A radar data blind area filling method and system based on 3D-GANs

Abstract

The application discloses a radar data blind area filling method and system based on 3D-GANs, and the method comprises the following steps: based on the hierarchical characteristics of three-dimensional radar data, a simulated blind area mask is constructed and spliced with corresponding radar reflectivity data to form an input tensor; the input tensor is input into a generator network to output the completed radar volume data; the generator network adopts an improved three-dimensional U-Net structure based on a three-dimensional attention mechanism, and in the encoding layer downsampling, the whole is compressed first, then horizontally compressed to reserve the vertical structure; in the process of the skip connection, a three-dimensional attention module is introduced to weight and control the low-order features output by the encoder; the real radar data and the completion result are compared and discriminated through a three-dimensional discriminator network, and the generator network performance is optimized in combination with the adversarial loss and the reconstruction loss. The application can fill the blind area of three-dimensional radar observation data, keep the echo structure consistency, and improve the detail restoration ability of three-dimensional radar data.

Description

Technical Field

[0001] This invention relates to the fields of meteorological data processing and artificial intelligence, and in particular to a method and system for filling radar data blind spots based on 3D-GANs. Background Technology

[0002] Weather radar, as an active remote sensing tool, utilizes electromagnetic waves to detect the spatial distribution and temporal evolution of precipitation particles in the atmosphere. It has become a crucial means in modern meteorological systems for monitoring severe convective weather, tracking storm structures, estimating precipitation intensity, and identifying extreme disasters. Represented by the Severe Weather Automatic Nowcasting (SWAN) system, this system integrates multi-source data from radar network mosaics, automatic weather stations, satellite remote sensing, and NWP numerical weather prediction. Employing classic algorithms such as COTREC, TITAN, and optical flow, it achieves intelligent early warning capabilities for severe convective weather on a timescale of 0–6 hours and a spatial scale of 1–5 kilometers. However, this system is highly dependent on radar observation data; the quality of the radar data directly affects forecast accuracy and warning timeliness. In practical applications, due to the combined effects of various factors such as terrain obstruction, equipment layout, elevation angle setting, signal attenuation, and abnormal propagation, radar beams cannot achieve seamless coverage of the entire atmospheric volume during propagation, thus forming "radar blind spots" or "beam blockages." This results in problems such as missing, distorted, or even incorrectly estimated values ​​in the echo images, significantly restricting the monitoring and early warning capabilities of the SWAN system.

[0003] Traditional methods for completing radar blind zones mainly fall into two categories: one is estimation methods based on physical modeling and spatial reasoning, such as reconstructing vertical profiles (VPR) using radar data at different elevation angles to estimate the echo intensity of the missing lower-level areas; or combining digital elevation models (DEMs) to calculate beam obstruction and power attenuation, thereby estimating the echo value of the obscured area. The other category is image processing and statistical interpolation methods, including inverse range weighted (IDW), Kriging interpolation, multivariate regression, co-kriging, etc., which use surrounding observations in radar images to interpolate and estimate the missing area. These methods have certain practical effects in echo repair tasks, but they also have obvious limitations, such as strong reliance on prior assumptions, difficulty in handling large-scale complex blind zones, especially insufficient ability to model abrupt structural changes under strong convective weather fields, and susceptibility to texture artifacts and morphological distortions. In addition, some methods are computationally complex, rely on expert experience, and are difficult to deploy operationally, which also limits their application breadth in high-frequency, high-time-efficiency nowcasting systems.

[0004] In recent years, with the rapid development of deep learning in the field of computer vision, breakthroughs have been made in image inpainting and semantic completion tasks, giving rise to the research direction of applying them to radar data completion. However, most current research still focuses on the completion of two-dimensional radar planar images (such as PPI maps and elevation cross-sectional maps). Although it has shown significant potential, it still faces technical bottlenecks such as the difficulty of modeling three-dimensional volumetric data, high training costs, and weak model stability and generalization ability. How to process the real three-dimensional volumetric data formed by radar observations and effectively model its inter-layer correlations, such as the structural evolution of echoes with height, remains the core challenge of current research. Summary of the Invention

[0005] Purpose of the invention: The present invention aims to provide a radar data blind zone filling method and system based on 3D-GANs, which can fill the blind zone region in radar observation data and improve the detail restoration capability of three-dimensional radar data while maintaining the consistency of echo structure.

