A structure-guided and frequency-aware radar downscaling method

Abstract

The application discloses a structure-guided and frequency-aware radar downscaling method, comprising the following steps: obtaining low-resolution radar echo data and corresponding waveband identification information, and extracting shallow network features; constructing a deep feature extraction network, dynamically calibrating the mean value and variance of the features according to the waveband identification through a dual-waveband adaptive modulation mechanism, and realizing cross-waveband feature alignment; introducing Laplace edge physical priori by using a radar echo structure-guided attention mechanism, generating an attention mask to guide the network to focus on the effective precipitation area; through a double-flow architecture and a gating strategy of a high-frequency excitation feedforward mechanism, local high-frequency gradient features are explicitly extracted and enhanced; and high-resolution radar echo images are generated by combining up-sampling and frequency domain residual enhancement processing. The application overcomes the problem of high-frequency texture loss caused by the sparsity of radar echo and the spectral bias of deep network, effectively restores the fine structure of the strong convective core, and realizes unified modeling of cross-waveband data through waveband adaptive modulation.

Description

Technical Field

[0001] This invention relates to a downscaling method for structure-guided and frequency-aware radar, belonging to the field of meteorological radar signal processing technology. Background Technology

[0002] With the deepening of modern meteorological research, acquiring high spatiotemporal resolution weather radar observation data plays a crucial role in quantitative precipitation estimation, short-term forecasting of extreme weather, and initialization of numerical weather prediction models. Accurately capturing microscale convective structure features such as hook echoes, narrow rainbands, and shear lines is a vital prerequisite for early warning of flash floods and severe storms. However, limited by physical constraints such as scanning strategies, angular resolution, and underlying hardware bandwidth, current S-band radar reflectivity data typically uses an azimuth sampling interval of approximately 1° and a range sampling interval of 1 km. Its polar coordinate sampling is relatively sparse, making it difficult to fully characterize the fine spatial structure in strong convective echoes. Utilizing super-resolution reconstruction techniques to recover high-resolution signals from low-resolution observations has become a feasible alternative.

[0003] Given the urgent need to improve the spatial resolution of weather radar observations, some studies have attempted to introduce deep learning super-resolution models into the meteorological field. However, existing weather radar downscaling methods still have the following shortcomings: (1) Radar reflectivity echo maps have significant spatial sparsity, with a high proportion of areas without echo background. Existing models do not fully consider this sparsity feature, and indiscriminate attention allocation will lead to a large amount of ineffective computational overhead. Moreover, the model attention is easily interfered with by background noise, making it difficult to focus on the real precipitation target. (2) Existing general frameworks usually rely on pixel-level loss functions such as L1 or L2 for optimization. This tendency of statistical mean regression can easily lead to over-smoothing, causing precipitation extreme values ​​to be underestimated and edges to be blurred, thereby weakening the authenticity and clarity of the echo structure. (3) There are significant differences in the actual operational radar network. There are significant physical differences in signal attenuation and noise distribution between S-band radars responsible for wide-area monitoring and X-band radars that undertake local blind spot tasks. Existing models lack cross-frequency band adaptability, and their generalization performance often declines when faced with such differences, requiring separate training for different types of radars. (4) Existing networks are prone to high-frequency distortion and loss of details during the upsampling stage, and rely only on local spatial information for modeling, making it difficult to capture global meteorological features with long-range dependence. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a structure-guided and frequency-aware radar downscaling method that overcomes the problem of high-frequency texture loss caused by radar echo sparsity and deep network spectrum deviation, effectively restoring the fine structure of strong convection cores; and to achieve unified modeling of cross-band data through band adaptive modulation, which significantly improves the universality and physical fidelity of radar echo reconstruction.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A structure-guided and frequency-aware radar downscaling method includes the following steps: Step 1: Collect the dataset for downscaling modeling of meteorological radar echoes and perform quality control. The dataset includes S-band low-resolution radar echo data and X-band high-resolution radar echo data. Set discrete radar band identifiers for the quality-controlled S-band low-resolution and X-band high-resolution radar echo data respectively, and perform time alignment and spatial registration, and then linearly normalize them to the [0,1] interval. Step 2: Establish a weather radar echo downscaling model, including convolutional layers, an echo structure extraction module, a dual-band adaptive modulation module, a deep feature extraction network, and a frequency domain annular residual refinement module. The convolutional layers extract shallow network features from the input radar echo data; the echo structure extraction module extracts multi-component radar structure prior features from the input radar echo data; and the dual-band adaptive modulation module generates scaling and translation factors based on the discrete radar band identifiers corresponding to the input radar echo data. These scaling and translation factors are then used for band-adaptive affine modulation within the deep feature extraction network. A deep feature extraction network is used to extract deep features from shallow network features by combining multi-component radar structure priors, scaling factors, and translation factors. The deep features are mapped to a channel dimension that matches the target upsampling factor, and sub-pixel rearrangement is used to generate an initial high-resolution prediction tensor. A frequency domain annular residual refinement module is used to perform frequency domain residual enhancement processing on the initial high-resolution prediction tensor to generate spatial residual features. The initial high-resolution prediction tensor and spatial residual features are fused to obtain the final high-resolution radar echo reconstruction result. Step 3: Perform the first stage of training on the model established in Step 2, using X-band high-resolution radar echo data. X-band low-resolution radar echo data is generated by bicubic interpolation downsampling. ,by As model input As a supervisory signal, it is used for same-band supervisory training, while setting discrete radar band identifiers. ; Step 4: Based on Step 3, conduct the second phase of training to unify the S-band low-resolution radar echo data at the altitude level. X-band high-resolution radar echo data was used as model input. Cross-band fine-tuning training is performed using this as a supervisory signal, while discrete radar band identifiers are set. Thus, the trained model is obtained; Step 5: During the inference process, input the S-band low-resolution radar echo data at the uniform height layer to be processed into the trained model, and set the discrete radar band identifier. This yields high-resolution radar echo reconstruction results at a unified height level.

