream international fusion products of satellite altimeters (DUACS, also known as AVISO products), there are still gaps. To better provide high-precision data support for scientific research, it is necessary to use mainstream international high-precision satellite altimeter products as a benchmark to correct the deviations of the domestically produced ALT MUL. This invention will use deep learning technology to correct the deviation of domestic satellite altimeter fusion products, improve the accuracy of the products, thereby further broadening their application and enhancing the understanding of domestic ocean satellite altimeter fusion products.

[0003] In the field of deep learning image reconstruction (IR), CNN- or Transformer-based frameworks are used to reconstruct input images, achieving high-quality reconstruction effects such as image denoising and detail restoration. This paper proposes a new deep learning framework combining CNN and Transformer, which can capture both local and global features and increase the range of reference information for bias correction. At the same time, it innovatively proposes a land filtering mechanism to block interference from land area information in global attention calculations, thereby improving the accuracy and efficiency of attention calculations. Furthermore, this paper integrates multi-source remote sensing sea surface observation data, all from publicly available channels, based on an open-source deep learning framework. This method rapidly corrects the bias of domestic satellite altimeter fusion products, outputting higher-precision global absolute dynamic altitude and sea surface height anomaly data. The model is easy to deploy and has fast inference speed, which is conducive to widespread application and provides a new and effective approach to improving the accuracy of domestic satellite altimeter fusion products. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for correcting the deviation of domestic satellite altimeter fusion products based on deep learning. The present invention is based on a domestic satellite altimeter fusion product deviation correction network A2A (Altimeter to Altimeter Network) based on deep learning, which aims to solve the problem of low accuracy of current domestic satellite altimeter fusion products. A2A can fuse five types of satellite sea surface observation data, including global high-precision satellite sea surface temperature, sea surface wind speed (longitudinal and latitudinal), longitude, and latitude, with the absolute dynamic altitude and sea surface height anomaly data of domestic satellite altimeter fusion products, quickly correct the deviation of domestic satellite altimeter fusion products, and output higher-precision global absolute dynamic altitude and sea surface height anomaly data, providing a new and effective way to improve the accuracy of domestic satellite altimeter fusion products.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for correcting the deviation of domestic satellite altimeter fusion products based on deep learning includes the following steps:

[0007] S1. Obtain satellite remote sensing data from open source channels as input and label data, integrate the obtained input and label data sets into a certain organizational structure and divide them into subsets for model training, validation and testing;

[0008] S2. Build a deep learning network A2A. The expression of A2A is as follows:

[0009]

[0010] in, The mapping relationship between the input and output of the deep learning network A2A is shown. The input is a normalized multivariate input batch. The output is the absolute dynamic altitude after deviation correction and the sea surface height anomaly data batch, Indicates the first Layer weights and biases;

[0011] S3. Alternately train and validate the deep learning network A2A obtained in step S2 using the data set obtained in step S1;

[0012] S4. Use the trained deep learning network A2A to correct the domestic satellite altimeter fusion product, that is, to obtain higher-precision domestic satellite altimeter data.

[0013] Preferably, step S1 specifically includes the following steps:

[0014] S11. Obtain seven variables as input, namely, satellite sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, global longitude grid field, global latitude grid field, and absolute dynamic altitude and sea surface height anomaly of domestic satellite altimeter fusion products; obtain two variables as labels, namely, absolute dynamic altitude and sea surface height anomaly of foreign high-precision satellite altimeter fusion products, and require the temporal resolution of the above variables to be consistent;

[0015] S12: spatially pair the input data obtained in step S11 with the label data, and use the bilinear interpolation method to unify the spatial resolution of all variables to ;

[0016] S13, convert the longitude and latitude sequence described in S11 into a two-dimensional grid field with the same resolution as the grid field in S12, ;

[0017] S14: Pair the inputs obtained in S12 and S13 with the label data in time and space to obtain a data pair matrix that matches time and space. , data pair matrix The organization is:

[0018]

[0019] Among them, the data pair matrix Each input variable is denoted as , expressed as The first sample variables, Indicates the arrive samples, one sample is a matrix In a row, due to the same time resolution, one moment corresponds to one sample, each row contains 9 variables, of which the input variables are the first 7, which is a matrix The first 7 columns of data are: sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, longitude field, latitude field, and absolute dynamic altitude and sea surface height anomaly data of domestic satellite altimeter fusion products. The label variables are the last two, which are matrices. The last two columns of data are the absolute dynamic altitude and sea level height anomaly data of the fusion products of foreign high-precision satellite altimeters. The label of each sample corresponds to the input variable one by one, and is consistent in time and space.

[0020] S15. Pair the paired data into a matrix according to , divided into training set , validation set , test set ,Right now The first 80% of the rows are used as the training set The remaining 20% of the rows are split in half as a validation set and test set ;

[0021] S16, the training set obtained from step S15 or validation set or test set Randomly pick samples, get a certain input batch , normalize the sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, longitude field, latitude field, absolute dynamic height and sea surface height anomaly data to interval, and obtain a batch of data after normalization , the normalization formula is expressed as follows:

[0022]

[0023] in, Represents a variable in the batch The sample set, For the training set The minimum value of a variable in For the training set The maximum value of a variable in the batch, a variable after normalization of the batch is recorded as ;

[0024] S17, a batch of data after normalization Converted to a 4D gridded tensor format , abbreviated as , Indicates the horizontal latitudinal span of the area where the deviation correction is performed. Indicates the longitudinal span of the area for deviation correction in the horizontal direction, and then takes the first 7 variables as the input batch , the last two variables are used as label batches .

