A weather radar reflectivity synthesis method and system based on synchronous stationary satellite

Abstract

The application relates to the technical field of weather radar synthesis, in particular to a weather radar reflectivity synthesis method and system based on a synchronous stationary satellite. The method comprises the following steps: acquiring satellite multi-channel observation data and weather radar data; performing data preprocessing based on the acquired satellite multi-channel observation data and weather radar data; constructing a weather radar synthesis model based on an improved Swin-Transformer and U-Net; performing model training on the constructed weather radar synthesis model; performing forward inference by using the trained model and obtaining weather radar reflectivity synthesis. The application provides high-quality weather observation data for the regions, solves the problem of weather radar data vacancy, and significantly improves the weather radar observation range, especially the monitoring capability for large-scale weather systems and strong convective weather.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather radar synthesis, in particular to a weather radar reflectivity synthesis method and system based on synchronous geostationary satellite. BACKGROUND

[0002] Severe convective weather has the characteristics of suddenness, locality and destructiveness, and accurate monitoring of its development is of great significance for disaster prevention and reduction. Ground-based weather radar, as a key device for identifying and monitoring severe convective weather, can directly reflect the precipitation structure, echo intensity and convective core position within the cloud body. However, the existing ground-based weather radar station layout is limited by factors such as topography, site construction conditions, power supply and communication, and there are obvious weather radar observation blind areas in complex regions such as oceans, plateaus and mountainous areas, resulting in that some areas cannot obtain continuous and comprehensive weather radar observation data. Especially in remote or complex terrain areas, the coverage capability of weather radar is greatly limited, which brings difficulties to the timely monitoring and early warning of severe convective weather.

[0003] Unlike ground-based weather radar, synchronous geostationary meteorological satellites (such as Himawari-8) are not limited by terrain and can provide extensive coverage and high temporal resolution. However, although satellite observations can reflect cloud thickness, ice water phase, cloud top temperature and other characteristics, since satellite observations are usually located at the top of the cloud layer, and it is difficult to accurately capture the details of the convective core, it is difficult for satellite observations to directly replace weather radar echoes, especially in the monitoring of severe convective systems, there are still problems such as incomplete physical information, invisible structure, and noise sensitivity.

[0004] In recent years, deep learning technology has made significant progress in remote sensing image processing and meteorological prediction. Deep learning can learn complex nonlinear mapping relationships through data-driven methods, especially the Transformer model, whose self-attention mechanism makes it exhibit superior performance in long-distance dependence and spatiotemporal feature modeling tasks. However, severe convective weather systems have significant multiscale characteristics, and existing techniques only rely on stacking independent satellite spectral channels to describe convective structures, which cannot fully capture these multiscale characteristics, resulting in limited ability to capture key features such as convective cores, boundary layer cloud systems, and deep convective cooling.

[0005] Therefore, the existing combination of satellites and weather radar has obvious deficiencies in severe convective weather monitoring and needs to be further improved. Existing techniques cannot effectively solve the problem of generating echo data consistent with real weather radar in weather radar blind areas or areas that cannot be covered, and the effective fusion and modeling capability of multi-modal satellite data is still insufficient.

[0006] In addition, existing ground-based weather radars are limited by site construction and maintenance conditions, and there are significant observation gaps in the ocean, plateau and remote areas. In the areas covered by the existing network, the resolution is deteriorated due to terrain obstruction and increased distance, and the ground clutter interference makes it difficult to continuously and finely depict the three-dimensional structure of deep convection and the evolution of echoes. In contrast, synchronous geostationary meteorological satellites have the advantage of wide-range and quasi-continuous observation, but the visible / infrared channels only reflect the cloud top radiation characteristics, which have significant cross-modal differences in physical dimension and information sensitivity compared with the ground-based weather radar composite reflectivity (CREF, dBZ). Directly using stacked independent satellite spectral channels cannot restore the strong echo center, convection boundary and precipitation continuity. At the same time, the inconsistency between satellites and weather radars in time sampling, projection geometry and spatial resolution makes it difficult to sample and align in time and space, further exacerbating the instability and bias accumulation of cross-modal mapping. The traditional scheme based on convolutional neural network is limited by fixed receptive field, and has insufficient ability to model multi-scale organizational structures (from convective cells to mesoscale convective systems) at the same time, which is prone to problems such as strong echo position drift, texture over-smoothing and inaccurate depiction of strong convective core (such as ≥35 dBZ region). SUMMARY

[0007] In order to solve the above-mentioned problems, the present application provides a weather radar reflectivity synthesis method and system based on synchronous geostationary satellite.

[0008] In the first aspect, the present application provides a weather radar reflectivity synthesis method based on synchronous geostationary satellite, which adopts the following technical scheme:

[0009] A weather radar reflectivity synthesis method based on synchronous geostationary satellite, comprising:

[0010] Obtaining satellite multi-channel observation data and weather radar data;

[0011] Performing data preprocessing based on the obtained satellite multi-channel observation data and weather radar data;

[0012] Constructing a weather radar synthesis model based on an improved Swin-Transformer and U-Net;

[0013] Model training is performed on the constructed weather radar synthesis model;

[0014] Forward inference is performed using the trained model to obtain weather radar reflectivity synthesis.

[0015] Further, the data preprocessing based on the obtained satellite multi-channel observation data and weather radar data includes unifying the spatial resolution and adjusting the size of the input synchronous static meteorological satellite multi-channel observation data and ground-based weather radar comprehensive reflectivity data, first, the spatial resolution of the satellite and weather radar data is unified to grid points , assuming that the original image is , and after being projected to the target grid, it is recorded as , for each target pixel center , a resampling operator is used for bilinear interpolation, which is represented as:

[0016]

[0017] wherein, is the interpolation block average weight, and then the sub-blocks are cut according to a fixed window, which is represented as:

[0018] ; then, the two types of data are time-registered, assuming that the satellite timestamp sequence is , and the weather radar timestamp sequence is , first, the threshold determination nearest neighbor pairing is performed: and for each , the following is selected:

[0019] At this time, if , a pair of is formed, when higher time consistency is required, linear time interpolation is performed on the satellite frame at the weather radar time , assuming that is the time of the nearest two satellite frames, at this time:

[0020] , considering the apparent drift caused by the translation of the cloud system, a simplified advection correction is added before interpolation, assuming that the related matching displacement is , at this time, the matching result is:

[0021] ,

[0022] To ensure spatial consistency, the satellite and weather radar are projected to a unified grid , the projection mapping is recorded as , and it is required that the paired blocks satisfy the upper limit of geometric error:

[0023] .

