A cloud layer prediction method based on kupfmann constraint and frequency domain state space

Abstract

The application discloses a cloud layer prediction method based on Kupman constraint and frequency domain state space, comprising the following steps: acquiring a static meteorological satellite multi-channel high-resolution cloud image sequence and constructing a cross-season dataset; a symmetric four-stage encoder-decoder network is constructed, the bottleneck layer of which is composed of a selective state space model time series modeling path, a frequency domain adaptive gating path and a Kupman physical constraint path in parallel, the outputs of the three paths are adaptively weighted and fused, and then the decoder is used to reconstruct a pixel-level cloud layer prediction mask; an end-to-end training is carried out by using a joint loss of mask mean square error and Kupman evolution constraint; and the trained network is deployed on a ground station of a space-ground laser communication to realize real-time prediction. The application realizes long-range modeling of high-resolution cloud images with linear time complexity, suppresses multi-step prediction error accumulation through evolution constraint, and improves the cross-season generalization ability by injecting seasonal prior information, thereby providing a decision basis for dynamic scheduling of a ground station of a space-ground laser link.

Description

Technical Field

[0001] This invention belongs to the field of satellite-to-ground laser communication technology, and particularly relates to a cloud prediction method based on Koopman constraints and frequency domain state space. Background Technology

[0002] Space-to-ground laser communication, due to its high speed, large bandwidth, and strong anti-interference capabilities, has become a core technology for building future integrated space-ground information networks. However, space-to-ground laser links need to pass through the atmosphere, and the absorption and scattering of laser signals by clouds can lead to severe signal attenuation or even communication interruption. Knowing the state and movement trends of clouds above ground stations in advance can provide crucial prior information for establishing space-to-ground laser links and for the dynamic scheduling of ground stations. Therefore, accurate and stable short-term forecasting of clouds above ground stations is a core prerequisite for ensuring reliable space-to-ground laser communication network deployment.

[0003] Currently, cloud forecasting technologies mainly fall into three categories. Traditional numerical weather prediction relies on solving atmospheric physical equations, resulting in enormous computational overhead and difficulty meeting minute-level real-time requirements. Optical flow methods, based on the "freeze flow" assumption, estimate cloud movement by shifting pixel values ​​between adjacent frames. While computationally fast, they neglect cloud formation and dissipation processes, leading to significant long-term prediction error accumulation. Deep learning-based methods have become mainstream in recent years, encompassing architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers.

[0004] However, existing deep learning methods still have the following significant shortcomings:

[0005] (1) Computational complexity bottleneck: When processing high-resolution cloud image sequences, the computational complexity of the Transformer model based on the self-attention mechanism increases quadratically with the sequence length. The resulting memory overhead and inference latency make it difficult to deploy on edge devices (such as ground stations).

[0006] (2) Lack of physical laws: The pure data-driven modeling method lacks explicit representation of the physical laws of the cloud system. The model only fits the mapping relationship between input and output at the pixel level. In the process of multi-step iterative prediction, the error is amplified step by step, and the predicted cloud map becomes blurred and distorted, resulting in insufficient physical rationality.

[0007] (3) Weak cross-seasonal generalization ability: The evolution characteristics of cloud layers are significantly affected by seasonal changes and geographical differences. When the training data and test data come from different seasons or different regions, the prediction performance of existing models often deteriorates significantly and cannot adapt to seasonal distribution shifts.

[0008] In summary, existing technologies cannot simultaneously meet the requirements of real-time performance, physical rationality, and seasonal adaptability. There is an urgent need to propose a new cloud forecasting method to overcome these shortcomings. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes a cloud prediction method based on Koopman constraints and frequency domain state space. The aim is to organically integrate selective state space modeling, frequency domain adaptive gating, and Koopman physical constraints to enhance the physical rationality of the prediction results while maintaining linear computational complexity, and to improve the generalization performance of the model under different seasonal conditions. This provides accurate and stable short-term cloud prediction results for the dynamic scheduling and seamless switching of ground stations in the satellite-to-ground laser link.

