Meteorological satellite cloud image extrapolation method based on geographical attention mechanism

Abstract

The application discloses a meteorological satellite cloud map extrapolation method based on a geographical attention mechanism, characterized by a deep learning framework constructed by GeoAttX, the introduction of a geographical attention mechanism to enhance the modeling ability of spatial features, and the synergistic effect of geographical semantic modeling, prediction path optimization and distribution correction mechanism, using the observation data of meteorological satellites to realize high-precision and low-error accumulation cloud prediction. The method uses a reverse greedy algorithm to realize flexible forecast time by reducing the autoregressive step, and proposes a second-order standardization method to normalize the pixel distribution, to maintain the consistency of brightness statistics and reduce error propagation. Compared with the prior art, the application has the advantages of using real-time stationary meteorological satellite observation data, having error suppression capability, geographical structure modeling capability, long-term stable prediction capability, high precision and stability, and can effectively improve the cloud motion extrapolation capability, and has important application value for weather forecasting and disaster warning.

Description

Technical Field

[0001] This invention relates to the fields of satellite remote sensing and meteorological and hydrological technology, and in particular to a meteorological satellite cloud image extrapolation method based on geographic attention mechanism for high spatiotemporal resolution geostationary meteorological satellite data. Background Technology

[0002] Geostationary meteorological satellites provide continuous observational data with high temporal resolution, making them a crucial data source for short-term weather forecasting. Currently, numerical weather prediction relies on solving complex physical equations, resulting in high computational costs and difficulty in meeting real-time operational needs. Existing artificial intelligence models largely depend on reanalysis data, which suffers from data acquisition delays and is difficult to use directly for real-time forecasting. In multi-step autoregressive forecasting, errors gradually accumulate, leading to a significant decrease in forecast accuracy. In long-term forecasts, the statistical distribution of images shifts, manifesting as extreme value compression and image blurring, affecting the usability of the results.

[0003] In summary, existing meteorological satellite cloud image extrapolation methods suffer from the following drawbacks: 1) Real-time observation-driven methods rely on reanalysis data or complex physical solution processes, making it difficult to effectively characterize the spatial continuity and evolution of cloud systems, and failing to meet the operational requirements for high timeliness in short-term weather forecasts; 2) Error accumulation inhibits the stability and reliability of long-term forecasts. Therefore, it is necessary to propose a cloud image extrapolation method that can directly utilize real-time geostationary meteorological satellite observation data, while possessing error suppression capabilities, geographic structure modeling capabilities, and long-term stable forecasting capabilities. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a meteorological satellite cloud image extrapolation method based on a geographic attention mechanism. This method utilizes the synergistic effect of geographic semantic modeling, prediction path optimization, and distribution correction mechanisms to achieve high-precision cloud image prediction with low error accumulation. This method uses continuous observation data from geostationary meteorological satellites for modeling, eliminating the need for reanalysis data or complex physical solution processes. Data acquisition latency can be controlled within 15 minutes, and single-step inference time can be controlled within seconds, significantly improving the stability and reliability of medium- and long-term forecasts. This greatly meets the operational requirements for high timeliness in short-term weather forecasting. Furthermore, the model exhibits good inference efficiency and scalability in high-resolution cloud image prediction tasks, providing data support for meteorological early warning, flood control, agricultural management, and other fields, demonstrating promising application prospects and commercial development value.

[0005] The specific technical solution to achieve the purpose of this invention is: a meteorological satellite cloud image extrapolation method based on the geographic attention mechanism, which includes the following steps:

[0006] (1) Data preprocessing

[0007] Acquire multi-band observation data from meteorological satellites, perform geometric correction on the original full-disk image and project it onto a unified coordinate system, select the target area and resample it into a fixed-size image sequence. The multi-band observation data from meteorological satellites is geostationary meteorological satellite observation data. Long-wave infrared bands are selected for modeling to achieve continuous all-weather observation.

[0008] (2) Multi-frame input construction

[0009] Satellite observation frames at consecutive time points are selected as input sequences to capture the temporal evolution characteristics of atmospheric motion;

[0010] (3) Feature extraction and dimensionality reduction

[0011] The embedding module performs spatiotemporal feature fusion on multi-band data, and the head module performs downsampling to reduce resolution and suppress noise. The embedding module adopts a structure combining three-dimensional convolution and two-dimensional convolution to perform joint feature encoding on multi-band and multi-temporal data, thereby obtaining a multi-channel feature tensor with spatiotemporal semantic expression capabilities.