[0006] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. This invention extracts structural priors in the radar physical quantity domain. After converting the dBZ domain echo into a linear reflectivity factor, it constructs multi-component radar echo structural prior features by combining the non-negative reflectivity base, edge high-frequency features, strong echo core confidence, and no-echo background confidence. This can more accurately distinguish precipitation boundaries, strong convection cores, and no-echo background areas, thereby effectively suppressing invalid calculations and false texture generation caused by large-scale no-echo backgrounds and improving the model's ability to focus on effective precipitation areas and strong echo areas.

[0007] 2. This invention employs a paired structure biased attention mechanism, which incorporates the structural similarity, edge differences, strong echo core consistency, and background suppression relationship between any two spatial locations within the window into the attention calculation process. This makes the attention response no longer solely dependent on ordinary feature similarity, but rather explicitly guided by the physical structure of the radar echo. This enhances feature aggregation within the strong echo core, weakens erroneous associations between the echo-free background and precipitation areas, and improves the structural preservation capability and boundary clarity of the reconstruction results.

[0008] 3. This invention introduces a dual-band adaptive modulation mechanism that integrates radar band identification and beam geometry parameters. By jointly generating scaling factors, translation factors, and range attenuation compensation factors through learnable frequency band embedding and geometric embedding, it achieves adaptive calibration of S-band and X-band radar echoes in terms of statistical distribution, spatial resolution, detection range, and beamwidth differences, thereby improving the model's adaptability in heterogeneous radar, multi-system radar, and cross-band migration scenarios.

[0009] 4. The present invention further sets up a frequency domain ring band residual refinement module, which performs residual correction on the mid-to-high frequency spectrum components in the frequency domain according to the radar spatial sampling characteristics, and ensures the real number consistency of the result after the inverse Fourier transform through Hermitian symmetry constraint, thereby realizing the synergistic optimization of spatial domain local detail enhancement and frequency domain global spectrum correction, which can reduce high frequency artifacts and improve the reconstruction quality of echo edges, textures and strong echo core areas.

[0010] 5. This invention optimizes the network through L1-frequency domain ring band consistency loss, so that the reconstruction results are constrained in terms of overall echo intensity, precipitation boundary, mid-to-high frequency spectrum and background suppression. This avoids the model from only pursuing the minimum pixel error and generating results that do not conform to meteorological and physical characteristics, thereby improving the authenticity, stability and operational availability of radar echo reconstruction results. Attached Figure Description

[0011] Figure 1 This is a flowchart of the structure-guided and frequency-aware radar downscaling method of the present invention; Figure 2 This is a diagram of the weather radar echo downscaling model architecture proposed in this invention; Figure 3 This is a diagram of the echo structure-guided attention module architecture proposed in this invention; Figure 4 This is a diagram of the high-frequency excitation feedforward enhancement module architecture proposed in this invention; Figure 5 This is a diagram of the frequency domain ring band residual refinement module architecture proposed in this invention. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] like Figure 1 As shown, this invention provides a method for downscaling structure-guided and frequency-aware radar, with the following specific steps: Step 1, construct a dual-band paired sample set: obtain a high-resolution X-band radar echo map at a uniform height layer; based on the high-resolution X-band radar echo map, generate a low-resolution X-band radar echo map through bicubic downsampling; obtain an S-band low-resolution radar echo map located at the same height layer as the high-resolution X-band radar echo map and complete time alignment, horizontal spatial registration, and reflectivity calibration; and configure a discrete radar band identifier for the input low-resolution radar echo map, wherein the discrete radar band identifier is used to distinguish between the X-band low-resolution radar echo map and the S-band low-resolution radar echo map.

[0014] In meteorological radar echo downscaling studies, data collection is a fundamental step for model training, validation, and testing. To ensure the representativeness, reliability, and applicability of the data for subsequent model training, this invention uses heterogeneous radar observation data from a certain region from December 2021 to June 2024 as the experimental data source. This dataset contains 580 different precipitation events, with the duration of each continuous observation sequence controlled within 24 hours, thus covering various typical weather scenarios such as severe convection, short-duration heavy precipitation, and general precipitation processes. A paired observation dataset is constructed using S-band radar data responsible for wide-area monitoring as low-resolution input and X-band phased array radar data responsible for local blind spot filling as high-resolution ground truth. The S-band data utilizes the CAPPI contour plane position indicator, covering an area of ​​112.3°E–113.5°E and 22.6°N–23.6°N, with a spatial resolution of 0.01°. The X-band data covers an area of ​​112.71°E–113.35°E and 22.83°N–23.45°N, with a spatial resolution of 0.0025°, achieving a four-fold improvement in spatial resolution. To ensure the validity of the paired samples, this invention performs quality control and preprocessing on the raw radar observation data, removing outliers and missing values. Strict temporal alignment and spatial registration are performed on the S-band and X-band radar data to ensure a one-to-one correspondence between the paired samples in both time and space, thus meeting the requirements of subsequent training and testing. After data processing and spatial alignment, the dataset was finally divided into a training set containing 14,310 image pairs and a test set containing 2,046 image pairs, with strict no overlap between the training and test sets to objectively evaluate the model's generalization ability. Before inputting the data into the model, radar reflectivity values ​​with a physical range between [0, 70] dBZ were linearly normalized to [0, 1] to improve the numerical stability of the model during training.

