A method for completing polar orbit satellite meteorological observation data

Abstract

This invention discloses a method for completing meteorological observation data from polar-orbiting satellites, belonging to the fields of space science and remote sensing technology. The invention employs a self-supervised training strategy, where the training mask and the missing regions have similar distribution patterns, specifically improving the network's completion results for missing polar-orbiting satellite observation data and reducing the network's requirement for label data. The network can be trained without introducing any other source data or performing spatiotemporal matching. Based on the characteristic of brightness temperature data being affected by land surface radiation, prior knowledge of continental contours is introduced to improve the completion effect of large areas of continuous missing data and missing data at the land-sea boundary. Finally, a gated convolutional network is used to complete the observation data from polar-orbiting satellites, obtaining gridded data under the complete Mercator projection. This invention, by controlling the number of deep learning model parameters and data requirements, utilizes the spatial consistency of the data and combines it with continental contours to achieve highly accurate prediction and completion of observation data.

Description

Technical Field

[0001] This invention relates to the application of artificial intelligence in the fields of space science and remote sensing, and in particular to a method for supplementing meteorological observation data from polar-orbiting satellites. Background Technology

[0002] Meteorological research is crucial for understanding and predicting weather and climate change, significantly impacting agriculture, transportation, disaster management, and many other fields. In modern meteorology, satellite observation data provides large-scale, high-resolution Earth observation information, vital for weather forecasting and climate change monitoring. However, due to limitations of polar-orbiting satellite sensors, signal interference, and satellite orbits, polar-orbiting satellite observation data often suffers from missing data, typically manifesting as olive-shaped missing areas perpendicular to the scan line or scattered missing points on both sides of the scan width. This data gap severely impacts the accuracy of data and the effectiveness of meteorological research. For example, it leads to inaccurate predictions of extreme weather events, affecting disaster response strategies, and impacting the accuracy of climate models, ultimately affecting long-term predictions of climate change trends. Traditional satellite data completion methods mainly include statistical models and interpolation techniques. While these methods can alleviate this problem to some extent, they rely on existing data points to estimate or infer missing data. Therefore, they are ineffective when the missing data area is large or unevenly distributed. Furthermore, these methods often fail to accurately capture the dynamic changes within the complex climate system, resulting in inaccurate and unreliable recovered data. With the development of deep learning technology, using advanced algorithms to reconstruct and predict missing satellite data has become an effective way to improve data integrity and accuracy. The effective application of these methods depends on multiple aspects such as algorithm design, the quality and quantity of training data, and the availability of computing resources.

[0003] Current deep learning-based satellite data completion methods can be divided into two categories: traditional convolution and partial convolution. Vanilla convolution does not distinguish between existing data and the parts to be completed. It works well for small or less complex missing data regions, but its accuracy and efficiency are affected when dealing with large-scale or complex background missing areas. Partial convolution focuses on feature extraction of effective data by performing convolution operations only on existing pixel regions, resulting in better performance in complex terrain environments. However, partial convolution requires a complex mask update mechanism, and insufficient background information can limit the completion effect. Gated convolution improves upon the mechanism of partial convolution by using a soft training mask update mechanism, further optimizing the network's ability to extract missing data and achieving better results in data completion tasks. Summary of the Invention

[0004] This invention proposes a method for completing meteorological observation data from polar-orbiting satellites. Based on a self-supervised framework, it combines gated convolution, attention mechanism, and can introduce prior knowledge to fully utilize the spatial consistency of satellite observation data and combine prior knowledge of continental contours to predict and complete missing observation data from polar-orbiting satellites with high accuracy, while controlling the number of parameters and data requirements of deep learning models.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for supplementing meteorological observation data from polar-orbiting satellites, comprising the following steps:

[0007] Step 1: Obtain satellite observation brightness temperature data for the target area;

[0008] Step 2: Convert the satellite observation brightness temperature data into point-by-point scan data to obtain point-by-point brightness temperature data; and each row of the point-by-point brightness temperature data contains information about one field of view.