[0012] (4) Geographic semantic modeling

[0013] The multi-channel feature tensor obtained in step (3) is input into the geographic attention module. Through long-range attention calculation in the latitude and longitude directions, semantic information extraction across spatial scales is realized. The geographic attention module adopts a long-range attention mechanism to decompose the two-dimensional spatial attention into a series calculation process of longitude direction attention and latitude direction attention in order to reduce computational complexity and realize global spatial dependency modeling.

[0014] (5) Predicted path optimization

[0015] The target prediction timeliness is decomposed based on the reverse greedy algorithm, and the multi-timescale model is called to perform autoregressive prediction.

[0016] (6) Distribution correction

[0017] A second-order normalization algorithm is used to perform statistical characteristic alignment distribution correction on the autoregressive prediction results to maintain the consistency of image distribution. The second-order normalization algorithm dynamically updates statistical parameters during the multi-step prediction process to maintain the image brightness distribution and extreme value characteristics.

[0018] (7) Result reconstruction

[0019] Spatial resolution is restored by upsampling and channel mapping, and a predicted cloud map is output. The upsampling uses a sub-pixel rearrangement method to spatially reconstruct the features in order to restore the original resolution.

[0020] (8) Model training and prediction output

[0021] The GeoAttX model is trained using historical meteorological satellite observation data to construct training samples. The training samples consist of a series of multi-band cloud images at consecutive time points. The cloud images of the previous and current time points are used as input, and the cloud images of the corresponding next time point are used as supervision labels to optimize the model parameters. In the inference phase, the cloud image data of the current time point and the previous time point are input into the trained model. Multi-step iterative inference is performed by combining a single-step prediction model with a reverse greedy strategy to gradually generate cloud image results with the target prediction timeliness. The GeoAttX model is a deep learning network structure that includes an embedding module, a head module, a backbone network composed of geographic attention modules, an upsampling module, an upsampling module, and a tail module.

[0022] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0023] 1) High-timeliness forecasting capability driven by real-time observation: This invention directly uses continuous observation data from geostationary meteorological satellites for modeling, without relying on reanalysis data or complex physical solution processes. The data acquisition delay can be controlled within 15 minutes, and the model's single-step inference time can be controlled within seconds. Compared with traditional numerical forecasting methods, which usually require several hours of data preparation and calculation, the overall timeliness is improved by more than an order of magnitude, which can meet the operational requirements for high timeliness in short-term weather forecasting. Compared with existing CNN or standard Vision Transformer models, this invention significantly enhances the ability to jointly model large-scale spatial migration patterns and local strong convective structures.

[0024] 2) Improved Long-Term Prediction Stability through Error Accumulation Suppression: By optimizing the prediction path using a reverse greedy algorithm, the traditional frame-by-frame prediction (e.g., 9 steps for 135 minutes) is reduced to approximately 3 steps, a reduction of about 60%–70% in the number of prediction steps, thus suppressing the cascading amplification of errors at the source. In long-term prediction (≥120 minutes), compared to traditional methods, this invention reduces the decrease in the Structural Similarity in Simplification (SSIM) index by about 30%–50%, significantly improving the stability and reliability of medium- to long-term predictions. Compared to current mainstream frame-by-frame recursive prediction methods, this invention significantly reduces the error propagation chain length and improves the stability of medium- to long-term predictions.

[0025] 3) Long-distance spatial dependency modeling capability of geographic perception: By decomposing two-dimensional spatial attention into long-range attention calculations in the longitude and latitude directions, the computational complexity of traditional global attention is reduced from O(N²) to O(2N√N), significantly reducing computational overhead while achieving cross-regional information interaction. Experimental results show that, compared with traditional convolutional models, this invention improves the SSIM index by approximately 0.02–0.04 in large-scale cloud structure modeling, and can more effectively characterize the spatial continuity and evolution characteristics of cloud systems.

[0026] 4) Multi-scale feature collaborative representation capability: Based on a multi-head attention mechanism, the global background field, local structure, and high-value regions are modeled in parallel, improving the ability to depict details in complex cloud scenarios. Compared with traditional convolutional networks such as U-Net, this invention improves PSNR by about 1–2 dB and reduces mean squared error (MSE) by about 10%–20% in short-term prediction, significantly improving the accuracy of local structure prediction while maintaining overall spatial consistency.

[0027] 5) Image quality preservation under statistical distribution constraints: The prediction results are dynamically corrected using a second-order normalization method, ensuring that the output remains consistent with the input in terms of mean and variance. In long-term prediction, compared to models without this method, the brightness shift of the predicted image is reduced by approximately 20%–40%, and the extreme value compression phenomenon is significantly weakened, thereby effectively suppressing the brightness drift problem and improving the stability of the prediction results.

