Meteorological satellite real-time precipitation retrieval method based on deep learning geographical attention mechanism

Abstract

This invention discloses a real-time precipitation retrieval method for meteorological satellites based on a deep learning geographic attention mechanism. The method includes: constructing a training dataset of FY-4B AGRI infrared cloud images and IMERG Final Run labels, preprocessing it, and inputting it into a U-shaped encoder-decoder network; achieving deep fusion of multi-scale features by embedding a GeoAB geographic attention module; using a dual-objective loss function to complete rain area classification and precipitation intensity regression training; and outputting precipitation data with 0.05° spatial resolution and 15-minute temporal resolution through T-transform decomposition of the cloud image, independent inference, T-inverse transform fusion, and average pooling to eliminate grid effects. Compared with existing technologies, this invention solves the technical problems of balancing real-time performance and accuracy in satellite precipitation retrieval, and the blurred distinction between rain areas and clear skies. This invention achieves a CSI of 0.693, a 3.3% improvement over IMERG Final Run, with a total retrieval time of less than 20 seconds. It balances high accuracy and real-time performance, providing data support for disaster early warning, agricultural management, and other applications, and has promising application prospects.

Description

Technical Field

[0001] This invention relates to the fields of satellite remote sensing and meteorological hydrology, specifically to a method for real-time precipitation inversion from meteorological satellites based on a deep learning geographic attention mechanism. Background Technology

[0002] Climate change has exacerbated the frequency and intensity of extreme precipitation and flooding disasters worldwide. Accurate and real-time satellite precipitation inversion data has become a core support for large-scale hydro-meteorological analysis and disaster monitoring and early warning. Existing precipitation monitoring methods have significant limitations: ground-based weather radars are limited by spatial coverage, making it difficult to achieve comprehensive monitoring in remote areas and ocean regions; traditional satellite precipitation inversion methods rely on physical models such as Zr-R relationships, resulting in low accuracy and susceptibility to observation conditions.

[0003] In satellite-based precipitation retrieval technologies, NASA's IMERG Final Run product integrates multi-source satellite microwave observation data, achieving high precipitation retrieval accuracy. However, it suffers from a delay of approximately 3.5 months, failing to meet the demands of real-time operational applications. Existing deep learning-based infrared satellite precipitation retrieval methods often suffer from problems such as blurred distinction between rain areas and clear skies, large estimation biases for light rainfall, and insufficient fusion of cross-scale features. Furthermore, they struggle to balance spatiotemporal resolution and real-time performance. Additionally, training with a single loss function can lead to near-zero rather than precise zero values ​​in neural network outputs, further reducing the accuracy of precipitation retrieval. The Fengyun-4B satellite (FY-4B), as a second-generation geostationary meteorological satellite, carries the Advanced Geostationary Radiometric Imager (AGRI), which provides infrared cloud imagery data with a 15-minute temporal resolution and a 4km spatial resolution, offering a data source advantage for real-time precipitation retrieval. However, currently, there is a lack of precipitation retrieval methods based on this satellite's infrared cloud imagery that balance high accuracy and low latency.

