Satellite cloud image infrared data extrapolation prediction method based on gan and xlstm self-learning gate

Abstract

The application discloses a satellite cloud picture infrared data extrapolation prediction method based on GAN and xLSTM self-learning gate, relates to the field of meteorological satellite data processing, and comprises the following steps: acquiring satellite infrared cloud picture original data and performing pretreatment, and constructing a training set and a test set; a generative adversarial network composed of a generator and a discriminator is constructed; the generator comprises a fusion feature extraction module, an xLSTM self-learning gate module and a separation reconstruction module connected in sequence; the training set is used for alternately training the generator and the discriminator; sample modes and data modes are provided through a graphical interface; the trained generative adversarial network is used for performing infrared data extrapolation prediction on a cloud picture sequence to be predicted; and a predicted cloud picture result is output. The cloud picture prediction is clearer, has stronger time sequence consistency, and is closer to business operation, and is suitable for meteorological short-term prediction, real-time satellite data analysis and disaster weather monitoring.

Description

Technical Field

[0001] This invention relates to the field of meteorological satellite data processing technology, and more specifically to a method for extrapolating and predicting satellite cloud image infrared data based on GAN and xLSTM self-learning gating. Background Technology

[0002] Satellite infrared cloud images are key foundational data for short-term forecasting and disaster monitoring. Time-series cloud image prediction can anticipate cloud evolution in advance, supporting meteorological early warning, aviation meteorological support, and other services. It is a core technical means for intelligent weather forecasting.

[0003] However, existing satellite infrared cloud image prediction methods suffer from the following problems: traditional LSTM networks are prone to gradient vanishing in long-term cloud image prediction, and their ability to model cloud clusters over time is insufficient; while the predicted cloud images generated by ordinary GAN networks have defects such as cloud image blurring, edge distortion, and easy pattern collapse, which cannot meet the needs of quantitative meteorological interpretation. Most existing algorithms only provide model inference and lack a complete business process from raw data preprocessing, model training, model prediction to result visualization output; most studies do not have a supporting visualization interface, making it difficult to achieve batch data processing and intuitive comparison of prediction results; at the same time, existing models generally only support single-duration prediction and do not support multi-duration hierarchical prediction (10min / 30min / 1h / 2h / 3h), which is difficult to meet business needs.

[0004] Therefore, how to provide a set of satellite cloud infrared data extrapolation prediction methods that are high-definition, time-series stable, easy to operate, and batch-processable, from data preprocessing to training and visualization prediction, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a satellite cloud image infrared data extrapolation prediction method based on GAN and xLSTM self-learning gating, which includes complete data preprocessing, model training, multi-duration prediction, and GUI visualization system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention discloses a method for extrapolating and predicting satellite cloud imagery infrared data based on GAN and xLSTM self-learning gating, comprising: Acquire raw satellite infrared cloud image data and preprocess it to construct training and testing sets; A generative adversarial network consisting of a generator and a discriminator is constructed; the generator includes a fusion feature extraction module, an xLSTM self-learning gating module, and a separation and reconstruction module connected in series. The generator and the discriminator are trained alternately against each other using the training set. The system provides sample and data modes through a graphical interface. It uses a trained generative adversarial network to extrapolate and predict the infrared data of the cloud image sequence to be predicted, and outputs the predicted cloud image results.

[0007] Furthermore, the preprocessing specifically includes: Extract the longest continuous time period data from the raw satellite infrared cloud image data; Convert the data of the longest continuous time period from an NPY file to a PNG image; The PNG images were divided into training and testing sets in an 8:2 ratio. In the training and testing sets, an input sequence consisting of six consecutive cloud images is constructed. The input sequence is checked and missing data is filled in.