[0025] Preferably, step S2 specifically includes the following steps:

[0026] S21, build SF (Shallow Feature) shallow feature extraction module, used to preliminarily extract input variable features and preliminarily allocate attention weights to different channels, including 1 layer containing A convolution layer with a 1×1 convolution kernel, a relu activation layer, and a channel attention weight calculation layer, namely the CA channel attention module, whose output feature map is , the expression is:

[0027]

[0028] in, Represents the output feature map of the SF shallow feature extraction module, with a size of This is because the 1st layer contains Convolutional layer with 1×1 convolution kernel , which will map the number of channels to higher dimensions to improve times; Represents the mapping function of the SF shallow feature extraction module, Describe the relu activation function, represents the CA channel attention module;

[0029] S22. Construct the LFGA (Land Filtering Global Attention) module for filtering land global ocean attention. The expression of the LFGA module for filtering land global ocean attention is:

[0030]

[0031] in, Output feature map of the global ocean attention module representing LFGA filtering of land, The mapping function of the global ocean attention module representing LFGA filtering of land;

[0032] S23. Construct the RF (Reconstructing Feature) reconstruction feature module. The RF module reconstructs the features extracted by the LFGA attention mechanism so that the number of layers of the output feature map is consistent with the number of layers of the label feature. It outputs high-precision satellite altimeter data after bias correction, and passes through 5×5, 3×3 and 1×1 convolution kernels in sequence, with relu activation in between. The final output channel number is 2, corresponding to the two bias-corrected variables, namely the absolute dynamic altitude and the sea level height anomaly data. The RF output module expression is as follows:

[0033]

[0034] in, Represents the output feature map of the RF reconstruction feature module, with a size of , Represents the mapping function of the RF reconstruction feature module, and Represents convolutional layers with 3×3 and 5×5 convolution kernel sizes;

[0035] S24: Connect the SF shallow feature extraction module obtained in step S21, the LFGA land-filtered global ocean attention module obtained in step S22, and the RF reconstruction feature module obtained in step S23 in sequence to obtain a deep learning network A2A, which is expressed as follows:

[0036]

[0037] Preferably, step S21 specifically includes the following steps:

[0038] S211, SF shallow feature extraction module contains 1 layer The convolutional layer with a 1×1 convolution kernel, a relu activation layer, and a CA channel attention module are respectively passed through a global maximum pooling layer and a global average pooling layer on a two-dimensional plane composed of length and width dimensions, and then input into a multi-layer perceptron MLP containing three fully connected layers. The output results are then added to the corresponding grids and then passed through a sigmoid activation layer to output the channel attention feature map. The CA channel attention module can be expressed as:

[0039]

[0040] in, Represents the output feature map of the CA channel attention module, represents the multi-layer perceptron mapping function, Represents the sigmoid activation layer mapping function, Represents the feature map of the input CA channel attention module. The above channel attention feature map The size is With size Multiplying the corresponding dimension of the shallow feature map is equivalent to giving all grid points on each channel of the input feature map the same weight value;

[0041] Represents the mapping function of the global average pooling layer in the length and width dimensions, using the function in the Pytorch function library To achieve this, the expression is as follows:

[0042]

[0043] in, First, along the length dimension, corresponding to The "2" in the function is averaged, and then along the width dimension, which corresponds to The "3" in the function is averaged to obtain the feature map , the input feature map size is , the size first becomes , and then becomes , Indicates that the sizes of the remaining dimensions remain unchanged; similarly, Represents the mapping function of the global maximum pooling layer in length and width dimensions, and the process is the same as Class, change the average to the maximum value, by the function in the Pytorch function library To achieve this, the expression is:

[0044]

[0045] The size of the input and output feature maps of the MLP multilayer perceptron remains unchanged. It contains three layers: a fully connected layer, a relu activation layer, and a fully connected layer. In essence, the fully connected layer is a matrix multiplication operation between the input feature map and the neuron weight matrix. The expression is:

[0046]

[0047] in, is the mapping function of the MLP multi-layer perceptron, and is the weight matrix of the two fully connected layers, and is the bias matrix of the two fully connected layers; Represents the sigmoid activation function layer, and the expression is as follows:

[0048] ;

[0049] Among them, since the sigmoid function activates the input tensor grid by grid, it activates each grid independently. Represents the value of an input grid;

[0050] S212, connect one 1×1 convolution layer, one ReLU activation layer, and one CA channel attention module in sequence to form the SF shallow feature extraction module, which is expressed as:

[0051]

[0052] Among them, the output shallow extraction feature map The size is ;

[0053] Preferably, step S22 specifically includes the following steps:

[0054] S221, LFGA global ocean attention module for land filtering, which includes five steps: feature map expansion, land filtering, Transformer multi-head self-attention mechanism to capture global attention, land information restoration, and feature map folding;

[0055] First, perform the Unfold operation to flatten the input feature map to the size required by the Transformer encoder module. Sliding window, according to the step size , respectively scan the input size is Feature map , scanned independently The size of the sample The feature map of the sliding window is On the feature map, the size of the window coverage area after each sliding is The feature map of is flattened, that is, the size of the same channel is The feature map of is flattened into a one-dimensional vector, and then indivual The feature maps are spliced together in sequence to obtain a length of The long vector, due to the total number of slides , you will get The length is vector, since The samples are scanned independently at the same time, and finally the feature map after Unfolding is obtained , size is , , ;

[0056] S222. Use the LFM (Land Filtering Mechanism) land filtering mechanism to filter land information worldwide. This is because, unlike the atmospheric space, the ocean is divided into discontinuous blocks by continents, islands, etc., and the global land surface area accounts for about 30%. If the grid on the land is directly assigned a fixed value and then the global attention calculation is performed directly, the attention module will be distracted by the land, which is not conducive to improving the accuracy of the deviation correction of the satellite altimeter data in the ocean. Therefore, it is necessary to filter the land information. The specific process is as follows: first, using the absolute dynamic altitude data in a label sample as a template, the grid value in the ocean is assigned to 1 and the grid value in the land is assigned to 0. Then, Unfold the sliding window with the same size and step size as in S221, and you will get a size of The sea and land coding information matrix Mask, since only the absolute dynamic altitude data in one label sample is Unfolded, , , the length of each line in Mask is That is, the size of the area covered by the sliding window in the originally expanded two-dimensional area. Since the land grid value is 0 and the ocean grid value is 1, if the grids covered by the sliding window are all on land, then the grids in this row in the Unfolded Mask are all 0. If they are in the ocean, they are all 1. If they are at the junction of land and sea, they contain both 0 and 1. Therefore, the land and sea coding information matrix Mask that can distinguish between land and sea is obtained. It is only necessary to mark the rows with all 0s in the Mask to obtain the vector Index that records the index of the row where the land is located. The length is The value of the land row is 0, and the rest are 1. The feature map after Unfold is retrieved according to Index. The row corresponding to the index position of , and the remaining rows are retained, that is, the feature map containing only the information of the ocean and the sea-land boundary area is obtained. , size is , The number of rows representing the ocean and the land-sea boundary, without having to process information about land areas, ensures that attention calculations are focused on the ocean;

