Satellite cloud image prediction method and system based on central region optimization

Abstract

The application discloses a satellite cloud picture prediction method and system based on center area optimization, and belongs to the field of digital image processing.In the application, historical satellite cloud pictures are respectively converted into original time features and original space features through reconstruction operation; the original time features are input into a time encoder to generate multi-scale time coding features, and the original space features are input into a space encoder to generate multi-scale space coding features; the coding features are combined into space-time features through splicing operation; the space-time features, the multi-scale time coding features and the multi-scale space coding features are input into a space-time decoder to generate multi-scale space-time decoding features for predicting cloud pictures; and the multi-scale space-time decoding features are input into a satellite cloud picture center area optimization module to generate a center area satellite cloud picture at a future moment.The application can predict future cloud pictures, and continuously optimize the center area of the predicted cloud pictures, thereby improving the prediction precision of the center area of the satellite cloud picture.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of digital image processing, and particularly relates to a satellite cloud image prediction method and system based on center region optimization. BACKGROUND

[0002] Developing cloud image prediction can improve the power prediction accuracy of photovoltaic bases. In existing cloud image prediction tasks, the area covered by the input cloud image not only includes photovoltaic bases but also surrounding areas, so that the model can learn the spatiotemporal evolution law of the target cloud layer. The size of the surrounding area is proportional to the historical cloud image frame number and the prediction scale. Therefore, when a large frame number is selected to obtain sufficient spatiotemporal information of the target cloud layer and hour-level power prediction is carried out, the range covered by the input cloud image is also large. The prediction cloud image generated by the encoder-decoder network not only includes the core area centered on the photovoltaic base but also a large number of irrelevant areas, resulting in a conflict between the overall cloud image prediction accuracy and the cloud image prediction accuracy of the core area. SUMMARY

[0003] Therefore, the present application provides a satellite cloud image prediction method and system based on center region optimization, proposes a new satellite cloud image prediction network and optimization framework, generates a high-precision initial prediction cloud image through the spatiotemporal separation encoding characteristics and spatiotemporal mixed decoding characteristics, and continuously optimizes the cloud image in the center region through a satellite cloud image center region optimization module, so as to improve the cloud image prediction accuracy in the center region and avoid the conflict between the overall cloud image prediction accuracy and the cloud image prediction accuracy of the core area.

[0004] In order to achieve the above-mentioned application purposes, the present application provides the following technical solutions:

[0005] According to a first aspect of the present application, a satellite cloud image prediction method based on center region optimization is provided, comprising:

[0006] Step S1: obtaining historical satellite cloud images, the size of the historical satellite cloud images being (B, C, T, H, W), wherein C, H and W are respectively the channel number, height and width of the satellite cloud image, B is the hyperparameter batch size during neural network training, and T is the frame number of the satellite cloud image;

[0007] Step S2: converting the historical satellite cloud images into original time features and original space features through reconstruction operations, wherein the size of the original time features and the original space features is respectively (B×C, T, H, W) and (B×T, C, H, W);

[0008] Step S3: inputting the original time features The input is fed into a temporal encoder to generate multi-scale temporal encoded features; the original spatial features are then processed. The input is fed into the spatial encoder to generate multi-scale spatial encoded features, specifically:

[0009] The temporal encoder and spatial encoder have identical structures. Both consist of three cascaded encoding modules with identical network architectures, with the output of one encoding module serving as the input of the next. (Original temporal features) The input is fed into the time encoder, and the three encoding modules in the time encoder output multi-scale time-coded features F1 according to the order of their positions. e,t , and F1 e,t , and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively, and The dimensions are (B×C,T3′,H / 8,W / 8), representing the original spatial features. The input is fed into the spatial encoder, and the three encoding modules of the spatial encoder output multi-scale spatial encoded features F1 according to the order of their positions. e,s , and F1 e,s , and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively, and The dimensions are (B×T,C3′,H / 8,W / 8);

[0010] Step S4: and F3 e,s Spatiotemporal features Fts are synthesized through reconstruction and splicing operations. The size of Fts is (B,CT3′+TC3′,H / 8,W / 8).