[0015] Furthermore, discrete radar band identifiers are configured for the input low-resolution radar echo map. This is used to distinguish the radar band to which the input data belongs. In this embodiment, This indicates that the input is a low-resolution X-band radar echo map. This indicates that the input is an S-band low-resolution radar echo map. By setting discrete radar band identifiers, the subsequent network can adaptively modulate according to the differences in reflectivity distribution of radar data in different bands.

[0016] Step 2: Establish a downscaling model of the weather radar echo, such as... Figure 2 As shown, it includes convolutional layers, an echo structure extraction module, a dual-band adaptive modulation module, a deep feature extraction network, and a frequency domain annular residual refinement module; For input low-resolution radar echo map First, shallow feature extraction is performed using convolutional layers with a kernel size of 3×3 to obtain shallow network features. And serve as input for subsequent deep feature extractors: (1) in, This indicates a convolutional layer with a kernel size of 3×3.

[0017] Meanwhile, in order to explicitly introduce the physical structure prior of the radar echo, the echo structure extraction module will extract the input low-resolution radar echo image. The linear reflectance factor map is obtained by converting the dBZ domain to the linear reflectance factor domain. Its expression is: (2) Linear reflectivity factor plot Threshold clipping and non-negative activation are performed to obtain a non-negative reflectivity substrate. : (3) in, Represents the ReLU activation function. This represents the threshold clipping function. This represents the minimum clipping threshold for the linear reflectivity factor. This represents the maximum clipping threshold for the linear reflectivity factor. Through the above processing, the influence of outlier extreme values ​​and negative background noise on subsequent prior structure extraction can be reduced.

[0018] Subsequently, a fixed Laplace structure prior kernel matrix was adopted. For non-negative reflectivity substrates Perform two-dimensional convolution and take the absolute value of the convolution response to obtain the high-frequency feature map of the edge. A fixed Laplace structure prior kernel matrix is ​​used to highlight high-frequency structural features such as the edges of the echo region, gradient abrupt changes, and the outer edge of the convection kernel. (4) (5) in, Represents the prior kernel matrix based on the Laplace structure Two-dimensional convolution.

[0019] Based on nonnegative reflectivity substrate Generate a strong echo core confidence map Confidence plot with no echo background Strong echo core confidence map Confidence map without echo background used to characterize the probability that a spatial location belongs to the core region of a strong precipitation echo. Used to characterize the probability that a spatial location belongs to an echo-free background region: (6) (7) in, This represents the Sigmoid activation function. This indicates the threshold for determining the core of a strong echo. The smoothing coefficient represents the core confidence plot of strong echoes. This indicates the threshold for determining the absence of background echo. This represents the smoothing coefficient of the confidence plot with no echo background.

[0020] Non-negative reflectivity substrate High-frequency feature map of the edge Strong echo core confidence map Confidence plot with no echo background Weighted channel splicing is performed to obtain multi-channel structural features. : (8) in, This indicates an operation that splices along the channel dimension. The weighting coefficients represent the weighting coefficients for high-frequency components at the edges. The weighting coefficients represent the core components of the strong echo. The weighting coefficients represent the weights of the echo-free background component; Multi-channel structural features Local normalization is performed to obtain the multi-component radar structure prior features. : (9) in, Indicates the direction of spatial height and width direction Local max pooling operation on, To prevent extremely small constants with a denominator of zero, the prior characteristics of multi-component radar structures are discussed. It includes not only the reflectivity intensity substrate, but also edge high frequency, strong echo core and background suppression information, which are used for subsequent radar structure-guided attention calculation and frequency domain residual refinement.

[0021] Step 3: The dual-band adaptive modulation module includes a learnable band embedding layer and a two-layer multilayer perceptron, which converts the input radar echo data into discrete radar band identifiers. Input a learnable band embedding layer to obtain band embedding features. : (10) in, It is a learnable band embedding layer, with band embedding features. This is used to characterize the statistical differences in reflectivity, spatial resolution, and imaging regime between X-band radar data and S-band radar data.

[0022] Subsequently, the band is embedded with features. The input consists of a two-layer multilayer perceptron composed of a fully connected layer and a GELU activation function. Generate modulation vectors that match the dimensions of deep feature channels. The data is then divided equally along the channel dimension to obtain the scaling factor. Translation factor : (11) (12) in, This indicates that the channel dimension is divided equally. Scaling factor. Translation factor This is used to perform band-adaptive affine modulation on shallow network features or intermediate features of deep networks, enabling the model to dynamically calibrate the statistical distribution of features according to different input radar bands, thereby improving the generalization ability in cross-band downscaling tasks.

[0023] Step 4, extract shallow network features The input is a deep feature extraction network. The deep feature extraction network consists of four cascaded radar structure Transformer groups (RSTGs). The output of the previous RSTG serves as the input to the next RSTG. For the first RSTG... RSTG Its input features are denoted as The output features are denoted as ,in, .

[0024] Each RSTG includes an echo structure-guided attention module and a high-frequency excitation feedforward enhancement module; in each RSTG, the input features are first processed... Perform layer normalization to obtain the first normalized feature. Then, the scaling factor generated by the dual-band adaptive modulation module is used. Translation factor For the first normalized feature Perform band-adaptive affine modulation to obtain the first modulation feature. : (13) in, This represents Hadamard's pixel-wise multiplication operation; The first modulation feature and multi-component radar structure prior features The local windows are divided into corresponding non-overlapping local windows to obtain their features. and local structural priors : (14) (15) in, This indicates a local window division.