[0006] Technical solution: The present invention provides a radar data blind spot filling method based on 3D-GANs, comprising the following steps:

[0007] (1) Based on the true blind zone morphology of the three-dimensional radar data in the training data, construct a simulated blind zone mask based on the hierarchical characteristics of the radar data, and embed it layer by layer into the non-missing region of the current radar data;

[0008] (2) The simulated blind zone mask and the corresponding radar reflectivity data are concatenated along the channel dimension to form the input tensor;

[0009] (3) Input the input tensor into the generator network and output the completed radar volume data; the generator network adopts an improved three-dimensional U-Net structure based on the three-dimensional attention mechanism. In the downsampling process of the encoding layer, the overall compression is performed first, and then the horizontal compression is performed to retain the vertical structure; in the process of skip connection between the decoding layer and the output of the encoder of the same scale, a three-dimensional attention module is introduced to perform weighted regulation on the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers;

[0010] (4) The real radar data and the completion results are compared and judged by the three-dimensional discriminator network, and the generator network performance is optimized by combining the adversarial loss and reconstruction loss.

[0011] Preferably, step (1) includes: for layers with large blind area shapes, i.e., layers with less effective data in the original data, generating random irregular shapes to simulate blind area masks and randomly embedding them into regions without missing data; for layers with moderate blind area shapes, constructing structural templates and using a rule of priority placement at the center or searching for feasible positions from the center for embedding; for layers with smaller blind area shapes, i.e., layers with more effective data, constructing structural templates and embedding them into the corresponding regions without missing data in the layers that need to be filled.

[0012] Preferably, the construction of the structural template includes extracting radar data with missing real data from the training data, obtaining the corresponding blind zone mask, extracting connected regions based on the mask using image processing methods, and statistically analyzing their shape and size to construct the structural template.

[0013] Preferably, in step (2), the radar reflectivity data is normalized.

[0014] Preferably, step (3) includes:

[0015] The input tensor is input, and preliminary feature extraction is performed through the initial convolution module to output a low-level feature map;

[0016] Several intermediate feature maps are generated by several downsampling modules. During the downsampling process, the three-dimensional space is first compressed as a whole based on the vertical layering characteristics of the three-dimensional radar data, and then horizontally compressed to retain the vertical structure.

[0017] The deepest feature map is upsampled by the deconvolution module, and a 3D attention module that models the inter-layer relationship is introduced to weight and regulate the low-order features in the skip connections. The weighted encoder features and the decoder upsampled features are concatenated and then input into the convolution module. This process is repeated for several decoding layers to generate a fused feature map.

[0018] The completed radar volume data is generated by the output convolution module;

[0019] The original blind zone mask is used to replace the non-blind zone area in the completed radar volume data to obtain the final output result.

[0020] Preferably, the calculation formula for the three-dimensional attention module is:

[0021]

[0022] in, For decoder upsampling features, For encoder features, , They are respectively Convolution mapping, for Convolution mapping, It is the Sigmoid activation function. For element-wise multiplication, BN is batch normalization.

[0023] Preferably, the adversarial loss in step (4) is a least squares loss function, and the reconstruction loss is an L1 loss function and / or an L2 loss function.

[0024] The present invention discloses a radar data blind spot filling system based on 3D-GANs, used to implement the aforementioned radar data blind spot filling method based on 3D-GANs, comprising:

[0025] The simulated blind zone mask construction module is used to construct a simulated blind zone mask based on the actual blind zone shape of the 3D radar data in the training data and the characteristics of the radar data hierarchy, and embed it into the missing area of ​​the current radar data.

[0026] The input tensor construction module is used to concatenate the simulated blind zone mask and the corresponding radar reflectivity data along the channel dimension to form the input tensor;

[0027] The data completion module is used to input the input tensor into the generator network and output the completed radar volume data. The generator network adopts an improved 3D U-Net structure based on a 3D attention mechanism. During the downsampling process of the encoding layer, overall compression is performed first, and then horizontal compression is performed to retain the vertical structure. During the skip connection between the decoding layer and the encoder output of the same scale, a 3D attention module is introduced to weight and regulate the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers.

