Weather radar reflectivity synthesis method and system based on synchronous stationary satellite

Abstract

The invention 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; the method comprises the following steps of: constructing a weather radar synthesis model based on an improved Swin-Transform and a U-Net; performing model training on the constructed weather radar synthesis model; and performing forward reasoning by using the trained model and obtaining weather radar reflectivity synthesis. According to the method, high-quality weather observation data of the areas are provided, the problem of data vacancy of the weather radar is solved, the observation range of the weather radar is remarkably expanded, and particularly, the monitoring capability of a large-scale weather system and severe convective weather is improved.

Description

Technical Field

[0001] This invention relates to the field of weather radar synthesis technology, and in particular to a weather radar reflectivity synthesis method and system based on geostationary satellites. Background Technology

[0002] Severe convective weather is characterized by its suddenness, locality, and destructiveness; accurate monitoring of its development is crucial for disaster prevention and mitigation. Ground-based weather radar, as a key device for identifying and monitoring severe convective weather, can directly reflect the precipitation structure, echo intensity, and location of the convective core within clouds. However, the existing layout of ground-based weather radar stations is limited by factors such as topography, site construction conditions, power supply, and communication. Significant blind spots exist in complex areas such as oceans, plateaus, and mountains, resulting in the inability to obtain continuous and comprehensive weather radar observation data in some areas. Especially in remote or complex terrain areas, the coverage capability of weather radar is severely limited, posing challenges to the timely monitoring and early warning of severe convective weather.

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

[0004] In recent years, deep learning technology has made significant progress in remote sensing image processing and weather forecasting. Deep learning can learn complex nonlinear mapping relationships in a data-driven manner, especially the Transformer model, whose self-attention mechanism demonstrates superior performance in long-range dependency and spatiotemporal feature modeling tasks. However, severe convective weather systems have significant multi-scale characteristics, and current technologies rely solely on stacking independent satellite spectral channels to describe convective structures, which cannot fully characterize these multi-scale features, resulting in limited ability to capture key features such as convective cores, boundary layer clouds, and deep convective cooling.

[0005] Therefore, existing methods combining satellites and weather radar have significant shortcomings in monitoring severe convective weather and require further improvement. Current technologies cannot effectively generate echo data that is highly consistent with real weather radar in areas not covered by weather radar, and their ability to effectively fuse and model multimodal satellite data remains insufficient.

[0006] Furthermore, existing ground-based weather radars suffer from significant observation gaps in ocean, plateau, and remote areas due to limitations in site construction and maintenance. Existing network areas are also hampered by terrain obstruction, resolution degradation due to increased distance, and ground clutter interference, making it difficult to continuously and accurately characterize the three-dimensional structure of deep convection and echo evolution. In contrast, while geostationary meteorological satellites offer advantages in wide-area and quasi-continuous observation, their visible / infrared channels only reflect cloud top radiation characteristics. This results in significant cross-modal differences in physical dimensions and information sensitivity compared to the comprehensive reflectance (CREF, dBZ) of ground-based weather radars. Directly utilizing stacked independent satellite spectral channels makes it difficult to recover strong echo centers, convective boundaries, and precipitation continuity. Simultaneously, inconsistencies in temporal sampling, projection geometry, and spatial resolution between satellites and weather radars make sample pairing and spatiotemporal alignment difficult, further exacerbating the instability and accumulated biases in cross-modal mapping. Traditional convolutional neural network-based solutions are limited by a fixed receptive field and lack the ability to simultaneously model multi-scale organizational structures (from convective cells to mesoscale convective systems). They are prone to problems such as strong echo position drift, excessive texture smoothing, and inaccurate characterization of strong convective cores (such as regions ≥35 dBZ). Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method and system for synthesizing weather radar reflectivity based on geostationary satellites.

[0008] In a first aspect, the present invention provides a weather radar reflectivity synthesis method based on geostationary satellites, which adopts the following technical solution: A method for synthesizing the reflectivity of weather radar based on geostationary satellites, comprising: 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.

[0009] Furthermore, 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: .