[0057] S223, adopt (Multi-Head Attention) The multi-head self-attention mechanism captures global attention and inputs the flattened feature map , after a normalization layer Norm, for each Normalize each row to improve training stability, and then expand the third dimension through three identical fully connected layers Linear times, size is , and use pytorch's reshape function to divide it into The sizes are of , , Matrix, then, perform SDPA scaling point-by-point multiplication operation, the specific steps are, First with Perform matrix multiplication and then scale to times, then along Softmax activation is performed on the dimension of Perform matrix multiplication and finally Segment merge restored to size , then through a Linear layer, restored to size , and then pass through the normalization layer Norm to get the feature map , the implementation process of the MHA multi-head attention module is expressed as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] in, represents the normalization layer, is the activation function of the fully connected layer, , , represents the neuron weight matrix of the three fully connected layers, , , Three fully connected layer bias values, Indicates that , , Divided into part, Indicates restoration to the original size. The specific expression of the Softmax activation function is:

[0064] ;

[0065] in, represents the input vector of the activation function, Represents the value of each grid on the vector; Finally, in order to reduce information loss and reduce the difficulty of training, and Perform residual connection, that is, add the grids at the same position in two feature maps of the same size;

[0066] S224, restore the land information, that is, refer to the vector Index that records the index of the row where the land is located, The length of the original land row to be deleted is filled in 0 vector, and obtain the feature map ;

[0067] S225, according to the reverse operation of Unfold operation, Fold operation, Feature map restored to size , size is ,This feature map records the attention relationship weights between different grids, and the attention weight of the land area is 0, thus shielding the interference of land information on attention calculation;

[0068] S226. Combine the above five steps: feature map expansion, land filtering, Transformer's multi-head self-attention mechanism to capture global attention, land information restoration, feature map folding and sequentially connect to construct the LFGA global ocean attention module for land filtering. The expression of the LFGA global ocean attention module for land filtering is:

[0069]

[0070] Preferably, step S23 specifically includes the following steps:

[0071] S231, input the multi-channel feature map After 5×5 convolution kernels, the dimension is reduced to channels;

[0072] S232, the result obtained in step S231 The feature maps of each channel are sequentially passed through a layer of ReLU activation layer and 3×3 convolution kernels, dimension reduction is channels;

[0073] S233, the result obtained in step S232 The feature map of each channel passes through one ReLU activation layer and two 1×1 convolution kernels in turn, reducing the dimension of the feature map to a two-channel output with a size of , which correspond to the two variables of absolute dynamic altitude and sea surface height anomaly data after deviation correction. The expression of RF reconstruction feature module is as follows:

[0074]

[0075] Preferably, step S24 specifically includes the following steps:

[0076] S241, the SF shallow feature extraction module obtained in step S21, the LFGA land-filtered global ocean attention module obtained in step S22, and the RF reconstruction feature module obtained in step S23 are sequentially connected to obtain a deep learning network A2A, which is expressed as follows:

[0077]

[0078] in, The output is the bias-corrected satellite altimeter data without denormalization, with a size of ;

[0079] Preferably, step 3 specifically includes the following steps:

[0080] S31. Select Charbonnier loss as the loss function for the deep learning network A2A obtained in training step S2. The expression of Charbonnier loss is as follows:

[0081]

[0082] in, Represents each grid in the output feature map and the label feature map, Is a very small constant, in order to make the value stable;

[0083] S32, based on step S31, the training set obtained in step S13 and validation set Input the deep learning network A2A obtained in step S2 in batches for training and fitting Parameters of each layer in ;

[0084] Preferably, step S32 specifically includes the following steps:

[0085] S321, set the open source deep learning framework, train the basic hyper parameters of the deep learning network A2A, and then based on step S31, the training set The data is input into the deep learning network A2A obtained in step S2 in batches for training. Therefore, the training set contain samples, so they can be divided into batches, and the output of each batch will calculate the loss in sequence , and uses the back propagation optimizer in the open source deep learning framework to back propagate the loss gradient and optimize the parameters , when all batches are trained for one round, it is called an epoch, and the total training epochs, at the end of each epoch, the validation set will be entered The loss of a validation set is calculated based on the data without backpropagation. The purpose is to observe the difference between the loss of the validation set and the loss of the training set, and to judge the model parameter fitting effect and generalization performance.

[0086] S322, after training, save the parameters , which covers all trainable parameters in the deep learning network A2A, and obtains the trained deep learning network A2A.

[0087] Preferably, step S4 specifically includes the following steps:

[0088] S41, the test set obtained in step S15 The data was normalized to the MinMax method. interval;

[0089] S42: The normalized test set data obtained in step S41 is input into the trained deep learning network A2A obtained in step S3 in batches. Each batch of output data is denormalized to obtain domestic satellite altimeter data with higher accuracy after bias correction. The denormalization formula is:

[0090]

[0091] in, represents the undenormalized test set output, represents the test set output after denormalization, For the training set The minimum value of the foreign satellite altimeter fusion product as a label, For the training set The maximum value of the foreign satellite altimeter fusion product used as a label.