[0010] The above objectives are achieved through the following technical solutions:

[0011] A cloud prediction method based on Koopman constraints and frequency domain state space includes the following steps:

[0012] S1. Obtain multi-channel high-resolution cloud image sequences from geostationary meteorological satellites, construct a dataset covering multiple climate regions by season and geographic area, and perform effective pixel filtering and standardization preprocessing;

[0013] S2. Construct a cloud prediction network, which adopts a symmetrical four-stage encoder-decoder architecture. The encoder compresses the input cloud map sequence into a multi-scale latent feature map through step-by-step downsampling. The shallow features of the multi-scale latent feature map are passed to the decoder via skip connections, and the deep compressed features of the multi-scale latent feature map are sent to the bottleneck layer. In the bottleneck layer, the deep compressed features serve as the common input of three parallel paths. The first path performs temporal modeling of the deep compressed features based on a selective state-space model. The second path performs frequency domain adaptive gating of the deep compressed features based on Fourier transform. The third path performs physical constraints on the deep compressed features based on the Koopman operator. The outputs of the three parallel paths are adaptively weighted and fused before being sent to the decoder. Together with the shallow features passed through skip connections, they are upsampled and fused step-by-step to reconstruct a pixel-level cloud prediction mask.

[0014] S3. Construct a joint loss function consisting of a weighted combination of mask mean square error loss and Koopman evolution constraint loss, and perform end-to-end training and optimization on the cloud prediction network;

[0015] S4. Deploy the trained cloud prediction network at the satellite-to-ground laser communication ground station, input the real-time satellite cloud image sequence, and output the pixel-level cloud prediction mask for future moments.

[0016] Furthermore, in the first path, the selective state-space model is described by continuous-time state-space equations: , , in For state variables, Let be the derivative of the state variable with respect to time. As input features, Here is the state transition matrix. and For the projection matrix, For feedforward parameters;

[0017] The continuous system is discretized using a zero-order hold scheme to obtain discrete parameters. and : , , in This is the discretized state transition matrix. The input projection matrix is ​​discretized, and Δ is the discretization step size, which is dynamically generated by a fully connected layer based on the input features.

[0018] The discretized recursive formula is , The computational complexity is O(N), where N is the sequence length. For the hidden state variables at the current moment, This refers to the hidden state variable from the previous time step. The output features at the current time. This represents the deep compression feature at the current moment.

[0019] Furthermore, in the second path, the input feature map... By applying a two-dimensional real fast Fourier transform to each channel, the complex spectral characteristics are obtained. Introducing a learnable complex weight matrix in the spectral space Element-wise gating is performed, and the output is then restored to the spatial domain using an inverse real fast Fourier transform. Specifically:

[0020] The complex spectrum features ,in This represents the two-dimensional real fast Fourier transform operation;

[0021] The element-by-element gating: ,in The spectral characteristics after gating. For Hadamard product, It is a non-linear activation function;

[0022] The learnable complex weight matrix Based on basic weights and month-specific modulation items Joint decision, namely: ,in This is the embedding vector for the month corresponding to the cloud map sequence. This represents a multilayer perceptron, which consists of two fully connected layers.

[0023] Furthermore, in the third pathway, the deep compression features input from the bottleneck layer are regarded as a state representation of cloud dynamics in the compression potential space, and its nonlinear evolution is expressed as follows: , This represents the deep compression characteristics at the current moment. Representing the deep compression features at the next time step; through a learnable observation function The system state is elevated to a higher-dimensional latent space, where the nonlinear evolution is characterized by a linear Koopman operator, whose action is as follows: ,in This is the matrix representation of the linear Koopman operator; in its finite-dimensional implementation, the observation function can be learned. Implemented by a fully connected layer. This is the learnable Koopman transition matrix; Koopman evolution loss is introduced. The cloud prediction network is constrained to learn an approximately linear cloud dynamics representation in the high-dimensional potential space.

[0024] Preferably, the adaptive weighted fusion is implemented through a lightweight gating network, which consists of a global average pooling layer and a fully connected layer. The lightweight gating network takes the deep compressed features as input, dynamically generates the fusion coefficients of three parallel paths, and fuses them in a weighted sum manner. The sum of the fusion coefficients of the three parallel paths is 1.