[0025] like Figure 3 As shown, the echo structure guides the attention module to perform linear projection on the features of each local window, resulting in the query matrix. Key matrix Sum matrix and will Divided along the channel dimension The first attention point, to get the first The query submatrix corresponding to each attention head Key matrix Sum submatrix : , , (16) , , (17) in, These are the learnable linear projection weights corresponding to the query matrix. These are the learnable linear projection weights corresponding to the key matrix. These are the learnable linear projection weights corresponding to the value matrix.

[0026] For any two spatial locations within the same local window and According to location and Corresponding multi-component radar structure prior features and strong echo core confidence map and the confidence plot of the background without echo Constructing a radar structure with paired offsets The paired offset of the radar structure includes a structural similarity enhancement term, a strong echo core consistency enhancement term, and a strong echo core and non-echo background error association suppression term: (18) in, Indicates position and Corresponding multi-component radar structure prior features and Structural similarity between them This indicates the weight of the strong echo core consistency enhancement term. Indicates position The core confidence level of the strong echo at the location, Indicates position The core confidence level of the strong echo at the location, This indicates the weight of the suppression term that correlates strong echo cores with unecho background errors. Indicates position Confidence level of the echo-free background at the location, Indicates position Confidence level of the echo-free background at the location; The radar structure is biased in pairs. By incorporating the attention score of the dot product between the query submatrix and the key submatrix, we obtain the first... The structure of each attention head guides attention scores : (19) in, Indicates transpose. It is the feature dimension of a single attention head in multi-head attention.

[0027] Combined with the first A dynamic discrete mask matrix transmitted by RSTG , obtained the The final attention response distribution of each attention head In this case, the dynamic discrete mask matrix of the first RSTG is set to an all-zero matrix: (20) in, It is normalization; Using the first The final attention response distribution of each attention head AND-valued submatrix Perform feature aggregation to obtain the first Attention output features of each attention head Then The attention output features of each attention head are concatenated through channels and then linearly mapped using a linear mapping matrix to obtain the window-level attention output features. : (twenty one) (twenty two) in, It is a linear mapping matrix.

[0028] Furthermore, based on the final attention response distribution Strong echo core confidence map Confidence plot of echo-free background and the A dynamic discrete mask matrix transmitted by RSTG Calculate the radar structure status score at each spatial location within the local window. And score the radar structural status. Perform TopK filtering to generate the first The dynamic discrete mask matrix passed from one RSTG to the next RSTG : (twenty three) (twenty four) in, This indicates averaging over the attention head dimension. , , , These represent the weighting coefficients of each item in the radar structural condition score. This indicates the extraction of the highest-rated item. The model can continuously focus on high-confidence precipitation structure regions and strong echo core regions in subsequent levels by passing the dynamic discrete mask matrix layer by layer.

[0029] Finally, the window-level attention output features are used. After window inverse rearrangement and linear mapping to restore the global permutation features, the attention output features are obtained. and with the Input features of RSTG Perform residual fusion to obtain intermediate features : (25) in, Indicates random depth.

[0030] Step 5, for intermediate features Perform layer normalization to obtain the second normalized feature. And a scaling factor generated by a dual-band adaptive modulation module. Translation factor For the second normalized feature Band-adaptive affine modulation is performed to obtain the second modulation feature. : (26) Then the second modulation feature Input high-frequency excitation feedforward enhancement module, such as Figure 4 As shown, channel dimensionality is first increased through channel convolution, and a smooth outer receptive field is constructed through reflection filling to obtain extended features. : (27) in, Represents up-dimensional channel convolution. This indicates a reflection fill operation.

[0031] Extracting local detail response features using depthwise separable convolution Local low-frequency smoothing features are extracted using average pooling. Based on local detail response characteristics With local low-frequency smoothing features The difference is used to obtain the high-frequency residual characteristics. : (28) (29) (30) in, This represents depthwise separable convolution. Indicates average pooling; Based on local detail response features Generate nonlinear gated weights : (31) in, This represents the Sigmoid activation function. This indicates the convolutional layer used to generate the gating weights.

[0032] Using nonlinear gated weights and learnable scaling parameters High-frequency residual characteristics Adaptive modulation is performed to obtain high-frequency excitation characteristics. High-frequency excitation characteristics By sequentially inputting the GELU activation function and channel-reduced convolution, feedforward enhanced features are obtained. : (32) (33) in, This represents channel-reduced convolution. Indicates the activation function; feedforward enhancement features With intermediate features Perform residual fusion to obtain the first Output characteristics of RSTG : (34) when At that time, output features As input features for the next RSTG; when At that time, output features As deep features .

[0033] Step 6, extract deep features The channel-expanded convolution is mapped to a channel dimension that matches the target upsampling factor, and then an initial high-resolution prediction tensor is generated through sub-pixel rearrangement. Initial high-resolution prediction tensor This represents the initial reconstruction result obtained by the model in the spatial domain. Subsequently, as... Figure 5 As shown, the frequency domain annular residual refinement module refines the initial high-resolution prediction tensor. Performing a two-dimensional discrete Fourier transform yields a complex frequency domain tensor. : (35) in, It is a two-dimensional discrete Fourier transform; The edge high-frequency feature map obtained in step 2 Strong echo core confidence map Confidence plot with no echo background Resampling to the initial high-resolution prediction tensor Spatial scales were used to obtain scale-aligned edge high-frequency maps. Scale-aligned strong echo core confidence map Scale-aligned echo-free background confidence plot : (36) (37) (38) in, It is a bilinear interpolation sizing operation. This indicates an operation to obtain spatial dimensions.