[0024] Further, the data preprocessing based on the obtained satellite multi-channel observation data and weather radar data further comprises, after completing the time matching of the satellite data and the weather radar data, further processing the blank area in the data set, by calculating the reflectivity value of each pixel in the weather radar data, and removing the pixel with reflectivity less than 10 dBZ as a blank area, assuming that the weather radar pixel reflectivity is Implementing weak echo elimination, the elimination method is as follows: , wherein is the training sample area, then the data is maximum-minimum normalized, and the normalization formula is:

[0025] , wherein is the normalized value, is the original data value, and are the minimum and maximum values in the data set, respectively, and finally after completing the preprocessing step, the obtained matching sample data set is proportionally divided into a training set, a validation set and a test set.

[0026] Further, the weather radar synthesis model based on the improved Swin-Transformer and U-Net is constructed, including that the input is composed of two types of satellite observations, visible light multi-channel VIS and infrared multi-channel IR, assuming that the preprocessed visible light and infrared observations are , the target weather radar reflectivity true value is , assuming that the patch side length is , the initial embedding channel number is D, the Swin layer number is , and the down-sampling rate set is ; for each input, the structure of patch embedding and shallow coding Swin-Transformer is adopted, and two modalities enter independent patch embedding and shallow encoder respectively to extract their own sensitive primary features, represented as:

[0027] , the shallow features are extracted by Swin blocks, by alternately using window self-attention and shift window self-attention to balance the details and receptive field, represented as: .

[0028] Further, the construction of the weather radar synthesis model based on improved Swin-Transformer and U-Net also includes aligning and fusing two-way features to form a unified representation of a single path after shallow representation of VIS and IR, avoiding deep redundant calculation and explicitly modeling complementarity, setting the alignment operator as , the fusion operator as :

[0029] , , , the implementation of the gated weighting is represented as: , wherein is channel-wise multiplication, is Sigmoid, is global average pooling, is splicing, while cross-modal attention is used to enhance mutual information, represented as:

[0030] , wherein is the key / query dimension, is the learnable projection parameter; the fused features enter the hierarchical backbone encoder, which gradually expands the receptive field and increases the channel number through multi-level down-sampling, and then stacks Swin blocks to jointly model local and global dependencies, represented as:

[0031] ,

[0032] ;

[0033] The bottleneck layer continues to apply self-attention to strengthen the large-scale context, obtaining ; finally, the decoder performs PatchExpanding multi-level up-sampling and cross-layer jump connection with the corresponding scale features of the encoder, fusing deep semantic and shallow edge texture details, represented as:

[0034]

[0035] , wherein is a lightweight convolution / attention fusion unit, which directly injects the same scale edges and details retained in the encoding stage into the current decoding layer, so that deep semantic information and shallow spatial details complete one fusion at each scale, and finally generate a linear projection to generate a synthetic weather radar reflectivity estimate: .

[0036] Further, the model training on the constructed weather radar synthesis model comprises adopting a weighted loss composed of four terms: absolute error L1, structural similarity SSIM, focal loss of strong echo mask, and gradient consistency term, for simultaneously optimizing numerical approximation, structural similarity, and key area of strong echo, wherein the L1 term is used to provide a pixel-level baseline constraint, the SSIM term is used to evaluate the similarity between the synthesized image and the true value from three dimensions of brightness, contrast, and structure, the focal loss is used to adjust the difficult sample weight and the class imbalance coefficient for the strong echo area, and the gradient consistency term is used to constrain the generated result from the edge and texture angles; the structure of the loss is as follows:

[0037] ,

[0038] wherein is a true value CREF, , is a binary mask of strong echo ; and is a loss weight, is a Focal loss hyperparameter, and through a four-dimensional coupled objective function design, the trained model obtains more balanced improvement in the CSI, POD strong convection index, and SSIM structure index.

[0039] Further, the forward inference using the trained model and obtaining weather radar reflectivity synthesis comprises: sending the preprocessed multi-channel satellite observation in batches into the trained model for forward inference according to a 512*512 standard sub-block, to avoid the stitching marks caused by block boundaries, the sub-block adopts a sliding window strategy with overlap, and after inference, the weighted fusion method is used to backfill to the full frame, for the case that there is a second-level difference between the satellite and the target time, a time registration frame is used as the input; if there is a missing frame at a certain time, the adjacent satellite data is used to obtain an approximate input through linear interpolation and simplified advection correction, to ensure the continuity on the time axis, all network outputs are first de-normalized and physical dimension restored, and then written back to the business product according to the unified projection grid, so as to be directly compared with the measured weather radar or numerical model field.

[0040] Further, the utilizing the trained model to perform forward inference and obtaining the weather radar reflectivity synthesis further comprises restraining isolated noise points and blocking effects under the premise of preserving details, introducing three types of lightweight post-processing, performing morphological denoising and small connected domain filtering on low-amplitude isolated pixels to avoid misjudgment of random noise points; using edge-preserving light smoothing to weaken the block splicing traces while not damaging the sharpness of strong echo boundaries; performing time consistency constraints, i.e., registering the synthesized echo of the previous time according to the data of adjacent times, and performing consistency test with the current output, and restraining or reverting to a conservative value for the region with non-physical mutation; to reduce the instantaneous flicker of adjacent time outputs, introducing a simplified time consistency constraint that does not depend on external platforms: using the estimated displacement between the adjacent two satellite fields to adjust the synthesized echo of the previous time to the current time, and performing difference with the current output and applying mild punishment or local smoothing to the abnormal mutation region, the constraint only works near the output layer of the model without changing the network structure, and improves the temporal smoothing and morphological continuity without damaging the details, to verify the achievement of the design goal, a minimum evaluation set is reserved without relying on the system: select several convective cases, calculate the overall error, structural similarity and simple detection index under the strong echo threshold.

[0041] Further, the utilizing the trained model to perform forward inference and obtaining the weather radar reflectivity synthesis further comprises obtaining a pixel-level confidence map at the same time as the synthesized reflectivity, which is used to indicate the reliable interval of the output and the uncertainty source, wherein a small number of times of forward inference is performed on the same time input by applying a lightweight disturbance, and the mean and standard deviation of the results are taken: the mean as the final synthesized echo, and the standard deviation or quantile difference as the uncertainty measure, and a strong convective probability map is regenerated for the strong echo threshold, which is convenient for subsequent decision-making to select different thresholds under different risk tolerance; to make the output meet the meteorological physical intuition, two types of posterior consistency constraints are applied to the results without changing the network structure: including a spatial continuity constraint for connected domain analysis of the synthesized echo to remove small, isolated and short-lived noise spots; and a strong echo edge sharpness constraint for using edge-preserving filtering only in the strong echo edge band to suppress the jaggedness caused by interpolation and splicing without damaging the sharpness of the peak value.