[0004] In summary, existing satellite precipitation retrieval methods suffer from several technical problems, including difficulty in balancing real-time performance and accuracy, ambiguity in distinguishing between rain areas and clear skies, and large errors in precipitation intensity estimation. Therefore, there is an urgent need to develop a deep learning-based precipitation retrieval method based on FY-4B infrared cloud images to address the issues of poor real-time performance, low accuracy, and ambiguous rain area identification in existing technologies, thereby achieving real-time precipitation retrieval with high spatiotemporal resolution. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a real-time precipitation inversion method for meteorological satellites based on a deep learning geographic attention mechanism. This method employs the construction of FY-4B AGRI infrared cloud images and embeds a GeoAB geographic attention module to achieve high spatiotemporal resolution real-time precipitation inversion based on these images. This method achieves high-precision inversion results using only infrared data sources, with convenient data source acquisition and low latency (10-minute cloud image delay). It effectively solves the problems of large latency, inability to be applied in real-time, blurred distinction between light rainfall and clear skies, and large precipitation estimation bias in existing multi-source fusion products. This invention uses a training dataset constructed from FY-4B AGRI infrared cloud images and IMERG Final Run labels, preprocessed and enhanced before inputting it into an improved U-shaped encoder-decoder network. Deep fusion of multi-scale features is achieved by embedding GeoAB geographic attention modules (Window Attention (WA), Sliding Window Attention (SWA), and Long Distance Attention (LRA)). A dual-objective loss function is used to complete rain area classification and precipitation intensity regression training. In the real-time inversion stage, cloud images are decomposed using T-transform, independent inference, inverse T-transform fusion, and average pooling to eliminate grid effects, outputting precipitation data with a spatial resolution of 0.05° and a temporal resolution of 15 minutes. This invention achieves a CSI of 0.693, a 3.3% improvement over IMERG Final Run, with a total inversion time of less than 20 seconds. Balancing high accuracy and real-time performance, it achieves high-precision real-time precipitation inversion with a spatial resolution of 0.05° and a temporal resolution of 15 minutes, providing data support for meteorological early warning, flood control, agricultural management, and other fields, demonstrating promising application prospects and commercial development value.

[0006] The specific technical solution to achieve the purpose of this invention is: a real-time precipitation inversion method for meteorological satellites based on a deep learning geographic attention mechanism, characterized by the following steps:

[0007] (1) Constructing the training dataset

[0008] FY-4B AGRI infrared cloud image data and IMERG Final Run precipitation label data were acquired. The cloud image data were recalibrated and the resolution was converted. After cropping, data augmentation, and sampling preprocessing, the model input data was obtained. The IMERG Final Run data was used as the label data GT after resolution matching.

[0009] (2) Feature extraction

[0010] The preprocessed FY-4B AGRI infrared cloud image data is input into an improved U-shaped encoder-decoder network. The encoder performs multi-scale downsampling, and the decoder performs upsampling. The GeoAB geographic attention module is embedded in each stage of the encoding and decoding process to extract the multi-scale spatial and spectral features of the cloud image.

[0011] (3) Geographic attention feature fusion

[0012] By employing the collaborative processing of window attention (WA), sliding window attention (SWA), and long-range attention (LRA) within the GeoAB module, deep fusion of local features, neighborhood features, and global long-range features of cloud images is achieved, capturing the complex correlation between infrared radiation and precipitation. The specific implementation of the GeoAB module is as follows:

[0013] 1) Window Attention (WA): The input feature map is divided into non-overlapping windows of a fixed size. Multi-head self-attention calculation is performed independently within each window. The calculation formula follows the classic paradigm of self-attention. Local calculation reduces model overhead and quickly captures the correlation features between infrared radiation and precipitation in the local cloud map.

[0014] 2) Sliding Window Attention (SWA): Based on WA, the window slides according to preset rules to break the isolation between windows and realize the feature interaction between adjacent windows. The calculation formula is: SWA(x(2))=UnShift(WA(Shift(x(2)))). The window slides through the Shift / UnShift operation to capture the cloud map feature dependency relationship across windows.

[0015] 3) Long Distance Attention (LRA): By simulating the information transmission along the latitude and longitude, the feature map is transformed by row and column dimensions to achieve feature interaction in the global scope, overcome the limitations of windows and local areas, capture the long-distance spatial dependence of large-scale precipitation systems, and enhance the model's ability to extract global precipitation features.

[0016] (4) Training of the dual-objective loss model

[0017] An improved U-shaped encoder-decoder network is trained using a dual-objective loss function of rain area classification and precipitation intensity regression. The training first achieves binary classification of rain areas and clear skies through cross-entropy loss, and then optimizes the precipitation intensity of rain areas through smoothing L1 loss, thereby improving the model's ability to identify rain areas and the accuracy of precipitation intensity estimation.