[0011] Step S5: Transfer the spatiotemporal features F ts Multi-scale temporal and spatial encoded features are input into the spatiotemporal decoder to generate multi-scale spatiotemporal decoding features for predicting cloud maps. Specifically:

[0012] The spatiotemporal decoder consists of three cascaded decoding modules, with the output of one module serving as the input to the next. The three modules have identical network architectures, and in the order of their positions, they output multi-scale spatiotemporal decoded features F1d,ts, etc. and F1 d,ts , and The resolutions are (H / 4, W / 4), (H / 2, W / 2), and (H, W), respectively.

[0013] The decoding steps are as follows:

[0014] Step S51: Input the spatiotemporal feature Fts into the spatiotemporal decoder. The first decoding module in the spatiotemporal decoder outputs F1d,ts, where the resolution of F1d,ts is F. ts Twice the resolution, F1 d,ts The dimensions are (B, C1′, H / 4, W / 4);

[0015] Step S52: Set F1 d,ts , as well as The features are concatenated together to form the second decoding module in the spatiotemporal decoder, which is then fed into the hybrid feature input. The output is...

[0016] Step S53: ... and These features are concatenated together to form the third decoding module in the spatiotemporal decoder, which then outputs the hybrid features. The resolution is 8 times that of the spatiotemporal feature Fts. The dimensions are (B, C3′, H, W);

[0017] Step S6: Set F1 d,ts , and Input the data into the satellite cloud image center region optimization module to generate a satellite cloud image of the center region for a future time.

[0018] The specific generation steps are as follows:

[0019] Step S61: ... The data is sent to the prediction layer to generate an initial satellite cloud image;

[0020] The prediction layer is a convolutional layer with a kernel size of 3×3 and a stride of 1.

[0021] Through the prediction layer The dimensions (B,C3′,H,W) are transformed to (B,1,H,W), and the resolution of the initial satellite cloud image is (H,W).

[0022] Step S62: Combine the initial satellite cloud image from step S61 with the image generated in step S52. Input into the satellite cloud image center region optimization module to optimize the central region cloud image;

[0023] The specific optimization steps are as follows:

[0024] Step S621: cropping the initial satellite cloud image to obtain a center region cloud image of the initial satellite cloud image; the resolution of the center region cloud image is (H / 2, W / 2);

[0025] Step S622: center region cloud image and splicing, using the residual generator to generate a center region cloud image optimization residual;

[0026] Step S623: summing the center region cloud image optimization residual generated in step S622 and the center region cloud image in S621 to obtain an optimized center region cloud image;

[0027] Step S624: summing the optimized center region cloud image in step S623 and F1 d,ts input to the satellite cloud image center region optimization module, and the optimization step is repeated to obtain a twice-optimized center region cloud image;

[0028] The optimization step of step S62 is repeated until the center region satellite cloud image at the future time is generated, and thus the satellite cloud image prediction based on center region optimization is completed.

[0029] According to the embodiment of the application, in step S2, the reshape operation uses Reshape in the Pytorch deep learning framework.

[0030] According to the embodiment of the application, in step S3, all encoding modules in the time encoder and the space encoder are composed of two convolution layers with a kernel size of 3x3, wherein the convolution step of the first convolution layer is 1 and the convolution step of the second convolution layer is 2.

[0031] According to the embodiment of the application, in step S4, the and are synthesized through the reconstruction and splicing operations to obtain the spatio-temporal feature F ts The process is as follows: the size of the is transformed from (BxC, T3', H / 8, W / 8) to (B, CT3', H / 8, W / 8) through the reshape operation, the size of the is transformed from (BxT, C3', H / 8, W / 8) to (B, TC3', H / 8, W / 8), and then the changed and are spliced to obtain the spatio-temporal feature F ts , and the size of F ts is (B, CT3'+TC3', H / 8, W / 8).

[0032] According to the embodiment of the present application, in step S5, the three decoding modules in the space-time decoder are each composed of two convolution layers with a convolution kernel size of 3*3, wherein the convolution step length of the first convolution layer is 2 and the convolution step length of the second convolution layer is 1.

[0033] According to the embodiment of the present application, in step S622, the residual generator is composed of two convolution layers with a convolution kernel size of 3*3, and the convolution step length is 1.