[0034] Perform two-dimensional discrete Fourier transforms on the scale-aligned structure diagrams and take the amplitudes to obtain the structure's frequency domain guidance features. : (39) in, It is a two-dimensional discrete Fourier transform. This represents the modulus of the complex frequency domain result obtained from the two-dimensional Fourier transform; for any complex frequency domain element... Its modulus is The structural frequency domain guidance features are used to guide the model in the frequency domain to identify spectral components associated with the edges, strong echo cores, and effective echo regions.

[0035] Furthermore, based on the structural frequency domain guidance characteristics Constructing the normalized radial frequency radius in the frequency domain coordinates : (40) in, Represents the frequency coordinates in the horizontal direction. Represents the frequency coordinates in the vertical direction. Indicates the Nyquist frequency in the horizontal direction. Indicates the Nyquist frequency in the vertical direction; A low-frequency protection mask is constructed based on the normalized radial frequency radius. and multiple radial frequency ring band masks : (41) (42) in, This represents a conditional function; if the condition is met, the value is 1, and if not, the value is 0. Indicates the first A radial frequency ring band mask Indicates the low-frequency protection threshold. Indicates the first The inner radius threshold of a radial frequency band. Indicates the first The outer radius threshold of a radial frequency loop Indicates the total number of radial frequency bands; low-frequency protection mask. Multiple radial frequency ring band masks are used to protect large-scale low-frequency intensity substrates in radar echoes. Used for ring band modeling of spectral components within different frequency radius ranges.

[0036] complex frequency domain tensor Real part, imaginary part, low frequency protection mask Radial frequency ring band mask to and structural frequency domain guidance features Input frequency domain loop-gated network Generate candidate frequency domain residuals : (43) in, Represents the complex frequency domain tensor The real part, Represents the complex frequency domain tensor The imaginary part; Then, using a low-frequency protection mask For candidate frequency domain residuals Low-frequency suppression is performed to prevent the frequency domain residual from damaging the low-frequency reflectivity substrate. Simultaneously, Hermitian conjugate symmetry constraints are applied to the candidate frequency domain residual after low-frequency suppression to obtain a frequency domain residual that satisfies the requirements for real-valued spatial domain reconstruction. : (44) in, This represents the Hermitian conjugate symmetry constraint. frequency domain residual The spatial residual features are obtained by mapping back to the spatial domain through a two-dimensional discrete Fourier inverse transform, and then performing local neighborhood fusion and cross-channel correction through 3×3 convolution. : (45) in, It is a two-dimensional discrete Fourier inverse transform; The initial high-resolution prediction tensor and spatial residual characteristics The data is then fused to obtain the final high-resolution radar echo reconstruction result. : (46) Through the refinement and fusion process of the frequency domain ring band residual features, the model can adaptively compensate for high-frequency errors in different radial frequency ring bands without destroying the large-scale low-frequency echo distribution, thereby improving the recovery ability of strong echo cores, precipitation edges and fine-grained texture structures.

[0037] Step 7: During the first phase of training, X-band high-resolution radar echo images at a unified altitude layer will be generated. Low-resolution X-band radar echo maps are generated using bicubic interpolation downsampling. and with As model input, As a supervisor, conduct supervised training in the same frequency band, while setting discrete radar band identifiers. Through this training phase, the model is able to learn the fundamental spatial reconstruction mapping of X-band radar echoes from low resolution to high resolution.

[0038] During the second phase of training, S-band low-resolution radar echo images at a unified altitude will be used. As input to the model, a high-resolution X-band radar echo map that has been time-aligned, horizontally spatially registered, and calibrated for reflectivity will be used. Cross-band fine-tuning training is performed using this as a supervisory signal, while discrete radar band identifiers are set. Through this training phase, the model can adapt to the differences in system architecture and reflectivity statistics between S-band and X-band radar data, achieving cross-band radar echo downscaling reconstruction.

[0039] During training, the L1-frequency domain annular band consistency loss function is used to optimize the model parameters. L1-frequency domain annular band consistency loss function Including L1 loss and frequency domain ring band consistency loss Its expression is: (47) in, It is frequency domain ring band consistency loss Weighting coefficients; L1 loss is used to constrain high-resolution radar echo reconstruction results. With X-band high-resolution radar echo data Pixel-level error in the spatial domain: (48) in, and These represent the height and width of the feature map, respectively. For position The model output value at that location, This represents the true high-resolution ground truth value at the corresponding location; Frequency domain ring band consistency loss is used to constrain the consistency of the amplitude spectrum between the high-resolution radar echo reconstruction results and the real X-band high-resolution radar echo data in different radial frequency ring bands: (49) in, Indicates the total number of radial frequency bands. Indicates the radial frequency band number. express Norm, Indicates the first A radial frequency ring band mask This indicates Hadamard's pixel-by-pixel multiplication operation.

[0040] During the inference process, the S-band low-resolution radar echo map at the uniform height layer to be processed is input into the trained model, and discrete radar band identifiers are set. This allows for the output of high-resolution radar echo reconstruction results at a uniform altitude layer. This result improves the spatial resolution of local precipitation structures, strong echo cores, and edge textures while maintaining the large-scale distribution of the original S-band radar echoes.