[0042] The second aspect is a weather radar reflectivity synthesis system based on a synchronous stationary satellite, comprising:

[0043] A data acquisition module configured to acquire satellite multi-channel observation data and weather radar data;

[0044] A preprocessing module configured to perform data preprocessing based on the acquired satellite multi-channel observation data and weather radar data;

[0045] The model construction module is configured to construct a weather radar synthesis model based on the improved Swin-Transformer and U-Net.

[0046] The model training module is configured to perform model training on the constructed weather radar synthesis model.

[0047] The synthesis module is configured to perform forward inference using the trained model and obtain weather radar reflectivity synthesis.

[0048] In a third aspect, the present application provides a computer-readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded and executed by a processor of a terminal device to implement the weather radar reflectivity synthesis method based on a synchronous stationary satellite.

[0049] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer-readable storage medium, the processor being configured to implement instructions; and the computer-readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded and executed by the processor to implement the weather radar reflectivity synthesis method based on a synchronous stationary satellite.

[0050] In summary, the present application has the following beneficial technical effects:

[0051] (1) Breakthrough in weather radar observation space limitation: The present application successfully fills the blind area of ground-based weather radar by using synchronous stationary satellite data, especially in the ocean, plateau and remote areas. Due to geographical conditions, traditional ground-based weather radar systems cannot cover these areas, resulting in missing or incomplete meteorological data. By combining satellite and weather radar data, the present application provides high-quality weather observation data for these areas, solving the problem of weather radar data gaps and significantly improving the weather radar observation range, especially for monitoring large-scale weather systems and severe convective weather.

[0052] (2) Use of Transformer global modeling capability to improve accuracy: Compared with traditional convolutional neural networks (CNN), the present application uses a multi-scale encoding structure based on Swin-Transformer, breaking through the limitations of the local receptive field of CNN models. Traditional CNN can only process image information in a local area, while Transformer can model the spatial dependence of cloud layer structure on a global scale through window self-attention and cross-window attention mechanisms. Through this method, the present application has significantly improved the generation of strong echoes in severe convective areas (especially ≥35 dBZ), enhancing the authenticity of weather radar echoes.

[0053] (3) Multi-modal fusion improves the synthesis accuracy of weather radar: The application innovatively processes visible light and infrared channels as two independent modalities, uses modal-specific encoders to extract differentiated information, and performs deep semantic fusion in the middle stage. Compared with the simple "stacked input channel" method, this multi-modal fusion method can significantly strengthen the mapping relationship between satellite data and weather radar composite reflectivity, thereby improving the accuracy of the synthesized weather radar. Through this method, the application improves the quality of weather radar reflectivity synthesis, especially in the generation of strong convective clouds and the recovery of weather radar reflectivity details, showing significant performance advantages. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The overall flowchart of the multi-scale artificial intelligence method based on the synchronous stationary satellite synthetic weather radar is implemented in the application;

[0055] Figure 2 The data preprocessing flowchart of the multi-scale artificial intelligence method for the synchronous stationary satellite synthetic weather radar is implemented in the application;

[0056] Figure 3 The model diagram of the multi-scale artificial intelligence model based on the synchronous stationary satellite synthetic weather radar is implemented in the application;

[0057] Figure 4 The effect diagram of the synthetic weather radar is implemented in the application. DETAILED DESCRIPTION

[0058] The application will be further described in detail below in combination with the drawings. EMBODIMENT

[0059] Glossary:

[0060] Synchronous stationary satellite: Geostationary Satellite.

[0061] Weather radar synthesis: weather radar synthesis.

[0062] Reference Figure 1 The weather radar reflectivity synthesis method based on the synchronous stationary satellite of the present embodiment comprises:

[0063] The synchronous stationary satellite synthetic weather radar of the present embodiment includes but is not limited to Chinese Fengyun 2, Chinese Fengyun 4, Japanese HINODE-8, American GOES-R series, etc.

[0064] S1 Data preprocessing stage

[0065] In the data preprocessing stage, the present application ensures the consistency and efficiency of the input data through multiple steps, so that the subsequent deep learning model can effectively learn and generate synthetic weather radar reflectivity data.

[0066] S1.1 Down-sampling and cutting step

[0067] Firstly, the input synchronized static meteorological satellite Himawari-8 (H8) multi-channel observation data and the ground-based weather radar composite reflectivity (CREF) data are unified in spatial resolution and size adjusted. Specifically, first, the spatial resolution of satellite and weather radar data is unified to grid , the resolution is 2km × 2km, which can effectively reduce the data size and ensure the spatial correspondence between data pixels. Next, let the original image be , and after being projected to the target grid, it is recorded as , for each target pixel center , the resampling operator (bilinear interpolation or block averaging) is adopted:

[0068]

[0069] , where is the interpolation block average weight. Then cut the sub-blocks in a fixed window:

[0070]

[0071] At this time, the actual coverage of each sub-block is 512 km × 512 km, corresponding to a 512 × 512 pixel matrix. This processing step ensures that large-scale meteorological observation data can be localized, and provides consistent resolution and efficient calculation input samples for subsequent deep learning model training.

[0072] S1.2 Time difference matching step

[0073] In order to ensure the consistency of satellite data and weather radar data in time, the present application performs time registration on the two types of data. Specifically, under the condition that the time error is less than 3 minutes, the satellite data frame and the weather radar echo frame are matched frame by frame to ensure the synchronization of the data set in time.

[0074] Let the satellite timestamp sequence be , and the weather radar timestamp sequence be . First, the threshold determination nearest neighbor pairing is performed:

[0075]

[0076] and for each Selection:

[0077]

[0078] At this time, if , a pair of is formed.

[0079] When higher temporal consistency is required, at the weather radar time , linear time interpolation is performed on the satellite frame: let be the latest two satellite times, at this time:

[0080]

[0081] ,

[0082] If the apparent drift caused by the translation of the cloud system is considered, a simplified advection correction is added before interpolation, and let the relevant matching displacement be .

[0083] At this time, the matching result is:

[0084] ,

[0085] To ensure spatial consistency, the satellite and weather radar are projected onto a unified grid , the projection mapping is denoted as , and it is required that the paired blocks satisfy the upper limit of geometric error:

[0086]

[0087] where is taken within pixels (about ).

[0088] The goal of this step is to minimize the cloud system position offset and precipitation evolution difference caused by the observation time difference, and to ensure the temporal consistency and physical comparability in the subsequent model learning process.

[0089] S1.3 Blank data removal step

[0090] After completing the time matching of satellite data and weather radar data, the blank areas in the data set are further processed. By calculating the reflectivity value (unit dBZ) of each pixel in the weather radar data, and regarding the pixels with reflectivity less than 10 dBZ as blank areas, they are removed.

[0091] We set the weather radar pixel reflectivity as (unit dBZ) to implement weak echo removal, and the removal method is as follows:

[0092] ,

[0093] wherein is the training sample region.