[0034] According to the second aspect of the present application, a satellite cloud image prediction system based on center region optimization is provided, comprising: a computer readable storage medium and a processor.

[0035] The computer readable storage medium is used to store executable instructions.

[0036] The processor is used to read the executable instructions stored in the computer readable storage medium and execute the steps of the satellite cloud image prediction method based on center region optimization.

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] 1) The time characteristics and spatial characteristics are separated and extracted through the parallel network structure, which can effectively avoid information confusion and improve the prediction accuracy of the cloud image prediction model.

[0039] 2) The satellite cloud image center region optimization module can continuously optimize the satellite cloud image center region, avoid the conflict between the overall prediction accuracy and the center region prediction accuracy, improve the prediction accuracy of the center region cloud image, and provide a reference for photovoltaic power prediction. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The flowchart of the method of the present application;

[0041] Figure 2 The overall structure diagram of the method of the present application;

[0042] Figure 3 The space-time feature F ts Splicing operation schematic diagram;

[0043] Figure 4 The structure diagram of the satellite cloud image center region optimization module in the method of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, features and advantages of the present application more obvious and easy to understand, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the present application is not limited by the following embodiments, and the specific implementation manners can be determined according to the technical solutions of the present application and the actual situation. In order to avoid confusing the essence of the present application, the well-known methods, processes and procedures are not described in detail.

[0045] The present application provides a satellite cloud image prediction method based on center region optimization, and a satellite cloud image prediction system based on center region optimization, which comprises a computer readable storage medium and a processor; the computer readable storage medium is used for storing executable instructions; the processor is used for reading the executable instructions stored in the computer readable storage medium and executing the steps of the satellite cloud image prediction method based on center region optimization.

[0046] Referring to Figure 1 and Figure 2 , the satellite cloud image prediction method based on center region optimization provided by the present application comprises the following steps:

[0047] Step S1: obtaining a historical satellite cloud image with a T-frame size of CxHxW;

[0048] The size of the historical satellite cloud image is (B, C, T, H, W), wherein C, H and W are respectively the channel number, height and width of the satellite cloud image, B is the hyperparameter batch size during neural network training, and T is the frame number of the historical satellite cloud image;

[0049] In this embodiment, B=32, C=1, T=8, H=256 and W=256;

[0050] Step S2: converting the input historical satellite cloud image into original time features and original space features through a reconstruction operation, specifically:

[0051] The reconstruction operation is realized by Reshape in the Pytorch deep learning framework, and the size of the original time features and the original space features is (BxC, T, H, W) and (BxT, C, H, W) respectively;

[0052] Step S3: inputting the original time features into a time encoder to generate multi-scale time encoding features, and inputting the original space features into a space encoder to generate multi-scale space encoding features;

[0053] The temporal encoder consists of three cascaded encoding modules, with the output of the previous encoding module serving as the input of the next encoding module. The three encoding modules have the same network architecture, each consisting of two 3×3 convolutional layers with a convolutional kernel size of 1. The first convolutional layer has a convolutional stride of 1, and the second convolutional layer has a convolutional stride of 2. The structure of the spatial encoder is completely identical to that of the temporal encoder.

[0054] The input features are from either a temporal encoder or a spatial encoder. Each encoding module outputs features, with the output of the previous module serving as the input for the next. Following the positional order, the three encoding modules in the temporal encoder output multi-scale temporal encoded features F1. e,t , and And F1 e,t , and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively; according to the order of their positions, the three coding modules in the spatial encoder output multi-scale spatial coding features F1. e,s , and And F1 e,s , and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively.

[0055] The specific encoding steps are as follows:

[0056] Step S31: Input to the time encoder, The output F1 is processed by the first encoding module in the time encoder. e,t F1 e,t resolution is Half the resolution, F1 e,t The dimensions are (B×C,T1′,H / 2,W / 2), and similarly... Input to the spatial encoder, The output F1 is processed by the first encoding module in the spatial encoder. e,s F1 e,s resolution is Half the resolution, F1 e,s The dimensions are (B×T,C1′,H / 2,W / 2).