[0092] Compared with the prior art, the present invention has the following beneficial effects:

[0093] (1) This invention provides a new means for correcting the deviation of domestic satellite altimeter fusion products. It is based on open source data and a deep learning framework. The operation is convenient and fast. The trained model has low computational load requirements and can be deployed in a lightweight manner. The inference speed is faster than traditional methods.

[0094] (2) The LFM land filtering mechanism proposed in this paper can effectively shield the interference of land information on ocean attention calculation. On the one hand, the land filtering mechanism filters the land information accounting for 30% of the global surface area, effectively reducing the computational complexity of the attention mechanism; on the other hand, it focuses the attention calculation on the land area, reduces the interference of irrelevant land information, and improves the accuracy of attention calculation.

[0095] (3) The A2A network framework of the present invention takes into account both CNN convolution and Transformer architectures, effectively integrating the two, so that it can capture both global and local features, so that the deviation correction has a wider range of surrounding information reference, thereby improving the accuracy of the correction;

[0096] (4) The present invention achieves the deviation correction of data from domestic satellite altimeter fusion products in global ocean areas, and has global applicability and global generalization; BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 It is the overall flow chart of the present invention.

[0098] Figure 2This is the structural diagram of the SF shallow feature extraction module.

[0099] Figure 3 This is the flowchart of the Unfold operation, with 2 Take 2 windows and a step size of 1 as an example.

[0100] Figure 4 It is a flow chart of the LFM land filtering mechanism to obtain the sea and land coding information matrix.

[0101] Figure 5 This is the structure diagram of the global ocean attention module of LFGA filtering land. DETAILED DESCRIPTION

[0102] In order to enable those skilled in the art to better understand and implement the technical solution of the present invention, the present invention is further described below with reference to specific embodiments, but the embodiments are not intended to limit the present invention.

[0103] like Figure 1 As shown, the embodiment of the present invention provides a method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning, which specifically includes the following steps:

[0104] S1. Obtain satellite remote sensing data from open source channels as input and label data, integrate the obtained input and label data sets into a certain organizational structure and divide them into subsets for model training, verification, and testing; specifically, the following steps are included:

[0105] S11. Get 7 variables as input, namely satellite sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, global longitude grid field, global latitude grid field, and absolute dynamic altitude and sea surface height anomaly of domestic satellite altimeter fusion products; get 2 variables as labels, namely absolute dynamic altitude and sea surface height anomaly data of foreign high-precision satellite altimeter fusion products; the time resolution of the above variables is consistent with that of downloading from 2021 to 2023 The average daily data for the year is Days, that is, each variable has The samples have the same coverage, covering the global ocean area. The specific description and source of the above data are as follows: The sea surface temperature data comes from the L4-level OISST satellite sea surface temperature optimal interpolation product provided by Remote Sensing Systems (reference website: https: / / www.remss.com / measurements / sea-surface-temperature / oisst-description / ), with a spatial resolution of approximately , the time resolution is the average of the day; the sea surface wind speed (longitudinal and latitudinal) data also comes from the L4 CCMP satellite wind vector analysis product provided by Remote Sensing Systems (reference website: https: / / www.remss.com / measurements / sea- surface-temperature / oisst-description / ), the sea surface 10m longitudinal and latitudinal wind speed data were selected, with a spatial resolution of approximately The time resolution is 6 hours. Therefore, in order to unify the time resolution, the data of four moments per day are averaged to obtain the average longitudinal and latitudinal wind speed data. The absolute dynamic altitude and sea surface height anomaly data of the domestic satellite altimeter fusion product are from the sea surface height fusion product ALT MUL developed by the National Satellite Ocean Application Center of my country (reference website: https: / / osdds.nsoas.org.cn / OceanDynamics ), the global data was selected, with a spatial resolution of approximately , the time resolution is the average of the day; the absolute dynamic altitude and sea surface height anomaly data of foreign satellite altimeter fusion products are from the fusion products provided by France AVISO, which integrates altimeter data from multiple satellites (reference website: https: / / marine.copernicus.eu / ), with a spatial resolution of approximately , the time resolution is the sky average;

[0106] S12: spatially pair the input data obtained in step S11 with the label data, and use the bilinear interpolation method to unify the spatial resolution of all variables to ;

[0107] S13, convert the longitude and latitude sequence described in S11 into a two-dimensional grid field with the same resolution as the grid field in S12, ;

[0108] S14: Pair the inputs obtained in S12 and S13 with the label data in time and space to obtain a data pair matrix that matches time and space. , data pair matrix The organization is:

[0109]

[0110] Among them, the data pair matrix Each input variable is denoted as , expressed as The first sample variables, Indicates the arrive samples, one sample is a matrix In a row, due to the same time resolution, one moment corresponds to one sample, each row contains 9 variables, of which the input variables are the first 7, which is a matrix The first 7 columns of data are: sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, longitude field, latitude field, and absolute dynamic altitude and sea surface height anomaly data of domestic satellite altimeter fusion products. The label variables are the last two, which are matrices. The last two columns of data are the absolute dynamic altitude and sea level height anomaly data of the fusion products of foreign high-precision satellite altimeters. The label of each sample corresponds to the input variable one by one, and is consistent in time and space.

[0111] S15. Pair the paired data into a matrix according to , divided into training set , validation set , test set ,Right now The first 80% of the rows are used as the training set The remaining 20% of the rows are split in half as a validation set and test set ;

[0112] S16, the training set obtained from step S15 or validation set or test set Randomly pick ( ) samples, and get a certain input batch , normalize the sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, longitude field, latitude field, absolute dynamic height and sea surface height anomaly data to interval, and obtain a batch of data after normalization , the normalization formula is expressed as follows:

[0113]

[0114] in, Represents a variable in the batch The sample set, For the training set The minimum value of a variable in For the training set The maximum value of a variable in the batch, a variable after normalization of the batch is recorded as ;

[0115] S17, a batch of data after normalization Converted to a 4D gridded tensor format , abbreviated as , Indicates the horizontal latitudinal span of the area where the deviation correction is performed. Indicates the longitudinal span of the area for deviation correction in the horizontal direction. Since the spatial resolution is , convert the spans in the longitude and latitude directions into the number of grids, so the size is , , 1440 , and then take the first 7 variables as the input batch , size is , the last two variables are used as label batches , size is .