[0008] Furthermore, the fusion feature extraction module is composed of five stacked two-dimensional convolutional layers, which independently extract spatial features from each frame of cloud image in the input sequence. Through the hierarchical representation capability of multi-level convolution, it captures the edge texture, brightness temperature gradient and multi-scale features of cloud spatial structure of cloud image step by step, and maps the original single-channel cloud image into a semantic representation in high-dimensional feature space. Each convolutional layer has a 3×3 convolutional kernel and is filled with normalized padding; The number of channels in each convolutional layer is 32, 64, 128, 256, and the number of hidden channels are respectively. Each convolutional layer is activated by the LeakyReLU function for nonlinear transformation.

[0009] Furthermore, the xLSTM self-learning gating module is composed of several identical XLSTMCell units cascaded together; Each XLSTMCell unit introduces learnable gating scaling parameters on top of the standard ConvLSTM (Convolutional Long Short-Term Memory) network unit, performing channel-level adaptive scaling of the activation values ​​of the input gate, forget gate, and output gate. The XLSTMCell unit constrains the spatial locality of state updates through convolution operations, so that the cell state and hidden state maintain a two-dimensional structure in the spatial dimension. At the same time, it models time-series dependencies and spatial contextual relationships, learning the motion trajectory, morphological changes, and brightness-temperature evolution characteristics of the cloud over time.

[0010] Furthermore, the separation and reconstruction module is composed of five stacked two-dimensional transposed convolutional layers, with the following channel mappings for each layer: first layer, hidden channel number → 256; second layer, 256 → 128; third layer, 128 → 64; fourth layer, 64 → 32; fifth layer, 32 → 1. The separation and reconstruction module progressively decodes and reconstructs the high-dimensional spatiotemporal features output by the XLSTMCell unit, performs deconvolution operations to map the feature space to the image space, and finally outputs a predicted cloud map with the same size as the input cloud map.

[0011] Furthermore, the discriminator consists of 5 convolutional layers and 3 fully connected layers connected in series. The 5 convolutional layers are used to extract the authenticity features of the cloud map layer by layer, and the 3 fully connected layers output the final discrimination probability based on the authenticity features of the cloud map.

[0012] Furthermore, during the alternating adversarial training process: a hybrid loss function consisting of MAE and SSIM is used; the generator uses the Adam optimizer and the discriminator uses the SGD optimizer; and adaptive learning rate scheduling is used to dynamically adjust the learning rates of the generator and discriminator; at the same time, an early stopping mechanism is introduced during the adversarial training process, and training stops if the validation set loss does not improve after 20 rounds.

[0013] Furthermore, the alternating adversarial training also includes training five independent generator models for different cloud map prediction time requirements, with corresponding prediction time ranges of 10 minutes, 30 minutes, 1 hour, 2 hours, and 3 hours, respectively.

[0014] Furthermore, the sample mode includes: loading a single set of sequences, visually comparing the original cloud map with the predicted cloud map, and supporting animation playback; The data model includes: batch reading of input folders, automatic batch prediction, and saving of results categorized by time.

[0015] Furthermore, the output prediction cloud map results support: automatic conversion of grayscale cloud maps to color mapping cloud maps; synthesis of GIF animations from multiple prediction results; and automatic categorization and storage of results into subfolders according to prediction duration.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention provides a satellite cloud image infrared data extrapolation prediction method based on GAN and xLSTM self-learning gating, which has the following beneficial effects: This invention addresses the task of satellite cloud image temporal prediction. It constructs a generator consisting of a FusionModule for spatial feature extraction, an xLSTM self-learning gated temporal modeling module, and a SeparationModule for image reconstruction, combined with a multi-layer convolutional discriminator to form an adversarial learning architecture. The system supports predictions for five durations: 10 minutes, 30 minutes, 1 hour, 2 hours, and 3 hours, and includes a complete preprocessing workflow, a GUI visualization interface, and batch prediction functionality. This invention ensures the fidelity of cloud image structure through a hybrid MAE+SSIM loss mechanism and improves training stability by employing an adaptive learning rate and early stopping mechanism. It can directly input PNG / JPG / NPY format cloud image sequences and output high-resolution predicted cloud images with GIF animation display support. Compared to traditional methods, this invention provides clearer cloud image predictions, stronger temporal consistency, and more operationally relevant features, making it suitable for short-term meteorological forecasting, real-time satellite data analysis, and severe weather monitoring. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a method for extrapolating and predicting satellite cloud imagery infrared data based on GAN and xLSTM self-learning gating, such as... Figure 1 As shown, it includes: Acquire raw satellite infrared cloud image data and preprocess it to construct training and testing sets; A generative adversarial network (GAN) consisting of a generator and a discriminator is constructed. The generator includes a fusion feature extraction module, an xLSTM self-learning gating module, and a separation and reconstruction module, which are connected in sequence. The generator and discriminator are trained adversarially using the training set. The system provides sample and data modes through a graphical interface. It uses a trained generative adversarial network to extrapolate and predict the infrared data of the cloud image sequence to be predicted, and outputs the predicted cloud image results.