[0025] Preferably, the encoder achieves hierarchical feature extraction through four feature evolution stages. Each stage uses depthwise separable convolutional residual blocks. Spatial downsampling and channel doubling are performed between adjacent stages through stride convolutional layers. The decoder reconstructs the cloud layer step by step through four symmetrical upsampling stages. Each stage first enlarges the spatial size through bilinear interpolation and then reduces the number of channels through convolution. The shallow features of the corresponding layer of the encoding path are fused with the upsampled features by channel splicing and then the cloud boundary is refined by residual blocks.

[0026] Preferably, the joint loss function is ,in, To compensate for the mask mean square error loss, For Koopman evolutionary constraint loss, and These are the weighting coefficients.

[0027] Furthermore, the dataset contains cloud image data for multiple geographical regions with significant differences in climate characteristics, covering cloud image sequences for each month of the year; the training set and test set are divided in chronological order and have seasonal distribution shifts.

[0028] Furthermore, the process includes evaluating the prediction results: comparing the predicted cloud mask with the actual cloud mask, and evaluating the results using five metrics: accuracy, precision, recall, F1 score, and intersection-over-union (IoU). Each metric is calculated based on the confusion matrix. , , , , , TP, TN, FP, and FN correspond to the number of pixels in the actual cloud cover predicted to be cloud cover, the actual cloudless cover predicted to be cloudless, the actual cloudless cover predicted to be cloud cover, and the actual cloud cover predicted to be cloudless, respectively.

[0029] The present invention also provides a cloud prediction system for performing a cloud prediction method based on Koopman constraints and frequency domain state space, comprising:

[0030] The data acquisition module is used to acquire multi-channel cloud image sequences from geostationary meteorological satellites and perform preprocessing.

[0031] The cloud prediction network module includes a symmetrical four-stage encoder-decoder structure. Its bottleneck layer consists of a time-series modeling path based on a selective state-space model, a frequency-domain adaptive gating path based on Fourier transform, and a physical constraint path based on Koopman operators in parallel. The outputs of the three paths are decoded and output as a cloud prediction mask after adaptive weighted fusion.

[0032] The training optimization module is used to perform end-to-end training of the network using a joint loss function of mask mean square error loss and Koopman evolution constraint loss;

[0033] The application deployment module is used to deploy the trained network at the satellite-to-ground laser communication ground station, inputting real-time cloud image sequences and outputting pixel-level cloud prediction masks for future moments.

[0034] The advantages of this invention compared to the prior art are:

[0035] (1) Linear computational complexity: The selective state space model is used to replace the traditional self-attention mechanism, which reduces the sequence modeling complexity from quadratic to linear, greatly reducing the memory overhead and inference latency of high-resolution cloud image processing, and meeting the real-time requirements of ground station edge deployment.

[0036] (2) Physical constraints suppress error accumulation: The nonlinear dynamics of the cloud system are elevated to a high-dimensional potential space by using the Koopman operator and evolutionary constraints are applied to maintain the physical consistency in multi-step prediction and effectively suppress the ambiguity and morphological distortion caused by the stepwise amplification of errors.

[0037] (3) Seasonal adaptive generalization: By injecting seasonal prior information, the network can adapt to the differences in atmospheric environment in different seasons, maintain stable prediction accuracy under the training-test seasonal distribution shift, and significantly improve cross-seasonal generalization ability.

[0038] (4) Synergistic gain of the three pathways: Ablation experiments show that the accuracy gain of the three pathways is approximately additive, with little functional overlap and significant synergistic benefits. The overall performance is better than any single pathway or any combination of two pathways. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the cloud prediction scenario for satellite-to-ground laser communication according to the present invention.

[0040] Figure 2 This is a schematic diagram of the overall architecture of the cloud prediction network of this invention.

[0041] Figure 3 Example diagrams comparing the prediction results of this invention with the actual cloud map (8 sets of prediction results).