[0094] This step effectively removes the areas without precipitation or weak echoes in weather radar data, suppresses ground noise, weak echoes and background blanking interference, thereby ensuring that the training data set mainly contains strong convective and precipitation structure information with greater meteorological significance, which is crucial for improving model training efficiency and convergence stability.

[0095] S1.4 Maximum-Minimum Standardization Step

[0096] In order to improve the stability and convergence of model training, the present application performs maximum-minimum standardization on the data, and the standardization formula is:

[0097] ,

[0098] wherein, is the normalized value, is the original data value, and are the minimum and maximum values in the data set, respectively. Through this standardization method, all input data will be unified to the same scale, ensuring the stability and accelerated convergence of the training process.

[0099] S1.5 Data Set Division Step

[0100] After completing all the above preprocessing steps, the matching sample data set obtained will be divided into training set, validation set and test set in proportion. The training set is used for model parameter learning, the validation set is used to monitor the generalization performance of the model and prevent overfitting, and the test set is used for final model effect evaluation. This division method can ensure the performance of the model on different data sets, and avoid overfitting phenomenon through the validation set and test set.

[0101] S2 Multi-scale Artificial Intelligence Model Building of Synchronous Static Satellite Synthetic Weather Radar

[0102] In the second phase of the present application, a multi-scale artificial intelligence model is established mainly through deep learning method, which synthesizes satellite observation data and weather radar data. The model is based on Swin-Transformer and U-Net structure, aiming to realize high-precision weather radar reflectivity synthesis.

[0103] S2.1 Input and Labeling

[0104] The input of this stage consists of two types of satellite observations: visible light (VIS) multi-channel and infrared (IR) multi-channel; both have natural complementarity in physical sensitivity - VIS is easier to depict cloud texture, optical thickness and boundary details, while IR is more sensitive to cloud top temperature and cold top deep convection. To make full use of this complementarity, we will input VIS and IR as two independent tensor models, and process them separately at an early stage of feature extraction.

[0105] Let the pre-processed visible light and infrared observations be The target weather radar reflectivity truth is (unit dBZ, linearly scaled to [0,1] during training).

[0106] Let the patch side length be , the initial embedding channel number be D, the Swin layer number be , and the down-sampling rate set be .

[0107] Such an organization can avoid mixing scale differences and noise characteristics completely at the beginning, reduce the learning burden of the underlying network, and make the network first extract discriminative features in the "familiar" modality, and then fuse at the appropriate level, thereby improving the stability and interpretability of cross-modal mapping, especially in the case of night IR dominance or daytime VIS dominance.

[0108] S2.2 Modality-specific encoding and patch embedding

[0109] For each input, we use the structure of "patch embedding + shallow Swin-Transformer encoding". Patch embedding divides large satellite images into fixed-size small blocks and linearly maps them to the feature space, which not only reduces computational overhead, but also provides a suitable granularity for subsequent self-attention mechanisms.

[0110] Two routes of modalities enter independent patch embedding (Patch Embedding, PE) and shallow Swin encoders to extract their sensitive primary features (VIS is sensitive to cloud texture / optical thickness, and IR is sensitive to cloud top temperature / height):

[0111]

[0112] ,

[0113] The shallow features are extracted by Swin blocks (window self-attention and shift window self-attention Alternation:

[0114]

[0115] ,

[0116] In the design, the method of alternately using window self-attention and shift window self-attention in the shallow Swin encoder is adopted: the former focuses on local details and edge contours, and the latter establishes long-range dependencies across windows by window displacement, gradually expanding the receptive field. Compared with convolutional networks with fixed receptive fields, this hierarchical attention achieves a better balance between "clearing local strong convective core details" and "perceiving wide-area weather system organization form", effectively alleviating the problems of strong echo position drift and texture over-smoothing, and also laying a more robust feature foundation for subsequent multi-scale modeling.

[0117] S2.3 Mid-term cross-modal fusion (explicit multi-modal)

[0118] After VIS and IR form representations with physical meaning in the shallow layer, we introduce cross-modal fusion in the middle term to align the features of the two routes and integrate them at the semantic level.

[0119] The mid-term fusion is exactly in the stage of "features have been shaped but not yet solidified", which can not only explicitly align important areas of the two routes (such as using cross-modal attention to align the cold top strong convective area of IR with the cloud tower texture edge of VIS), but also adaptively adjust the contribution of the two modalities at the sample level through gating weighting.

[0120] After the shallow representation, align and fuse the features of the two routes to form a unified representation of a single path, avoiding redundant calculations in the deep layer and explicitly modeling complementarity. Let the alignment operator be (e.g. channel calibration / position encoding alignment), and the fusion operator be :

[0121] ,

[0122] ,

[0123] ,

[0124] One implementation of gating weighting is:

[0125]

[0126] ,

[0127] where is the element-wise multiplication, is the Sigmoid, For global average pooling, For concatenation. Cross-Attention (CA) can also be used to enhance mutual information:

[0128]

[0129]

[0130]

[0131] ,

[0132] where is the key / query dimension, is the learnable projection parameter. After fusion, only goes into the deep backbone.

[0133] When the day-night transition, channel quality difference or weather type changes, the network can still maintain the robustness of cross-modal mapping and preferentially retain information more valuable for deep convection discrimination.

[0134] S2.4 Multi-scale representation of backbone encoder and bottleneck

[0135] The fused features enter the hierarchical backbone encoder, which gradually expands the receptive field and increases the channel number through multi-level patch merging, and then stacks Swin blocks (W-MSA + SW-MSA) to jointly model local and global dependencies:

[0136] ,

[0137] ,

[0138] The bottleneck layer continues to apply self-attention to strengthen the large-scale context (the organization structure of the frontal zone, rain band and mesoscale convective complex), and obtains .

[0139] This process is equivalent to building a "scale pyramid": shallow layers retain local textures of cloud edges and strong echoes, middle layers gradually gather the organizational relationships between convective cells, and deep layers and bottlenecks focus on larger-scale background fields and systematic structures (such as frontal zones, rain bands and mesoscale convective complexes). This bottom-up multi-scale representation can simultaneously stabilize the positioning of strong echoes and the coherence of large-scale patterns, avoiding the imbalance between "only fine and not holistic" or "only holistic and not fine", thereby providing semantic sufficient and hierarchical clear features for subsequent decoding reconstruction.

[0140] S2.5 Decoder and cross-scale skip connection

[0141] The decoder upsamples the feature maps by Patch Expanding and connects the corresponding scale features across layers (Skip-Connection) to fuse the deep semantic information and shallow edge / textural details:

[0142]

[0143]

[0144] where is the light-weight convolutional / attention fusion unit. Its role is to directly inject the same-scale edge and detail preserved in the encoding stage into the current decoding layer, so that the deep semantic information and shallow spatial details are fused at each scale.