[0041] In this embodiment, the following models are selected as baselines for comparison of model performance. (1) Enhanced Deep Super-Resolution Network (EDSR): By removing redundant modules in the traditional residual network and expanding the model capacity, an enhanced deep residual super-resolution network is constructed to improve the reconstruction capability of high-frequency details. (2) Residual Channel Attention Network (RCAN): Through the residual structure in the residual and the channel attention mechanism, different channel features are adaptively recalibrated, so that the network can more effectively focus on high-frequency information in super-resolution reconstruction. (3) Image Restoration Network Based on Swin Transformer (SwinIR): The image restoration model is constructed based on the hierarchical shift window transformer. The residual shift window transformer module is used to enhance the ability to characterize long-range dependencies while maintaining local modeling capabilities. (4) Hybrid Attention Transformer (HAT): Combining channel attention and window-based self-attention mechanism, the complementary advantages of the two types of attention are used to activate more effective input pixels, thereby enhancing the image detail restoration capability. (5) Image Processing Graph Neural Network (IPG): Construct a super-resolution model based on graph neural networks, and improve the representation ability of complex texture regions and unbalanced detail information through degree-variable graph modeling and local and global node connection strategies. (6) Adaptive Token Dictionary Transformer (ATD): Introduce a learnable adaptive lexicon dictionary, and utilize the typical image structure prior formed during training through dictionary cross-attention to achieve global dependency modeling and feature enhancement with complexity linearly related to image size. (7) Progressive Focusing Transformer (PFT): Progressively focus attention gradually associates the attention graphs in the network, making the model more focused on important lexicons related to the current query, improving the super-resolution reconstruction performance while reducing the overhead of irrelevant similarity calculation.

[0042] A quantitative comparison of signal fidelity between the proposed method and current state-of-the-art methods was conducted. Performance was evaluated using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Point of Detection (POD), False Alarm Rate (FAR), and Critical Success Index (CSI), as shown in Table 1.

[0043] Table 1. Quantitative Comparison of Signal Fidelity

[0044] As shown in Table 1, the method proposed in this invention outperforms all comparative baseline models on the heterogeneous radar test data, thus verifying the effectiveness of this invention.

[0045] Based on the same inventive concept, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned structure-guided and frequency-aware radar downscaling method.

[0046] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned structure-guided and frequency-aware radar downscaling method.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This invention is described with reference to flowchart illustrations of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each step in the flowchart, and combinations of steps in the flowchart, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the steps in the flowchart. Figure 1 A device for a function specified in one or more processes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1Steps of a specified function in one or more processes.

[0051] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for downscaling structure-guided and frequency-aware radar, characterized in that, Includes the following steps: Step 1: Collect the dataset for downscaling modeling of meteorological radar echoes and perform quality control. The dataset includes S-band low-resolution radar echo data and X-band high-resolution radar echo data. Set discrete radar band identifiers for the quality-controlled S-band low-resolution and X-band high-resolution radar echo data respectively, and perform time alignment and spatial registration, and then linearly normalize them to the [0,1] interval. Step 2: Establish a weather radar echo downscaling model, including convolutional layers, an echo structure extraction module, a dual-band adaptive modulation module, a deep feature extraction network, and a frequency domain annular residual refinement module. The convolutional layers extract shallow network features from the input radar echo data; the echo structure extraction module extracts multi-component radar structure prior features from the input radar echo data; and the dual-band adaptive modulation module generates scaling and translation factors based on the discrete radar band identifiers corresponding to the input radar echo data. These scaling and translation factors are then used for band-adaptive affine modulation within the deep feature extraction network. A deep feature extraction network is used to extract deep features from shallow network features by combining multi-component radar structure priors, scaling factors, and translation factors. The deep features are mapped to a channel dimension that matches the target upsampling factor, and sub-pixel rearrangement is used to generate an initial high-resolution prediction tensor. A frequency domain annular residual refinement module is used to perform frequency domain residual enhancement processing on the initial high-resolution prediction tensor to generate spatial residual features. The initial high-resolution prediction tensor and spatial residual features are fused to obtain the final high-resolution radar echo reconstruction result. Step 3: Perform the first stage of training on the model established in Step 2, using X-band high-resolution radar echo data. X-band low-resolution radar echo data is generated by bicubic interpolation downsampling. ,by As model input As a supervisory signal, it is used for same-band supervisory training, while setting discrete radar band identifiers. ; Step 4: Based on Step 3, conduct the second phase of training to unify the S-band low-resolution radar echo data at the altitude level. X-band high-resolution radar echo data was used as model input. Cross-band fine-tuning training is performed using this as a supervisory signal, while discrete radar band identifiers are set. Thus, the trained model is obtained; Step 5: During the inference process, input the S-band low-resolution radar echo data at the uniform height layer to be processed into the trained model, and set the discrete radar band identifier. This yields high-resolution radar echo reconstruction results at a unified height level.