[0010] Furthermore, the data preprocessing based on the acquired satellite multi-channel observation data and weather radar data also includes, after completing the time matching of satellite data and weather radar data, further processing of blank areas in the dataset. This is achieved by calculating the reflectivity value of each pixel in the weather radar data and removing pixels with a 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.

[0011] Furthermore, 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: .

[0012] Furthermore, the construction of the weather radar synthesis model based on the improved Swin-Transformer and U-Net also includes aligning and fusing the two features after shallow representation of VIS and IR 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:

[0013] ,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 PatchExpanding 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. .

[0014] Furthermore, 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.

[0015] Furthermore, the process of using the trained model for forward inference to obtain the weather radar reflectivity synthesis 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 registration 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.

[0016] Furthermore, 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. This involves introducing three types of lightweight post-processing: performing morphological denoising and small connected component filtering on low-amplitude isolated pixels to avoid misjudging random noise; employing edge-preserving smoothing to reduce block stitching artifacts without compromising the sharpness of strong echo boundaries; and implementing temporal consistency constraints, i.e., registering the synthesized echo from the previous time interval based on data from adjacent time intervals and performing consistency checks with the current output, suppressing or reverting to conservative values ​​for regions exhibiting non-physical abrupt changes. To reduce the instantaneous flicker of outputs in adjacent time intervals, a simplified temporal consistency constraint that does not rely on external platforms is introduced: by using the estimated displacement between satellite fields in two adjacent time intervals, the synthetic echo from the previous time interval is adjusted to the current time interval, 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 relying on the system: several convection cases are selected, and the overall error, structural similarity, and simple detection indices under the strong echo threshold are calculated.

[0017] Furthermore, the process of using the trained model for forward inference to obtain the weather radar reflectivity synthesis also includes acquiring a pixel-level confidence map simultaneously with the synthesized reflectivity. This map is used to indicate 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 strong echo edge band to suppress jaggedness caused by interpolation and splicing without affecting the peak value.

[0018] Secondly, a weather radar reflectivity synthesis system based on geostationary satellites includes: 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.

[0019] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned weather radar reflectivity synthesis method based on geostationary satellites.

[0020] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned weather radar reflectivity synthesis method based on geostationary satellites.

[0021] In summary, the present invention has the following beneficial technical effects: (1) Overcoming the spatial limitations of weather radar observation: This invention successfully fills the gaps in ground-based weather radar coverage by utilizing geostationary satellite data, especially in ocean, plateau, and remote areas. Traditional ground-based weather radar systems cannot cover these areas due to geographical limitations, resulting in missing or incomplete meteorological data. By combining satellite and weather radar data, this invention provides high-quality weather observation data for these areas, solving the problem of missing weather radar data and significantly improving the observation range of weather radar, especially its ability to monitor large-scale weather systems and severe convective weather.

[0022] (2) Improving accuracy by leveraging the global modeling capabilities of Transformer: Compared with traditional convolutional neural networks (CNNs), this invention adopts a multi-scale encoding structure based on Swin-Transformer, breaking through the limitation of the local receptive field of CNN models. Traditional CNNs can only process image information in local areas, while Transformer, through window self-attention and cross-window attention mechanisms, can model the spatial dependencies of cloud structures at a global scale. Through this method, this invention significantly improves the generation of strong echoes in strong convection regions (especially ≥35 dBZ), enhancing the realism of weather radar echoes.

[0023] (3) Multimodal fusion improves the accuracy of weather radar synthesis: This invention innovatively processes the visible light and infrared channels as two independent modes, uses a mode-specific encoder to extract differentiated information, and performs deep semantic fusion in the middle stage. Compared with the simple "stacked input channels" method, this multimodal fusion method can significantly strengthen the mapping relationship between satellite data and weather radar composite reflectivity, thereby improving the accuracy of synthetic weather radar. Through this method, this invention 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. Attached Figure Description

[0024] Figure 1 This is an overall flowchart of the multi-scale artificial intelligence method based on geostationary satellite synthetic weather radar implemented in this invention; Figure 2 This is a schematic diagram of the data preprocessing process for implementing the multi-scale artificial intelligence method of geostationary satellite synthetic weather radar according to the present invention; Figure 3 A model diagram illustrating the multi-scale artificial intelligence model based on geostationary satellite synthetic weather radar for implementing this invention; Figure 4 This is a rendering of the synthetic weather radar implemented in this invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings. Example

[0026] Definitions: Geostationary satellite.