[0028] 6) Computational efficiency advantages in high-resolution scenarios: By adopting the latitude and longitude decomposition attention calculation method, under the condition of 800×800 high-resolution input, the computation amount is reduced by more than 50% compared with the traditional global self-attention model, the memory usage is significantly reduced, while maintaining the characteristics of a large receptive field, so that the model still has good inference efficiency and scalability in high-resolution cloud map prediction tasks. Attached Figure Description

[0029] Figure 1 A schematic diagram of the overall structure of the meteorological satellite cloud image data preprocessing and GeoAttX cloud image extrapolation model;

[0030] Figure 2 This is a schematic diagram of the cubic embedding module, head module, and tail module structure in the GeoAttX model.

[0031] Figure 3 A schematic diagram illustrating the structure and visualization of the geographic attention mechanism in the GeoAB module;

[0032] Figure 4 This is a diagram comparing the reverse greedy prediction strategy and predictions with different time lengths.

[0033] Figure 5 This is a schematic diagram of the second-order normalization method. Detailed Implementation

[0034] The following detailed implementation further elaborates on the steps described above, including data preprocessing, multi-frame input construction, feature extraction and dimensionality reduction, geographic semantic modeling, prediction path optimization, distribution correction, result reconstruction, and model training and prediction output. This embodiment is only for explaining the present invention and should not be construed as limiting the scope of protection of the present invention.

[0035] This invention includes the following steps:

[0036] (1) Data preprocessing

[0037] Acquire multi-band observation data from meteorological satellites, perform geometric correction, projection, cropping, and resampling on the original full-disk imagery, and form a fixed-size multi-band cloud image sequence.

[0038] (2) Multi-frame input construction

[0039] Satellite observation frames at consecutive time points are selected as input sequences to capture the temporal evolution characteristics of atmospheric motion.

[0040] (3) Feature extraction and dimensionality reduction

[0041] The spatiotemporal features of multi-band observation data are fused by the embedding module and downsampled by the head module to obtain a multi-channel feature tensor with spatiotemporal semantic expression capabilities.

[0042] (4) Geographic semantic modeling

[0043] By decomposing two-dimensional spatial attention into a series of computational processes along the longitude and latitude directions, long-distance spatial dependency modeling is achieved, reducing computational complexity while maintaining the receptive field. The geographic attention mechanism includes longitude-direction attention computation units and latitude-direction attention computation units. The longitude-direction attention computation unit establishes pixel dependencies along the same latitude direction, while the latitude-direction attention computation unit establishes pixel dependencies along the same longitude direction. These two units are connected in series to achieve global spatial dependency modeling. Simultaneously, a multi-head attention structure is introduced to perform parallel modeling of cloud structures at different scales, enabling a hierarchical representation of complex atmospheric structures.

[0044] (5) Predicted path optimization

[0045] By decomposing the target prediction timeline into a combination of multiple different time intervals and prioritizing the use of models with larger time steps for prediction, the number of autoregressive iterations is reduced, thereby decreasing error accumulation. The reverse greedy prediction strategy decomposes the data in reverse from the target prediction time to the reporting time, forming a prediction sequence that satisfies the conditions for obtaining the input frame during forward inference.

[0046] (6) Distribution correction

[0047] This invention proposes a second-order normalization method to dynamically constrain the statistical distribution of the prediction results, thereby stabilizing the output image in terms of mean and variance. This method combines global statistics from the training data with dynamic statistics from the input cloud map during the prediction stage to constrain distribution drift in the autoregressive prediction process, effectively suppressing brightness drift, extreme value compression, and contrast degradation in long-term prediction.

[0048] (7) Result reconstruction

[0049] Spatial resolution is restored by upsampling and channel mapping, and a predicted cloud map is output. The upsampling uses a sub-pixel rearrangement method to spatially reconstruct the features in order to restore the original resolution.

[0050] (8) Model training and prediction output

[0051] Training samples are constructed using historical meteorological satellite observation data to train the GeoAttX model. During the inference phase, cloud image data of the current moment and the previous moment are input into the trained model. Multi-step iterative inference is performed by combining a single-step prediction model with a reverse greedy strategy to gradually generate cloud image results with the target prediction timeliness.

[0052] Example 1

[0053] See Figure 1 This embodiment uses FY-4B geostationary meteorological satellite cloud image data to achieve cloud image extrapolation prediction by following the steps of data preprocessing, multi-frame input construction, feature extraction and dimensionality reduction, geographic semantic modeling, prediction path optimization, distribution correction, result reconstruction, and model training and prediction output. The specific steps are as follows:

[0054] (1) Data preprocessing

[0055] 1-1: Acquire multispectral remote sensing data from the FY-4B geostationary meteorological satellite AGRI (Advanced Geostationary Radiation Imager). The AGRI sensor contains 15 spectral bands with a spatial resolution of 0.5 km to 4 km and a temporal resolution of 15 minutes. This invention selects 7 long-wave infrared bands as model input data for time-series extrapolation modeling of cloud images.