[0009] Acquire the continental outline image of the target area and convert it into a single-channel grayscale matrix;

[0010] Step 3, set up the brightness temperature grid dataset:

[0011] For point-by-point scanning data, based on the scanning time of each field of view, the daily observation data is divided into several groups, with each group representing a sample;

[0012] The point-by-point brightness temperature data of the target area is mapped to the grid points of the single-channel grayscale image matrix to obtain the brightness temperature grid point data of each sample. During the mapping, for a point containing multiple observation data, it is fused to obtain the unique observation data of that grid point. For grid points without observation data, a value of 0 is set. 0 grid points are used to represent the area where data needs to be completed.

[0013] Step 4: Generate a self-supervised training mask. This training mask is an olive-shaped mask and / or a random mask. The training mask is a binary mask, where a value "1" indicates that grid points used as input data are retained, and a value "0" indicates that grid points to be filled in are to be filled in. During the training of the data completion network model, the training mask can be used to actively cover some existing data as labels for self-supervised learning.

[0014] Step 5: Train the data completion network model based on the training samples to obtain the data completion network model for completing polar-orbiting satellite meteorological observation data.

[0015] The data completion network model consists of two parts: a contour feature extraction network and a data repair network. The input of the contour feature extraction network is a single-channel grayscale image matrix, which is used to obtain the overall distribution pattern of the continental contour image and serve as the coarse repair result.

[0016] The data inpainting network adopts an autoencoder architecture, which includes an encoder and a decoder based on a gated convolutional network. The encoder's input includes the coarse inpainting result output by the contour feature extraction network, and the image data obtained by multiplying the training mask by the brightness temperature grid data of each training sample. The encoder and decoder include two branches: branch one is used to extract the global features of the feature map output by the encoder, and branch two focuses on the attention feature map of similar blocks in the feature map output by the encoder based on the context attention mechanism module. The global features of branch one and the attention feature map of branch two are concatenated and then input into the decoder. The completed data, i.e., the completed brightness temperature grid data, is obtained based on the decoder's output.

[0017] Furthermore, the resolution of the grid points in the single-channel grayscale image matrix is ​​set to 0.5°×0.5°.

[0018] Furthermore, in step 3, for each point containing multiple observation data, outliers are removed and the average is taken.

[0019] Furthermore, the olive-shaped mask is: the numerical distribution is similar to the data-free olive-shaped region between two adjacent scan swaths of a polar-orbiting satellite; the random mask is: the brightness temperature grid data is randomly masked according to a predetermined ratio (preferably 10%) to obtain the random mask.

[0020] Furthermore, step 4 specifically includes the following steps:

[0021] Step 401: For each sample X used for training, generate a grid mask M corresponding to its brightness temperature grid data. grid Grid mask M grid This is a binary mask where a value of "1" represents grid points with observed data and a value of "0" represents grid points with missing observed data.

[0022] Step 402, use the connected component algorithm to process the grid mask M. grid Find and number all spatially contiguous missing data regions. Set a threshold based on the statistical results of the number of grid points in the olive-shaped missing regions, and filter out several olive-shaped missing regions whose number of missing grid points exceeds the threshold. Define k to represent the number of olive-shaped missing regions selected.

[0023] Step 403, Generate olive-shaped mask: Randomly select n from the selected k olive-shaped missing regions, and combine them with the grid mask M. gridIn an olive-shaped mask of the same size, the corresponding grid points of n regions are set to 0, and the rest are set to 1. The resulting olive-shaped mask is then cyclically shifted in the horizontal direction, with the displacement distance being half the width of the cut; where n is a preset integer greater than 1.

[0024] Step 404, Generate a random mask: Based on the grid mask M grid In a binary matrix of the same size, 0 and 1 are randomly distributed, wherein the proportion of 1 to the total number of points is a preset value, preferably 10%;

[0025] Step 405, Generate training mask: Combine the olive-shaped mask and the random mask with the grid mask M respectively. grid The training mask is obtained by bitwise multiplication.

[0026] In step 5, when training the data completion network model, a point-to-point reconstruction loss L is used. g and smoothing loss L tv The weighted sum is used as the total loss L of the model. Two sub-losses (L...) g L tv The weighting coefficients can be set based on the actual application scenario.