[0018] (5) Real-time precipitation inversion and post-processing

[0019] The real-time acquired FY-4B AGRI infrared cloud images are recalibrated, cropped, and preprocessed with mean pooling before being input into the trained deep learning prediction model (i.e., the improved U-shaped encoder-decoder network). The cloud images are decomposed into multiple sub-images through T-transform and inferred independently. Then, the precipitation results of the sub-images are fused through inverse T-transform. Finally, the grid effect is eliminated through average pooling, and the final high-resolution real-time precipitation data is output, realizing real-time precipitation inversion with high spatiotemporal resolution.

[0020] In step (1), the FY-4B AGRI infrared cloud image data used is water vapor channel data in bands 9-15. The nadir resolution of these 7 bands is 4km. After recalibration, it is converted into a 2400×2400 pixel image with a spatial resolution of 0.05°. Then, it is randomly cropped into 800×800 pixel image blocks. After 2×2 mean pooling, a 400×400 pixel model input data is obtained.

[0021] The data augmentation methods in step (1) include: random flipping, random rotation at 90° / 180° / 270°, where the probability of horizontal and vertical flipping is 50% and the probability of rotation at each angle is 25%; at the same time, samples with a rainfall area ratio of less than 2% are filtered out to avoid model training bias caused by class imbalance.

[0022] The tag data GT in step (1) is precipitation data with a spatial resolution of 0.1° from IMERG Final Run, which is upsampled and matched to a spatial resolution of 0.05°, and is consistent with the spatiotemporal range of FY-4B AGRI cloud map data.

[0023] In step (2), the encoder and decoder of the U-shaped encoder-decoder network are each equipped with 5 GeoAB geographic attention modules. The encoder performs 4 times 2x downsampling in sequence to gradually downsample the 400×400 pixel input feature map to 50×50 pixels. The decoder recovers the feature map to 400×400 pixels through 4 times 2x upsampling, thereby realizing the fusion and recovery of multi-scale features.

[0024] The window attention (WA) in step (3) is implemented by dividing the feature map into multiple non-overlapping windows, performing self-attention calculation independently in each window, reducing the overhead of global calculation, and quickly capturing the local infrared radiation-precipitation correlation features of the cloud map.

[0025] The Sliding Window Attention (SWA) in step (3) breaks the isolation of windows through the sliding window strategy. It slides the window according to the rules in different calculation steps to realize feature interaction and fusion between adjacent windows and capture the feature dependency relationship of cloud map across windows.

[0026] In step (3), the Long Distance Attention (LRA) simulates the information transmission in the latitude and longitude directions, transcends the limitations of windows and local areas, captures long-distance spatial dependencies in the global cloud map, and improves the model's ability to extract features from large-scale precipitation systems.

[0027] The bi-objective loss function loss in step (4) is calculated by the following formula:

[0028] loss=SmoothL1(Pt¹,Yt¹)+CrossEntropy(Pt³,Yt²,Yt³).

[0029] Wherein, SmoothL1 is the smoothing L1 loss for precipitation intensity regression, CrossEntropy is the cross-entropy loss for rain / clear sky classification, Pt¹ is the target precipitation rate, Yt¹ is the model output precipitation rate, Yt² and Yt³ are the classification labels for clear sky and rain, respectively, and Pt³ is the model output classification result.

[0030] The T-transform in step (5) is implemented as follows: the 800×800 pixel cloud map with a resolution of 0.05° is decomposed into four non-overlapping 400×400 pixel sub-images according to the odd and even row and column indices; the T-inverse transform is to stitch the precipitation inference results of the four sub-images into an 800×800 pixel precipitation map according to the original index.