[0056] 1-2: Extract Digital Number (DN) values ​​from the raw AGRI remote sensing imagery. The DN values ​​range from 0 to 4095. The DN values ​​serve as a discrete representation of the raw radiometric information and are directly input into subsequent models.

[0057] 1-3: The original FY-4B full-disk imagery is recalibrated. This recalibration process includes geometric correction and projection onto a unified spatial reference to obtain full-disk cloud imagery data under the unified spatial reference. The target study area is then spatially cropped to form a unified 7×800×800 multi-band cloud imagery for model input. The cropped area prioritizes covering the target study area to reduce computational complexity and maintain spatial consistency.

[0058] 1-4: Data Standardization. To ensure model input consistency, the original DN values ​​are directly used as input features without radiometric scaling to avoid information loss and maintain end-to-end learning characteristics. Simultaneously, distribution constraints are applied during model training and prediction using a subsequent second-order standardization method.

[0059] (2) Multi-frame input construction

[0060] 2-1: Construct a time-series input sample using remote sensing images from two consecutive time points, specifically including: the current time point cloud image. The previous time-series cloud image is X_(t-Δt); where Δt represents the time interval, with a minimum time resolution of 15 minutes. It should be noted that the current time-series cloud image and the previous time-series cloud image are not limited to fixed 15-minute consecutive frames; non-equidistant time-series samples can be constructed based on data availability.

[0061] (3) Feature extraction and dimensionality reduction

[0062] 3-1: See Figure 1 and Figure 2 The GeoAttX cloud map extrapolation model includes, in sequence according to the data flow direction, an input layer, a cube embedding module, a head module, a geographic attention backbone network composed of multiple GeoAB modules, an upsampling module, a tail module, and an output layer; wherein, the cube embedding module corresponds to the Cube Embedding module, the head module corresponds to the HeadBlock, the tail module corresponds to the Tail Block, and the GeoAB module is a specific implementation of the geographic attention module.

[0063] In terms of model connectivity, the continuous time-series cloud map data first form a dual-temporal input. The current time-series cloud map and the previous time-series cloud map are jointly input into the cubic embedding module, and the initial feature tensor is obtained through joint encoding of the time dimension and the channel dimension. At the same time, the other input head module of the current time-series cloud map is processed by convolution, normalization, nonlinear activation and downsampling to obtain low-resolution spatial features, which are used to retain the current input information in the back end of the network through residual connections.

[0064] The initial feature tensor is input into a geographic attention backbone network composed of multiple GeoAB modules. This geographic attention backbone network comprises three stages: the first stage includes four GeoAB modules for extracting spatially dependent features at the original backbone scale; the output features of the first stage are then input into the second stage after the spatial resolution is reduced by a downsampling module, and the second stage also includes four GeoAB modules for extracting spatial semantic features with a larger receptive field at a low-resolution scale; the output features of the second stage are then input into the third stage after the spatial scale is restored by an upsampling module, and the third stage also includes four GeoAB modules for fusing multi-scale features and further extracting long-range spatial dependency information.

[0065] The features processed by the GeoAB module in the third stage are input into the tail module. The tail module performs channel mapping and spatial reconstruction on the output features of the backbone network, and performs residual connection or feature fusion with the low-resolution spatial features output by the head module to retain the input information of the current time-stack cloud map. Finally, the original spatial resolution is restored through sub-pixel rearrangement, and the predicted cloud map at the target time is output. Among them, the cubic embedding module is used to encode the multi-band cloud map data of continuous time into feature tensors with spatiotemporal semantic expression capabilities; the head module is used to extract the spatial features of the current time-stack cloud map; the GeoAB module is used to extract long-range spatial dependency features through longitude and latitude attention calculation units; the upsampling and downsampling modules are used to realize multi-scale feature expression; and the tail module is used to map the upsampled features to the output channels of the predicted cloud map.

[0066] 3-2: See Figure 2 (a) The cubic embedding module structure is shown in the figure. This cubic embedding module is used to fuse and encode the dual-temporal input consisting of the current time-phase cloud map X_t and the previous time-phase cloud map X_{t-Δt}. The size of both input cloud maps is 7×800×800, where 7 represents the seven selected infrared bands. The cubic embedding module sequentially uses three-dimensional convolution, batch normalization, SiLU activation function, and two-dimensional convolution for feature transformation; the kernel size of the three-dimensional convolution is (2, 2, 2), and the stride is (2, 2, 2); the kernel size of the two-dimensional convolution is (3, 3), and the stride is (1, 1), which are used to form the initial fused features in the time dimension, band dimension, and spatial dimension.