[0116] S2. Build a deep learning network A2A. The expression of A2A is as follows:

[0117]

[0118] in, The mapping relationship between the input and output of the deep learning network A2A is shown. The input is a normalized multivariate input batch. The output is the absolute dynamic altitude after deviation correction and the sea surface height anomaly data batch, Indicates the first The weights and biases of the layer are as follows:

[0119] S21, build SF (Shallow Feature) shallow feature extraction module, such as Figure 2 As shown, it is used to preliminarily extract the input variable features and preliminarily allocate the attention weights to different channels, including 1 layer containing A convolution layer with a 1×1 convolution kernel, a relu activation layer, and a channel attention weight calculation layer, namely the CA channel attention module, whose output feature map is , the expression is:

[0120]

[0121] in, Represents the output feature map of the SF shallow feature extraction module, with a size of This is because the 1st layer contains A convolution layer with a 1×1 convolution kernel will map the number of channels to a higher dimension, improving times; Represents the mapping function of the SF shallow feature extraction module, Describe the relu activation function, represents the CA channel attention module;

[0122] The specific steps of step S21 are as follows:

[0123] S211, SF shallow feature extraction module contains 1 layer The convolutional layer with a 1×1 convolution kernel, a relu activation layer, and a CA channel attention module are respectively passed through a global maximum pooling layer and a global average pooling layer on a two-dimensional plane composed of length and width dimensions, and then input into a multi-layer perceptron MLP containing three fully connected layers. The output results are then added to the corresponding grids and then passed through a sigmoid activation layer to output the channel attention feature map. The CA channel attention module can be expressed as:

[0124]

[0125] in, Represents the output feature map of the CA channel attention module, represents the multi-layer perceptron mapping function, Represents the sigmoid activation layer mapping function, Represents the feature map of the input CA channel attention module. The above channel attention feature map The size is With size Multiplying the corresponding dimension of the shallow feature map is equivalent to giving all grid points on each channel of the input feature map the same weight value;

[0126] Represents the mapping function of the global average pooling layer in the length and width dimensions, using the function in the Pytroch function library To achieve this, the expression is as follows:

[0127]

[0128] in, First, along the length dimension, corresponding to The "2" in the function is averaged, and then along the width dimension, which corresponds to The "3" in the function is averaged to obtain the feature map , the input feature map size is , the size first becomes , and then becomes , Indicates that the sizes of the remaining dimensions remain unchanged; similarly, Represents the mapping function of the global maximum pooling layer in length and width dimensions, and the process is the same as Class, change the average to the maximum value, by the function in the Pytroch function library To achieve this, the expression is:

[0129]

[0130] The size of the input and output feature maps of the MLP multilayer perceptron remains unchanged. It contains three layers: a fully connected layer, a relu activation layer, and a fully connected layer. In essence, the fully connected layer is a matrix multiplication operation between the input feature map and the neuron weight matrix. The expression is:

[0131]

[0132] in, is the mapping function of the MLP multi-layer perceptron, and is the weight matrix of the two fully connected layers, and is the bias matrix of the two fully connected layers; Represents the sigmoid activation function layer, and the expression is as follows:

[0133] ;

[0134] Among them, since the sigmoid function activates the input tensor grid by grid, it activates each grid independently. Represents the value of an input grid;

[0135] S212, connect one 1×1 convolution layer, one ReLU activation layer, and one CA channel attention module in sequence to form the SF shallow feature extraction module, which is expressed as:

[0136]

[0137] Among them, the output shallow extraction feature map The size is ;

[0138] S22. Construct the LFGA (Land Filtering Global Attention) module for filtering land global ocean attention. The expression of the LFGA module for filtering land global ocean attention is:

[0139]

[0140] in, Output feature map of the global ocean attention module representing LFGA filtering of land, The mapping function of the global ocean attention module representing LFGA filtering of land;

[0141] The specific steps of step S22 are as follows:

[0142] S221, LFGA filters the global ocean attention module of land, such as Figure 5As shown, it includes five steps: feature map expansion, land filtering, Transformer's multi-head self-attention mechanism to capture global attention, land information restoration, and feature map folding;

[0143] First, perform the Unfold operation, such as Figure 3 As shown (the figure is 2 2 windows with a step size of 1 as an example), flatten the input feature map to a size that meets the input requirements of the Transformer encoder module, using a size of Sliding window, according to the step size , respectively scan the input size is Feature map , scanned independently The size of the sample The feature map of the sliding window is On the feature map, the size of the window coverage area after each sliding is The feature map of is flattened, that is, the size of the same channel is The feature map is flattened into a one-dimensional vector , then indivual The feature maps are spliced together in sequence to obtain a length of The long vector, due to the total number of slides , you will get The length is vector, since The samples are scanned independently at the same time, and finally the feature map after Unfolding is obtained , size is , , ;

[0144] S222, using LFM (Land Filtering Mechanism) land filtering mechanism to filter global land information, such as Figure 4 As shown in the figure, this is because, unlike the atmospheric space, the ocean is divided into discontinuous blocks by continents, islands, etc., and the global land surface area accounts for about 30%. If the grid on land is directly assigned a fixed value and the global attention calculation is performed directly, the attention module will be distracted by the land, which is not conducive to improving the accuracy of the deviation correction of the satellite altimeter data in the ocean. Therefore, it is necessary to filter the land information. The specific process is as follows: first, the absolute dynamic altitude data in a label sample is used as a template, the grid value in the ocean is assigned to 1, and the grid value in the land is assigned to 0, and then Unfold according to the sliding window of the same size and step size in S221, and a size of The sea and land coding information matrix Mask, since only the absolute dynamic altitude data in one label sample is Unfolded, , , each row in the Mask is 16 in length, which corresponds to the size of the area covered by the sliding window in the originally expanded two-dimensional area. Since the land grid value is 0 and the ocean grid value is 1, if the grids covered by the sliding window are all on land, then the grids of this row in the Unfolded Mask are all 0, if they are in the ocean, they are all 1, and if they are at the junction of land and sea, they contain both 0 and 1. Therefore, the land and sea coding information matrix Mask that can distinguish between land and sea is obtained. It is only necessary to mark the rows with all 0s in the Mask to obtain the vector Index that records the index of the row where the land is located, with a length of The value of the land row is 0, and the rest are 1. The feature map after Unfold is retrieved according to Index. The row corresponding to the index position of , and the remaining rows are retained, that is, the feature map containing only the information of the ocean and the sea-land boundary area is obtained. , size is , The number of rows representing the ocean and the land-sea boundary, without having to process information about land areas, ensures that attention calculations are focused on the ocean;