[0031] In step (5), the grid effect is eliminated by 3×3 average pooling. The internal pixels of the spliced ​​precipitation map are pooled and smoothed, while the edge pixels retain their original values. This eliminates the grid effect caused by block decomposition and splicing, thereby improving the spatial continuity of the precipitation map and ensuring its spatial resolution and feature integrity.

[0032] The input to the deep learning prediction model (i.e., the improved U-shaped encoding and decoding network) in step (5) is a real-time FY-4B AGRI infrared cloud image. After recalibration, cropping, and mean pooling, the inference time of a single sub-image is less than 0.6s, and the total time of the entire inversion process (including data preprocessing, inference, and postprocessing) is less than 20s, which meets the real-time requirements.

[0033] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0034] 1) This invention achieves precipitation inversion based on FY-4B AGRI infrared cloud image. It can achieve high-precision inversion effect using only infrared data source. The data source is easy to obtain and has low latency (cloud image latency is 10 minutes), which solves the problems of large latency and inability to be applied in real time in existing multi-source fusion products.

[0035] 2) Introducing the GeoAB geographic attention module, which uses WA, SWA, and LRA to collaboratively capture local, neighborhood, and global features of cloud images, fully explores the complex relationship between infrared radiation and precipitation, and improves the model's ability to extract features from precipitation systems.

[0036] 3) By adopting a dual-objective loss function, accurate classification of rain areas and clear skies and accurate regression of precipitation intensity in rain areas are achieved, which solves the problems of ambiguous distinction between light rain and clear skies and large precipitation estimation bias in existing methods;

[0037] 4) Design T-transform and inverse transform to achieve high-resolution precipitation inversion. Combine 3×3 average pooling to eliminate grid effect. While ensuring 0.05° spatial resolution and 15-minute temporal resolution, the spatial continuity of precipitation map is improved.

[0038] 5) The model has high inference efficiency, with a total inversion time of less than 20 seconds on a regular personal computer. It can realize real-time precipitation inversion for operational purposes, providing high-precision and high spatiotemporal resolution precipitation data support for meteorological early warning, flood disaster prevention and control, agricultural water resource management and other fields. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0040] Figure 2 This is a schematic diagram of the overall structure of a U-shaped encoding / decoding network;

[0041] Figure 3 A schematic diagram of the training logic for a dual-objective loss function;

[0042] Figure 4 This is a schematic diagram of the inference acceleration process for concatenating T-transform decomposition and inverse T-transform. Detailed Implementation

[0043] The embodiments of the present invention are described in detail below. These embodiments are based on FY-4B AGRI infrared cloud images and IMERG FinalRun data, and utilize a deep learning framework to achieve precipitation inversion. The software and hardware used are commercially available or commonly used in the field. Unless otherwise specified, the methods in these embodiments are conventional methods in the art. These embodiments are only for explaining the present invention and should not be construed as limiting the present invention.

[0044] Example 1

[0045] See Figure 1 The specific steps to achieve high spatiotemporal resolution real-time precipitation inversion are as follows:

[0046] Step 1: Training Dataset Construction

[0047] 1-1: Data Source Acquisition

[0048] 1) Infrared cloud image data: Acquire water vapor channel data in bands 9-15 of FY-4B AGRI (center wavelength 6.25~13.3μm, nadir resolution 4km, time resolution 15min). The data time range is 2023-2024 for model training and July-December 2022 for model testing and validation.

[0049] 2) Precipitation label data: Obtain NASA's IMERG Final Run V07B data (spatial resolution 0.1°, temporal resolution 30min) as the GT label data for model training;

[0050] 3) Validation data: Data from global surface meteorological stations in HadISD v3.4.3.2025f were obtained, and 3-hour precipitation data from 444 stations in the study area were selected as ground validation data for model accuracy.

[0051] 1-2: Data Preprocessing

[0052] 1) Cloud image recalibration: The original 2746×2746 pixel full disk image of FY-4B AGRI is recalibrated to 2400×2400 pixels, corresponding to a spatial resolution of 0.05° and a latitude and longitude range of 120°×120°.