[0145] Finally, the synthetic weather radar reflectivity estimate is generated by linear projection:

[0146]

[0147] Compared with simple upsampling, this approach can significantly reduce the blurring of boundaries and the loss of texture, maintain the intensity gradient of strong echo centers and the sharp transition of convective boundaries, and at the same time, allow continuous structures such as rain bands and arcs to be preserved during upsampling, thereby improving the visual and structural consistency of the final synthetic weather radar echoes.

[0148] S2.6 Training objective and loss function

[0149] To optimize the numerical approximation, structural similarity, and the depiction of key areas such as strong echoes, we use a weighted loss composed of four terms: absolute error (L1), structural similarity (SSIM), focal loss for strong echo masks (Focal), and gradient consistency term. The L1 term provides a robust pixel-level baseline constraint that is not easily swayed by a small number of extreme errors; the SSIM term evaluates the similarity between the synthetic image and the ground truth from three dimensions: brightness, contrast, and structure, allowing the network to focus on "form and organization" rather than "individual pixels," which is particularly important for continuous structures such as rain bands and arcs; the focal loss targets the sparse but business-sensitive area of strong echoes (threshold set to ≥ 35 dBZ), adjusting the difficulty sample weight and class imbalance coefficient to appropriately "press" the model's learning focus towards the strong convective core, improving recognition rate and positioning accuracy; the gradient consistency term constrains the generated results from the edge and texture perspective, suppressing excessive smoothing and making strong echo centers and convective boundaries clearer. In actual training, we provide a set of practical weight configurations and adaptively fine-tune them based on the validation set performance: when the business places more emphasis on strong convective detection, increase the Focal weight or moderately lower the strong echo threshold; when the structural continuity is insufficient, appropriately increase the proportion of SSIM and gradient terms. The structure of this loss is shown in the following formula:

[0150] ,

[0151] where is the true value CREF (dBZ, linearly scaled to training), , is the binary mask of strong echoes ; is the loss weight, is the Focal loss hyperparameter.

[0152] With this four-dimensional coupling of the objective function design, the model can achieve more balanced improvement in strong convective indicators such as CSI and POD, and structural indicators such as SSIM, while maintaining good generalization stability across regions and seasons.

[0153] S2.7 Model adjustment and correction based on the validation set

[0154] To balance the synthetic echoes in the three aspects of “numerical approximation, structural consistency, and strong echo identification”, this method performs constrained model adjustment and post-processing on the validation set. First, take CSI@35 dBZ as the main indicator, and link POD, FAR, SSIM, and RMSE as auxiliary indicators. The model selection criteria are “main indicator first, auxiliary indicator bottom”: when CSI is close to 1, prefer the model with higher SSIM and lower FAR to avoid false alarm expansion or structural degradation caused by only pursuing detection rate. Under this criterion, we do a narrow search on a small number of key hyperparameters, and set the range based on prior experience to avoid overfitting the validation set.

[0155] At the decision level, we calibrate the correspondence between the “strong echo threshold and probability threshold” on the validation set. The specific method is to scale the network output of the strong echo mask to make the output probability consistent with the true frequency, and then select the probability threshold corresponding to 35 dBZ based on the reliability curve (usually between 0.4-0.6, with slight changes depending on the sample distribution). This way, we can stabilize the meaning of “probability as confidence” without changing the order, making the threshold migration more stable across different months / regions. If the calibration sample is small, use order-preserving regression instead to reduce the risk of overfitting.

[0156] For uncertainty, we evaluate the benefit and cost of light TTA (test-time augmentation) on the validation set: 4-6 weakly perturbed forward passes on the same input, averaging the output and standard deviation as the confidence. If we observe stable improvement in CSI and SSIM with acceptable latency, we keep this step in the inference pipeline; otherwise, we turn off TTA and only keep the pixel-level confidence estimation based on a single forward pass to hint low-confidence regions. For patterns identified as "out-of-distribution" in the validation set, we only enable hinting and mild shrinking, without strong modification to avoid hurting real severe developing processes.

[0157] Finally, to avoid "overfitting" on the validation set, we use a narrow range and low degree of freedom strategy for all hyperparameter tuning: first, only touch a small number of knobs that are highly related to business performance (loss weights, strong echo threshold, amplitude mapping, temporal consistency threshold); second, only give 2-3 candidate values for each, and check the consistency of the results through a small validation slice; third, repeat 2-3 times to confirm that the variance of the indicators is not abnormal. After tuning, we use the test set that did not participate in the tuning to make a final evaluation to ensure that the benefits of the changes can be extrapolated, rather than fitting the validation set by chance.

[0158] S3 Synthetic Weather Radar Phase

[0159] S3.1 Inference Pipeline and Timing Assembly

[0160] During business operation, the multi-channel satellite observations pre-processed by S1 are sent to the trained model in batches according to the 512x512 standard sub-block for forward inference. To avoid the stitching marks caused by block boundaries, the sub-block uses a sliding window strategy with an overlap (128-pixel "guard band"), and after inference, it is filled back to the full-frame picture by weighted fusion or center cropping. For cases where there is a second-level difference between the satellite and the target time, the time registration frame obtained in S1.2 is preferred as the input; if there is a missing frame at a certain time, the adjacent satellite data is used to obtain an approximate input through linear interpolation and simplified advection correction, ensuring the continuity on the time axis. All network outputs will first be de-normalized and physically dimensioned (mapped back to dBZ), and then written back to the business product according to the unified projection grid for direct comparison and use with the measured weather radar or numerical model field.

[0161] S3.2 Post-processing and Consistency Constraints This step should give the calculation process, give an example

[0162] Model output needs to suppress isolated noise and blocking effects while preserving details. To this end, the process introduces three types of light post-processing:

[0163] First, perform morphological denoising and small connected component filtering on low-amplitude isolated pixels to avoid false positives. Set the amplitude threshold to For each pixel (i represents the pixel position), the following operations are performed:

[0164] ,

[0165] where NA represents the pixel excluding the pixel.

[0166] Secondly, a mild smoothing with edge preservation is adopted, which uses limited times of bilateral filtering to weaken the block stitching marks while not destroying the sharpness of strong echo boundaries. The formula of bilateral filtering is as follows:

[0167] ,

[0168] where, is the amplitude of pixel i, is the amplitude of pixel j. controls the influence range of the spatial neighborhood, controls the influence of the amplitude difference.