[0067] 3-3: See also Figure 2 (b) The schematic diagram of the head module structure is shown in the figure. This head module is used to extract spatial features from the current time-state cloud map X_t. The head module sequentially uses two-dimensional convolution, batch normalization, SiLU activation function, and two-dimensional convolution to perform feature transformation; wherein the kernel size of the first two-dimensional convolution is (3,3) and the stride is (2,2), and the kernel size of the second two-dimensional convolution is (3,3) and the stride is (1,1), which is used to extract the spatial features of the current time-state cloud map while reducing the spatial scale.

[0068] See Figure 2 (c) The tail module structure is shown in the figure. The tail module sequentially uses two-dimensional convolution, batch normalization, SiLU activation function, two-dimensional convolution and PixelShuffle sub-pixel rearrangement operation to map the output features of the backbone network and restore the predicted image with the same number of bands as the input cloud image. Figure 2In this context, Conv3D represents 3D convolution, Conv2D represents 2D convolution, BatchNorm represents batch normalization, SiLU represents activation function, PixelShuffle represents subpixel rearrangement, H and W represent feature map height and width, and C represents the number of hidden channels in the network.

[0069] 3-4: Input the initial feature tensor extracted in 3-2 into the backbone network composed of multiple GeoAB modules. The cloud map extrapolation model has a total of 12 GeoAB modules, and multi-scale feature representation is achieved through downsampling and upsampling at different stages. The specific geographic attention calculation process of the GeoAB module is shown in step (4).

[0070] (4) Geographic semantic modeling

[0071] 4-1: See Figure 3 The GeoAB module is the geographic attention module in the GeoAttX model, used to extract geographic semantic information from cloud map features.

[0072] See Figure 3 (a) This figure shows the decomposition calculation of longitude direction attention and latitude direction attention. Each pixel or region feature corresponds to a geographical location determined by latitude and longitude. The geographic attention mechanism adopts the idea of ​​long-range attention, decomposing the two-dimensional spatial attention into two consecutive steps: first, calculate the longitude direction attention at the same latitude, and then calculate the latitude direction attention at the same longitude.

[0073] 4-2: Longitude Direction Attention Calculation. For any region A_{i,j} in the feature map, where i represents the row position and j represents the column position, first calculate its relationship with other regions A_{i,m} in the same latitude direction to obtain the longitude direction aggregation result A′_{i,j}, calculated as follows:

[0074] .

[0075] Where A_{i,j} represents the region feature in the i-th row and j-th column of the feature map; m represents the column index in the same row; M represents the number of columns in the feature map; LonAtt(·) represents the longitude direction attention function, used to represent the contribution of A_{i,j} to the relationship between A_{i,j} and A_{i,m} in the same latitude row; A′_{i,j} represents the intermediate feature after aggregating longitude-related information in the same latitude direction. Through this step, the target region can aggregate cloud semantic information in the same latitude direction.

[0076] 4-3: Latitude Direction Attention Calculation. After obtaining the longitude direction aggregation result A′_{i,j}, the relationship between it and other regions A′_{n,j} in the same longitude direction is calculated to obtain the spatial attention feature A″_{i,j} that simultaneously contains longitude and latitude dependencies. The calculation formula is:

[0077] .

[0078] Where n represents the row index in the same column; N represents the number of rows in the feature map; LatAtt(·) represents the latitudinal attention function, used to represent the contribution of A′_{i,j} to the relationship between A′_{n,j} in the same longitude column; A″_{i,j} represents the spatial attention feature after the concatenated attention calculation in the longitude and latitudinal directions. This feature integrates spatial dependency information in the same latitude and longitude directions, and is used for further feature extraction in subsequent GeoAB modules or backbone network stages.

[0079] 4-4: The A″_{i,j} mentioned above is not a standalone attention map for display, but rather a spatial dependency enhancement feature output by the GeoAB module. For each region A_{i,j}, spatial attention information is extracted through consecutive steps of longitude and latitude attention, and the extracted A″_{i,j} is used as the geographic semantic feature representation of that region, thereby enabling the target pixel to simultaneously capture long-distance spatial information in both longitude and latitude directions.