[0042] Figure 4 This is a schematic diagram of the confusion matrix described in this invention. Detailed Implementation

[0043] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0044] like Figure 1 As shown, this method is applied to satellite-to-ground laser communication scenarios. Using multi-channel cloud image sequences from geostationary meteorological satellites as input, it predicts the cloud cover over ground stations at future times, providing a basis for link switching decisions. For example... Figure 2 As shown, the cloud prediction network adopts a symmetric four-stage encoder-decoder architecture, with the bottleneck layer consisting of three parallel paths. The specific steps are as follows:

[0045] S1. Obtain multi-channel high-resolution cloud image sequences from geostationary meteorological satellites, construct datasets covering multiple climate regions by season and geographic area, and perform effective pixel filtering and standardization preprocessing.

[0046] For example, the Fengyun-4 geostationary meteorological satellite is used, with an image spatial resolution of 512×512 pixels and a temporal resolution of 15 minutes / frame. The data covers multiple geographical regions with significant differences in climate characteristics, such as Beijing, Nanjing, and Bangkok, and includes cloud image sequences for all months of the year. Valid pixel filtering removes invalid fill areas at the edges of the satellite images and bad pixels from the sensor. Standardization preprocessing normalizes the pixel values ​​of each band to the [0,1] interval. The training and test sets are divided chronologically and have seasonal distribution shifts; for example, the training set is taken from winter months, and the test set is taken from the following spring months. A sliding window sampling method is used to construct input-output sample pairs. For example, using 10 consecutive frames of cloud images as input, the cloud mask for the next 2 frames is predicted, with a sliding step size of 1 frame.

[0047] S2. Construct a cloud prediction network, which adopts a symmetrical four-stage encoder-decoder architecture. The encoder compresses the input cloud image sequence into a multi-scale latent feature map through step-by-step downsampling. The shallow features of the multi-scale latent feature map are passed to the decoder via skip connections, and the deep compressed features of the multi-scale latent feature map are sent to the bottleneck layer. In the bottleneck layer, the deep compressed features serve as the common input of three parallel paths. The first path performs temporal modeling of the deep compressed features based on a selective state-space model. The second path performs frequency domain adaptive gating of the deep compressed features based on Fourier transform. The third path applies physical constraints to the deep compressed features based on the Koopman operator. The outputs of the three parallel paths are adaptively weighted and fused before being sent to the decoder. Together with the shallow features passed through skip connections, they are upsampled and fused step-by-step to reconstruct a pixel-level cloud prediction mask.

[0048] like Figure 2 As shown, the encoder achieves hierarchical feature extraction through four feature evolution stages. Each stage employs depthwise separable convolutional residual blocks, and adjacent stages are spatially downsampled and the number of channels is doubled through stride convolutional layers. For example, the encoder first maps the input cloud image frame to a high-dimensional feature space using overlapping convolutional patch embedding. The patch size is 4×4 pixels, and the convolution stride is 2. After four stages, the spatial resolution is halved step by step (512→256→128→64→32 pixels). The decoder reconstructs the cloud layer through four symmetrical upsampling stages. Each stage first enlarges the spatial size through bilinear interpolation, then reduces the number of channels through convolution. The shallow features corresponding to the encoding path are fused with the upsampled features through channel concatenation, and then the cloud boundaries are refined through residual blocks. The resolution is gradually restored from 32 pixels to 512 pixels, and a pixel-level cloud prediction mask is output by a 1×1 convolution at the end.

[0049] The first path is a selective state-space temporal modeling path. The selective state-space model is described by continuous-time state-space equations: , , in For state variables, Let be the derivative of the state variable with respect to time. As input features, Here is the state transition matrix. and For the projection matrix, For feedforward parameters;

[0050] The continuous system is discretized using a zero-order hold scheme to obtain discrete parameters. and : , , in This is the discretized state transition matrix. The input projection matrix is ​​discretized, and Δ is the discretization step size, which is dynamically generated by a fully connected layer based on the input features.

[0051] The discretized recursive formula is , The computational complexity is O(N), where N is the sequence length. For the hidden state variables at the current moment, This refers to the hidden state variable from the previous time step. The output features at the current time. This represents the deep compression feature at the current moment. The feature output by this pathway is denoted as... .