[0027] Furthermore, the reconstruction loss L g and smoothing loss L tv Specifically set as follows:

[0028]

[0029]

[0030] Where X represents the sample used for training, and Y represents the completed data output by the data completion network model corresponding to sample X. (i,j) This indicates the value of data Y at grid point (i,j). c This indicates the coarse repair result, and the evaluation mask M is used. l =M grid -M in M grid The grid mask M represents the brightness temperature grid data of sample X. grid Grid mask M grid For the binary mask, the value "1" is used to represent grid points with observed data, and the value "0" is used to represent grid points with missing observed data. N is the number of grid points, and λ1 to λ4 are four preset coefficients.

[0031] w(x) is a point-to-point reconstruction loss function containing constants ω, ε, and C, and its expression is:

[0032]

[0033] Here, x is the input to the reconstruction loss function w(x).

[0034] Furthermore, the network structure of the contour feature extraction network of the data completion network model includes, in sequence: m1 stacked lower convolutional layers, several stacked residual modules, m1 stacked upper convolutional layers, and one channel convolutional layer. The coarse repair result output by the channel convolutional layer of the contour feature extraction network is consistent with the data dimension of the single-channel grayscale matrix of the input continental contour image, where m1 is a preset integer greater than 1.

[0035] Furthermore, the encoder and decoder of the data repair network are both multi-layered gated convolutional layers, and both include the same number of gated convolutional layers, for example, a 6-layer gated convolutional layer.

[0036] Furthermore, the two branches between the encoder and decoder of the data repair network are specifically set as follows:

[0037] Branch 1 consists of several stacked dilated convolutional layers and m² stacked gated convolutional layers.

[0038] Branch 1 consists of m2 stacked gated convolutional layers, a context attention module, and m2 stacked gated convolutional layers, where m2 is a preset integer greater than 1.

[0039] The technical solution provided by this invention brings at least the following beneficial effects:

[0040] This invention employs a self-supervised training strategy, where the training mask and the missing regions share a similar distribution pattern. This specifically improves the network's completion results for missing polar-orbiting satellite observation data and reduces the network's need for label data. The network can be trained without introducing any other source data or performing spatiotemporal matching. Furthermore, based on the characteristic of brightness temperature data being affected by land surface radiation, prior knowledge of continental contours is introduced to improve the completion effect for large areas of continuous missing data and missing data at land-sea boundaries. Finally, a gated convolutional network is used to complete the polar-orbiting satellite observation data, obtaining gridded data under the complete Mercator projection. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of a method for supplementing meteorological observation data from polar-orbiting satellites, provided as an embodiment of the present invention;

[0043] Figure 2 A flowchart for generating a self-supervised training mask;

[0044] Figure 3 The diagrams show the original grid mask (3-a), olive-shaped mask (3-d), random mask (3-c), and training mask (3-b).

[0045] Figure 4 Repair the neural network model structure diagram for gated convolutional data;

[0046] Figure 5 This is a schematic diagram of the data completion result of a method for completing meteorological observation data from polar-orbiting satellites provided in an embodiment of the present invention, wherein (5-a) is the data to be completed, and (5-b) is the completion result. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0048] This invention provides a method for completing meteorological observation data from polar-orbiting satellites, based on a self-supervised framework, combined with gated convolution, attention mechanisms, and the introduction of prior knowledge. It fully utilizes the spatial consistency of satellite observation data and combines prior knowledge of continental contours to predict and complete missing observation data from polar-orbiting satellites with high accuracy, while controlling the number of parameters and data requirements of the deep learning model.

[0049] As one possible implementation, such as Figure 1 As shown, the specific implementation process of a method for supplementing meteorological observation data from polar-orbiting satellites provided in this embodiment of the invention includes:

[0050] S11: Acquire satellite observation brightness temperature data.

[0051] Download AMSU-A observational brightness temperature data (in sv format) from the National Oceanic and Atmospheric Administration (NOAA) data platform.

[0052] S12: Data preprocessing.