[0169] Thirdly, the time consistency constraint is performed, that is, the synthetic echo of the previous time is registered according to the data of the adjacent time, and a consistency test is performed with the current output, and the area with non-physical mutation is inhibited or returned to the conservative value. For the time consistency registration of the synthetic echo of the adjacent time, first, the translation vector between the adjacent frames is calculated , assuming that the synthetic echo of the previous time is , the synthetic echo of the current time is , and the translation vector is used to register the previous time echo to obtain the registered echo :

[0170] ,

[0171] Then, the output of the current time is compared with the registered previous time echo:

[0172] ,

[0173] If the difference is greater than the set threshold, the abnormal area is inhibited or returned:

[0174] ,

[0175] For extreme echo values, reasonable upper and lower limits of reflectivity are set (the lower limit is 0-10 dBZ, and the safe upper limit is 60-65 dBZ), which avoids pollution of abnormally high values and retains the peak structure of strong convective cores.

[0176] S3.3 Quality control and uncertainty annotation This step should give the calculation process, give an example

[0177] To reduce the instantaneous flicker of adjacent time outputs, a simplified temporal consistency constraint independent of external platforms can be introduced: using the estimated displacement between the satellite fields of adjacent two time instances, the synthetic echoes of the last time instance are adjusted to the current time, and the difference with the current output is calculated and a mild penalty or local smoothing is applied to the abnormal abrupt change area:

[0178] First, set the satellite fields of adjacent two time instances as and and get the displacement vector :

[0179] ,

[0180] where is the result of the last time instance echo after displacement adjustment.

[0181] Then, we calculate the difference between the current time output and the adjusted last time instance echo :

[0182] ,

[0183] If , a smoothing or mild penalty is applied to these areas with large differences.

[0184] For areas with large differences, local smoothing can be performed, which can be implemented by a simple weighted average:

[0185] ,

[0186] where is a smoothing coefficient, usually taking a value between 0.5 and 0.7, used to control the weight of the current time output and the adjusted last time instance echo.

[0187] This constraint only works near the model output layer and does not change the network structure, which can improve the temporal smoothness and morphological continuity without compromising the details.

[0188] S3.4 Error review and minimum evaluation set This step should give the calculation process, give an example

[0189] To verify the achievement of the design goals in S2, it is recommended to reserve a minimum evaluation set under the condition of not relying on the system: select several convective cases, calculate the overall error (MAE / RMSE), structural similarity (SSIM), and simple detection indicators under the strong echo threshold (hit rate and false alarm rate of ≥ 35 dBZ).

[0190] First, the synthetic weather radar reflectivity output by the model is given , and the measured weather radar reflectivity is . The entire evaluation process can be described as the following steps:

[0191] (1) Error calculation. First, calculate the error between the model output and the measured weather radar reflectivity, here including MAE and RMSE, respectively:

[0192] ,

[0193] ,

[0194] where N is the number of pixels in the image, and are the model output and measured value of the i-th pixel, respectively.

[0195] (2) Structural similarity (SSIM) calculation. Calculate the structural similarity between the model output and the measured reflectivity, the formula is:

[0196] ,

[0197] where and are the mean values of the model output and the measured value, and are their variances. is the covariance, and are constant terms to avoid zero denominator.

[0198] (3) Strong echo detection (POD and FAR). Calculate the hit rate (POD) and false alarm rate (FAR) for the strong echo area (≥ 35 dBZ).

[0199] (4) Comprehensive evaluation formula. Considering MAE, RMSE, SSIM, POD, FAR, the overall performance of the model can be evaluated by weighted average:

[0200]

[0201] where , , , , are weight parameters reflecting the contribution of each index to the final evaluation result. These weights can be adjusted according to the specific requirements of the task.

[0202] This type of assessment only uses real-time weather radar fields as a reference, and does not require online monitoring or productization processes, making it easy to quickly check whether the model maintains a balance in the three aspects of "numerical, structural, and strong echo" during the research phase.

[0203] S4 Methodology Supplements and Usability Assurance

[0204] S4.1 Uncertainty Estimation and Confidence Level Plotting: This step should provide the calculation process and an example.

[0205] In addition to the synthesized reflectance, a pixel-level confidence map is also provided to indicate the confidence interval of the output and the source of uncertainty. Specifically, a small number of forward inferences are performed by applying a light perturbation to the input at the same time, and the mean and standard deviation of the results are taken: the mean is used as the final synthesized echo, and the standard deviation or quantile difference is used as the uncertainty measure.

[0206] In this process, we lightly perturb the input data at the same time and use these perturbed inputs for a small number of forward inferences (e.g., 4–6). Each forward inference generates a synthetic echo.

[0207] for It will be different.

[0208] Next, we use all the forward inference results to calculate the mean and standard deviation:

[0209] ,

[0210] ,

[0211] in, It is the final synthesized echo. It is a measure of the model's uncertainty at that pixel. This will be the final output. The confidence level will be used to represent the model.

[0212] For our output and confidence level, we will calculate the probability of greater than 35 dBZ in each forward inference:

[0213] ,

[0214] in, (x) is an indicator function, representing whether the current pixel exceeds the threshold of 35 dBZ. The final strong convection probability map. This represents the probability that each pixel belongs to a region of strong convection.

[0215] On this basis, we can distinguish the uncertain areas and make subsequent decisions. Through the confidence map and the strong convection probability map, we can significantly distinguish the areas where the model is "very confident" and the areas where the model is "not very confident". For the areas where the model is not very confident, the subsequent decision can adopt a more conservative strategy, such as increasing the threshold or manual review. For the areas where the model is confident, the threshold can be reduced, and more aggressive inference can be made.

[0216] S4.2 Physical consistency and posterior constraints This step should give the computational procedure, give an example

[0217] To make the output more consistent with meteorological physical intuition, two types of posterior consistency constraints are imposed on the results without changing the network structure:

[0218] First, spatial continuity constraint - perform connected component analysis on the synthesized echo, remove small, isolated and short-lived noise spots, and at the same time perform weak morphological closing operation on narrow and small breaks to restore the continuity of rain bands / arc lines:

[0219] (1) First, we perform connected component analysis on the synthesized echo to remove small and isolated noise regions. Define the minimum connected region size of the synthesized echo as , if the area of a connected region is smaller than the threshold, the region is considered to be noise and is removed. The specific operation is:

[0220] ,

[0221] where is the echo after noise removal.

[0222] (2) For narrow and small echo breaks, morphological closing operation is used to restore continuity. Morphological closing operation can be realized by first expanding and then eroding the image, the formula is:

[0223] where denotes the expansion operation, denotes the erosion operation, is the structure element (such as circle or rectangle), which is used to connect small broken areas.

[0224] Second, strong echo edge sharpness constraint - only use edge preserving filter in the strong echo edge band to suppress the jaggedness caused by interpolation / splicing without damaging the peak. In this process, bilateral filter (Bilateral Filter) is used to preserve the details of the strong echo edge, the formula is:

[0225] ,

[0226] where, Ri, the reflectivity value of the current pixel, Rj, the reflectivity value of the neighboring pixel, with Ri, the reflectivity value of the current pixel,

[0227] Both aim to make the structure more weather radar-like, rather than simply smooth.

