A multi-band radar heavy rain data correction method based on deep learning and physical guided feature fusion

Abstract

This invention belongs to the field of meteorological observation data processing and discloses a method for correcting multi-band radar heavy rainfall data based on deep learning and physical guidance feature fusion. The method first performs spatiotemporal registration and preprocessing on S-band and X-band radar data; then, it constructs a composite attenuation characterization quantity to generate a physical guidance map representing the attenuation risk level; finally, it constructs a deep learning model, partitions the attenuation risk area according to the physical guidance map, and outputs the correction results using a differentiated cross-band fusion strategy. During training, content loss, structural constraint loss, and regional adaptive structure preservation loss are introduced, and the original observation consistency term in the regional adaptive structure preservation loss is used to constrain the output of low-attenuation areas to be consistent with the original observations. This method can improve the structural reconstruction quality and data reliability of X-band radar attenuation areas under heavy rainfall conditions.

Description

Technical Field

[0001] This invention belongs to the field of meteorological observation data processing, specifically involving a method for correcting multi-band radar heavy rainfall data based on the fusion of deep learning and physical guidance features. Background Technology

[0002] X-band weather radars offer advantages such as small size, low cost, and high spatial resolution, making them suitable for detailed monitoring of heavy rainfall at the city level. However, their electromagnetic waves attenuate significantly in the path of heavy rainfall, often resulting in problems such as a sharp drop in echo energy, no-data voids downstream of the rain core, and echo structure distortion. This leads to increased bias in quantitative precipitation estimation and reduced operational availability. In contrast, S-band radars have stronger resistance to attenuation and can stably characterize the macroscopic structure of heavy rainfall, but their spatial resolution is lower, making it difficult to provide detailed structural information at the city scale. Existing X-band attenuation compensation methods mainly include: (1) attenuation correction methods based on radar physical equations, such as compensation ideas using parameters like phase / differential propagation phase shift. When strong attenuation or complete attenuation leads to the loss of echo information, there are problems such as insufficient available observations and unstable compensation results; (2) methods based on interpolation or image morphology, which focus on restoring spatial continuity and are difficult to ensure that the completed structure is consistent with the actual precipitation echo organization morphology; (3) single-band super-resolution / repair methods based on deep learning, which usually lack calculable physical indication signals for attenuation regions and lack constraints on the statistical consistency of non-attenuation regions. They are prone to generating false details at strong convection boundaries or strong echo cores, reducing the reliability of the results. Therefore, there is an urgent need for a radar data correction method that can utilize the complementary advantages of S-band structural integrity and X-band high-resolution details, introduce a calculable attenuation guidance mechanism in the cross-band fusion process, and impose physical consistency constraints on the correction results, so as to achieve reliable completion and structural restoration of X-band attenuation hole regions in heavy rainfall scenarios. Summary of the Invention

[0003] Objective: This invention addresses the problems of echo attenuation, local voids, structural distortion, and decreased reliability of quantitative applications in X-band weather radar under heavy rainfall conditions. It proposes a multi-band radar data correction method based on deep learning and physical guidance feature fusion. This invention leverages the advantages of S-band radar (lower attenuation and more stable echo structure under heavy rainfall conditions) and X-band radar (higher spatial resolution and richer local detail information). By constructing a composite physical guidance map, it characterizes the potential attenuation risks at different spatial locations. Based on this, it achieves partitioned adaptive cross-band feature fusion and regional adaptive training constraints, thereby improving the structural recovery capability, detail preservation capability, and result reliability of X-band radar data under heavy rainfall scenarios.