[0053] The downloaded AMSU-A data in SV format was converted into point-by-point scan data in TXT format. Each line contains the latitude and longitude of a field of view, the ascending and descending orbit markers, the scan time, the brightness temperature values ​​of 15 channels (in K), the number of scan lines and fields of view, and the land surface type information. The "drawcoastlines" function from the Basemap 1.3.7 function library in Python 3.8 was used to generate a continental contour image of the 60°N-60°S region. After removing the bounding boxes and scales of the image, it was saved as a single-channel grayscale matrix with values ​​in the range [0:255].

[0054] S13: Create a brightness temperature grid dataset.

[0055] The point-by-point data obtained in step 2 is mapped to the grid points of a matrix (single-channel grayscale matrix) to obtain grid point data X, with a grid point resolution of 0.5°×0.5. Based on the scanning time of each field of view, the daily observation data is divided into two groups: 6:00±7h and 18:00±7h. Each point-by-point data is assigned to a corresponding grid point according to the center position of its field of view, retaining only the data with observation values ​​in the range [0,400]K. For grid points containing one or more data points, the average value is taken; for grid points with no observation data, a value of 0 is taken. Ultimately, each day contains two matrix samples of size 241×720, namely 6:00±7h and 18:00±7h. Grid points with 0 values ​​in each matrix sample need to be padded with data; these padded grid points account for approximately 30%-50% of the total grid points. After all samples are processed, the dataset is divided into training:validation:test sets in a ratio of 5:2:1.

[0056] S14: Generate a self-supervised training mask.

[0057] Generate a corresponding original binary mask M for each sample. grid This distinguishes between grid points with data ("1") and missing grid points ("0"). For each sample used for training, some existing data is actively masked and used as a label during training. This label is treated as missing data during training and used to calculate training loss and parameter metrics during validation. The generated binary training mask M is used. inTo achieve this function, the training mask consists of an olive-shaped mask, a random mask, and the original mask. The olive-shaped mask's numerical distribution pattern is similar to the data-free region between two adjacent scan swaths of a polar-orbiting satellite, but the missing region is located in the horizontal middle of the swath. The method for obtaining this mask is as follows: The connected component algorithm is used to extract the olive-shaped missing regions from the grid data to be repaired. n regions are randomly selected (where n is an empirical value; in this embodiment, n=4 for the training phase and n=2 for the verification phase) and marked as "0" in the binary mask, with the rest marked as "1". The resulting matrix is ​​then cyclically shifted horizontally by a distance equal to half the swath width. The random mask has "0" and "1" randomly distributed across all grid points. The method for obtaining this mask is as follows: A binary random matrix of the same size as the grid data is generated. The grid point values ​​are set according to probability, P(grid point value 1) = 0.9, P(grid point value 0) = 0.1.

[0058] S15: Build and train the data completion network.

[0059] This invention's embodiments construct a data completion network based on gated convolution, the network structure of which is as follows: Figure 4 As shown, the network consists of two parts: ① a contour feature extraction network, including 3 lower convolutional layers, 5 residual modules, 3 upper convolutional layers, and 1 channel convolutional layer, which generates a data distribution pattern Y after inputting a land contour image. c As a coarse repair result; ② Data repair network, inputting samples of the masked grid points. and training mask M in The algorithm employs an autoencoder (AE) architecture, replacing convolutional layers with gated convolutional layers. The encoder and decoder consist of two branches. The upper branch has four stacked dilated convolutional layers with dilation rates of 2, 4, 8, and 16, which helps extract global features. The lower branch introduces a context attention mechanism module, which improves the inpainting effect by focusing on similar regions in the feature maps. The feature maps output from the two branches are concatenated and input into the decoder to obtain the completed data Y.

[0060] The data repair network is connected to the data repair network via a single-channel convolutional layer.