[0228] S4.3 Failure modes and boundary conditions

[0229] The method is at higher risk in the following scenarios: VIS texture distortion due to extreme low solar elevation angle, exacerbated IR parallax at edge viewing angles, lack of comparable historical patterns in strongly occluded terrain regions, and rare cloud system organization outside the training distribution. To mitigate the risk, we recommend raising the uncertainty threshold to trigger the “low confidence” flag in these areas (or time periods), and preferentially relying on IR-dominated information.

[0230] Experimental validation

[0231] (1) Data selection. We selected paired samples of Himawari-8 (visible / infrared multi-channel) and ground-based composite reflectivity (CREF) during June-September 2023, unified to 2 km × 2 km resolution, and cut into 512 × 512 sub-blocks. Time and region were strictly split (training / validation / testing ~6:2:2) to ensure no “same process” leakage between different sets. The strong echo evaluation uniformly adopted the operational criterion of ≥ 35 dBZ. For comparability, the three models (UNet, GAN, Swin-UNet) used the same pre-processing and data division; no amplitude posterior correction (such as quantile mapping) was made, only a unified scaling calibration was performed on the strong echo probability to ensure the consistent meaning of the threshold. Each index was reported on the test set with 3 independent repetitions of mean ± standard deviation.

[0232] (2) Index selection.

[0233] CSI@35 dBZ (critical success index, higher is better): comprehensive hit and false alarm.

[0234] POD@35 dBZ (probability of detection, higher is better) and FAR@35 dBZ (false alarm rate, lower is better).

[0235] SSIM (structural similarity, higher is better): measures echo morphology and continuity.

[0236] RMSE (dBZ) (lower is better): measures amplitude error.

[0237] MAE (dBZ) (lower is better): measures the average deviation of amplitude, more robust to outliers than RMSE.

[0238] PSNR (dB) (higher is better): measures the quality of the synthesized echoes, relative to the peak of the real weather radar, the higher / lower of PSNR reflects the relationship between amplitude error and signal-to-noise ratio.

[0239] Overall, “Proposed (ours)” is superior to the three types of baselines in CSI and SSIM, and further reduces FAR while improving POD; MAE / RMSE is the lowest in amplitude error, resulting in higher PSNR. Compared with UNet and GAN, the advantages mainly come from explicit multi-modal mid-fusion and hierarchical self-attention multi-scale representation; compared with Swin-UNet, further fusion / weight design and edge detail modeling make the strong echo core positioning and structural continuity more stable. As shown in Table 1.

[0240] Table 1: Synthesis result indicators

[0241] Model CSI@35 ↑ POD@35 ↑ FAR@35 ↓ SSIM ↑ MAE (dBZ) ↓ RMSE (dBZ) ↓ PSNR (dB) ↑ UNet 0.41 ± 0.01 0.68 ± 0.02 0.32 ± 0.01 0.78 ± 0.01 5.1 ± 0.1 7.2 ± 0.1 20.9 ± 0.1 GAN 0.44 ± 0.01 0.66 ± 0.02 0.29 ± 0.01 0.80 ± 0.01 4.9 ± 0.1 6.9 ± 0.1 23.5 ± 0.1 Swin-UNet 0.53 ± 0.01 0.74 ± 0.01 0.24 ± 0.01 0.86 ± 0.01 4.3 ± 0.1 6.1 ± 0.1 26.6 ± 0.1 Proposed(ours) 0.57 ± 0.01 0.76 ± 0.01 0.22 ± 0.01 0.88 ± 0.01 4.0 ± 0.1 5.7 ± 0.1 27.1 ± 0.1

[0242] Embodiment 2

[0243] The embodiment provides a weather radar reflectivity synthesis system based on a synchronous stationary satellite.

[0244] A computer readable storage medium, wherein a plurality of instructions are stored, the instructions are suitable for being loaded and executed by a processor of a terminal device, and the instructions implement the weather radar reflectivity synthesis method based on the synchronous stationary satellite.

[0245] A terminal device, comprising a processor and a computer readable storage medium, the processor is used to implement instructions; the computer readable storage medium is used to store a plurality of instructions, the instructions are suitable for being loaded and executed by the processor, and the instructions implement the weather radar reflectivity synthesis method based on the synchronous stationary satellite.

[0246] The above are preferred embodiments of the present application, not limited to the protection scope of the present application, therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered in the protection scope of the present application.

Claims

1. A method for synthesizing the reflectivity of weather radar based on geostationary satellites, characterized in that, include: Acquire multi-channel satellite observation data and weather radar data; Data preprocessing is performed based on the acquired satellite multi-channel observation data and weather radar data; Construct a weather radar synthetic model based on an improved Swin-Transformer and U-Net; The constructed weather radar synthetic model is trained. The trained model is used for forward inference to obtain the weather radar reflectivity synthesis. The construction of the weather radar synthetic model based on the improved Swin-Transformer and U-Net includes input consisting of two types of satellite observations: visible light multi-channel VIS and infrared multi-channel IR. The preprocessed visible light and infrared observations are respectively... The true value of the target weather radar reflectivity is Let the side length of the patch be... The initial number of embedded channels is D, and the number of Swin layers is [number missing]. The set of downsampling magnifications is For each input, a Patch Embedding and Swin-Transformer structure is adopted. The two modalities are fed into independent Patch Embedding and Swin-Transformer to extract their respective sensitive primary features, as shown below: Shallow features are formed by Extracting Swin blocks by alternating window self-attention Self-attention of shifted windows To balance details and receptive domain, it can be represented as: ; The construction of the weather radar synthesis model based on the improved Swin-Transformer and U-Net also includes aligning and fusing the features of VIS and IR after shallow representation to form a unified representation of a single path, avoiding deep redundant computation and explicitly modeling complementarity. Let the alignment operator be... The fusion operator is : , , The implementation of gated weighting is represented as: ， in To multiply by each channel, It is Sigmoid. For global average pooling, To perform the splicing, and simultaneously utilize cross-modal attention to enhance mutual information, it is represented as: , ,in For key / query dimensions, The projected parameters are learnable; the fused features enter a hierarchical backbone encoder, which gradually expands the receptive field and increases the number of channels through multi-level downsampling, and then stacks Swing blocks to jointly model local and global dependencies, represented as: , The bottleneck layer continues to apply self-attention to enhance the broader context, resulting in... Finally, the decoder uses patch expanding to upsample step by step and connects across layers with the corresponding scale features of the encoder, fusing deep semantics with shallow edge texture details, as shown below: in As a lightweight convolutional / attention fusion unit, it directly injects the same-scale edges and details preserved in the encoding stage into the current decoding layer, enabling deep semantic information and shallow spatial details to be fused once at each scale. Finally, a synthetic weather radar reflectivity estimate is generated by linear projection. .

2. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 1, characterized in that, The data preprocessing based on the acquired satellite multi-channel observation data and weather radar data includes unifying the spatial resolution and adjusting the size of the input geostationary meteorological satellite multi-channel observation data and the ground-based weather radar integrated reflectivity data. First, a downsampling method is used to unify the spatial resolution of the satellite and weather radar data to the grid level. Let the original image be After being uniformly projected onto the target mesh, it is denoted as For the center of each target pixel Employing resampling operators Bilinear interpolation is expressed as: ， Where, is the interpolation. The average weight of the blocks is then used to cut sub-blocks according to a fixed window, as follows: Then, time registration is performed on the two types of data, assuming the satellite timestamp sequence is... The weather radar timestamp sequence is First, perform nearest neighbor pairing based on threshold determination: And for each choose: At this time, if Then they form a pair When higher time consistency is required, at the weather radar time... Perform linear time interpolation on satellite frames: Let , representing the two most recent satellite frames, at this time: Considering the apparent drift caused by cloud system translation, a simplified advection correction is added before interpolation, assuming the relevant matching displacement is phase. The matching result at this time is: ， To ensure spatial consistency, satellites and weather radars are projected onto a unified grid. Projection mapping is denoted as And requires that the mating blocks meet the upper limit of geometric error: 。 3. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 2, characterized in that, The data preprocessing based on the acquired satellite multi-channel observation data and weather radar data also includes further processing of blank areas in the dataset after completing the time matching of satellite data and weather radar data. This is done by calculating the reflectivity value of each pixel in the weather radar data and removing pixels with reflectivity less than 10 dBZ as blank areas. Let the reflectivity of a weather radar pixel be... Weak echo rejection is implemented using the following formula: ,in To determine the training sample region, the data is then standardized using the maximum-minimum standard. The standardization formula is: ,in, It is the normalized value. These are the original data values. and These are the minimum and maximum values ​​in the dataset, respectively. Finally, after completing the preprocessing steps, the obtained matching sample dataset will be divided into training set, validation set, and test set proportionally.

4. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 3, characterized in that, The constructed weather radar synthetic model is trained using a weighted loss consisting of four terms to simultaneously optimize numerical approximation, structural similarity, and key regions with strong echoes: absolute error L1, structural similarity SSIM, focal loss for strong echo masks, and gradient consistency term. The L1 term provides pixel-level baseline constraints, the SSIM term evaluates the similarity between the synthetic image and the ground truth from three dimensions: brightness, contrast, and structure, the focal loss adjusts the weights of hard samples and the class imbalance coefficient for strong echo regions, and the gradient consistency term constrains the generated results from the perspectives of edges and texture. The structure of the loss is as follows: ， in The true value is CREF. , For strong echo A binary mask; To lose weight, By designing a four-dimensional coupled objective function for the Focal loss hyperparameter, the trained model achieves a more balanced improvement in CSI, POD strong convection index, and SSIM structure index.

5. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 4, characterized in that, The process of using the trained model for forward inference to obtain the composite weather radar reflectivity includes feeding pre-processed multi-channel satellite observations into the trained model in batches according to 512×512 standard sub-blocks for forward inference. To avoid stitching marks caused by block boundaries, the sub-blocks adopt an overlapping sliding window strategy. After inference, the data is backfilled into the full frame using a weighted fusion method. For cases where there is a second-level difference between satellite and target times, time-registered frames are used as input. If a frame is missing at a certain time, the satellite data from adjacent time points is used to obtain an approximate input through linear interpolation and simplified advection correction to ensure continuity on the time axis. All network outputs are first denormalized and physical dimension restored, and then written back to the business product according to a unified projection grid so that they can be directly compared with measured weather radar or numerical model fields.

6. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 5, characterized in that, The process of using the trained model for forward inference to obtain the weather radar reflectivity synthesis also includes suppressing isolated noise and block artifacts based on the model output while preserving details. Three types of lightweight post-processing are introduced: morphological denoising and small connected component filtering are performed on low-amplitude isolated pixels to avoid misclassifying random noise; edge-preserving light smoothing is used to reduce block stitching artifacts without compromising the sharpness of strong echo boundaries; temporal consistency constraints are applied, i.e., the synthesized echo from the previous time period is registered based on data from adjacent time periods, and consistency is checked with the current output, suppressing or reverting to conservative values ​​for regions with non-physical abrupt changes; and measures are taken to reduce phase transitions. The instantaneous flickering of adjacent time-series outputs introduces a simplified time consistency constraint that is independent of external platforms: by utilizing the estimated displacement between satellite fields of two adjacent time-series, the synthetic echo of the previous time-series is distorted to the current time, the difference is calculated with the current output, and a mild penalty or local smoothing is applied to anomalous abrupt regions. The constraint only works near the model output layer without changing the network structure, improving temporal smoothness and morphological continuity without compromising details. To verify the achievement of the design goals, a minimal evaluation set is retained without system dependence: several convection cases are selected, and the overall error, structural similarity, and simple detection indicators under the strong echo threshold are calculated.

7. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 6, characterized in that, The process of using the trained model for forward inference to obtain weather radar reflectivity synthesis also includes acquiring a pixel-level confidence map simultaneously with the synthesized reflectivity. This map indicates the confidence interval and uncertainty sources of the output. A small number of forward inferences are performed on the input at the same time, with the mean and standard deviation of the results taken: the mean is used as the final synthesized echo, and the standard deviation or quantile difference is used as an uncertainty measure. A strong convection probability map is generated for the strong echo threshold, facilitating subsequent decision-making to select different thresholds under different risk tolerances. To ensure the output conforms to meteorological physical intuition, two types of posterior consistency constraints are applied to the results without altering the network structure: spatial continuity constraints, used for connected component analysis of the synthesized echo to remove extremely small, isolated, and short-lived noise specks; and strong echo edge sharpness constraints, used only within the edge band of strong echoes to suppress jaggedness caused by interpolation and splicing without affecting peak values.

8. A weather radar reflectivity synthesis system based on geostationary satellites, executing the weather radar reflectivity synthesis method based on geostationary satellites as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire satellite multi-channel observation data and weather radar data; The preprocessing module is configured to perform data preprocessing based on the acquired satellite multi-channel observation data and weather radar data; The model building module is configured to build a weather radar synthetic model based on the improved Swin-Transformer and U-Net; The model training module is configured to train the constructed weather radar synthetic model. The synthesis module is configured to perform forward inference using the trained model and obtain the weather radar reflectivity synthesis.

Images