[0004] The method of the present invention includes the following steps: Step 1: Acquire S-band radar observation data and X-band radar observation data for the same precipitation process used in training. The S-band radar observation data includes reflectivity factors for two elevation or height layers. Perform time synchronization and spatial registration on the S-band radar observation data and X-band radar observation data used in training, and resample them to the same grid to obtain S-band grid data and X-band grid data used in training. Step 2: Construct a training composite physical guidance map A* based on the training S-band grid data and the training X-band grid data. The training composite physical guidance map A* is used to characterize the potential attenuation risk level of each spatial location in the training samples. Step 3: Construct a deep learning correction model, which includes a semantic structure feature extraction module, a detail feature extraction module, a physically guided feature fusion module, and a high-resolution reconstruction output module. The semantic structure feature extraction module is used to extract multi-scale semantic structure features F from S-band grid data. s ; The detail feature extraction module is used to extract multi-scale local detail features F from X-band grid data. x ; The physical guidance feature fusion module is used to partition the spatial location into attenuation risk zones based on the composite physical guidance map A*, and to apply differentiated cross-band fusion strategies to different zones to obtain the fused feature F. fuse ; The high-resolution reconstruction output module is used to determine the fusion feature F. fuse Output the corrected X-band reflectivity factor map X0; Step 4: Use the training samples consisting of training S-band grid data, training X-band grid data, and training composite physical guidance map A* to train the deep learning correction model. The training loss function includes content loss, structural constraint loss, and region adaptive structure preservation loss. The region adaptive structure preservation loss is used to apply differentiated supervision constraints to the high attenuation risk region, the transition fusion region, and the low attenuation reliable region based on the different attenuation risk regions represented by the composite physical guidance map A*. Step 5: Obtain the S-band radar observation data and X-band radar observation data to be corrected, preprocess them in the same way as in Step 1 and Step 2, and construct the composite physical guidance map A* to be corrected. Input the S-band grid data to be corrected, the X-band grid data to be corrected, and the composite physical guidance map A* to be corrected into the trained deep learning correction model to obtain the corrected X-band reflectivity factor map X0.

[0005] In step 1, the time synchronization involves matching S-band radar observation data and X-band radar observation data according to the same observation time or a preset time window; the spatial registration involves converting data from different radar coordinate systems to a unified geographic coordinate system and cropping it to a common coverage area; and the resampling involves unifying S-band radar observation data and X-band radar observation data to grid data with the same spatial resolution.

[0006] Step 2 includes: performing differential or gradient calculations on the reflectivity factors of adjacent elevation angles or height layers in the S-band to obtain the first attenuation characterization quantity; The second attenuation characterization quantity is obtained based on the spatial connectivity, morphological distribution, or inconsistency with the S-band high echo region in the low-value region or missing region of X-band. The first attenuation characterization quantity and the second attenuation characterization quantity are weighted and combined, and then segmented and mapped to obtain the composite physical guidance diagram A*.

[0007] The X-band low-value region is a connected region where the X-band reflectivity factor is lower than a preset low-value threshold; the missing region is a region in the X-band radar data where there is no effective reflectivity factor value; the S-band strong echo region is a region where the S-band reflectivity factor is higher than a preset strong echo threshold. The preset low-value threshold can be determined based on the X-band radar noise level, historical precipitation sample statistics, or operational experience, preferably between 10 dBZ and 20 dBZ; the preset strong echo threshold can be determined based on the S-band radar heavy rainfall echo statistics or operational experience, preferably between 35 dBZ and 45 dBZ.

[0008] In step 2, the first attenuation characterization quantity includes at least one of the following: the difference in reflectivity factor between adjacent elevation angles in the S-band; the difference in reflectivity factor between adjacent height layers in the S-band; the local vertical gradient in the S-band; and the rate of change of reflectivity factor between adjacent grids or adjacent height layers within the strong echo region of the S-band. The second attenuation characterization quantity includes at least one of the following: the area of ​​the connected domain, boundary morphology, aspect ratio, or hole distribution characteristics of the low-value region or missing region in the X-band; the spatial overlap difference between the low-value region or missing region in the X-band and the strong echo region in the S-band; and the difference in reflectivity factor between the S-band and X-band at the corresponding grid positions.

[0009] In step 2, the expression for the weighted combination of the first attenuation characteristic and the second attenuation characteristic is: G*=αG1+βG2+γG3+δG4, Wherein, G* is the composite attenuation characterization quantity, G1 is the difference in reflectivity factor between adjacent elevation angles or adjacent height layers in the S-band, G2 is the change rate of local vertical gradient or strong echo in the S-band, G3 is the spatial connectivity structure term of low-value area or missing area in the X-band, G4 is the difference in reflectivity factor between the S-band and X-band at corresponding grid positions, and α, β, γ, and δ are weighting coefficients. The composite attenuation characterization quantity G* is obtained by performing piecewise normalization mapping or partition threshold mapping to obtain the composite physical guidance graph A*.

[0010] In step 3, the attenuation risk partitioning includes dividing the spatial location into a high attenuation risk zone, a transition fusion zone, and a low attenuation reliable zone according to the composite physical guidance map A*; wherein, in the high attenuation risk zone, the contribution of S-band semantic structure features is enhanced, in the low attenuation reliable zone, the preservation of X-band local detail features is enhanced, and in the transition fusion zone, the collaborative fusion of S-band semantic structure features and X-band local detail features is performed.