[0080] 4-5: See reference Figure 3 (b) The GeoAB module adopts a multi-head attention structure. Figure 3 (b) This figure shows the visualization results of the four attention heads in the GeoAttX_60 model, each attention head corresponding to a different attention map; a to g in the figure are seven reference points, based on Figure 3 (a) A coordinate system with the origin at the bottom left corner, with positions ranging from (100, 700) to (700, 100); red indicates high-attention areas, blue indicates the current cloud map outline, and green indicates the cloud map outline one hour later. Different attention heads reflect different attention patterns: Head 1 focuses on a larger range of global information, Head 3 is more focused on the local area near the reference point, Head 2 mainly focuses on high-value cloud areas, and Head 4 focuses more on low-value areas. Through the multi-head structure, the GeoAB module can express the global structure, local details, and intensity features of different cloud areas in parallel within the same network layer.

[0081] (5) Predicted path optimization

[0082] See Figure 4To achieve long-term, multi-step cloud map extrapolation prediction, this invention proposes a reverse greedy prediction strategy, abbreviated as RGA, which combines a multi-timescale model to achieve autoregressive prediction. The specific steps are as follows:

[0083] 5-1: Multi-timescale model construction, which involves building multiple cloud image extrapolation models with different time resolutions to form a model set M, including: GeoAttX-15 (time step of 15 minutes); GeoAttX-60 (time step of 60 minutes); GeoAttX-180 (time step of 180 minutes). Each model is used as input for current and historical cloud images and outputs prediction results for the corresponding time step.

[0084] 5-2: Decomposition of Prediction Duration. For the target prediction duration T, it is decomposed into the sum of multiple sub-time steps represented by the following formula:

[0085] .

[0086] Where: T is the target prediction duration; Δt_i is the i-th prediction step size; k is the number of prediction steps;

[0087] The time step sequence {Δt_1…, Δt_k} should satisfy the following constraints:

[0088] 1) Each Δt_i is an integer multiple of 15 minutes;

[0089] 2) The total time constraint shown in the following formula must be satisfied:

[0090] ;

[0091] 3) Each time step corresponds to calling a unique model in the model set M.

[0092] 5-3: Reverse Greedy Sequence Generation. This involves generating a time step sequence according to the reverse greedy principle, specifically including:

[0093] 1) When decomposing backward from the target prediction time to the reporting time, the maximum time step that does not exceed the remaining duration is selected first; when executing forward inference, the time step sequence is constructed in reverse order of the backward decomposition results.

[0094] 2) The generated time step sequence satisfies the monotonically non-decreasing constraint shown in the following equation:

[0095] ≥ ;

[0096] 3) When Δt_i > Δt_{i-1}, the following equation must be satisfied to ensure that the required historical input frames can be obtained from the observation data:

[0097] ;

[0098] 5-4: Autoregressive prediction execution: Based on the time step sequence, multi-step prediction is executed sequentially. In each prediction step, the cloud map extrapolation model for the corresponding time step is called to generate the prediction result for the next time step, specifically including:

[0099] 1) When i=1, prediction is made directly using historical observation data X_(t-Δt) and current data X_t;

[0100] 2) When i≥2, obtain the input data according to the following two cases:

[0101] If Δt_i > Δt_{i-1}, then the required input data satisfies the following formula:

[0102] t+ .

[0103] The corresponding data can be obtained directly from historical observation data; if Δt_i = Δt_(i-1), then the previous step prediction result shown in the following formula needs to be input:

[0104] ;

[0105] Therefore, the recursion can be completed without additional calculations.

[0106] 5-5: Output the target time. After the above k-step prediction, we finally obtain X′_(t+T), which is the target time cloud map prediction result.

[0107] (6) Distribution correction

[0108] See Figure 5 ,in, Figure 5 (c) is a schematic diagram of the original autoregressive process without using SoS. Figure 5 (d) is a schematic diagram of the standardized autoregressive process using SoS.

[0109] See Figure 5 (c) and Figure 5 (d) To ensure the stability of the network output distribution during multi-step autoregressive prediction, a second-order normalization method, abbreviated as SoS, is proposed. Among them, Figure 5 (c) shows the original autoregressive process without SoS. In this process, each prediction step is standardized and destandardized using a fixed statistic. As the number of autoregressive iterations increases, the prediction error gradually accumulates, causing the pixel distribution of the output image to drift and exhibiting extreme value information decay and distribution distortion.

[0110] See Figure 5(d) This figure illustrates the standardized autoregressive process using SoS. In this process, the present invention introduces a second-order statistical constraint mechanism on the basis of traditional standardization. By combining the global statistics of the training data with the dynamic statistics of the input cloud map in the prediction stage, distribution consistency constraints are achieved, thereby suppressing the distribution drift problem in multi-step prediction. Specifically, it includes:

[0111] 6-1: Calculation of Global Statistics

[0112] The global mean and standard deviation are calculated based on the training dataset using the following formula:

[0113] ;

[0114] .