[0028] The adversarial training module is used to compare and distinguish real radar data with the completion results through a 3D discriminator network, and optimize the generator network performance by combining adversarial loss and reconstruction loss.

[0029] The present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned radar data blind spot filling method based on 3D-GANs.

[0030] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned radar data blind spot filling method based on 3D-GANs.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0032] 1. This invention transfers the blind zone morphology extracted from actual observation of three-dimensional radar volume data to non-blind zone areas for simulation training, making full use of the "pseudo-missing" data in the existing data as priors to improve the model's generalization ability and robustness to real blind zones.

[0033] 2. Use the missing mask as a conditional input, and after generating the output, fuse the observed values ​​and generated values ​​through a conditional selection method to ensure that the model only reconstructs and repairs in the blind area, effectively avoiding overfitting to the original observation data and unnecessary interference to non-missing areas, and reducing artifacts and structural distortion in the echo map.

[0034] 3. To address the hierarchical characteristics of 3D radar data, such as vertical structure sensitivity and differences in blind zone distribution at different levels, a strategy of constructing a layered simulated blind zone mask is adopted to enhance the diversity of spatial scale and morphological distribution.

[0035] 4. A 3D attention module, LR-Attention, is introduced into the improved 3D U-Net structure. This module is specifically designed for 3D structural fields. By introducing a dedicated height-oriented convolutional structure, it enhances the model's ability to model the physical continuity between layers, significantly reducing unnatural jumps or breaks between upper and lower layers in the completed region. It can fully capture the attenuation relationship of radar reflectivity in the vertical direction and the echo structure distribution, significantly improving the accuracy and spatial consistency of recovering echoes from radar data blind zones. Compared with traditional direct stitching, LR-Attention combines deep semantics with shallow structural attention guidance, which can dynamically suppress noisy regions and improve the accuracy and stability of feature fusion during the completion process. At the same time, the module has a simple structure, extremely low parameter overhead and memory consumption, and is suitable for embedding and deployment in practical radar application systems.

[0036] 5. Improve the effectiveness of nowcasting severe convective weather. This invention can be seamlessly integrated into nowcasting systems such as SWAN, enhancing the spatial integrity of radar data in the lower and middle layers, providing more accurate input for the identification and development trend judgment of small and medium-scale severe convective systems, and is expected to significantly improve the forecasting capability of severe weather such as heavy precipitation, thunderstorms and strong winds within 0-6 hours. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a flowchart illustrating the construction process of the structural template of the present invention.

[0039] Figure 3 This is a mask stitching diagram of the method of the present invention;

[0040] Figure 4 This is a diagram of the generator structure of the method of the present invention;

[0041] Figure 5 This is a structural diagram of the discriminator in the method of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0043] The SWAN system has constructed 21 layers of 3D radar mosaic volume data, which has rich vertical structural features. If this information can be fully utilized and combined with a depth generation model to fill in missing areas, it is expected to break through the existing bottleneck of blind zone processing and further improve the spatial accuracy and temporal stability of mesocyclone identification, echo extrapolation, and heavy precipitation area forecast.

[0044] This embodiment uses 21-layer radar volumetric mosaic data generated in the SWAN system to reasonably fill in observation blind spots caused by factors such as terrain obstruction and limited equipment viewing angle. Figure 1 As shown, the radar data blind spot filling method based on 3D-GANs described in this invention includes a mask-guided input tensor construction stage and a 3D-GANs model construction and training stage.