[0011] In step 3, the physical guidance feature fusion module includes a scale matching unit, a weight generation unit, a weighted fusion unit, and a cross-modal attention enhancement unit; wherein, the scale matching unit is used to map the composite physical guidance graph A* to a scale consistent with the current scale features; the weight generation unit is used to generate gating weights W based on the mapped composite physical guidance graph. l ; The gating weight W l Composite Physical Guidance Map A after Scale Matching l It is obtained through normalized mapping, and A l The larger the value of W, the better. l The larger the value, the greater the weight of the S-band semantic structure features in the fused features. The specific formula is as follows: W l =σ(Conv(A l )) , Where A l Let be the composite physical guidance map at the l-th scale, where Conv represents a convolutional mapping with a kernel size of 1×1, and σ represents the Sigmoid function. The weighted fusion unit is used to fuse features at the l-th scale in the following manner: F mix l =W l ⊙F s l +(1 W l )⊙F x l , Among them, F s l For the semantic structure features of the S-band at the l-th scale, Fx l For the local detail features of the X-band at the l-th scale, F mix l This represents the initial fusion feature at the l-th scale; ⊙ indicates element-wise multiplication; The cross-modal attention enhancement unit is used to enhance the preliminary fused features across modalities to obtain the fused feature F. fuse l .

[0012] In step 4, the content loss is used to constrain the numerical difference between the correction result and the reference supervision; the structural constraint loss is used to constrain the edge, gradient, or structural similarity between the correction result and the reference supervision; and the region adaptive structure preservation loss is used to apply different supervision rules to different decay risk regions according to the composite physics guidance graph A*. The regional adaptive structure preservation loss includes an S-band macroscopic structure consistency term, an X-band original observation consistency term, and a spatial continuity term. The weight of the S-band macroscopic structure consistency term is increased in high-attenuation-risk areas, the weight of the X-band original observation consistency term is increased in low-attenuation-reliable areas, and the weight of the spatial continuity term is increased in transition fusion areas.

[0013] Step 4 also includes: constructing an effective X-band observation mask M valid and attenuation region mask M att , of which M valid Used to mark grids with effective reflectivity factor values ​​in X-band radar data; M att The grid is used to mark grids that meet preset attenuation criteria. These preset attenuation criteria include: missing X-band data; or, an X-band reflectivity factor lower than a preset low threshold and a corresponding S-band reflectivity factor higher than a preset strong echo threshold. The non-attenuation region mask M is obtained by the following expression. non : M non =M valid ⊙(1 M att ), The non-attenuation region mask M non Used to constrain the consistency of correction results with the original X-band observations within the low-attenuation reliability zone during the training phase.

[0014] In step 4, the reference supervision X corresponding to the training sample consists of at least one of the following data: effective reflectivity factor grid data of the non-attenuation region of the X-band; high-quality X-band reflectivity factor grid data; X-band reflectivity factor reference field formed after manual quality control; X-band reflectivity factor reference field generated based on S-band mapping. The X-band reflectivity factor reference field formed after manual quality control is a two-dimensional reflectivity factor numerical field located in the same spatiotemporal grid as the corrected X-band reflectivity factor map X0 output by the deep learning correction model, and is used as a reference supervision X during the training phase.

[0015] In this regard, missing measurements are not directly used as supervision for attenuation hole regions; instead, indirect constraints are applied by combining the region adaptive structure preservation loss. The attenuation hole region refers to the region where X-band data is missing, or where the X-band reflectivity factor is lower than a preset low threshold and the corresponding S-band reflectivity factor is higher than a preset strong echo threshold.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By constructing a composite physical guidance map A* that combines the vertical structural features of the S-band, the cavity morphology features of the X-band, and cross-band difference information, this invention can more accurately characterize the potential attenuation risk level of different spatial locations and improve the ability to identify areas of heavy rainfall attenuation.

[0017] (2) Based on the composite physical guidance map A*, the present invention divides the spatial location into attenuation risk zones and adopts a differentiated cross-band fusion strategy for different regions, thereby realizing the partitioned adaptive fusion of S-band semantic structure information and X-band detail information, thereby improving the structural completion capability of attenuation region and the detail preservation capability of non-attenuation region.

[0018] (3) The present invention introduces regional adaptive structure preservation loss during the training phase and applies differentiated supervision constraints to regions with different decay risks, which can enhance the regional continuity and structural rationality of the output results and reduce false edges and unnatural transitions.

[0019] (4) The present invention can restore the spatial structure and local details of heavy rainfall echoes in the attenuation risk area and maintain the consistency of the original X-band observations in the low attenuation reliable area, thereby improving the reliability and application value of X-band radar data correction results in heavy rainfall scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the overall process of the multi-band radar heavy rainfall data correction method based on deep learning and physical guidance feature fusion, as presented in this invention. It shows the acquisition of S-band and X-band data, spatiotemporal registration and preprocessing, construction of the composite physical guidance map A*, A*-based partitioned adaptive cross-band feature fusion, model training, and consistent stitching output X. out The steps involved.