[0145] S223, adopt (Multi-Head Attention) The multi-head self-attention mechanism captures global attention and inputs the flattened feature map , after a normalization layer Norm, for each Normalize each row to improve training stability, and then expand the third dimension through three identical fully connected layers Linear times, size is , and use pytorch's reshape function to divide it into The sizes are of , , Matrix, then, perform SDPA scaling point-by-point multiplication operation, the specific steps are, First with Perform matrix multiplication and then scale to times, then along Softmax activation is performed on the dimension of Perform matrix multiplication and finally merge the 8 segments to restore the size , then through a Linear layer, restored to size , and then pass through the normalization layer Norm to get the feature map , the implementation process of the MHA multi-head attention module is expressed as follows:

[0146]

[0147]

[0148]

[0149]

[0150]

[0151] in, represents the normalization layer, is the activation function of the fully connected layer, , , represents the neuron weight matrix of the three fully connected layers, , , Three fully connected layer bias values, Indicates that , , Divided into part, Indicates restoration to the original size. The specific expression of the Softmax activation function is:

[0152] ;

[0153] in, represents the input vector of the activation function, Represents the value of each grid on the vector; Finally, in order to reduce information loss and reduce the difficulty of training, and Perform residual connection, that is, add the grids at the same position in two feature maps of the same size;

[0154] S224, restore the land information, that is, refer to the vector Index that records the index of the row where the land is located, The length of the original land row to be deleted is filled in 0 vector, and obtain the feature map ;

[0155] S225, according to the reverse operation of Unfold operation, Fold operation, Feature map restored to size , size is ,This feature map records the attention relationship weights between different grids, and the attention weight of the land area is 0, thus shielding the interference of land information on attention calculation;

[0156] S226. Combine the above five steps: feature map expansion, land filtering, Transformer's multi-head self-attention mechanism to capture global attention, land information restoration, feature map folding and sequentially connect to construct the LFGA global ocean attention module for land filtering. The expression of the LFGA global ocean attention module for land filtering is:

[0157]

[0158] S23. Construct the RF (Reconstructing Feature) reconstruction feature module. The RF module reconstructs the features extracted by the LFGA attention mechanism so that the number of layers of the output feature map is consistent with the number of layers of the label feature. It outputs high-precision satellite altimeter data after bias correction, and passes through 5×5, 3×3 and 1×1 convolution kernels in sequence, with relu activation in between. The final output channel number is 2, corresponding to the two bias-corrected variables, namely the absolute dynamic altitude and the sea level height anomaly data. The RF output module expression is as follows:

[0159]

[0160] in, Represents the output feature map of the RF reconstruction feature module, with a size of , Represents the mapping function of the RF reconstruction feature module, and Represents convolutional layers with 3×3 and 5×5 convolution kernel sizes;

[0161] The specific steps of step S23 are as follows:

[0162] S231, input the multi-channel feature map After 5×5 convolution kernels, the dimension is reduced to channels;

[0163] S232, the result obtained in step S231 The feature maps of each channel are sequentially passed through a layer of ReLU activation layer and 3×3 convolution kernels, dimension reduction is channels;

[0164] S233, the result obtained in step S232 The feature map of each channel passes through one ReLU activation layer and two 1×1 convolution kernels in turn, reducing the dimension of the feature map to a two-channel output with a size of , which correspond to the two variables of absolute dynamic altitude and sea surface height anomaly data after deviation correction. The expression of RF reconstruction feature module is as follows:

[0165]

[0166] S24, connect the SF shallow feature extraction module obtained in step S21, the LFGA land filtering global ocean attention module obtained in step S22, and the RF reconstruction feature module obtained in step S23 in sequence to obtain a deep learning network A2A, which is expressed as follows:

[0167]

[0168] are the parameters of each layer, and the expression is .

[0169] The specific steps of step S24 are as follows:

[0170] S241, sequentially connect the SF shallow feature extraction module obtained in step S21, the LFGA land-filtered global ocean attention modules obtained in step S22, and the RF reconstruction feature module obtained in step S23 to obtain a deep learning network A2A, which is expressed as follows:

[0171]

[0172] in, The output is the bias-corrected satellite altimeter data without denormalization, with a size of ;

[0173] S3. Alternately train and validate the deep learning network A2A obtained in step S2 using the data set obtained in step S1;

[0174] The specific steps of step S3 are as follows:

[0175] S31. Select Charbonnier loss as the loss function for the deep learning network A2A obtained in training step S2. The expression of Charbonnier loss is as follows:

[0176]

[0177] in, Represents each grid in the output feature map and the label feature map, Is a very small constant, in order to make the value stable;

[0178] S32, based on step S31, the training set obtained in step S13 and validation set Input the deep learning network A2A obtained in step S2 in batches for training and fitting Parameters of each layer in ;

[0179] The specific steps of step S32 are as follows:

[0180] S321. Set the open source deep learning Pytorch framework and set the number of initialization iterations to , the initial learning rate is , initialize the batch size to , use Kaiming initializer to randomly initialize the weights of each layer of the network, and then based on step S31, the training set The data is input into the deep learning network A2A obtained in step S2 in batches for training. Therefore, the training set Contains about 2922 samples, so it can be divided into batches, and the loss is calculated for each batch of output results The optimizer uses Adam, which has been encapsulated and inherited as the function torch.optim.Adam in the Pytorch framework. It can automatically solve the gradient and backpropagate according to the calculated loss, thereby optimizing the parameters of each layer of the deep learning network A2A , all batches are trained once and it is called an epoch, and the total training epochs, at the end of each epoch, the validation set will be entered The loss of a validation set is calculated based on the data without backpropagation. The purpose is to observe the difference between the loss of the validation set and the loss of the training set, and to judge the model parameter fitting effect and generalization performance.