[0027] Weather radar synthesis.

[0028] Reference Figure 1 This embodiment of a weather radar reflectivity synthesis method based on geostationary satellites includes: This embodiment uses a geostationary satellite synthetic weather radar. The meteorological geostationary satellites include, but are not limited to, China's Fengyun-2, China's Fengyun-4, Japan's Himawari-8, and the US GOES-R series.

[0029] S1 Data Preprocessing Stage In the data preprocessing stage, this invention 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.

[0030] S1.1 Downsampling and Cutting Steps First, the spatial resolution and size of the input multi-channel observation data from the geostationary meteorological satellite Himawari-8 (H8) and the ground-based weather radar composite reflectivity (CREF) data are unified. Specifically, a downsampling method is first used to unify the spatial resolution of the satellite and weather radar data to the grid level. At this point, the resolution is 2km × 2km. This operation can effectively reduce the data size while ensuring the spatial correspondence between data pixels. Next, 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 or block averaging): Where, is the interpolation. Average block weight. Then cut sub-blocks according to a fixed window: At this point, the actual coverage area 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 input samples with consistent resolution and efficient computation for subsequent deep learning model training.

[0031] S1.2 Time Difference Matching Steps To ensure temporal consistency between satellite data and weather radar data, this invention performs time registration on the two types of data. Specifically, under the condition that the time error is less than 3 minutes, satellite data frames are matched frame by frame with weather radar echo frames to ensure the synchronization of the dataset in time.

[0032] Let the satellite timestamp sequence be 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 .

[0033] 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: , If the apparent drift caused by cloud system translation is considered, a simplified advection correction is added before interpolation, assuming the relevant matching displacement is phase. .

[0034] At this point, the matching result 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: in Pick Within one pixel (approximately) ).

[0035] The goal of this step is to minimize the shifts in cloud position and differences in precipitation evolution caused by observation time differences, so as to ensure temporal consistency and physical comparability in the subsequent model learning process.

[0036] S1.3 Blank Data Removal Steps After completing the time matching of satellite data and weather radar data, the blank areas in the dataset are further processed. The reflectivity value (in dBZ) of each pixel in the weather radar data is calculated, and pixels with reflectivity less than 10 dBZ are considered blank areas and removed.

[0037] We assume the reflectivity of weather radar pixels is... (Unit: dBZ) Weak echo rejection is implemented, and the rejection method is as follows: , in This is the training sample region.

[0038] This step effectively removes areas with no precipitation or weak echoes from weather radar data, suppresses ground noise, weak echoes, and background blankness interference, thereby ensuring that the training dataset mainly contains strong convection and precipitation structure information with significant meteorological implications. This is crucial for improving model training efficiency and convergence stability.

[0039] S1.4 Maximum-Minimum Standardization Steps To improve the stability and convergence of model training, this invention performs maximum-minimum standardization on the data. The standardization formula is as follows: , in, It is the normalized value. These are the original data values. and These are the minimum and maximum values ​​in the dataset, respectively. This standardization method unifies all input data to the same scale, ensuring the stability of the training process and accelerating convergence.

[0040] S1.5 Dataset Partitioning Steps After completing all the preprocessing steps described above, the resulting matched sample dataset will be proportionally divided into training, validation, and test sets. The training set is used for model parameter learning, the validation set is used to monitor the model's generalization performance and prevent overfitting, and the test set is used for final model performance evaluation. This division ensures the model's performance on different datasets and avoids overfitting through the validation and test sets.

[0041] Building a multi-scale artificial intelligence model for the S2 geostationary satellite synthetic weather radar In the second phase of this invention, a multi-scale artificial intelligence model is primarily established using deep learning methods to synthesize satellite observation data with weather radar data. This model, based on the Swin-Transformer and U-Net architecture, aims to achieve high-precision weather radar reflectivity synthesis.