[0021] In one specific embodiment, the preprocessing specifically includes: Extract the longest continuous time period of data from the raw satellite infrared cloud image data to ensure temporal continuity; Convert the longest continuous time period data from NPY files (multi-process) into PNG images; The PNG images were divided into training and testing sets in an 8:2 ratio. In the training and testing sets, an input sequence consisting of six consecutive cloud images is constructed. The input sequence is checked and missing data is filled to ensure sequence integrity.

[0022] Specifically, detecting and filling missing data in the input sequence includes: Step 1, Raw Data Integration: Multi-phase cloud map data files scattered across different date subfolders are uniformly copied to the preprocessing folder, achieving standardized integration of data sources. Step 2, Time Series Integrity Detection: Based on the cloud map acquisition time interval (preferably 10 minutes), time series analysis is performed on the files in the preprocessing folder. Missing data acquisition nodes are detected by the difference in timestamps between adjacent files, and the longest continuous valid time period is identified as a high-quality training data source. Step 3, Adaptive Filling of Missing Data: For detected missing time points, an adjacent frame copying strategy is used for automatic filling. That is, when a cloud map file for a certain moment is missing, the valid cloud map file for the next adjacent time point is automatically copied and renamed to the filename corresponding to the missing moment. Data completion is achieved by utilizing the assumption of spatiotemporal continuity of cloud maps within a short period.

[0023] In a specific embodiment, the FusionModule is composed of five stacked two-dimensional convolutional layers. It independently extracts spatial features from each frame of cloud image in the input sequence. Through the hierarchical representation capability of multi-level convolution, it captures multi-scale features such as edge texture, brightness temperature gradient and cloud spatial structure of cloud image step by step, and maps the original single-channel cloud image into a semantic representation in high-dimensional feature space. Each convolutional layer has a 3×3 convolutional kernel and normalized padding (padding=1); The number of channels in each convolutional layer is 32, 64, 128, 256, and the number of hidden channels (preferably 64) are respectively. Each convolutional layer is activated by the LeakyReLU activation function (with a negative slope coefficient preferably of 0.2) for nonlinear transformation.

[0024] In one specific embodiment, the xLSTM self-learning gating module is composed of several identical XLSTMCell units (learnable gated convolutional long short-term memory units) cascaded together (preferably 4 layers); Each XLSTMCell unit introduces learnable gating scaling parameters (gate_i, gate_f, gate_o) on top of the standard ConvLSTM (Convolutional Long Short-Term Memory) network unit, corresponding to the input gate, forget gate, and output gate, respectively. These parameters are initialized as all-1 tensors in the form of (nn.Parameter) with a shape of (1, hidden_dim, 1, 1), and the activation values ​​of the input gate, forget gate, and output gate are adaptively scaled at the channel level. The XLSTMCell unit constrains the spatial locality of state updates through convolution operations, so that the cell state c and the hidden state h maintain a two-dimensional structure in the spatial dimension, thereby simultaneously modeling temporal series dependencies and spatial contextual relationships, and learning the motion trajectory, morphological changes, and brightness-temperature evolution characteristics of the cloud over time.