[0052] The second path is a frequency-domain adaptive gated path. For the input feature map... By applying a two-dimensional real fast Fourier transform to each channel, the complex spectral characteristics are obtained. Introducing a learnable complex weight matrix in the spectral space Element-wise gating is performed, and the output is then restored to the spatial domain using an inverse real fast Fourier transform. Specifically:

[0053] The complex spectrum features ,in This represents the two-dimensional real fast Fourier transform operation;

[0054] The element-by-element gating: ,in The spectral characteristics after gating. For Hadamard product, It is a non-linear activation function;

[0055] The learnable complex weight matrix Based on basic weights and month-specific modulation items Joint decision, namely: ,in This is the embedding vector for the month corresponding to the cloud map sequence. This represents a multilayer perceptron (MLP), which consists of two fully connected layers. The feature output by this pathway is denoted as... .

[0056] The third pathway is the Koopman physical constraint pathway. The deep compression characteristics of the bottleneck layer input are considered as a state representation of cloud dynamics in the compression potential space, and their nonlinear evolution is expressed as follows: , This represents the deep compression characteristics at the current moment. Representing the deep compression features of the next time step; through a learnable observation function The system state is elevated to a higher-dimensional latent space, where the nonlinear evolution is characterized by a linear Koopman operator, whose action is as follows: ,in This is the matrix representation of the linear Koopman operator; in its finite-dimensional implementation, the observation function can be learned. Implemented by a fully connected layer. This is the learnable Koopman transition matrix; Koopman evolution loss is introduced. The cloud prediction network is constrained to learn an approximately linear cloud dynamics representation in the high-dimensional latent space, and this loss is used as part of the joint loss function during training. The features output by this pathway are denoted as... .

[0057] Characteristics of the output of the three paths , and The data is integrated by an adaptive weighted fusion module. This module is a lightweight gated network consisting of a global average pooling layer and a fully connected layer. It takes the deep compressed features as input and outputs normalized fusion coefficients via a Softmax activation function. , , ( + + =1), merged using a weighted sum method: ,in This indicates channel-wise multiplication. Features after fusion. The data is fed into the decoder, which reconstructs the cloud layer through four symmetrical upsampling stages: each stage first enlarges the spatial size through bilinear interpolation, then reduces the number of channels through convolutional layers. Shallow features corresponding to the encoded path are fused with the upsampled features via channel concatenation, then refined by depth-separable convolutional residual blocks. Finally, a 1×1 convolutional layer outputs a pixel-level cloud prediction mask. For example, the gated network outputs normalized fusion coefficients via a Softmax activation function at the end.

[0058] S3. Construct a joint loss function consisting of a weighted combination of mask mean square error loss and Koopman evolution constraint loss, and perform end-to-end training and optimization on the cloud prediction network.

[0059] The joint loss function is: ,in, To compensate for the mask mean square error loss, For Koopman evolutionary constraint loss, and These are weighting coefficients. For example, A binary mask is used to eliminate interference from invalid edge regions and boundary registration artifacts in satellite images. In the binary mask, valid pixel positions are set to 1, and invalid regions are set to 0. The Adam optimizer is used during training, with an initial learning rate of... Train for 50 epochs, with a batch size of, for example, 4 or 8, and save the model weights with the lowest validation loss.

[0060] After training, the prediction results are evaluated: the predicted cloud mask is compared with the actual cloud mask, and the evaluation is based on five metrics: accuracy, precision, recall, F1 score, and intersection-over-union (IoU). Each metric is calculated based on the confusion matrix. Figure 4 As shown, the four quadrants of the confusion matrix correspond to the number of pixels in TP (actually cloudy, predicted cloudy), TN (actually cloudless, predicted cloudless), FP (actually cloudless, predicted cloudy), and FN (actually cloudy, predicted cloudless), respectively. , , , , ,

[0061] like Figure 3As shown in the figure, the comparison between cloud prediction results and actual cloud images is presented at eight consecutive time points, with each group spaced 15 minutes apart, covering a total prediction window of two hours. Each comparison includes two columns of images: the cloud mask predicted by this method and the actual cloud mask at the corresponding time point. Figure 3 It can be seen that during the continuous prediction process spanning 2 hours, the cloud prediction mask of this method maintains a high degree of consistency with the real mask in terms of the main position of the cloud area, the shape of the cloud cluster, and the boundary contour. The tracking of cloud movement, deformation, and formation and dissipation processes is continuous and stable, and no obvious blurring or shape distortion occurs as the prediction step length increases.