[0057] Step S32: Set F1 e,t As input, it is fed into the second encoding module of the time encoder, and after processing by the second encoding module, it is output. Resolution F1e,t Half the resolution, similarly for F1 e,s As input, it is fed into the second encoding module of the spatial encoder, and after processing by the second encoding module, it is output. Resolution F1 e,s Half the resolution;

[0058] Step S33: ... As input, it is fed into the third encoding module of the time encoder, and after processing by the third encoding module, it is output. resolution is 1 / 8 of the resolution The dimensions are (B×C,T3′,H′,W′), and similarly... As input, it is fed into the third encoding module of the spatial encoder, and after processing by the third encoding module, it is output. resolution is 1 / 8 of the resolution The dimensions are (B×T,C3′,H′,W′).

[0059] It should be noted that in this invention, T1′ and T3′ both represent the number of frame dimensions; C1′ and C3′ both represent the number of channels.

[0060] Step S4: and Spatiotemporal features F are synthesized through reconstruction and splicing operations. ts ;

[0061] The Reshape operation will The dimensions are transformed from (B×C,T3′,H / 8,W / 8) to (B,CT3′,H / 8,W / 8). The dimensions are transformed from (B×T,C3′,H / 8,W / 8) to (B,TC3′,H / 8,W / 8), and then further transformed by splicing. and Obtain the spatiotemporal feature F ts F ts The dimensions are (B,CT3′+TC3′,H / 8,W / 8);

[0062] The splicing operation is shown as follows Figure 3 As shown.

[0063] Step S5: Transfer the spatiotemporal features F ts Multi-scale temporal coding features and multi-scale spatial coding features are input into the spatiotemporal decoder to generate multi-scale spatiotemporal decoding features for predicting cloud maps;

[0064] The spatiotemporal decoder consists of three cascaded decoding modules, with the output of the previous decoding module serving as the input of the next decoding module. The three decoding modules have the same network architecture, each consisting of two convolutional layers with a kernel size of 3×3. The first convolutional layer has a stride of 2, and the second convolutional layer has a stride of 1.

[0065] The features input to the spatiotemporal decoder are used to output features after each decoding module, with the output of the previous encoding module serving as the input to the next. Following the order of their positions, the three decoding modules in the spatiotemporal decoder output multi-scale spatiotemporal decoded features F1. d,ts , and F1 d,ts , and The resolutions are (H / 4, W / 4), (H / 2, W / 2), and (H, W), respectively.

[0066] The specific decoding steps are as follows:

[0067] Step S51: Transfer the spatiotemporal features F ts The input is fed into the spacetime decoder, and the first decoding module in the spacetime decoder outputs F1. d,ts F1 d,ts The resolution is F ts Twice the resolution, decoding feature F1 d,ts The dimensions are (B, C1′, H / 4, W / 4);

[0068] Step S52: Convert the F1 output in step S51 to... d,ts The output of step S32 as well as The features are concatenated together and input into the second decoding module of the spatiotemporal decoder, and the output is... The resolution is F1 d,ts , Twice the resolution;

[0069] Step S53: ... The output of step S33 and The features are concatenated together and fed into the third decoding module of the spatiotemporal decoder, outputting F3. d,ts , The resolution is the spatiotemporal feature F ts 8 times, The dimensions are (B, C3′, H, W);

[0070] Step S6: Decode the multi-scale spatiotemporal features F1 d,ts , and The initial satellite cloud picture is input into the satellite cloud picture center area optimization module to generate a center area satellite cloud picture at a future time H / 4xW / 4;

[0071] The specific generation steps are as follows:

[0072] Step S61: The initial satellite cloud picture is cropped to obtain a center area cloud picture of the initial satellite cloud picture; The initial satellite cloud picture is input into the prediction layer to generate an initial satellite cloud picture;

[0073] The prediction layer is a convolutional layer with a convolution kernel size of 3x3 and a convolution step size of 1;

[0074] The size (B, C3', H, W) of the initial satellite cloud picture is transformed into (B, 1, H, W) through the prediction layer, and the resolution of the initial satellite cloud picture is (H, W);

[0075] Step S62: The initial satellite cloud picture in step S61 and the center area cloud picture generated in step S52 are input into the satellite cloud picture center area optimization module to optimize the center area cloud picture;

[0076] The satellite cloud picture center area optimization module is shown in FIG. 6. Figure 4