2. The downscaling method for structure-guided and frequency-aware radar according to claim 1, characterized in that, The echo structure extraction module converts the input radar echo data from the dBZ domain to the linear reflectivity factor domain, obtaining a linear reflectivity factor map. For linear reflectivity factor plot Threshold clipping and non-negative activation are performed to obtain a non-negative reflectivity substrate. : ， in, Represents the ReLU activation function. This represents the threshold clipping function. , These represent the minimum and maximum clipping thresholds for the linear reflectivity factor, respectively. Laplace structure prior kernel matrix For non-negative reflectivity substrates Perform two-dimensional convolution and take the absolute value of the convolution response to obtain the high-frequency feature map of the edge. The formula is as follows: ， ， in, Represents the prior kernel matrix based on the Laplace structure Two-dimensional convolution; Based on nonnegative reflectivity substrate Generate a strong echo core confidence map Confidence plot with no echo background : ， ， in, This represents the Sigmoid activation function. This indicates the threshold for determining the core of a strong echo. The smoothing coefficient represents the core confidence plot of strong echoes. This indicates the threshold for determining the absence of background echo. The smoothing coefficient represents the confidence plot of the echo-free background. Non-negative reflectivity substrate High-frequency feature map of the edge Strong echo core confidence map Confidence plot with no echo background Weighted channel splicing is performed to obtain multi-channel structural features. : ， in, This indicates an operation that splices along the channel dimension. The weighting coefficients represent the weighting coefficients of the high-frequency components at the edge. The weighting coefficients represent the core components of the strong echo. The weighting coefficients represent the weights of the echo-free background component; Multi-channel structural features Local normalization is performed to obtain the multi-component radar structure prior features. : ， in, Indicates the direction of spatial height and width direction Local max pooling operation on, To prevent extremely small constants with a denominator of zero.

3. The downscaling method for structure-guided and frequency-aware radar according to claim 1, characterized in that, The dual-band adaptive modulation module includes a learnable band embedding layer and a two-layer multilayer perceptron. The discrete radar band identifiers corresponding to the input radar echo data are input into the learnable band embedding layer to obtain band embedding features. Embedding band features The input consists of a two-layer multilayer perceptron composed of a fully connected layer and a GELU activation function. Generate modulation vectors that match the dimensions of deep feature channels. The data is then divided equally along the channel dimension to obtain the scaling factor. Translation factor : ， ， in, This indicates that the channel dimension is divided equally.

4. The structure-guided and frequency-aware radar downscaling method according to claim 2, characterized in that, The deep feature extraction network comprises four cascaded radar structure Transformer groups. The output of the previous radar structure Transformer group serves as the input to the next radar structure Transformer group. The input features of a group of radar structure Transformers are denoted as follows: The output features are denoted as , The input of the first radar structure Transformer group , These are characteristics of shallow networks; Each radar architecture Transformer group includes an echo structure-guided attention module and a high-frequency excitation feedforward enhancement module; A radar structure Transformer group pairs input features Perform layer normalization to obtain the first normalized feature. Scaling factor generated using dual-band adaptive modulation module Translation factor For the first normalized feature Perform band-adaptive affine modulation to obtain the first modulation feature. : ， in, This represents Hadamard's pixel-wise multiplication operation; The first modulation feature and multi-component radar structure prior features The local windows are divided into corresponding non-overlapping local windows to obtain their features. and local structural priors : ， ， in, Indicates a local window division; The echo structure guides the attention module to perform a linear projection on each local window feature to obtain the query matrix. Key matrix Sum matrix and will Divided along the channel dimension The first attention point, to get the first The query submatrix corresponding to each attention head Key matrix Sum submatrix : ， ， ， ， ， ， in, These are the learnable linear projection weights corresponding to the query matrix. These are the learnable linear projection weights corresponding to the key matrix. The learnable linear projective weights corresponding to the value matrix; For any two spatial locations within the same local window and ,according to and Corresponding multi-component radar structure prior features and strong echo core confidence map and the confidence plot of the background without echo Constructing a radar structure with paired offsets : ， in, Indicates position and Corresponding multi-component radar structure prior features and Structural similarity between them This indicates the weight of the strong echo core consistency enhancement term. Indicates position The core confidence level of the strong echo at the location, Indicates position The core confidence level of the strong echo at the location, This indicates the weight of the suppression term that correlates strong echo cores with unecho background errors. Indicates position Confidence level of the echo-free background at the location, Indicates position Confidence level of the echo-free background at the location; The radar structure is biased in pairs. By incorporating the attention score of the dot product between the query submatrix and the key submatrix, we obtain the first... The structure of each attention head guides attention scores : ， in, Indicates transpose. It is the feature dimension of a single attention head in multi-head attention; Combined with the The dynamic discrete mask matrix transmitted by the Transformer group of the radar structure , obtained the The final attention response distribution of each attention head In this case, the dynamic discrete mask matrix received by the first radar structure Transformer group is set to an all-zero matrix: ， in, It is normalization; Using the first The final attention response distribution of each attention head AND-valued submatrix Perform feature aggregation to obtain the first Attention output features of each attention head Then The attention output features of each attention head are concatenated through channels and then linearly mapped using a linear mapping matrix to obtain window-level attention output features. : ， ， in, It is a linear mapping matrix; Based on the final attention response distribution Strong echo core confidence map Confidence plot of echo-free background and the The dynamic discrete mask matrix transmitted by the Transformer group of the radar structure Calculate the radar structure status score at each spatial location within a local window. And score the radar structural status. Perform TopK filtering to generate the first The first radar structure Transformer group towards the first The dynamic discrete mask matrix transmitted by the Transformer group of the radar structure : ， ， in, This indicates averaging over the attention head dimension. , , , These represent the weighting coefficients of each item in the radar structural condition score. This indicates the extraction of the highest-rated item. A spatial location; Window-level attention output features The global permutation features are restored through window inverse rearrangement and linear mapping, resulting in the attention output features. and with the Input features of a radar structure Transformer group Perform residual fusion to obtain intermediate features : ， in, Indicates random depth.