[0061] This invention uses a large number of data samples to be repaired to train the model, and uses a combined loss function L of wing Loss (point-to-point reconstruction loss) and TV Loss (smoothing loss) to optimize the model training. Its expression is as follows:

[0062] L = L g +L tv

[0063] Among them, L g For point-to-point reconstruction loss, L tvFor smoothing the loss, the expressions are as follows:

[0064]

[0065]

[0066] Where λ1 to λ4 are four preset coefficients, in this example λ1 = 5, λ2 = λ3 = 10, λ4 = 60, Y (i,j) The evaluation mask M represents the value of the completed data Y output by the model at grid position (i,j). l =M grid -M in N is the number of grid points, and w(x) is the wing loss containing constants ω, ε, and C, expressed as follows:

[0067]

[0068] Here, x represents the input of the function w(x).

[0069] S16: Data completion test.

[0070] Load the trained model weights and input the data to be completed. In this embodiment, the number of olive-shaped regions to be covered is n=1, and 10% of the grid points are randomly covered. After the model completes the data, calculate the Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) for the actively covered grid points.

[0071] Further, see Figure 2 The specific implementation process for generating self-supervised training masks includes:

[0072] S21: Generate a grid mask. Generate a corresponding binary mask M for each sample X. grid The mask size is the same as the sample. The grid points with data in the sample correspond to "1" in the binary mask, and the grid points to be filled correspond to "0".

[0073] S22: Search for missing regions. Use the connected component algorithm to search M. grid Locate and number all spatially contiguous missing data regions. Based on the statistical results of the number of grid points in the olive-shaped missing regions, set a threshold and filter out k olive-shaped missing regions.

[0074] S23: Generate an olive-shaped mask. Randomly select n regions from the selected k olive-shaped missing regions, and combine them with M... gridIn an olive-shaped mask of the same size, set the corresponding grid points of n regions to 0, and the rest to 1. Then, perform a cyclic shift in the horizontal direction on the resulting olive-shaped mask, with the shift distance being half the width of the cut.

[0075] S24: Generate a random mask. (In conjunction with M) grid In a binary matrix of the same size, 0 and 1 are randomly distributed, with 1 accounting for 10% of the total number of points.

[0076] S25: Generate the training mask. The training mask consists of an olive-shaped mask, a random mask, and M. grid The result is obtained by bitwise multiplication. See the results below. Figure 3 , where (3-a) is the original grid mask, (3-b) is the training mask, (3-c) is the random mask, and (3-d) is the olive-shaped mask.