[0077] The specific optimization steps are as follows:

[0078] Step S621: The initial satellite cloud picture is cropped to obtain a center area cloud picture of the initial satellite cloud picture;

[0079] The resolution of the center area cloud picture is (H / 2, W / 2);

[0080] The cropping operation is shown in FIG. 5. Figure 4

[0081] Step S622: The center area cloud picture and the center area cloud picture generated in step S52 are spliced to generate a center area cloud picture optimization residual using a residual generator; The residual generator is composed of two convolutional layers with a convolution kernel size of 3x3, and the convolution step size is 1;

[0082] Step S623: The center area cloud picture optimization residual generated in step S622 is added to the center area cloud picture in step S621 to obtain an optimized center area cloud picture;

[0083] Step S63: The optimized center area cloud picture in step S623 and the F1 d,ts The optimized center area cloud picture is input into the satellite cloud picture center area optimization module, and the optimization steps described in step S62 are repeated to obtain a twice-optimized center area cloud picture;

[0084]

[0085] ​​​​​The resolution of the central region cloud map after secondary optimization is (H / 4, W / 4);

[0086] Table 1 shows the prediction error of the cloud map under different optimization times. EPE (EndPointError) is the evaluation index of prediction error, and it is calculated as follows:

[0087]

[0088] In the formula, ci and These are the actual cloud map and the predicted cloud map for the i-th optimization, respectively;

[0089] It is evident that the prediction errors of the overall and central regions are quite similar, indicating a conflict between local and overall prediction accuracy. The improvement in the prediction accuracy of the central region is limited by the overall cloud map. This method, through repeated optimization of the central region of the cloud map, significantly improves the prediction accuracy of the central region. Compared to the unoptimized scenario, the prediction error of the central region cloud map is reduced by 18.79% and 27.63% after one and two optimizations, respectively. This effectively resolves the contradiction between local and overall prediction accuracy.

[0090] Table 1

[0091]