[0042] S2.1 Input and Marking The input for this stage consists of two types of satellite observations: visible light (VIS) multichannel and infrared (IR) multichannel. These two types are naturally complementary in terms of physical sensitivity—VIS is better at characterizing cloud texture, optical thickness, and boundary details, while IR is more sensitive to cloud top brightness temperature and cold top depth convection. To fully utilize this complementarity, we treat VIS and IR as two independent tensor input models, processing them separately early in the feature extraction process.

[0043] Let the preprocessed visible light and infrared observations be respectively The true reflectivity of the target weather radar is... (Unit: dBZ, often linearly scaled to [0,1] during training).

[0044] 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 .

[0045] This organizational approach avoids completely mixing scale differences and noise characteristics at the beginning, reducing the learning burden on the underlying network. It allows the network to first extract highly discriminative features from its "familiar" modalities and then fuse them at appropriate levels, thereby improving the stability and interpretability of cross-modal mapping. It can adapt and play an advantage, especially in situations where nighttime IR is dominant or daytime VIS is dominant.

[0046] S2.2 Modality-Specific Coding and Patch Embedding For each input path, we employ a structure of "Patch Embedding + Shallow Swin-Transformer Encoding". Patch embedding divides the large-format satellite image into fixed-size blocks and linearly maps them to the feature space, which reduces computational overhead and provides appropriate granularity for the subsequent self-attention mechanism.

[0047] The two modalities are fed into independent Patch Embedding (PE) and shallow Swin encoders respectively to extract their respective sensitive primary features (VIS is sensitive to cloud texture / optical thickness, and IR is sensitive to cloud top temperature / altitude): , Shallow features are formed by Swin block extraction (window self-attention) Self-attention of shifted windows alternately): , The design employs a shallow Swin encoder that alternates between window self-attention and shifted window self-attention: the former focuses on local details and edge contours, while the latter establishes long-range dependencies across windows through window displacement, gradually expanding the receptive field. Compared to convolutional networks with fixed receptive fields, this hierarchical attention achieves a better balance between "clearly seeing the core details of local strong convection" and "perceiving the organizational morphology of wide-area weather systems," effectively alleviating the problems of strong echo position drift and excessive texture smoothing, and laying a more robust feature foundation for subsequent multi-scale modeling.

[0048] S2.3 Mid-term cross-modal fusion (explicit multimodal) After VIS and IR form physically meaningful representations at the shallow level, we introduce cross-modal fusion in the middle stage to integrate the two features at the semantic level after aligning them.

[0049] Mid-term fusion is precisely at the stage where "features have taken shape but have not yet solidified". It can explicitly align the important regions of the two paths (for example, using cross-modal attention to align the cold top strong convection region of IR with the cloud tower texture edge of VIS), and can also adaptively adjust the contribution of the two modes at the sample level through gating weighting.

[0050] After shallow representation, the two feature paths are aligned and fused to form a unified representation of a single path, avoiding deep redundant computation and explicitly modeling complementarity. Let the alignment operator be... (e.g., channel calibration / position encoding alignment), the fusion operator is : , , , One implementation of gated weighting is as follows: , in To multiply by each channel, It is Sigmoid. For global average pooling, For splicing. Cross-modal attention (CA) can also be used to enhance mutual information: , in For key / query dimensions, These are learnable projection parameters. Only those parameters are retained after fusion. Enter the deep trunk.

[0051] Even during day-night transitions, channel quality variations, or weather changes, the network can maintain robustness in cross-modal mapping and prioritize the retention of information more valuable for deep convection discrimination.

[0052] S2.4 Multi-scale characterization of the backbone encoder and bottleneck The fused features enter a hierarchical backbone encoder, where the receptive field is gradually expanded and the number of channels is increased through multi-level downsampling (Patch Merging). Subsequently, Swin blocks (W-MSA + SW-MSA) are stacked to jointly model local and global dependencies. , , The bottleneck layer continues to apply self-attention to enhance the large-scale context (organizational structure of frontal zones, rainbands, and mesoscale convective complexes), resulting in... .