[0077] After the above processing steps, the following can be obtained: Figure 5 The data completion network shown can complete the data of polar-orbiting satellite observations. In this network, (5-a) is the data to be completed and (5-b) is the completed result.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0079] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for supplementing meteorological observation data from polar-orbiting satellites, characterized in that, Includes the following steps: Step 1: Obtain satellite observation brightness temperature data for the target area; Step 2: Convert the satellite observation brightness temperature data into point-by-point scan data to obtain point-by-point brightness temperature data; and each row of the point-by-point brightness temperature data contains information about one field of view. Acquire the continental outline image of the target area and convert it into a single-channel grayscale matrix; Step 3, set up the brightness temperature grid dataset: For point-by-point scanning data, based on the scanning time of each field of view, the daily observation data is divided into several groups, with each group representing a sample; The point-by-point brightness temperature data of the target area is mapped to the grid points of the single-channel grayscale image matrix to obtain the brightness temperature grid point data of each sample. During the mapping, for grid points containing multiple observation data, they are fused to obtain the unique observation data of the grid point. For grid points without observation data, a value of 0 is set. 0 grid points are used to represent the area where data needs to be completed. Step 4: Generate a self-supervised training mask, which is an olive-shaped mask and / or a random mask. The training mask is a binary mask, where "1" indicates that grid points are retained as input data and "0" indicates that grid points are to be filled in. Step 5: Train the data completion network model based on the training samples to obtain the data completion network model for completing polar-orbiting satellite meteorological observation data. The data completion network model includes a contour feature extraction network and a data repair network. The input of the contour feature extraction network is a single-channel grayscale image matrix, which is used to obtain the overall distribution pattern of the continental contour image and serve as the coarse repair result. The data inpainting network consists of an encoder and a decoder based on a gated convolutional network. The encoder's input includes the coarse inpainting result output from the contour feature extraction network, and the image data obtained by multiplying the training mask by the brightness temperature grid data of each training sample. The encoder and decoder have two branches: branch one extracts the global features of the feature map output by the encoder, and branch two focuses on the attention feature maps of similar blocks in the feature map output by the encoder based on the context attention mechanism module. The global features of branch one and the attention feature maps of branch two are concatenated and then input into the decoder. The completed data is obtained based on the decoder's output. Step 4 specifically includes the following steps: Step 401: For each sample X used for training, generate a grid mask corresponding to its brightness temperature grid data. Grid mask This is a binary mask where a value of "1" represents grid points with observed data and a value of "0" represents grid points with missing observed data. Step 402, use the connected component algorithm to process the grid mask. Find and number all spatially contiguous missing data regions. Set a threshold based on the statistical results of the number of grid points in the olive-shaped missing regions, and filter out several olive-shaped missing regions whose number of missing grid points exceeds the threshold. Define k to represent the number of olive-shaped missing regions selected. Step 403, Generate olive-shaped mask: Randomly select n from the selected k olive-shaped missing regions, and combine them with the grid mask. In an olive-shaped mask of the same size, the corresponding grid points of n regions are set to 0, and the rest are set to 1. The resulting olive-shaped mask is then cyclically shifted in the horizontal direction, with the displacement distance being half the width of the cut; where n is a preset integer greater than 1. Step 404, Generate a random mask: in conjunction with the grid mask In a binary matrix of the same size, 0 and 1 are randomly distributed, where the proportion of 1 to the total number of grid points is a preset value; Step 405, Generate training mask: Combine the olive-shaped mask and the random mask with the grid mask respectively. The training mask is obtained by bitwise multiplication; The contour feature extraction network structure of the data completion network model includes, in sequence: m1 stacked lower convolutional layers, several stacked residual modules, m1 stacked upper convolutional layers, and one channel convolutional layer. The coarse repair result output by the channel convolutional layer of the contour feature extraction network is consistent with the data dimension of the single-channel grayscale matrix of the input continental contour image, where m1 is a preset integer greater than 1. The two branches between the encoder and decoder of the data repair network are specifically set as follows: Branch 1 consists of several stacked dilated convolutional layers and m² stacked gated convolutional layers. Branch 2 consists of m2 stacked gated convolutional layers, a context attention module, and m2 stacked gated convolutional layers, where m2 is a preset integer greater than 1.

2. The method as described in claim 1, characterized in that, The resolution of the grid points in the single-channel grayscale image matrix is ​​set to 0.5°×0.5°.

3. The method as described in claim 1, characterized in that, In step 3, for grid points containing multiple observation data, outliers are removed and the average is taken.

4. The method as described in claim 1, characterized in that, The olive-shaped mask is: the numerical distribution is similar to the data-free olive-shaped region between two adjacent scans of a polar-orbiting satellite; the random mask is: the brightness temperature grid data is randomly masked according to a predetermined ratio to obtain the random mask.

5. The method as described in claim 1, characterized in that, In step 5, when training the data completion network model, a point-to-point reconstruction loss is used. and smoothing loss The weighted sum is used as the total loss L of the model.

6. The method as described in claim 5, characterized in that, Reconstruction loss and smoothing loss Specifically set as follows: ； ； in, This represents the samples used for training. This indicates that the output of the data completion network model corresponds to the sample. Supplementary information, This indicates that the value of data Y at grid point (i,j) is to be completed. This indicates the coarse repair result, and the evaluation mask is used. , Indicates the training mask. Indicates sample The grid mask corresponding to the brightness temperature grid data Grid mask This is a binary mask where a value of "1" represents grid points with observed data and a value of "0" represents grid points with missing observed data. The number of grid points, ~ These are four preset coefficients; For included constants , , The point-to-point reconstruction loss function is expressed as follows: ； in, To reconstruct the loss function Input.

7. The method as described in claim 1, characterized in that, The encoder and decoder of the data repair network are both multi-layered gated convolutional layers, and both contain the same number of gated convolutional layers.

Images