[0181] S322, after training, save the parameters , which covers all trainable parameters in the deep learning network A2A, and obtains the trained deep learning network A2A.

[0182] S4. Use the trained deep learning network A2A to correct the domestic satellite altimeter fusion product, that is, to obtain higher-precision domestic satellite altimeter data.

[0183] The specific steps of step S4 are as follows:

[0184] S41, the test set obtained in step S15 The data was normalized to the MinMax method. interval;

[0185] S42: The normalized test set data obtained in step S41 is input into the trained deep learning network A2A obtained in step S3 in batches. Each batch of output data is denormalized to obtain domestic satellite altimeter data with higher accuracy after bias correction. The denormalization formula is:

[0186]

[0187] in, represents the undenormalized test set output, represents the test set output after denormalization, For the training set The minimum value of the foreign satellite altimeter fusion product as a label, For the training set The maximum value of the foreign satellite altimeter fusion product used as a label.

[0188] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for correcting the deviation of domestic satellite altimeter fusion products based on deep learning, characterized by: The following steps are involved: S1. Obtain satellite remote sensing data as input and label data, integrate the obtained input and label data sets into an organizational structure and divide them into subsets for model training, validation, and testing; S2. Build a deep learning network A2A. The specific process is as follows: S21, build SF shallow feature extraction module, output feature map is ; S22. Constructing a global ocean attention module for filtering land by LFGA, including the following steps: S221, the input feature map is flattened to the size required by the Transformer encoder module input, using a size of Sliding window, according to the step size , respectively scan the input size is Feature map , Format of 4D gridded tensors , scanned independently The size of the sample The feature map of the sliding window is On the feature map, the size of the window coverage area after each sliding is Flatten the feature map and then indivual The feature maps are spliced together in sequence to obtain a length of The long vector, due to the total number of slides ,get The length is vector, since The samples are scanned independently at the same time, and finally the feature map after Unfolding is obtained , size is , , ; S222, using the LFM land filtering mechanism to filter global land information, first, using the absolute dynamic altitude data in a label sample as a template, assigning the grid value in the ocean to 1 and the grid value in the land to 0, then expanding the sliding window with the same size and step size as in S221 to obtain a size of The sea-land coding information matrix Mask, so , , the length of each line in Mask is That is, the size of the area covered by the sliding window in the originally expanded two-dimensional area. Since the land grid value is 0 and the ocean grid value is 1, if the grids covered by the sliding window are all on land, then the grids in this row in the Unfolded Mask are all 0. If they are in the ocean, they are all 1. If they are at the junction of land and sea, they contain both 0 and 1. The land and sea coding information matrix Mask that distinguishes the ocean from the land is obtained. It is only necessary to mark the rows with all 0s in the Mask to obtain the vector Index that records the index of the row where the land is located. The length is The value of the land row is 0, and the rest are 1. The feature map after Unfold is retrieved according to Index. The row corresponding to the index position of , and the remaining rows are retained, that is, the feature map containing only the information of the ocean and the sea-land boundary area is obtained. , size is , The number of rows representing the ocean and the boundary between land and sea; S223, adopt The multi-head self-attention mechanism captures global attention and inputs the flattened feature map , after a normalization layer Norm, for each Normalize each row by the dimension, and then expand the third dimension by three identical fully connected layers Linear times, and use pytorch's reshape function to divide it into Segment, get , , The matrix,then, performs SDPA scaling point-by-point multiplication; S224, restore the land information, that is, refer to the vector Index that records the index of the row where the land is located, The length of the original land row to be deleted is filled in 0 vector, and obtain the feature map ; S225, according to the reverse operation Fold folding operation of Unfold operation, the feature map Restore to feature map ; S226. The expression of the global ocean attention module for LFGA land filtering is constructed by expanding the feature map, filtering land, using the Transformer's multi-head self-attention mechanism to capture global attention, restoring land information, and folding the feature map. ； in, Output feature map of the global ocean attention module representing LFGA filtering of land, The mapping function of the global ocean attention module representing LFGA filtering of land; S23. Construct the RF reconstruction feature module. The RF output module expression is as follows: ； in, Represents the output feature map of the RF reconstruction feature module, Represents the mapping function of the RF reconstruction feature module, and Represents convolutional layers with 3×3 and 5×5 convolution kernel sizes; S24, connect the SF shallow feature extraction module obtained in step S21, the LFGA land filtering global ocean attention module obtained in step S22, and the RF reconstruction feature module obtained in step S23 in sequence to obtain a deep learning network A2A, which is expressed as follows: ； in, The mapping relationship between the input and output of the deep learning network A2A is shown. The input is a normalized multivariate input batch. The output is the absolute dynamic altitude after deviation correction and the sea surface height anomaly data batch, Indicates the first Layer weights and biases; S3, using the data set obtained in step S1 to alternately train and verify the deep learning network A2A obtained in step S2, the parameters of each layer of the deep learning network A2A ; S4. Use the trained deep learning network A2A to correct the domestic satellite altimeter fusion product; S41, the test set obtained in step S15 The data was normalized to the MinMax method. interval; S42: The normalized test set data obtained in step S41 is input into the trained deep learning network A2A obtained in step S3 in batches. Each batch of output data is denormalized to obtain bias-corrected domestic satellite altimeter data. The denormalization formula is: ； in, represents the undenormalized test set output, represents the test set output after denormalization, For the training set The minimum value of the foreign satellite altimeter fusion product as a label, For the training set The maximum value of the foreign satellite altimeter fusion product used as a label.

2. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 1 is characterized in that: Step S21: constructing a SF shallow feature extraction module, including the following sub-steps; S211, CA channel attention module is expressed as: in, Represents the output feature map of the CA channel attention module, represents the multi-layer perceptron mapping function, Represents the sigmoid activation layer mapping function, Represents the feature map of the input CA channel attention module; Mapping function of the global average pooling layer in length and width dimensions The expression is: ； in, First, along the length dimension, corresponding to The "2" in the function is averaged, and then along the width dimension, which corresponds to The "3" in the function is averaged to obtain the feature map , Represents the mapping function of the global maximum pooling layer in length and width dimensions, expressed as: ； The mapping function expression of the MLP multi-layer perceptron is: ； in, is the mapping function of the MLP multi-layer perceptron, and is the weight matrix of the two fully connected layers, and is the bias matrix of the two fully connected layers; Represents the sigmoid activation function layer, and the expression is as follows: ; Among them, since the sigmoid function activates the input tensor grid by grid, Represents the value of an input grid; The expressions of S212 and SF shallow feature extraction modules are: ； in, Represents the output feature map of the SF shallow feature extraction module; Represents the mapping function of the SF shallow feature extraction module, Describe the relu activation function, Represents the CA channel attention module.

3. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 1 is characterized in that: Step S223 specifically includes the following steps: First with Perform matrix multiplication and then scale to times, then along Softmax activation is performed on the dimension of Perform matrix multiplication and finally Segment merge restored to size , then through a Linear layer, restored to size , and then pass through the normalization layer Norm to get the feature map , the implementation process of the MHA multi-head attention module is expressed as follows: ， ， ， ， ， in, represents the normalization layer, is the activation function of the fully connected layer, , , represents the neuron weight matrix of the three fully connected layers, , , Three fully connected layer bias values, Indicates that , , Divided into part, Indicates restoration to the original size. The specific expression of the Softmax activation function is: ; in, represents the input vector of the activation function, Represents the value of each grid on the vector; Finally, in order to reduce information loss and reduce the difficulty of training, and Perform residual connection.

4. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 1 is characterized in that: Step S1 specifically includes the following steps: S11. Obtain seven variables as input, namely, satellite sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, global longitude grid field, global latitude grid field, and absolute dynamic altitude and sea surface height anomaly of domestic satellite altimeter fusion products; obtain two variables as labels, namely, absolute dynamic altitude and sea surface height anomaly of foreign high-precision satellite altimeter fusion products, with consistent temporal resolution of the variables; S12: spatially pair the input data obtained in step S11 with the label data, and use the bilinear interpolation method to unify the spatial resolution of all variables to ; S13, convert the longitude and latitude sequence described in S11 into a two-dimensional grid field with the same resolution as the grid field in S12, ; S14: Pair the inputs obtained in S12 and S13 with the label data in time and space to obtain a data pair matrix that matches time and space. , data pair matrix The organization is: ， Among them, the data pair matrix Each input variable is denoted as , expressed as The first sample variables, Indicates the arrive samples; S15. Pair the paired data into a matrix according to , divided into training set , validation set , test set ; S16, the training set obtained from step S15 or validation set or test set Randomly pick samples, get a certain input batch , normalize the sea surface temperature, sea surface meridional wind speed, sea surface zonal wind speed, longitude field, latitude field, absolute dynamic height and sea surface height anomaly data to interval, and obtain a batch of data after normalization , the normalization formula is expressed as follows: ， in, Represents a variable in the batch The sample set, For the training set The minimum value of a variable in For the training set The maximum value of a variable in the batch, a variable after normalization of the batch is recorded as ; S17, a batch of data after normalization Converted to a 4D gridded tensor format , abbreviated as , Indicates the horizontal latitudinal span of the area where the deviation correction is performed. Indicates the longitudinal span of the area for deviation correction in the horizontal direction, and then takes the first 7 variables as the input batch , the last two variables are used as label batches .

5. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 1 is characterized in that: Step S23 specifically includes the following steps: S231, input the multi-channel feature map After 5×5 convolution kernels, the dimension is reduced to channels; S232, the result obtained in step S231 The feature maps of each channel are sequentially passed through a layer of ReLU activation layer and 3×3 convolution kernels, dimension reduction is channels; S233, the result obtained in step S232 The feature map of each channel is sequentially passed through a ReLU activation layer and two 1×1 convolution kernels to reduce the dimension of the feature map to the output of two channels, corresponding to the two variables of absolute dynamic altitude and sea level height abnormal data after deviation correction. The expression of the RF reconstruction feature module is as follows: 。 6. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 1 is characterized in that: Step 3 specifically includes the following steps: S31. Select Charbonnier loss as the loss function for the deep learning network A2A obtained in training step S2. The expression of Charbonnier loss is as follows: in, Represents each grid in the output feature map and the label feature map, is a constant; S32, based on step S31, the training set and validation set Input the deep learning network A2A obtained in step S2 in batches for training and fitting Parameters of each layer in .

7. The method for correcting the deviation of a domestic satellite altimeter fusion product based on deep learning according to claim 6 is characterized in that: Step S32 specifically includes the following steps: S321, set the open source deep learning framework, train the basic hyper parameters of the deep learning network A2A, and then based on step S31, the training set The data is input into the deep learning network A2A obtained in step S2 in batches for training, and the output results of each batch are calculated in turn. , using the backpropagation optimizer in the open source deep learning framework to backpropagate the gradient of the loss and optimize the parameters , when all batches are trained for one round, it is called an epoch, and the total training epochs, where at the end of each epoch, the validation set is input , calculate the loss of a validation set without backpropagation, observe the difference between the loss of the validation set and the training set, and judge the model parameter fitting effect and generalization performance; S322, after training, save the parameters , which covers all trainable parameters in the deep learning network A2A, and obtains the trained deep learning network A2A.

Images