[0053] This process is equivalent to constructing a "scale pyramid": the shallow layer preserves the cloud edges and local textures of strong echoes; the middle layer gradually converges the organizational relationships between convective cells; and the deep layer and bottlenecks focus on the background field and systemic structures over a larger area (such as the morphology of fronts, rainbands, and mesoscale convective complexes). This bottom-up multi-scale representation can simultaneously stabilize the localization of strong echoes and the coherence of large-scale morphology, avoiding the imbalance of "only detailed but not holistic" or "only holistic but not detailed," thus providing semantically sufficient and hierarchically clear features for subsequent decoding and reconstruction.

[0054] S2.5 decoder and cross-scale jump connection The decoder uses patch expanding to upsample step by step and skips connections across layers with corresponding scale features from the encoder to fuse deep semantics with shallow edge / texture details. in This is a lightweight convolutional / attention fusion unit. Its function is to directly inject the same-scale edges and details preserved in the encoding stage into the current decoding layer, so that deep semantic information and shallow spatial details are fused at each scale.

[0055] Finally, a synthetic weather radar reflectivity estimate is generated from linear projection: Compared to simple upsampling, this method can significantly reduce boundary blurring and texture loss, maintain the intensity gradient of the strong echo center and the sharp transition of the convective boundary, and at the same time ensure that continuous structures such as rainbands and arc lines are not destroyed during the upsampling process, thereby improving the visual and structural consistency of the final synthetic weather radar echo.

[0056] S2.6 Training Objective and Loss Function To simultaneously optimize the characterization of key regions such as numerical approximation, structural similarity, and strong echoes, we employ a weighted loss consisting of four terms: absolute error (L1), structural similarity (SSIM), focal loss for strong echo masks, and gradient consistency. The L1 term provides robust pixel-level baseline constraints, making it less susceptible to being swayed by small extreme errors. The SSIM term evaluates the similarity between the synthesized image and the ground truth from three dimensions: brightness, contrast, and structure, shifting the network's focus from "single pixels" to "morphology and organization," which is particularly crucial for continuous structures such as rainbands and curved lines. The focal loss targets the sparse yet most business-sensitive region of strong echoes (threshold set at ≥35 dBZ), adjusting the weights of hard samples and the class imbalance coefficient to appropriately "push" the model's learning focus towards the core of strong convection, improving recognition rate and localization accuracy. The gradient consistency term constrains the generated results from the perspectives of edges and texture, suppressing excessive smoothing and making the center of strong echoes and convection boundaries clearer. In actual training, we provide a workable weight configuration and adaptively fine-tune it based on validation set performance: when the business prioritizes strong convection detection, we increase the Focal weight or moderately lower the strong echo threshold; when structural continuity is insufficient, we appropriately increase the proportion of SSIM and gradient terms. The structural representation of this loss is as follows: , in The true value CREF(dBZ, linearly scaled to) train), , For strong echo ( A binary mask; To lose weight, The Focal loss hyperparameter.

[0057] By designing an objective function that couples numerical values, structure, key regions, and details, the model can achieve a more balanced improvement in strong convection indicators such as CSI and POD, as well as structural indicators such as SSIM, while maintaining good cross-regional and cross-seasonal generalization stability.

[0058] S2.7 Model Adjustment and Correction Based on Validation Set To achieve a balance between numerical approximation, structural consistency, and strong echo identification in the synthesized echo, this method performs constrained model tuning and posterior correction on the validation set. First, CSI@35 dBZ is used as the primary metric, while POD, FAR, SSIM, and RMSE are used as secondary metrics. A model selection criterion of "primary metric priority, secondary metric fallback" is adopted: when CSI values ​​are close, models with higher SSIM and lower FAR are prioritized to avoid false alarm inflation or structural degradation caused by solely pursuing detection rate. Under this criterion, a narrow-range search is performed on a few key hyperparameters, with the range set based on prior experience to avoid overfitting the validation set.

[0059] At the decision-making level, we perform validation set calibration on the correspondence between the "strong echo threshold and probability threshold." Specifically, we scale the network output of the strong echo mask to ensure the output probability matches the true frequency. Then, based on the reliability curve, we select the probability threshold corresponding to 35 dBZ (typically between 0.4 and 0.6, varying slightly with sample distribution). This stabilizes the meaning of "probability equals confidence" without altering the ranking, resulting in more stable threshold migration across different months / regions. If the calibration sample size is small, we use ordinal-preserving regression instead to reduce the risk of overfitting.