[0025] In one specific embodiment, the SeparationModule consists of five stacked 2D transposed convolutional layers (ConvTranspose2d), with the following channel mappings for each layer: first layer, hidden channels → 256; second layer, 256 → 128; third layer, 128 → 64; fourth layer, 64 → 32; fifth layer, 32 → 1, forming a symmetrical decoding structure with the FusionModule. The SeparationModule progressively decodes and reconstructs the high-dimensional spatiotemporal features output by the XLSTMCell unit, mapping the feature space to the image space through deconvolution operations, and finally outputting a predicted cloud map with the same size as the input cloud map. This invention achieves high-precision prediction from continuously observed cloud maps to cloud maps at future times through the synergistic effect of feature extraction from the FusionModule, spatiotemporal evolution modeling of the XLSTMCell, and image reconstruction from the SeparationModule. In particular, the learnable gating mechanism enhances the model's adaptive ability to capture the spatiotemporal evolution of cloud maps.

[0026] In one specific embodiment, the discriminator consists of 5 convolutional layers and 3 fully connected layers connected in series. The 5 convolutional layers are used to extract the authenticity features of the cloud image layer by layer, and the 3 fully connected layers output the final discrimination probability based on the authenticity features of the cloud image. The discriminator is trained adversarially against the generator to improve the realism and detail fidelity of the cloud image.

[0027] In a specific embodiment, during the alternating adversarial training process: a hybrid loss function consisting of MAE and SSIM is used; the generator uses the Adam optimizer and the discriminator uses the SGD optimizer; and adaptive learning rate scheduling is used to dynamically adjust the learning rates of the generator and discriminator; at the same time, an early stopping mechanism is introduced during the adversarial training process, and training is stopped if the validation set loss does not improve after 20 rounds.

[0028] Specifically, the total loss function is: ; in, Total loss; For MAE loss (mean absolute error), The SSIM loss (structural similarity loss) is calculated using the following formula: ; ; in, For the real cloud map The brightness temperature value of each pixel. To predict the cloud map The brightness temperature value of each pixel. The total number of pixels in the image; the formula for calculating SSIM is: ; in, and These are the local means of the real image and the predicted image, respectively; and These are the local variances of the real image and the predicted image, respectively. The local covariance between the real image and the predicted image; and All are stable constants to avoid the denominator being zero.

[0029] MAE loss ensures pixel-level brightness temperature accuracy in the predicted cloud map, while SSIM loss ensures visual structural clarity and texture realism. The two work together to overcome the problems of overly smoothed prediction results and loss of texture details caused by MAE loss alone, while also overcoming the problems of brightness temperature value shift and impact on quantitative inversion caused by SSIM loss alone, thus achieving a balance between numerical accuracy and visual quality.

[0030] Heterogeneous optimizer configuration: The generator uses the Adam optimizer for parameter updates, while the discriminator uses the SGD (Stochastic Gradient Descent) optimizer. The Adam optimizer utilizes an adaptive learning rate adjustment mechanism based on first-order and second-order moment estimation to adapt to the optimization requirements of the non-convex loss surface in the generation task. β 1.β 2 represents the exponential decay coefficients of the first and second moments of the Adam optimizer, controlling the forgetting rate of the historical gradient and its square; the generator calls the built-in Adam optimizer in PyTorch and does not explicitly reset the betas parameters, so the library defaults β1=0.9 and β2=0.999 are actually used; its parameter updates satisfy: m_t= β 1·m_{t-1}+(1- β 1) g_t, v_t = β 2·v_{t-1}+(1- β 2)·g_t², after deviation correction, is calculated as θ_t=θ_{t-1}-η· _t / (√ The generator parameters are updated using _t+ε), where η is the learning rate. The SGD optimizer updates the discriminator parameters using gradient descent with a fixed learning rate, preventing the discriminator from converging too early and becoming too strong, which would cause the generator gradient to vanish, thus maintaining the dynamic balance of GAN training.