[0045] The masked guided input tensor construction phase specifically includes the following steps:

[0046] (1) Constructing a simulated blind zone mask. For the real blind zone areas formed in 3D radar data due to external factors such as terrain obstruction and equipment viewing angle limitations, the true mask shape corresponding to the layer to be filled is extracted based on the hierarchical characteristics of the radar data. This shape is then embedded into the non-blind zone of the current research layer to construct a simulated blind zone mask with a spatial morphology consistent with the real blind zone mask and a stable contour structure. This operation is effective for radar blind zones to be filled at different height levels and can utilize the vertical structural features of SWAN 3D data. For example... Figure 2 As shown, this step specifically includes:

[0047] (1.1) For large-area blind zone morphology, that is, the level with less effective data in the original data, such as the first and second layers with less effective data, it is impossible to construct a simulated blind zone with a real shape in the effective data area. Therefore, a random irregular shape blind zone mask is generated to construct a small-scale simulated blind zone mask. The simulated blind zone mask is randomly embedded into the non-missing region to enhance the richness of the training samples in terms of spatial scale and morphological distribution.

[0048] (1.2) For moderate and small blind zone patterns, extract radar data with missing real data from the training data and obtain the corresponding real blind zone mask. Here, 1 represents the missing region (blind zone), 0 represents the normally observed region (non-blind zone), and (D, H, W) represent the number of layers, height, and width of the mask matrix, respectively. Connected regions of the real blind zone mask are extracted using image processing methods, and their shape and size are statistically analyzed to construct a structural template. This is used to construct a simulated blind zone mask for subsequent description of the spatial distribution characteristics of the blind zone. .

[0049] For layers with moderate blind zone morphology, such as layer 3, the structural template embedding process prioritizes center placement. If this is not feasible, feasible positions are searched sequentially from the center to the left, while minimizing overlap with the original blind zone area, to construct simulated blind zone masks for training. These simulated blind zone masks maintain the structural outline and size of the original blind zone to the greatest extent possible. This operation ensures the consistency of the simulated blind zone with the real blind zone in spatial morphology, helping 3D-GANs models learn the structural distribution characteristics of the real blind zone.

[0050] For layers with smaller blind zone shapes and more effective data, such as the 4th layer and above, because of the large amount of effective data, it is possible to support the construction of simulated blind zones with the same shape and size as the blind zone in the non-blind zone area, and embed the corresponding structural template into the corresponding non-missing area in the layer that needs to be filled.

[0051] (2) Simulate the blind zone mask With corresponding radar reflectivity data The input tensor is constructed by concatenating the components along the channel dimension. , where (D, H, W) represent the number of layers, height and width of the mask matrix, respectively, and 2 represents the mask channel and the radar data channel.

[0052] Preferably, to enhance the model's numerical stability and training efficiency, the radar data undergoes numerical normalization and is combined with an explicit mask channel to achieve unified perception of spatial structure and blind zone information, thereby improving the model's blind zone completion capability and guidance effect. Specifically, as follows... Figure 3 As shown:

[0053] First, the radar reflectivity data The following normalization formula is used to uniformly map the input values ​​to the closed interval [-0.5, 1.5], in order to stabilize the input range, accelerate the convergence speed, and enhance the ability to perceive echoes of different intensities during training:

[0054]

[0055] Here, `clip` represents a truncation operation, ensuring that the normalized data falls within the expected range. The normalization function is implemented using the `normalize()` method, with a default maximum value of 70 and a minimum value of 0, adapting to the current data distribution.

[0056] Then the normalized radar data Compared with the simulated blind zone mask generated in step (1.2) The components are concatenated along the channel dimension to form the final network input tensor:

[0057]

[0058] The first channel is the normalized radar reflectivity, and the second channel is the simulated blind zone mask. This input not only preserves complete vertical structure information but also explicitly embeds the location features of missing regions, providing clear conditional guidance for subsequent model generation.