[0062] S4. Deploy the trained cloud prediction network at the satellite-to-ground laser communication ground station, input the real-time satellite cloud image sequence, and output the pixel-level cloud prediction mask for future moments.

[0063] For example, during deployment, the network operates in inference mode on the ground station computing platform, receiving continuous cloud image sequences transmitted from satellites in real time. These sequences are buffered according to a preset input window (e.g., 10 frames) and then sent to the network. After one forward inference operation, a pixel-level cloud prediction mask for future moments (e.g., 30 minutes in the future) is output. The ground station scheduling system uses this cloud prediction mask to determine the cloud cover status above the station—if the cloud coverage exceeds a preset threshold, the communication link is switched to a backup ground station in advance; if it is below the threshold, the current link is maintained, achieving intelligent proactive link scheduling.

[0064] Correspondingly, a cloud prediction system for executing the above method includes: a data acquisition module for acquiring and preprocessing multi-channel cloud image sequences from geostationary meteorological satellites; a cloud prediction network module, including a symmetrical four-stage encoder-decoder structure, the bottleneck layer of which is composed of a time-series modeling path based on a selective state-space model, a frequency-domain adaptive gating path based on Fourier transform, and a physical constraint path based on the Koopman operator in parallel, and the outputs of the three paths are decoded and output as a cloud prediction mask after adaptive weighted fusion; a training and optimization module for end-to-end training of the network using a joint loss function of mask mean square error loss and Koopman evolution constraint loss; and a deployment and application module for deploying the trained network at a satellite-to-ground laser communication ground station, inputting real-time cloud image sequences and outputting pixel-level cloud prediction masks for future times.

[0065] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cloud prediction method based on Koopman constraints and frequency domain state space, characterized in that, Includes the following steps: S1. Obtain multi-channel high-resolution cloud image sequences from geostationary meteorological satellites, construct a dataset covering multiple climate regions by season and geographic area, and perform effective pixel filtering and standardization preprocessing; S2. Construct a cloud prediction network, which adopts a symmetrical four-stage encoder-decoder architecture. The encoder compresses the input cloud map sequence into a multi-scale latent feature map through step-by-step downsampling. The shallow features of the multi-scale latent feature map are passed to the decoder via skip connections, and the deep compressed features of the multi-scale latent feature map are sent to the bottleneck layer. In the bottleneck layer, the deep compressed features serve as the common input of three parallel paths. The first path performs temporal modeling of the deep compressed features based on a selective state-space model. The second path performs frequency domain adaptive gating of the deep compressed features based on Fourier transform. The third path performs physical constraints on the deep compressed features based on the Koopman operator. The outputs of the three parallel paths are adaptively weighted and fused before being sent to the decoder. Together with the shallow features passed through skip connections, they are upsampled and fused step-by-step to reconstruct a pixel-level cloud prediction mask. S3. Construct a joint loss function consisting of a weighted combination of mask mean square error loss and Koopman evolution constraint loss, and perform end-to-end training and optimization on the cloud prediction network; S4. Deploy the trained cloud prediction network at the satellite-to-ground laser communication ground station, input the real-time satellite cloud image sequence, and output the pixel-level cloud prediction mask for future moments.

2. The method according to claim 1, characterized in that, In the first path, the selective state-space model is described by continuous-time state-space equations: ， ， in For state variables, Let be the derivative of the state variable with respect to time. For input features, Here is the state transition matrix. and Let be the projection matrix. For feedforward parameters; The continuous system is discretized using a zero-order hold scheme to obtain discrete parameters. and : ， ， in This is the discretized state transition matrix. The input projection matrix is ​​discretized, and Δ is the discretization step size, which is dynamically generated by a fully connected layer based on the input features. The discretized recursive formula is , The computational complexity is O(N), where N is the sequence length. For the hidden state variables at the current moment, This refers to the hidden state variable from the previous time step. The output features at the current time. This represents the deep compression feature at the current moment.