[0060] To address uncertainty, we evaluate the benefits and costs of lightweight TTA (Test-Time Augmentation) on the validation set: perform 4–6 weakly perturbed forward passes on the same input, taking the mean as the final output and the standard deviation as the confidence level. If a stable improvement in CSI and SSIM is observed with acceptable latency, this step is retained in the inference process; otherwise, TTA is disabled, and only the cell-level confidence estimate based on a single forward pass is retained to indicate low-confidence regions. For morphologies identified as "out-of-distribution" in the validation set, only hints and mild contraction are enabled, without strong morphological corrections, to avoid unnecessarily affecting the actual dramatic development process.

[0061] Finally, to avoid "over-correction" on the validation set, all parameter tuning employed a narrow-range, low-degree-of-freedom strategy: first, only a few knobs highly correlated with business performance were touched (loss weight, strong echo threshold, amplitude mapping, time consistency threshold); second, only 2-3 candidate values ​​were given for each item, and the consistency of the results was checked once through a small validation shard; third, the indicator variance was repeated 2-3 times with a fixed random seed to confirm that it was not abnormal. After tuning, an overall evaluation was performed using a test set that was not involved in parameter tuning to ensure that the benefits of the changes could be extrapolated, rather than being a random fit to the validation set.

[0062] S3 Synthetic Weather Radar Phase S3.1 Reasoning Flow and Timing Assembly During operation, multi-channel satellite observations preprocessed by S1 are fed into the trained model in batches according to standard 512×512 sub-blocks for forward inference. To avoid stitching marks caused by block boundaries, the sub-blocks adopt an overlapping sliding window strategy (a 128-pixel "protective band"), and are then backfilled into the full frame using weighted fusion or center cropping after inference. For cases where there is a second-level difference between satellite and target times, the time registration frame obtained in S1.2 is used as the input first; if a frame is missing at a certain time, the approximate input is obtained by linear interpolation and simplified advection correction using satellite data from adjacent time points to ensure continuity on the time axis. All network outputs are first denormalized and physical dimension restored (mapped back to dBZ), and then written back to the operational product according to a unified projection grid for direct comparison with measured weather radar or numerical model fields.

[0063] S3.2 Post-processing of results and consistency constraints: This step should provide the calculation process and an example. The model output needs to suppress isolated noise and block artifacts while preserving details. To this end, the workflow introduces three types of lightweight post-processing: Firstly, morphological denoising and small connected component filtering are performed on low-amplitude isolated pixels to avoid misclassifying random noise points. An amplitude threshold is set at... For each pixel (Where i represents the cell position), perform the following operations: , Where NA indicates that the pixel has been removed.

[0064] Secondly, a slight smoothing with edge preservation is employed, utilizing a finite number of bilateral filters to reduce block splicing artifacts without compromising the sharpness of strong echo boundaries. The formula for bilateral filtering is as follows: , in, Let i be the amplitude of pixel i. Let j be the amplitude of pixel j. Control the influence range of the spatial neighborhood. The impact of controlling amplitude differences.

[0065] Third, temporal consistency constraints are applied. This involves registering the synthesized echo from the previous time frame with data from adjacent time frames and performing a consistency check with the current output. Regions exhibiting non-physical abrupt changes are suppressed or reverted to conservative values. For temporal consistency registration of synthesized echoes from adjacent time frames, the translation vector between adjacent frames is first calculated. Assuming the synthesized echo of the previous time step is The synthesized echo of the current time is Translation vector Used to register the previous echo, resulting in the registered echo. : , Then, compare the output of the current time step with the registered echo of the previous time step: , If differences If the value exceeds a set threshold, the abnormal region will be suppressed or rolled back. , For extreme echo values, set reasonable upper and lower limits for reflectivity (lower limit 0–10 dBZ, safe upper limit 60–65 dBZ) to avoid abnormally high value contamination while preserving the peak structure of the strong convection core.