[0053] 2) Resolution matching: The precipitation data with a resolution of 0.1° from the IMERG Final Run was upsampled to 0.05° to match the spatiotemporal range of the cloud image data;

[0054] 3) Cropping and pooling: The recalibrated 2400×2400 pixel cloud map is randomly cropped into 800×800 pixel image blocks, and after 2×2 mean pooling, a 400×400 pixel model input feature map is obtained.

[0055] 1-3: Data Augmentation and Sampling

[0056] 1) Data augmentation: Randomly flip the 800×800 pixel cloud tiles horizontally / vertically (50% probability) and rotate them randomly by 90° / 180° / 270° (25% probability each).

[0057] 2) Sample filtering: Filter out samples with a rainfall area of ​​less than 2% to solve the class imbalance problem;

[0058] 3) Time sampling: The training set is sampled at 6-hour intervals, and the test set is sampled at 49-hour intervals to avoid feature redundancy in continuous cloud maps and ensure the diversity of the dataset.

[0059] The final result was 23,165 training samples and 728 validation samples. Each sample was a 7×400×400 feature map, where 7 represents the AGRI 9-15 band number.

[0060] Step 2: Feature Extraction Network Construction

[0061] See Figure 2An improved U-shaped encoder-decoder network is used as the basic framework for feature extraction. This network mainly includes a head module, encoder, decoder, and tail module. The core is the multi-scale feature transformation of the encoder and decoder, and the specific settings are as follows:

[0062] 1) Header module: The input feature map of 7×400×400 is adjusted by a 1×1 convolutional layer, and the number of feature channels is set to 90 to obtain an initial feature map of 90×400×400.

[0063] 2) Encoder: Set up 5 GeoAB geographic attention modules and perform 4 times 2x downsampling in sequence (achieved by convolution with stride of 2) to gradually downsample the initial feature map from 400×400 to 200×200, 100×100 and 50×50. In each downsampling stage, multi-scale features are extracted by GeoAB module.

[0064] 3) Decoder: Five GeoAB geographic attention modules are set up, symmetrical to the encoder, and four times 2x upsampling is performed sequentially (achieved through deconvolution) to gradually restore the 50×50 feature map to 400×400. The corresponding scale features of the encoder are fused in each upsampling stage to achieve complementarity of multi-scale features.

[0065] 4) Tail module: The number of channels of the 90×400×400 feature map output by the decoder is adjusted to 3 through a 1×1 convolutional layer, which correspond to the precipitation rate, clear sky area mask and rain area mask, respectively, to provide output for subsequent dual-target loss training.

[0066] Step 3: GeoAB Geographic Attention Feature Fusion

[0067] The GeoAB module is adopted as the core feature fusion unit of the network, which includes three sub-modules: Window Attention (WA), Sliding Window Attention (SWA), and Long Distance Attention (LRA). These modules work together to fuse local, neighborhood, and global features of the cloud image. The specific implementation method is as follows:

[0068] 1) Window Attention (WA): The input feature map is divided into non-overlapping windows of a fixed size. Multi-head self-attention calculation is performed independently within each window. The calculation formula follows the classic paradigm of self-attention. Local calculation reduces model overhead and quickly captures the correlation features between infrared radiation and precipitation in the local cloud map.

[0069] 2) Sliding Window Attention (SWA): Based on WA, the window slides according to preset rules to break the isolation between windows and realize the feature interaction between adjacent windows. The calculation formula is: SWA(x(2))=UnShift(WA(Shift(x(2)))). The window slides through the Shift / UnShift operation to capture the cloud map feature dependency relationship across windows.

[0070] 3) Long Distance Attention (LRA): By simulating the information transmission along the latitude and longitude, the feature map is transformed by row and column dimensions to achieve feature interaction in the global scope, overcome the limitations of windows and local areas, capture the long-distance spatial dependence of large-scale precipitation systems, and enhance the model's ability to extract global precipitation features.