5. The structure-guided and frequency-aware radar downscaling method according to claim 4, characterized in that, For intermediate features Perform layer normalization to obtain the second normalized feature. And a scaling factor generated by a dual-band adaptive modulation module. Translation factor For the second normalized feature Band-adaptive affine modulation is performed to obtain the second modulation feature. : ， The second modulation feature The high-frequency excitation feedforward enhancement module is input, and channel dimensionality is increased through channel convolution and extended features are obtained through reflection filling. : ， in, Represents up-dimensional channel convolution. Indicates a reflection fill operation; Extracting local detail response features using depthwise separable convolution Local low-frequency smoothing features are extracted using average pooling. Based on local detail response characteristics With local low-frequency smoothing features The difference is used to obtain the high-frequency residual characteristics. : ， ， ， in, This represents depthwise separable convolution. Indicates average pooling; Based on local detail response features Generate nonlinear gated weights : ， in, This represents the Sigmoid activation function. This indicates the convolutional layer used to generate the gate weights; Using nonlinear gated weights and learnable scaling parameters High-frequency residual characteristics Adaptive modulation is performed to obtain high-frequency excitation characteristics. High-frequency excitation characteristics By sequentially inputting the GELU activation function and channel-reduced convolution, feedforward enhanced features are obtained. : ， ， in, This represents channel-reduced convolution. Indicates the activation function; feedforward enhancement features With intermediate features Perform residual fusion to obtain the first Output characteristics of a radar structure Transformer group : ， when At that time, output features As deep features .

6. The structure-guided and frequency-aware radar downscaling method according to claim 5, characterized in that, deep features The channel-expanded convolution is mapped to a channel dimension that matches the target upsampling factor, and then an initial high-resolution prediction tensor is generated through sub-pixel rearrangement. The frequency domain annular residual refinement module refines the initial high-resolution prediction tensor. Performing a two-dimensional discrete Fourier transform yields a complex frequency domain tensor. ; High-frequency feature maps of the edges Strong echo core confidence map Confidence plot with no echo background Resampling to the initial high-resolution prediction tensor Spatial scales were used to obtain scale-aligned edge high-frequency maps. Scale-aligned strong echo core confidence map Scale-aligned echo-free background confidence plot : ， ， ， in, It is a bilinear interpolation sizing operation. This indicates an operation to obtain spatial dimensions; Perform two-dimensional discrete Fourier transforms on the scale-aligned structure diagrams and take the amplitudes to obtain the structure's frequency domain guidance features. : ， in, It is a two-dimensional discrete Fourier transform. This represents the modulus of the complex frequency domain result obtained from the two-dimensional Fourier transform; Based on structural frequency domain guidance characteristics Constructing the normalized radial frequency radius in the frequency domain coordinates : ， in, Represents the frequency coordinates in the horizontal direction. Represents the frequency coordinates in the vertical direction. Indicates the Nyquist frequency in the horizontal direction. Indicates the Nyquist frequency in the vertical direction; Construct a low-frequency protection mask and multiple radial frequency ring band masks based on the normalized radial frequency radius: ， ， in, This represents a conditional function; if the condition is met, the value is 1, and if not, the value is 0. Indicates a low-frequency protection mask. Indicates the first A radial frequency ring band mask Indicates the low-frequency protection threshold. Indicates the first The inner radius threshold of a radial frequency band. Indicates the first The outer radius threshold of a radial frequency loop Indicates the total number of radial frequency bands; complex frequency domain tensor Real part, imaginary part, low frequency protection mask Radial frequency ring band mask to and structural frequency domain guidance features Input frequency domain loop-gated network Generate candidate frequency domain residuals : ， in, Represents the complex frequency domain tensor The real part, Represents the complex frequency domain tensor The imaginary part; Using low-frequency protection mask For candidate frequency domain residuals Low-frequency suppression is performed, and Hermitian conjugate symmetric constraints are applied to the candidate frequency domain residuals after low-frequency suppression to obtain frequency domain residuals that meet the requirements of real-valued spatial domain reconstruction. : ， in, This represents the Hermitian conjugate symmetry constraint. frequency domain residual The spatial residual features are obtained by mapping back to the spatial domain through a two-dimensional discrete Fourier inverse transform, and then performing local neighborhood fusion and cross-channel correction through 3×3 convolution. The initial high-resolution prediction tensor and spatial residual characteristics The data is then fused to obtain the final high-resolution radar echo reconstruction result. .

7. The downscaling method for structure-guided and frequency-aware radar according to claim 1, characterized in that, In both the first and second phases of training, the L1-frequency domain annular band consistency loss function was used to optimize the model parameters. Including L1 loss and frequency domain ring band consistency loss The expression is: ， in, It is frequency domain ring band consistency loss Weighting coefficients; L1 loss is used to constrain high-resolution radar echo reconstruction results. With X-band high-resolution radar echo data Pixel-level error in the spatial domain: ， in, and These represent the height and width of the feature map, respectively. For position The model output value at that location, This represents the true high-resolution ground truth value at the corresponding location; Frequency domain ring band consistency loss is used to constrain the consistency of the amplitude spectrum between the high-resolution radar echo reconstruction results and the real X-band high-resolution radar echo data in different radial frequency ring bands: ， in, Indicates the total number of radial frequency bands. Indicates the radial frequency band number. express Norm, Indicates the first A radial frequency ring band mask This represents a two-dimensional Fast Fourier Transform. Indicates taking the amplitude spectrum. This indicates Hadamard's pixel-by-pixel multiplication operation.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the structure-guided and frequency-aware radar downscaling method as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the structure-guided and frequency-aware radar downscaling method as described in any one of claims 1 to 7.

Images