[0059] The 3D-GANs model construction and training phase specifically includes the following steps:

[0060] (3) The input tensor is input into the generator network, and the completed radar volume data is output. The generator network adopts an improved 3D U-Net structure based on a 3D attention mechanism. It includes 5 encoding layers and 5 decoding layers; wherein, the encoding layer is used for spatial compression and feature extraction through 3D convolutional blocks, BatchNorm and LeakyReLU activation functions; the decoding layer is used for upsampling through deconvolution operation and skip connections with the encoder output of the same scale to fuse features and restore spatial details; a 3D attention module is introduced in the skip connections to weight and regulate the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers to improve the ability to restore blind zone edges and details. Specifically, as follows Figure 4 As shown, it includes:

[0061] (3.1) Input the input tensor constructed in step (2.2). (in Indicates the size of the batch. (Representing reflectance and mask channels), preliminary feature extraction is performed through the initial convolutional module INC, outputting a low-level feature map, denoted as . It can be used to capture low-level texture and edge information of radar data; in this embodiment, the number of low-level feature map channels is 64, and the initial convolution module INC contains two consecutive 3D convolutional layers (kernel=3, padding=1), followed by BatchNorm3d and LeakyReLU activation.

[0062] (3.2) Input the low-level feature map x1 obtained in step (3.1) and generate intermediate feature maps through the downsamplingBlock module. In this embodiment, intermediate feature maps x2-x5 are generated sequentially through four downsampling modules. During the downsampling process, the first two levels of downsampling use a voxel stride of (2,2,2) to compress the entire three-dimensional space, while the last two levels use a voxel stride of (1,2,2) to compress only horizontally to preserve the vertical structure. The number of channels in the feature maps is increased sequentially to 128, 256, 512, and 1024, thereby improving the network's ability to express spatial structure and semantic information. Increasing the number of channels in the feature maps helps the network extract deeper and more abstract radar data features. The downsampling module includes a max pooling operation and a ConvBlock3D convolution stack structure.

[0063] (3.3) The deepest feature map is upsampled by a deconvolution module, and a LayerRelationAttention3D (LR-Attention) module based on inter-layer relationships is introduced to weight and regulate low-order features in skip connections. This LR-Attention module, based on the conventional gating attention mechanism, introduces encoding and modeling of the continuity structure of the 3D radar volume data in the height (inter-layer) direction, to enhance the information selection capability of skip connections and the consistency of cross-layer structures, thereby improving the data completion effect in radar blind zones. The calculation formula of the LR-Attention module is as follows:

[0064]

[0065] in, For decoder upsampling features, For encoder features, , They are respectively Convolution mapping, for Convolution mapping, It is the Sigmoid activation function. For element-wise multiplication, BN is batch normalization.

[0066] This is a high-level semantic guidance path used to extract high-level semantic information from deeper layers. After processing with a 1×1×1 3D convolution and a BatchNorm layer, the feature is aligned with the current shallow layer structure in terms of feature dimension and resolution. The goal of this path is to guide shallow features selectively through, suppressing redundant or invalid parts.

[0067] This is the current layer structure path, used to obtain the structural features of the layer where the current skip connection resides. It is compressed and normalized using a 1×1×1 3D convolution and BatchNorm for subsequent fusion. This path preserves the original shallow spatial information and is the direct carrier of the output.

[0068] Encoding path for inter-layer relationships: Use a (3,1,1) three-dimensional convolutional kernel to encode the current skip feature along the height dimension, and extract the continuous features between the current layer and the adjacent layers above and below;

[0069] Attention Fusion and Generation Module: The features of the three paths are concatenated along the channel dimension, and then processed by two layers of 1×1×1 three-dimensional convolution, BatchNorm and ReLU to generate a single-channel attention map. After being normalized by the Sigmoid activation function, it is applied to the current skip input to achieve accurate feature weighting.

[0070] Compared to existing mainstream attention mechanisms such as the Squeeze-and-Excitation module in channel attention, CBAM combining spatial and channel attention, or Attention Gate, most existing mainstream attention mechanisms focus on learning attention weights in the channel dimension or two-dimensional spatial dimension, lacking the ability to model inter-layer information in volumetric data and failing to reflect the key role of hierarchical structure in 3D perception tasks. While self-attention mechanisms have good performance, they are computationally intensive and consume a lot of memory. Since radar volumetric data has clear physical meaning in the height dimension, such as echoes, reflectivity, or meteorological structures between different layers, it is difficult for common models to capture this continuity. In particular, blind spots or missing measurement areas are more likely to cause discontinuities and breaks in the upper and lower layer structures, affecting the naturalness and physical consistency of reconstruction. The 3D attention module LR-Attention of this invention introduces the encoding and modeling of the continuity structure of 3D radar volumetric data in the height (inter-layer) direction on the basis of conventional gated attention mechanisms. This is used to enhance the ability to select skip connection information and the consistency of cross-layer structure, and improve the data completion effect of 3D radar blind spots.