[0071] Step 4: Training the dual-objective loss model

[0072] 4-1: Loss Function Setting

[0073] See Figure 3 A bi-objective loss function is adopted, which combines rain area classification and precipitation intensity regression. Cross-entropy loss is used for binary classification of rain areas / clear skies, and SmoothL1 loss is used for regression of precipitation intensity in rain areas. The loss function is calculated by the following formula:

[0074] loss=SmoothL1(Pt¹,Yt¹)+CrossEntropy(Pt³,Yt²,Yt³).

[0075] In the formula, Pt¹ is the target precipitation rate of the IMERG Final Run; Yt¹ is the precipitation rate output by the model; Yt² and Yt³ are the binary classification labels for clear skies and rainy areas, respectively; and Pt³ is the classification probability output by the model.

[0076] 4-2: Training Parameter Settings

[0077] The model is built on the PyTorch framework and uses the AdamW optimizer with an initial learning rate of 0.001, which is halved every 10 epochs. The batch size is set to 32, and the model is trained on 4 NVIDIA RTX 6000 Ada GPUs for a total of 50 epochs. The total number of parameters in the model is 241.78M. The model is fine-tuned based on the pre-trained model, and the total training time is about 1 day.

[0078] 4-3: Model Validation

[0079] The trained model was evaluated using a validation set. The model with the lowest SmoothL1 loss in the 46th epoch was selected as the final model. The critical success index (CSI) of the model on the validation set reached 0.46 and the R² reached 0.45, which is far better than the model trained with a single loss function.

[0080] Step 5: Real-time precipitation inversion and post-processing

[0081] The real-time precipitation inversion process of this invention is based on FY-4B AGRI real-time infrared cloud imagery, and the entire process takes less than 20 seconds, meeting the requirements of operational real-time applications. The specific steps are as follows:

[0082] 5-1: Real-time cloud image preprocessing

[0083] The real-time acquired FY-4B AGRI 9-15 band cloud image (delayed by 10 minutes) was recalibrated and converted into a 2400×2400 pixel image with a resolution of 0.05°. It was then cropped into an 800×800 pixel image block and subjected to 2×2 mean pooling to obtain a 400×400 pixel inference input.

[0084] 5-2: T-transform decomposition

[0085] See Figure 4 The 800×800 pixel cloud patch is decomposed into four 400×400 pixel sub-images according to odd and even row and column indices using the T-transform. The decomposition formula is: T(S)=S1,S2,S3,S4; E={0,2,...,798}, O={1,3,...,799}; S1={Tᵢⱼ=Sᵢⱼ | i∈E,j∈E}, S2={Tᵢⱼ=Sᵢⱼ | i∈O,j∈E}; S3={Tᵢⱼ=Sᵢⱼ | i∈E,j∈O}, S4={Tᵢⱼ=Sᵢⱼ | i∈O,j∈O}; where S is the original cloud patch, S1~S4 are the decomposed sub-images, E is the even index set, and O is the odd index set.

[0086] 5-3: Model Reasoning

[0087] The four sub-images are input into the trained model. The inference time for a single sub-image is 0.6s, and the model outputs four 400×400 pixel precipitation sub-results.

[0088] 5-4: T-inverse transform fusion

[0089] The four precipitation sub-results were stitched together according to their original odd and even row and column indices to form an 800×800 pixel precipitation map, restoring the precipitation data to a resolution of 0.05°.

[0090] 5-5: Elimination of the grid effect

[0091] The stitched precipitation map is processed by 3×3 average pooling, and the internal pixels are processed by 1 / 9 Σ (m=i-1 to i+1) Σ.

[0092] (n = j-1 to j+1)P mn Smoothing is performed, edge pixels retain their original values, and the mesh effect caused by block decomposition and stitching is eliminated.