3. The method according to claim 1, characterized in that, In the second path, the input feature map By applying a two-dimensional real fast Fourier transform to each channel, the complex spectral characteristics are obtained. Introducing a learnable complex weight matrix in the spectral space Element-wise gating is performed, and the output is then restored to the spatial domain using an inverse real fast Fourier transform. Specifically: The complex spectrum features ,in This represents the two-dimensional real fast Fourier transform operation; The element-by-element gating: ,in The spectral characteristics after gating. For Hadamard product, It is a non-linear activation function; The learnable complex weight matrix Based on basic weights and month-specific modulation items Joint decision, namely: ,in This is the embedding vector for the month corresponding to the cloud map sequence. This represents a multilayer perceptron, which consists of two fully connected layers.

4. The method according to claim 1, characterized in that, In the third pathway, the deep compression features input from the bottleneck layer are regarded as a state representation of cloud dynamics in the compression potential space, and its nonlinear evolution is expressed as follows: , This represents the deep compression characteristics at the current moment. This indicates the deep compression characteristics at the next moment; Through learnable observation functions The system state is elevated to a higher-dimensional latent space, where the nonlinear evolution is characterized by a linear Koopman operator, whose action is as follows: ,in For the matrix representation of the linear Koopman operator; in the finite-dimensional implementation, the observation function can be learned. Implemented by a fully connected layer. This is the learnable Koopman transition matrix; Koopman evolution loss is introduced. The cloud prediction network is constrained to learn an approximately linear cloud dynamics representation in the high-dimensional potential space.

5. The method according to claim 1, characterized in that, The adaptive weighted fusion is implemented through a lightweight gating network, which consists of a global average pooling layer and a fully connected layer. The lightweight gating network takes the deep compressed features as input, dynamically generates the fusion coefficients of three parallel paths, and fuses them in a weighted sum manner. The sum of the fusion coefficients of the three parallel paths is 1.

6. The method according to claim 1, characterized in that, The encoder achieves hierarchical feature extraction through four feature evolution stages. Each stage uses depthwise separable convolutional residual blocks. Spatial downsampling and channel doubling are performed between adjacent stages through stride convolutional layers. The decoder reconstructs the cloud layer step by step through four symmetrical upsampling stages. Each stage first enlarges the spatial size through bilinear interpolation and then reduces the number of channels through convolution. The shallow features of the corresponding layer of the encoding path are fused with the upsampled features by channel splicing and then the cloud boundary is refined by residual blocks.

7. The method according to claim 1, characterized in that, The joint loss function is: ,in, To account for the mask mean square error loss, For Koopman evolutionary constraint loss, and These are the weighting coefficients.

8. The method according to claim 1, characterized in that, The dataset contains cloud image data for multiple geographical regions with significant differences in climate characteristics, covering cloud image sequences for each month of the year; the training set and test set are divided in chronological order and have seasonal distribution shifts.

9. The method according to claim 1, characterized in that, It also includes a step of evaluating the prediction results: comparing the predicted cloud mask with the actual cloud mask, and evaluating the results using five metrics: accuracy, precision, recall, F1 score, and intersection-over-union (IoU). Each metric is calculated based on the confusion matrix. ， ， ， ， ， TP, TN, FP, and FN correspond to the number of pixels in the actual cloud cover predicted to be cloud cover, the actual cloudless cover predicted to be cloudless, the actual cloudless cover predicted to be cloud cover, and the actual cloud cover predicted to be cloudless, respectively.

10. A cloud prediction system employing the method described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire multi-channel cloud image sequences from geostationary meteorological satellites and perform preprocessing. The cloud prediction network module includes a symmetrical four-stage encoder-decoder structure. Its bottleneck layer consists of a time-series modeling path based on a selective state-space model, a frequency-domain adaptive gating path based on Fourier transform, and a physical constraint path based on Koopman operators in parallel. The outputs of the three paths are decoded and output as a cloud prediction mask after adaptive weighted fusion. The training optimization module is used to perform end-to-end training of the network using a joint loss function of mask mean square error loss and Koopman evolution constraint loss; The application deployment module is used to deploy the trained network at the satellite-to-ground laser communication ground station, inputting real-time cloud image sequences and outputting pixel-level cloud prediction masks for future moments.

Images