[0066] S3.3 Quality Control and Uncertainty Labeling: This step should provide the calculation process and an example. To reduce transient flicker in outputs between adjacent time intervals, a simplified time consistency constraint independent of external platforms can be introduced: using the estimated displacement between satellite fields of two adjacent time intervals, the synthesized echo from the previous time interval is adjusted to the current time, the difference is calculated with the current output, and a mild penalty or local smoothing is applied to regions of anomalous abrupt change. First, define the satellite fields of two adjacent time periods as... and The displacement vector is obtained through optical flow estimation. : , in, This is the result of the previous echo after displacement adjustment.

[0067] Then, we calculate the difference between the current output and the adjusted previous echo. : , if Then, a smooth or mild penalty is applied to these areas with large differences.

[0068] For regions with significant differences, local smoothing can be performed, specifically using a simple weighted average: , in, It is a smoothing coefficient, usually between 0.5 and 0.7, used to control the weight of the current output and the adjusted previous echo.

[0069] This constraint only works near the model output layer, does not change the network structure, and can improve temporal smoothness and morphological continuity without compromising details.

[0070] S3.4 Error Inspection and Minimum Evaluation Set: This step should provide the calculation process and an example. To verify the achievement of the design objectives in S2, it is recommended to retain a minimal evaluation set without relying on the system: select several convection cases and calculate the overall error (MAE / RMSE), structural similarity (SSIM), and simple detection metrics (hit rate and false alarm rate at ≥35 dBZ) under the strong echo threshold.

[0071] First, given the synthetic weather radar reflectivity output by the model. and measured weather radar reflectivity The entire evaluation process can be described in the following steps: (1) Error Calculation. First, calculate the error between the model output and the measured weather radar reflectivity, which includes MAE and RMSE, respectively: , , Where N is the number of pixels in the image. and Let be the model output and the measured value of the i-th pixel, respectively.

[0072] (2) Structural Similarity (SSIM) Calculation. The structural similarity between the model output and the measured reflectance is calculated using the following formula: , in, and The average of the model output and the measured values. and Let be their variance. For covariance, and This is a constant term to avoid the denominator being zero.

[0073] (3) Strong echo detection (POD and FAR). The hit rate (POD) and false alarm rate (FAR) are calculated for the strong echo region (≥35 dBZ).

[0074] (4) Comprehensive evaluation formula. Taking into account MAE, RMSE, SSIM, POD, and FAR, the overall performance of the model can be evaluated by weighted averaging: in, , , , , These are weighting parameters that reflect the contribution of each indicator to the final evaluation result. These weights can be adjusted according to the specific needs of the task.

[0075] 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.

[0076] S4 Methodology Supplements and Usability Assurance S4.1 Uncertainty Estimation and Confidence Level Plotting: This step should provide the calculation process and an example. 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.

[0077] 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. for It will be different.

[0078] Next, we use all the forward inference results to calculate the mean and standard deviation: , , 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.

[0079] For our output and confidence level, we will calculate the probability of greater than 35 dBZ in each forward inference: , 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.

[0080] Based on this, we differentiate between uncertainties and make subsequent decisions. Using confidence plots and strong convection probability plots, we can clearly distinguish between regions where the model is "very confident" and regions where the model is "not very confident." For regions where the model is not very confident, subsequent decisions can adopt more conservative strategies, such as increasing the threshold or performing manual verification. Conversely, for regions where the model is confident, the threshold can be lowered to make more aggressive inferences.

[0081] S4.2 Physical Consistency and Posterior Constraints: This step should provide the calculation process and an example. To make the output more consistent with meteorological physics intuition, two types of posterior consistency constraints are applied to the results without changing the network structure: Firstly, spatial continuity constraints—connected-domain analysis is performed on the synthetic echo to remove extremely small, isolated, and short-lived noise specks, while extremely weak morphological closing operations are applied to narrow breaks to restore the continuity of rainbands / arc lines: (1) First, we analyze the synthesized echo. Perform connected component analysis to remove extremely small and isolated noise regions. Define the minimum connected component size of the synthesized echo as... If the area of ​​a connected region is smaller than this threshold, the region is considered noise and is removed. The specific operation is as follows: , in, This is the echo after noise removal.