[0071] The weighted encoder features are concatenated with the corresponding decoder upsampled features and then input into the convolutional module Conv. This process is repeated across four decoding layers to generate a fused feature map. The convolutional module Conv consists of maxpooling, a 3D convolutional block, BatchNorm, and LeakyReLU activation functions.

[0072] (3.4) The completed radar volume data is generated through the output convolution module Out. The output convolution module Out contains two convolutional layers: the first layer uses kernel=3 convolution to extract spatial features and compress the number of channels, and the second layer is a 1×1×1 convolution to generate the final output channels, and uses BatchNorm3d and Tanh activation functions to make the output value fall between [-1, 1]. In order to meet the actual physical range of radar reflectivity data, the output is finally forcibly restricted to the interval [-0.5, 1.5].

[0073] (3.5) Using the original blind zone mask Replace the non-blind areas in the completed data to obtain the final output. The formula for mask region replacement is as follows:

[0074]

[0075] Where M is the original blind zone mask. For radar reflectivity data, This is the complete radar data.

[0076] (4) The real radar data and the completion results are compared and judged by a three-dimensional discriminator network, and the generator network performance is optimized by combining adversarial loss and reconstruction loss. The three-dimensional discriminator network includes four 3D convolutional layers and two fully connected layers. Figure 5 As shown, step (4) specifically includes:

[0077] (4.1) The real SWAN radar data tensor and the model output data tensor are input into the 3D discriminator. A convolutional module is used to extract high-level semantic information by compressing the spatial dimension layer by layer and increasing the number of channels. To adapt to the structural characteristics of radar volumetric data, the discriminator is designed with a spatial compression ratio of z:xy = 1:2, which retains more information in the height direction and is beneficial for the discriminator to learn the vertical interlayer variation trend. Unlike the traditional uniform compression strategy, this structure is more suitable for radar completion tasks that are sensitive to vertical changes.

[0078] (4.2) The output features of the convolutional module are flattened into vectors and then fed into two fully connected layers. The true and false probabilities are obtained through the Sigmoid activation function:

[0079]

[0080] Where W1 and W2 are the weights of the fully connected layer, respectively. Let Z be the Sigmoid function, and Z be the true sample. This represents the probability that the discriminator will judge a real sample.

[0081] (4.3) During training, the generator and discriminator establish an adversarial game relationship, alternating updates through a minimization-maximization objective function. The discriminator aims to maximize its ability to distinguish between real and generated samples, while the generator optimizes its generation performance by minimizing the overall loss function. The overall optimization objective can be expressed as:

[0082]

[0083] in, Indicates the portion of the loss to combat. This represents the output of the generator, i.e., the filled-in radar data. This represents the probability that the discriminator classifies a true sample. It represents the mathematical expectation (average of the sample distribution). Represents the distribution of real data. This represents the data distribution of the conditional input X.

[0084] Specifically, the performance of the generator network can be further optimized by combining adversarial loss and reconstruction loss. The adversarial loss measures the degree to which the generated sample is judged as "real" by the discriminator, thereby guiding the generator to produce completion results that more closely match the actual radar echo distribution. Preferably, the adversarial loss can be a least-squares loss function.

[0085] Reconstruction loss is used in radar infilling tasks to evaluate the difference in distribution and detail between the generator output and the real target, and backpropagates this difference to the model for optimization training. Preferably, the reconstruction loss can be an L1 loss function, an L2 loss function, or both L1 and L2 loss functions.