[0115] Where μ represents the mean; σ represents the standard deviation; L represents the number of training samples; x_(n,i,j) represents the pixel value in the i-th row and j-th column of the n-th sample; 800×800 is the image space size, which can be generalized to W×H.

[0116] 6-2: Calculation of dynamic statistics during the forecasting phase

[0117] For the current input cloud map x_t, its local statistics are calculated using the following formula:

[0118] ;

[0119] .

[0120] Where μ_t and σ_t are used to describe the local distribution characteristics of the current input.

[0121] 6-3: Standardize the input data using dynamic statistics according to the following formula.

[0122] .

[0123] Where Xdot_t represents the standardized input features, which are used to input the GeoAttX model for prediction.

[0124] 6-4: Model Prediction Output

[0125] The standardized input Xdot_t is input into the GeoAttX model to obtain the predicted output Xdot_(t+Δt), which is the prediction result in the standardized space.

[0126] Then, destandardization is performed using global statistics from the training data to restore the prediction results to the original data distribution space:

[0127] = · + .

[0128] Among them, see Figure 5 (c) Without SoS, the prediction results are only constrained by a fixed statistic in each autoregression step, which can easily lead to distribution drift as the number of prediction steps increases; see [reference missing]. Figure 5 (d) When using SoS, dynamic statistics are further introduced during the prediction process to perform second-order standardization on the intermediate prediction results, so that subsequent inputs maintain a more stable statistical distribution.

[0129] 6-5: Distribution Consistency Constraint

[0130] By introducing a two-layer processing mechanism of "dynamic standardization + global destandardization" in each prediction step, the predicted output maintains the same statistical distribution characteristics as the training data during the multi-step autoregression process. Figure 5 (c) reflects the original autoregressive process without the use of SoS, which has weak distribution stability; Figure 5 (d) reflects the standardized autoregressive process after adopting SoS, which can effectively suppress the distribution drift caused by error accumulation and improve the stability and consistency of long-term series prediction.

[0131] (7) Result reconstruction

[0132] 7-1: See Figure 1 and Figure 2 The features processed by the GeoAB module are then input into the upsampling module, and the original spatial resolution is restored through spatial reconstruction. Subsequently, the features are input into the tail module for channel mapping, and combined with residual connections to output the predicted cloud map of the target time.

[0133] 7-2: The tail module uses two-dimensional convolution, batch normalization, SiLU nonlinear activation function, two-dimensional convolution and sub-pixel rearrangement method to map the upsampled features to the output channel of the predicted cloud map and restore the predicted cloud map result to 7×800×800.

[0134] (8) Model training and prediction output

[0135] 8-1: Construct training and testing datasets to validate the model's temporal generalization ability and cross-year prediction performance. The training dataset uses cloud map data from the entire year of 2023 FY-4B, while the testing dataset uses cloud map data from 2024 FY-4B. Training samples use the cloud map data from the previous and current time points as input, with the corresponding future time point cloud map data as supervision labels to optimize the GeoAttX model parameters.

[0136] 8-2: The single-step prediction process of the model is expressed by the following formula:

[0137] .

[0138] Where X_(t-Δt) represents the cloud image observation data at time t−Δt (the previous historical frame); X_t represents the cloud image observation data at time t (the current frame); Δt represents the time interval (i.e., the time step between two adjacent frames); GeoAttX_I(⋅) represents the cloud image extrapolation model based on the geographic attention mechanism (the mapping function of the GeoAttX model); input (X_t, X_t-Δt) indicates that the model uses two consecutive cloud images as time-series input information; X′_(t+Δt) represents the cloud image result (extrapolation result) predicted by the model at time t+Δt; I represents the prediction time-scale model number called; the symbol “′” indicates that the result is a model simulation / prediction value, not a real observation value.

[0139] 8-3: In the inference phase, the cloud map data of the current time and the previous time are input into the trained GeoAttX model. Multi-step iterative inference is performed by combining the single-step prediction model with the reverse greedy strategy to gradually generate the cloud map results of the target prediction time.