[0093] Improve the spatial continuity of precipitation maps.

[0094] 5-6: Result Output

[0095] The post-processed precipitation maps are stitched together to form precipitation data within a 120°×120° latitude and longitude range, outputting real-time precipitation rates (mm / hr) with a spatial resolution of 0.05° and a temporal resolution of 15 minutes.

[0096] Step 6: Model Accuracy Verification

[0097] The precipitation retrieval model of this invention was validated in multiple dimensions using HadISD ground station data and mainstream satellite precipitation products. The results show that:

[0098] 1) Ground station validation: In the validation of the model with 3-hour precipitation data from 444 HadISD stations, the critical success index (CSI) reached 0.693, which is 3.3% higher than IMERG Final Run (CSI=0.671), the correlation coefficient (CC) reached 0.444, and the root mean square error (RMSE) was 2.634 mm / hr. All indicators are better than mainstream real-time precipitation products such as FY-4B QPE, PERSIANN-CCS, and GSMAP_NOW.

[0099] 2) Adaptability to precipitation intensity: The model has good inversion effect for precipitation of different intensities. The CSI of light rainfall (0~3mm / hr), moderate rainfall (3~6mm / hr) and heavy rainfall (≥6mm / hr) reached 0.44, 0.54 and 0.57 respectively. The inversion accuracy gradually improves with the increase of precipitation intensity, and it can effectively capture extreme heavy rainfall events.

[0100] 3) Typhoon scenario verification: In typhoon-induced precipitation events, the model can accurately reconstruct the spatial distribution pattern of typhoon precipitation, has a strong ability to identify typhoon precipitation systems, and only slightly underestimates the precipitation intensity, meeting the application requirements for typhoon disaster early warning.

[0101] The above embodiments are merely illustrative of the present invention and are not intended to limit the scope of the patent. All equivalent implementations of the present invention should be included within the scope of the claims of this patent. Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions, and variations can be made to the above embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism, characterized in that, The real-time precipitation inversion specifically includes the following steps: Step 1: Construct the training dataset Infrared cloud image data from the FY-4B AGRI geostationary radiometric imager on Fengyun-4B satellite and precipitation tag data from the IMERG Final Run precipitation retrieval product of the Global Precipitation Measurement Program were acquired. The cloud image data were recalibrated and the resolution was converted. Then, after cropping, data augmentation and sampling, the model input data was obtained. The IMERG Final Run precipitation tag data was used as the tag data GT after resolution matching. Step 2: Feature Extraction The preprocessed FY-4B AGRI infrared cloud image data is input into a U-shaped encoder-decoder network. The encoder performs multi-scale downsampling, and the decoder performs upsampling. The GeoAB geographic attention module is embedded in each stage of the encoder-decoder to extract the multi-scale spatial and spectral features of the cloud image. Step 3: Geographic Attention Feature Fusion By employing the collaborative processing of window attention (WA), sliding window attention (SWA), and long-range attention (LRA) in the GeoAB module, deep fusion of local features, neighborhood features, and global long-range features of the cloud image is achieved. The specific implementation method is as follows: 1) Window Attention (WA): The input feature map is divided into non-overlapping windows of a fixed size. Multi-head self-attention calculation is performed independently within each window. The calculation formula follows the classic paradigm of self-attention. Local calculation reduces model overhead and quickly captures the correlation features between infrared radiation and precipitation in the local cloud map. 2) Sliding Window Attention (SWA): Based on WA, the window slides according to preset rules to break the isolation between windows and realize the feature interaction between adjacent windows. The calculation formula is: SWA(x(2))=UnShift(WA(Shift(x(2)))). The window slides through the Shift / UnShift operation to capture the cloud map feature dependency relationship across windows. 3) Long Distance Attention (LRA): By simulating the information transmission along the latitude and longitude, the feature map is transformed by row and column dimensions to achieve feature interaction in the global scope, overcome the limitations of windows and local areas, capture the long-distance spatial dependence of large-scale precipitation systems, and enhance the model's ability to extract global precipitation features. Step 4: Training the dual-objective loss model A dual-objective loss function of rain area classification and precipitation intensity regression is used to train the U-shaped encoder-decoder network. The training first achieves binary classification of rain areas and clear sky through cross-entropy loss, and then optimizes the precipitation intensity of rain areas through smoothing L1 loss. Step 5: Real-time precipitation inversion and post-processing The real-time acquired FY-4B AGRI infrared cloud image is preprocessed and input into the trained model. The cloud image is decomposed into multiple sub-images by T-transform and inferred independently. The precipitation results of the sub-images are fused by inverse T-transform. Then, the grid effect is eliminated by average pooling operation. Finally, high-resolution real-time precipitation data is output to achieve real-time precipitation inversion with high spatiotemporal resolution.

2. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The FY-4B AGRI infrared cloud image data in step 1 uses water vapor channel data in the 9-15 band. It is recalibrated to a 2400×2400 pixel image with a spatial resolution of 0.05°, and randomly cropped into 800×800 pixel image blocks. Then, it is subjected to 2×2 mean pooling to obtain 400×400 pixel model input data.

3. The real-time precipitation inversion method for meteorological satellites based on deep learning geographic attention mechanism according to claim 1, characterized in that, The data augmentation in step 1 includes: random flipping and random rotation at 90° / 180° / 270°, with a 50% probability of horizontal and vertical flipping and a 25% probability of rotation at each angle; at the same time, samples with a rainfall area ratio of less than 2% are filtered out.

4. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The U-shaped encoding and decoding network in step 2 includes a head module, an encoder, a decoder, and a tail module. The encoder and decoder each have 5 GeoAB geographic attention modules. The encoder performs 4 times of downsampling by 2x, and the decoder performs 4 times of upsampling by 2x to achieve the fusion and recovery of multi-scale features.

5. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The window attention (WA) in step 3 divides the feature map into multiple non-overlapping windows, and performs self-attention calculation independently within each window; the sliding window attention (SWA) breaks the isolation of windows through the sliding window strategy, realizing feature interaction and fusion between adjacent windows; the long-distance attention (LRA) captures long-distance spatial dependencies in the global range of the cloud map by simulating information transmission in the latitude and longitude directions, transcending the limitations of windows and local areas.

6. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The bi-objective loss function loss in step 4 is calculated by the following formula: loss=SmoothL1(Pt¹,Yt¹)+CrossEntropy(Pt³,Yt²,Yt³); Wherein, SmoothL1 is the smoothing L1 loss for precipitation intensity regression; CrossEntropy is the cross-entropy loss for rain / clear sky classification; Pt¹ is the target precipitation rate; Yt¹ is the model output precipitation rate; Yt² and Yt³ are the classification labels for clear sky and rain, respectively; and Pt³ is the model output classification result.

7. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The T-transform in step 5 decomposes the 800×800 pixel cloud image with a resolution of 0.05° into four non-overlapping 400×400 pixel sub-images according to the odd and even row and column indices; the inverse T-transform stitches the precipitation inference results of the four sub-images together into an 800×800 pixel precipitation map according to the original index.

8. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The grid effect in step 5 is eliminated by 3×3 average pooling. Specifically, the internal pixels of the stitched precipitation map are smoothed by pooling, while the edge pixels retain their original values, thus eliminating the grid effect caused by block decomposition and stitching.

9. The method for real-time precipitation retrieval from meteorological satellites based on a deep learning geographic attention mechanism according to claim 1, characterized in that, The total time for the real-time precipitation inversion process in step 5 is less than 20 seconds, and the output data is precipitation rate data with a spatial resolution of 0.05° and a temporal resolution of 15 minutes.

Images