Claims

1. A method for satellite cloud map prediction based on central region optimization, characterized in that, Comprise: Step S1: acquire historical satellite cloud images, the size of the historical satellite cloud images is (B, C, T, H, W), wherein C, H and W are respectively the channel number, height and width of the satellite cloud images, B is the hyperparameter batch size during neural network training, and T is the frame number of the satellite cloud images; Step S2: converting the historical satellite cloud images into original temporal features respectively through a reconstruction operation and original spatial features The sizes of the original temporal features and original spatial features are (BxC, T, H, W) and (BXT, C, H, W) respectively; Step S3: input the original time feature to the time encoder to generate multi-scale time encoding features; input the original spatial feature to the spatial encoder to generate multi-scale spatial encoding features, specifically: The temporal encoder and spatial encoder have identical structures. Both consist of three cascaded encoding modules with identical network architectures, with the output of one encoding module serving as the input of the next. (Original temporal features) The input is fed into the time encoder, and the three encoding modules of the time encoder output multi-scale time-coded features in the order of their positions. and and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively, and The dimensions are (B×C,T3′,H / 8,W / 8), representing the original spatial features. The input is fed into the spatial encoder, and the three encoding modules of the spatial encoder output multi-scale spatial encoded features in the order of their positions. and and The resolutions are (H / 2, W / 2), (H / 4, W / 4), and (H / 8, W / 8), respectively, and The dimensions are (B×T,C3′,H / 8,W / 8); Step S4: and Spatiotemporal features Fts are synthesized through reconstruction and splicing operations. The size of Fts is (B,CT3′+TC3′,H / 8,W / 8). Step S5: inputting the spatiotemporal feature F ts , the multi-scale time encoding feature and the multi-scale space encoding feature into a spatiotemporal decoder to generate a multi-scale spatiotemporal decoding feature for predicting the cloud image, specifically: The space-time decoder is composed of three decoding modules in series, and the output of the previous decoding module is taken as the input of the next decoding module, the three decoding modules have consistent network architecture, and the three decoding modules in the space-time decoder output multi-scale space-time decoding features in the order of position and and The resolutions of the three decoding modules are (H / 4, W / 4), (H / 2, W / 2) and (H, W) respectively. The decoding steps are as follows: Step S51: input the spatio-temporal feature Fts into the spatio-temporal decoder, and output the first decoded feature Fts1 from a first decoding module in the spatio-temporal decoder with a resolution of F ts times the resolution of F with a size of (B, C1', H / 4, W / 4) Step S52: concatenating and together to form a mixed feature input into a second decoding module of the spatio-temporal decoder, outputting Step S53: concatenating and together to form a mixed feature input into a third decoding module of the spatio-temporal decoder, outputting with a resolution of 8 times that of the spatio-temporal feature Fts, with a size of (B, C3', H, W). Step S6: inputting the center region satellite cloud image of the future time into the center region satellite cloud image optimization module to generate the center region satellite cloud image of the future time. and inputting the center region satellite cloud image of the future time into the center region satellite cloud image optimization module to generate the center region satellite cloud image of the future time. The specific generation steps are as follows: Step S61: generating an initial satellite cloud image from the input prediction layer the input prediction layer The prediction layer is a convolution layer with a convolution kernel size of 3*3 and a convolution step of 1; by the prediction layer the dimensions (B, C3', H, W) of the initial satellite cloud map to (B, 1, H, W) with a resolution of (H, W); Step S62: input the initial satellite cloud picture in step S61 and the cloud picture generated in step S52 into the satellite cloud picture center region optimization module to optimize the center region cloud picture. Step S62: input the initial satellite cloud picture in step S61 and the cloud picture generated in step S52 into the satellite cloud picture center region optimization module to optimize the center region cloud picture. The specific optimization steps are as follows: Step S621: crop the initial satellite cloud image to obtain a center region cloud image of the initial satellite cloud image; the resolution of the center region cloud image is (H / 2, W / 2); Step S622: The center region cloud map and stitching, generating the center region cloud map optimization residual by using the residual generator; Step S623: sum the center region cloud image optimization residual generated in step S622 and the center region cloud image in S621 to obtain the optimized center region cloud image. Step S624: inputting the center region cloud picture optimized in step S623 and the center region cloud picture in step S51 into the satellite cloud picture center region optimization module, repeating the optimization step to obtain a center region cloud picture optimized twice. Step S624: inputting the center region cloud picture optimized in step S623 and the center region cloud picture in step S51 into the satellite cloud picture center region optimization module, repeating the optimization step to obtain a center region cloud picture optimized twice. Repeat the optimization steps of step S62 until the future time center region satellite cloud image is generated, and thus the center region optimization based satellite cloud image prediction is completed.

2. The center region optimization based satellite cloud map prediction method according to claim 1, characterized in that, In step S2, the reshape operation adopts Reshape in the Pytorch deep learning framework.

3. The center region optimization based satellite cloud map prediction method according to claim 1, characterized in that, In step S3, all encoding modules in the time encoder and the space encoder are composed of two convolution layers with a convolution kernel size of 3*3, wherein the convolution step of the first convolution layer is 1, and the convolution step of the second convolution layer is 2.

4. The center region optimization based satellite cloud map prediction method of claim 1, wherein, In step S4, and The spatiotemporal feature F ts is synthesized through a reshaping and splicing operation. The process is as follows: the size of F is changed from (B×C, T3', H / 8, W / 8) to (B, CT3', H / 8, W / 8) through a Reshape operation, the size of F is changed from (B×T, C3', H / 8, W / 8) to (B, TC3', H / 8, W / 8), and then the changed F and F are spliced to obtain the spatiotemporal feature F ts , and the size of F ts is (B, CT3'+TC3', H / 8, W / 8).

5. The center region optimization based satellite cloud map prediction method according to claim 1, wherein, In step S5, the three decoding modules in the space-time decoder are composed of two convolution layers with a convolution kernel size of 3*3, wherein the convolution step of the first convolution layer is 2, and the convolution step of the second convolution layer is 1.

6. The center region optimization based satellite cloud map prediction method according to claim 1, wherein, In step S622, the residual generator is composed of two convolution layers with a convolution kernel size of 3*3, and the convolution step is 1.

7. A satellite cloud map prediction system based on center region optimization, characterized by, Comprise: Computer readable storage medium and processor; The computer readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer readable storage medium, and execute the steps of the center region optimization based satellite cloud image prediction method in any one of claims 1-6.

Images