[0086] This invention optimizes generator network performance by combining adversarial loss and reconstruction loss, including constructing the generator's final loss function. :

[0087]

[0088] in,

[0089]

[0090]

[0091]

[0092] in, This represents the generator's final loss function. The adversarial loss function ensures that the spatial distribution and texture details of the completion result are closer to real radar observations, improving visual and physical credibility; the reconstruction loss function... (Mean Absolute Error) ensures the accuracy of overall trend and spatial field intensity reconstruction, suppressing large local deviations; reconstructs the loss function. (Root mean square error) is used to more strictly penalize large deviations and improve the numerical consistency of the completion results; , , These represent the weighting coefficients; This indicates the generator output. This represents the probability that the discriminator judges a real sample; n represents the total number of data points in a sample; Oi represents the i-th data point in the generator's output; and Ti represents the corresponding target data point. Through weighted combination, both numerical accuracy and the adversarial nature of the generative model are ensured, avoiding the blurring and oversmoothing problems caused by using only pixel-level loss.

[0093] Based on the same inventive concept, the radar data blind spot filling system based on 3D-GANs described in this invention includes:

[0094] The simulated blind zone mask construction module is used to construct a simulated blind zone mask based on the actual blind zone shape of the 3D radar data in the training data and the characteristics of the radar data hierarchy, and embed it into the missing area of ​​the current radar data.

[0095] The input tensor construction module is used to concatenate the simulated blind zone mask and the corresponding radar reflectivity data along the channel dimension to form the input tensor;

[0096] The data completion module is used to input the input tensor into the generator network and output the completed radar volume data. The generator network adopts an improved 3D U-Net structure based on a 3D attention mechanism. During the downsampling process of the encoding layer, overall compression is performed first, and then horizontal compression is performed to retain the vertical structure. During the skip connection between the decoding layer and the encoder output of the same scale, a 3D attention module is introduced to weight and regulate the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers.

[0097] The adversarial training module is used to compare and distinguish real radar data with the completion results through a 3D discriminator network, and optimize the generator network performance by combining adversarial loss and reconstruction loss.

[0098] Based on the same inventive concept, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned radar data blind spot filling method based on 3D-GANs.

[0099] Based on the same inventive concept, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned radar data blind spot filling method based on 3D-GANs.

[0100] To verify the effectiveness of the method of this invention in radar blind zone filling, four indicators—POD (Point of Detection), FAR (False Alarm Rate), FSS (Fractional Skill Score), and TS (Score of Success)—were used to evaluate the model performance. Comparative experiments were conducted with the current representative method Unet and the original GAN ​​(VGAN) on the same dataset. The test dataset consisted of 6562 data points from June 1st to June 30th, 2024, generated from a 21-layer radar volume mosaic dataset in the SWAN system. Due to the large number of research layers, three representative height layers (layers 3 to 5 of the SWAN radar data, 1.5km to 2.5km altitude) were selected for presentation. The experimental results are shown in Tables 1-3. As can be seen from the tables, among the four indicator scores, the method of this invention exhibits higher detection capabilities in POD and TS, greater stability in FAR control, and a higher FSS score, all of which outperform the baseline model in conventional radar blind zone filling work.

[0101] Table 1. Filling results of the present invention in the third layer (1.5km height) of SWAN data.

[0102]

[0103] Table 2. Filling results of the present invention in the 4th layer (2km height) of SWAN data.

[0104]

[0105] Table 3. Filling results of the present invention in the 5th layer (2.5km height) of SWAN data.

[0106]

[0107] Therefore, the radar blind zone intelligent completion method and system proposed in this invention can integrate the structural characteristics of the SWAN system, adapt to the distribution characteristics of radar network mosaic data, and have three-dimensional perception and generation capabilities, providing solid support for accurate early warning and near-term monitoring of extreme weather.

Claims

1. A radar data blind zone filling method based on 3D-GANs, characterized in that, Includes the following steps: (1) Based on the true blind zone morphology of the three-dimensional radar data in the training data, construct a simulated blind zone mask based on the hierarchical characteristics of the radar data, and embed it layer by layer into the non-missing region of the current radar data; (2) The simulated blind zone mask and the corresponding radar reflectivity data are concatenated along the channel dimension to form the input tensor; (3) Input the input tensor into the generator network and output the completed radar volume data; the generator network adopts an improved three-dimensional U-Net structure based on the three-dimensional attention mechanism. In the downsampling process of the encoding layer, the overall compression is performed first, and then the horizontal compression is performed to retain the vertical structure; in the process of skip connection between the decoding layer and the output of the encoder of the same scale, a three-dimensional attention module is introduced to perform weighted regulation on the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers; (4) The real radar data and the completion results are compared and judged by the three-dimensional discriminator network, and the generator network performance is optimized by combining the adversarial loss and reconstruction loss.