[0140] The above embodiments are merely preferred embodiments of the present invention. Equivalent transformations or substitutions made by those skilled in the art without departing from the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A meteorological satellite cloud image extrapolation method based on a geographic attention mechanism, characterized in that, The method includes the following steps: (1) Data preprocessing Acquire multi-band observation data from meteorological satellites, perform geometric correction on the original full-disk image and project it onto a unified coordinate system, select the target area and resample it into a fixed-size image sequence. The multi-band observation data from meteorological satellites is geostationary meteorological satellite observation data. Long-wave infrared bands are selected for modeling to achieve continuous all-weather observation. (2) Multi-frame input construction Satellite observation frames at consecutive time points are selected as input sequences to capture the temporal evolution characteristics of atmospheric motion; (3) Feature extraction and dimensionality reduction The embedding module performs spatiotemporal feature fusion on multi-band observation data, and the head module performs downsampling to reduce resolution and suppress noise. The embedding module adopts a structure combining three-dimensional convolution and two-dimensional convolution to perform joint feature encoding on multi-band and multi-temporal data, thereby obtaining a multi-channel feature tensor with spatiotemporal semantic expression capabilities. (4) Geographic semantic modeling The multi-channel feature tensor obtained in step (3) is input into the geographic attention module. Through long-range attention calculation in the latitude and longitude directions, semantic information extraction across spatial scales is achieved. The geographic attention module adopts a long-range attention mechanism to decompose the two-dimensional spatial attention into a series calculation process of longitude direction attention and latitude direction attention, so as to reduce the computational complexity and realize global spatial dependency modeling. The geographic attention module includes a longitude direction attention calculation unit and a latitude direction attention calculation unit. The longitude direction attention calculation unit is used to establish the dependency relationship between pixels in the same latitude direction, and the latitude direction attention calculation unit is used to establish the dependency relationship between pixels in the same longitude direction. The long-range attention mechanism refers to the attention calculation method that establishes the dependency relationship between pixels in the entire spatial range. (5) Predicted path optimization The target prediction timeliness is decomposed based on the reverse greedy algorithm, and the multi-timescale model is called to perform autoregressive prediction. (6) Distribution correction A second-order normalization algorithm is used to perform statistical characteristic alignment distribution correction on the autoregressive prediction results to maintain the consistency of image distribution. The second-order normalization algorithm dynamically updates statistical parameters during the multi-step prediction process to maintain the image brightness distribution and extreme value characteristics. (7) Result reconstruction Spatial resolution is restored by upsampling and channel mapping, and a predicted cloud map is output. The upsampling uses a sub-pixel rearrangement method to spatially reconstruct the features in order to restore the original resolution. (8) Model training and prediction output The GeoAttX model is trained using historical meteorological satellite observation data to construct training samples. The training samples consist of a series of multi-band cloud images at consecutive time points. The cloud images of the previous and current time points are used as input, and the cloud images of the corresponding next time point are used as supervision labels to optimize the model parameters. During the inference phase, the GeoAttX model inputs the cloud image data of the current time point and the previous time point into the trained model. It performs multi-step iterative inference by combining a single-step prediction model with a reverse greedy strategy to gradually generate cloud image results with the target prediction timeliness. The GeoAttX model is a deep learning network structure that includes an embedding module, a head module, a backbone network composed of geographic attention modules, an upsampling module, an upsampling module, and a tail module.

2. The meteorological satellite cloud image extrapolation method based on geographic attention mechanism according to claim 1, characterized in that, The geographic attention module in step (4) adopts a multi-head attention structure to perform parallel modeling of different spatial patterns, thereby enhancing the ability to express features of global structure, high-value areas and local details.

3. The meteorological satellite cloud image extrapolation method based on geographic attention mechanism according to claim 1, characterized in that, The reverse greedy algorithm in step (5) decomposes the target prediction timeliness, constructs the prediction sequence in non-decreasing order of time intervals, and calls pre-trained models of different time scales for combined prediction.

4. The meteorological satellite cloud image extrapolation method based on geographic attention mechanism according to claim 3, characterized in that, The time intervals are integer multiples of the base time resolution, and the sum of the time intervals equals the target prediction lead time, to ensure that the input data at each step of the prediction process can be obtained from historical frames or previous outputs.

5. The method according to claim 1 or claim 3, characterized in that, The reverse greedy algorithm reduces the number of autoregressive iterations to decrease error accumulation and improve the stability of long-term predictions.

6. The meteorological satellite cloud image extrapolation method based on geographic attention mechanism according to claim 1, characterized in that, The second-order normalization algorithm in step (6) calculates the mean and standard deviation of the input image and performs inverse normalization on the prediction results to achieve consistency constraints between the output distribution and the input distribution.

7. The meteorological satellite cloud image extrapolation method based on geographic attention mechanism according to claim 1 or claim 6, characterized in that, The second-order normalization algorithm is used to suppress distribution distortion and extreme value compression in autoregressive prediction, thereby improving pixel-level prediction accuracy.

Images