[0082] (2) For narrow echo breaks, morphological closing operations are used to restore continuity. Morphological closing operations can be achieved by dilating the image followed by erosion, as shown in the formula:

[0083] in, This indicates an expansion operation. This indicates a corrosion operation. These are structural elements (such as circles or rectangles) used to connect small, broken areas.

[0084] Secondly, strong echo edge sharpness constraint—edge-preserving filtering is used only within the strong echo edge band to suppress jaggedness caused by interpolation / stitching without compromising peak values. In this process, a bilateral filter is used to preserve details at the strong echo edges, as shown in the formula: , in, This represents the reflectance value of the current pixel. The reflectance value of the neighboring pixels. and Parameters used to control amplitude differences and spatial neighborhood effects.

[0085] Both share the goal of creating a structure that is "more like a weather radar," rather than a simple, smooth one.

[0086] S4.3 Failure Modes and Boundary Conditions The method carries a higher risk in the following situations: VIS texture distortion caused by extremely low solar elevation angles, increased IR parallax at edge viewpoints, lack of comparable historical patterns in areas with strong topographic occlusion, and rare cloud formations outside the training distribution. To mitigate this risk, it is recommended to increase the uncertainty threshold to trigger a "low confidence" flag in these areas (or time periods) and prioritize IR-dominated information.

[0087] Experimental verification (1) Data Selection. We selected paired samples of Himawari-8 (visible / infrared multichannel) and ground-based combined reflectance (CREF) from June to September 2023, uniformly scaled to 2 km × 2 km resolution according to S1, and divided into 512 × 512 sub-blocks. Strict temporal and regional separation (approximately 6:2:2 for training / validation / testing) ensured no leakage of "same process" between different sets. Strong echo assessment uniformly adopted an operational caliber of ≥35 dBZ. For comparability, the three models (UNet, GAN, Swin-UNet) used the same preprocessing and data partitioning; no posterior amplitude correction (such as quantile mapping) was performed, only a uniform scaling calibration was applied to the strong echo probability to ensure consistent threshold meaning. The mean ± standard deviation of each metric was reported on the test set for three independent replicates.

[0088] (2) Selection of indicators.

[0089] CSI@35 dBZ (Critical Success Index, the higher the better): Combines hit rate and false alarms.

[0090] POD@35 dBZ (detectability, the higher the better) and FAR@35 dBZ (false alarm rate, the lower the better).

[0091] SSIM (Structural Similarity, the higher the better): Measures the echo pattern and continuity.

[0092] RMSE (dBZ) (lower is better): measures amplitude error.

[0093] MAE (dBZ) (lower is better): measures the average deviation of amplitude, and is more robust to outliers than RMSE.

[0094] PSNR (dB) (higher is better): Measures the quality of synthetic echo. Compared to the peak value of real weather radar, the level of PSNR reflects the relationship between amplitude error and signal-to-noise ratio.

[0095] Overall, "Proposed(ours)" outperforms the three baselines in both CSI and SSIM, and further reduces FAR while improving POD; it also has the lowest MAE / RMSE in terms of amplitude error, resulting in higher PSNR. Compared to UNet and GAN, its advantages mainly come from explicit multimodal mid-term fusion and hierarchical self-attention multi-scale representation; compared to Swin-UNet, further fusion / weight design and edge detail modeling make strong echo core localization and structural continuity more stable. As shown in Table 1.

[0096] Table 1 Synthesis Result Indicators 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

[0097] Example 2 This embodiment provides a weather radar reflectivity synthesis system based on geostationary satellites; A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned weather radar reflectivity synthesis method based on geostationary satellites.

[0098] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the method for synthesizing weather radar reflectivity based on geostationary satellites.

[0099] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

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.

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 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: .

5. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 4, characterized in that, 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. .

6. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 5, 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.

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 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.

8. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 7, 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.

9. The weather radar reflectivity synthesis method based on geostationary satellites according to claim 8, 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.

10. A weather radar reflectivity synthesis system based on geostationary satellites, 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