[0031] Adaptive learning rate scheduling: The ReduceLROnPlateau (plateau learning rate decay) strategy is used to dynamically adjust the learning rates of the generator and discriminator, with the validation set loss as the monitoring metric. _ When the validation set loss is continuous The relative improvement of the number of wheels (preferably 4 wheels) is less than the threshold. When the value is preferably 0.001, the learning rate is multiplied by the decrease factor. (Preferred value: 0.8); after the learning rate decreases, a cooldown period (preferred to be 3 rounds) is set, during which loss changes are not monitored; the lower limit of the learning rate min_lr is set to 0.0001 to avoid parameter updates from stalling due to an excessively small learning rate.

[0032] Early stopping mechanism: An early stopping patience parameter (preferably 20 epochs) is set. After each training epoch, the loss value is calculated on the validation set. If the current validation loss is better than the historical best validation loss, the current model parameters are saved to the best model file, and the no-improvement counter is reset to 0. If the current validation loss is not better than the historical best validation loss, the no-improvement counter is incremented by 1. When the no-improvement counter reaches the patience parameter, training is terminated, and the best model is output. This invention effectively prevents model overfitting through the early stopping mechanism, and obtains the optimal model parameters for generalization ability early in the training process.

[0033] In one specific embodiment, the alternating adversarial training further includes training five independent generator models for different cloud map prediction time requirements, with corresponding prediction time ranges of 10 minutes, 30 minutes, 1 hour, 2 hours, and 3 hours, respectively.

[0034] Specifically, five independent generator models are trained to meet different cloud map prediction time requirements, corresponding to prediction time ranges of 10 minutes, 30 minutes, 1 hour, 2 hours, and 3 hours, respectively. The training labels for each model are constructed by using the cloud map of the Nth frame after the input sequence (N=1, 3, 6, 12, 18 corresponding to the five time ranges mentioned above) as the prediction target for that input sequence. After training, the five pre-trained model files are saved separately, allowing for automatic loading of the corresponding model based on the user-selected prediction time range during subsequent prediction stages. This invention, through parallel training and independent storage of multi-timescale models, enables users to flexibly select prediction models for different lead times according to actual business needs, improving the system's practicality and applicability.

[0035] In one specific embodiment, the sample mode includes: loading a single set of sequences, visually comparing the original cloud map with the predicted cloud map, and supporting animation playback; The data processing modes include: batch reading of input folders, automatic batch prediction, and saving results categorized by time.

[0036] Supported input files: .png, .jpg, .npy.

[0037] The output of predicted cloud map results supports: automatic conversion of grayscale cloud maps to color-mapped cloud maps; synthesis of GIF animations from multiple frames of prediction results; and automatic categorization and storage of results into subfolders according to prediction duration.

[0038] Specifically, in terms of prediction applications, this invention provides a graphical user interface prediction system built on the PyQt6 framework, supporting both sample mode and data mode. In sample mode, users can load a sequence of six consecutive cloud images, visually compare the original cloud images with the prediction results through a two-column layout, and switch to animation mode to generate GIF animations containing multi-temporal prediction results. In data mode, all cloud image data in a folder can be processed in batches, automatically completing prediction, color mapping, and classification saving. The system supports PNG, JPG, and NPY input formats, and uses a custom 22-level color mapping scheme to output the prediction results as color cloud images. This color mapping is optimized for the brightness and temperature characteristics of satellite cloud images, and enhances cloud image contrast through 5%~95% quantile normalization, facilitating manual interpretation of cloud structure. Through the synergistic effect of the aforementioned technical solutions, this invention achieves high-precision multi-temporal prediction of satellite cloud images, possessing advantages such as high prediction accuracy, strong ability to capture spatiotemporal evolution patterns, and convenient operation. It can be widely applied in fields such as weather forecasting and disaster early warning.

[0039] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0040] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for extrapolating and predicting satellite cloud imagery infrared data based on GAN and xLSTM self-learning gating, characterized in that, include: Acquire raw satellite infrared cloud image data and preprocess it to construct training and testing sets; A generative adversarial network consisting of a generator and a discriminator is constructed; the generator includes a fusion feature extraction module, an xLSTM self-learning gating module, and a separation and reconstruction module connected in series. The generator and the discriminator are trained alternately against each other using the training set. The system provides sample and data modes through a graphical interface. It uses a trained generative adversarial network to extrapolate and predict the infrared data of the cloud image sequence to be predicted, and outputs the predicted cloud image results.

2. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, The preprocessing specifically includes: Extract the longest continuous time period data from the raw satellite infrared cloud image data; Convert the data of the longest continuous time period from an NPY file to a PNG image; The PNG images were divided into training and testing sets in an 8:2 ratio. In the training and testing sets, an input sequence consisting of six consecutive cloud images is constructed. The input sequence is checked and missing data is filled in.

3. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, The fusion feature extraction module consists of five stacked two-dimensional convolutional layers. It independently extracts spatial features from each frame of cloud image in the input sequence. Through the hierarchical representation capability of multi-level convolution, it captures the edge texture, brightness temperature gradient and multi-scale features of cloud spatial structure of cloud image step by step, and maps the original single-channel cloud image into a semantic representation in high-dimensional feature space. Each convolutional layer has a 3×3 convolutional kernel and is filled with normalized padding; The number of channels in each convolutional layer is 32, 64, 128, 256, and the number of hidden channels are respectively. Each convolutional layer is activated by the LeakyReLU function for nonlinear transformation.

4. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, The xLSTM self-learning gating module is composed of several identical XLSTMCell units cascaded together. Each XLSTMCell unit introduces learnable gating scaling parameters on top of the standard ConvLSTM (Convolutional Long Short-Term Memory) network unit, performing channel-level adaptive scaling of the activation values ​​of the input gate, forget gate, and output gate. The XLSTMCell unit constrains the spatial locality of state updates through convolution operations, so that the cell state and hidden state maintain a two-dimensional structure in the spatial dimension. At the same time, it models time-series dependencies and spatial contextual relationships, learning the motion trajectory, morphological changes, and brightness-temperature evolution characteristics of the cloud over time.

5. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, The separation and reconstruction module consists of five stacked two-dimensional transposed convolutional layers, with the following channel mappings for each layer: first layer, hidden channel number → 256; second layer, 256 → 128; third layer, 128 → 64; fourth layer, 64 → 32; fifth layer, 32 → 1. The separation and reconstruction module progressively decodes and reconstructs the high-dimensional spatiotemporal features output by the XLSTMCell unit, performs deconvolution operations to map the feature space to the image space, and finally outputs a predicted cloud map with the same size as the input cloud map.

6. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating according to claim 1, characterized in that, The discriminator consists of 5 convolutional layers and 3 fully connected layers connected in series. The 5 convolutional layers are used to extract the authenticity features of the cloud map layer by layer, and the 3 fully connected layers output the final discrimination probability based on the authenticity features of the cloud map.

7. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, During the alternating adversarial training process: a hybrid loss function consisting of MAE and SSIM is used; the generator uses the Adam optimizer and the discriminator uses the SGD optimizer; and adaptive learning rate scheduling is used to dynamically adjust the learning rates of the generator and discriminator; at the same time, an early stopping mechanism is introduced during the adversarial training process, and training is stopped if the validation set loss does not improve after 20 rounds.

8. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating as described in claim 1, characterized in that, The alternating adversarial training also includes training five independent generator models for different cloud map prediction time requirements, with corresponding prediction time ranges of 10 minutes, 30 minutes, 1 hour, 2 hours and 3 hours respectively.

9. The satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating according to claim 1, characterized in that, The sample modes include: loading a single set of sequences, visually comparing the original cloud map with the predicted cloud map, and supporting animation playback; The data model includes: batch reading of input folders, automatic batch prediction, and saving results categorized by time.

10. A satellite cloud image infrared data extrapolation and prediction method based on GAN and xLSTM self-learning gating according to claim 1, characterized in that, The output prediction cloud map results support: automatic conversion of grayscale cloud maps to color mapping cloud maps; synthesis of GIF animations from multiple prediction results; and automatic categorization and storage of results into subfolders according to prediction duration.

Images