2. The radar data blind spot filling method based on 3D-GANs according to claim 1, characterized in that, Step (1) includes: for large blind area shapes, i.e., layers with less effective data in the original data, generating random irregular shapes to simulate blind area masks and randomly embedding them into regions without missing data; for layers with moderate blind area shapes, constructing structural templates and using a rule of priority placement at the center or searching for feasible positions from the center for embedding; for smaller blind area shapes, i.e., layers with more effective data, constructing structural templates and embedding them into the corresponding regions without missing data in the layers that need to be filled.

3. The radar data blind spot filling method based on 3D-GANs according to claim 2, characterized in that, The construction of the structural template includes extracting radar data with missing real data from the training data, obtaining the corresponding blind zone mask, extracting connected regions based on the mask using image processing methods, and statistically analyzing their shape and size to construct the structural template.

4. The radar data blind spot filling method based on 3D-GANs according to claim 1, characterized in that, In step (2), the radar reflectivity data is normalized.

5. The radar data blind spot filling method based on 3D-GANs according to claim 1, characterized in that, Step (3) includes: The input tensor is input, and preliminary feature extraction is performed through the initial convolution module to output a low-level feature map; Several intermediate feature maps are generated by several downsampling modules. During the downsampling process, the three-dimensional space is first compressed as a whole based on the vertical layering characteristics of the three-dimensional radar data, and then horizontally compressed to retain the vertical structure. The deepest feature map is upsampled by the deconvolution module, and a 3D attention module that models the inter-layer relationship is introduced to weight and regulate the low-order features in the skip connections. The weighted encoder features and the decoder upsampled features are concatenated and then input into the convolution module. This process is repeated for several decoding layers to generate a fused feature map. The completed radar volume data is generated by the output convolution module; The original blind zone mask is used to replace the non-blind zone area in the completed radar volume data to obtain the final output result.

6. The radar data blind spot filling method based on 3D-GANs according to claim 1, characterized in that, The calculation formula for the three-dimensional attention module is as follows: in, For decoder upsampling features, For encoder features, , They are respectively Convolution mapping, for Convolution mapping, It is the Sigmoid activation function. For element-wise multiplication, BN is batch normalization.

7. The radar data blind spot filling method based on 3D-GANs according to claim 1, characterized in that, The adversarial loss mentioned in step (4) is the least squares loss function, and the reconstruction loss is the L1 loss function and / or the L2 loss function.

8. A radar data blind spot filling system based on 3D-GANs, used to implement the radar data blind spot filling method based on 3D-GANs as described in any one of claims 1-7, characterized in that, include: The simulated blind zone mask construction module is used to construct a simulated blind zone mask based on the actual blind zone shape of the 3D radar data in the training data and the characteristics of the radar data hierarchy, and embed it into the missing area of ​​the current radar data. The input tensor construction module is used to concatenate the simulated blind zone mask and the corresponding radar reflectivity data along the channel dimension to form the input tensor; The data completion module is used to input the input tensor into the generator network and output the completed radar volume data. The generator network adopts an improved 3D U-Net structure based on a 3D attention mechanism. During the downsampling process of the encoding layer, overall compression is performed first, and then horizontal compression is performed to retain the vertical structure. During the skip connection between the decoding layer and the encoder output of the same scale, a 3D attention module is introduced to weight and regulate the low-order features of the encoder output, so as to realize the weighted fusion of the features of the encoding and decoding layers. The adversarial training module is used to compare and distinguish real radar data with the completion results through a 3D discriminator network, and optimize the generator network performance by combining adversarial loss and reconstruction loss.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the radar data blind spot filling method based on 3D-GANs as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the radar data blind spot filling method based on 3D-GANs as described in any one of claims 1-7.

Images