[0021] Figure 2This is a schematic diagram of the multi-band radar data preprocessing process of the present invention, illustrating the time synchronization, spatial registration and coordinate transformation, common coverage area clipping, grid resampling, quality control, and training / inference sample construction process of S-band and X-band radar data.

[0022] Figure 3 This is a schematic diagram of the overall structure of the deep learning correction model of the present invention. The model consists of a semantic structure feature extraction module, a detail feature extraction module, a physical guidance feature fusion module, and a high-resolution reconstruction output module. The physical guidance feature fusion module performs partitioned adaptive fusion of cross-band features based on the composite physical guidance map A*.

[0023] Figure 4 This is a schematic diagram of the semantic structure feature extraction module of the present invention, showing a multi-scale coding structure based on S-band input data, used to extract the macroscopic organization morphology and boundary structure features F of heavy rainfall echoes. s l .

[0024] Figure 5 This is a schematic diagram of the physical guidance feature fusion module of the present invention, showing the scale-gated weights W formed by the composite physical guidance graph A* after scale matching and weight generation. l And achieve F through weighted fusion unit mix l =W l ⊙F s l +(1 W l )⊙F x l The feature fusion method is then combined with a cross-modal attention enhancement unit to obtain the fused feature F. fuse l .

[0025] Figure 6 This is a schematic diagram of the model training and loss function design process of the present invention, showing the model output X0, the reference supervision X, and the non-decaying region mask M. non The loss is calculated by including content loss, structural constraint loss, and region adaptive structure preservation loss, and the total loss L is obtained by weighted summation. total The process.

[0026] Figure 7 This diagram illustrates the correction effect of the method of the present invention, showing how the original X-band radar reflectivity factor map is corrected by the trained model to generate an X-band radar reflectivity factor map X0. It demonstrates that the present invention can fill in low-value, missing, or echo-hole areas caused by heavy rainfall attenuation, improve echo continuity, restore the main shape and edge details of rain clusters, and enhance the usability and reliability of X-band radar data. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments.

[0028] This invention provides a method for correcting heavy rainfall data from multi-band radar based on deep learning and physical guidance feature fusion. It is primarily used to correct echo holes, local missing data, structural distortion, and strong echo region distortion caused by precipitation attenuation in X-band weather radar under heavy rainfall conditions. This method leverages the characteristics of S-band radar (low attenuation and stable macroscopic structure under heavy rainfall conditions) and X-band radar (high spatial resolution and strong ability to represent local details). By constructing a composite physical guidance map A*, it achieves cross-band feature adaptive fusion oriented towards attenuation risk zones, and improves the reliability and structural rationality of the output results through regional adaptive supervision constraints. In this embodiment, S-band radar refers to a weather radar operating in the S-band, and X-band radar refers to a weather radar operating in the X-band; A* represents the composite physical guidance map, G* represents the composite attenuation characterization quantity, and F... s F x F represents the features extracted from the S-band and X-band, respectively. fuse X0 represents the fused features; X0 represents the corrected X-band reflectance factor map directly output by the deep learning correction model. out This represents the final output result after consistent splicing of the non-attenuation regions; M valid M att M non These represent the effective observation mask, attenuation region mask, and non-attenuation region mask for the X-band, respectively. The gradient operator is represented here. The letters, subscripts, and mathematical symbols retained in the accompanying figures are symbolic representations of the aforementioned technical variables, used to maintain consistency with the formulas in the specification.

[0029] like Figure 1 As shown, the overall process of this invention includes data preprocessing, construction of a composite physics guidance graph, construction of a deep learning correction model, model training, model correction output, and consistency splicing output. Figure 1 Each processing box in the diagram corresponds one-to-one with the steps described below in this embodiment, wherein the correction result X0 is the direct output of the model, and the final output result X... out This is to provide the operationally usable correction results obtained by combining the non-attenuation region mask.

[0030] (I) Data Acquisition and Preprocessing like Figure 2As shown, the data acquisition and preprocessing stage simultaneously receives S-band radar observation data and X-band radar observation data, and sequentially completes time synchronization, spatial registration and coordinate transformation, grid resampling, data quality control, S-band multi-layer reflectivity field extraction, X-band effective observation area identification, and attenuation area mask construction, ultimately forming training or inference samples.

[0031] First, acquire S-band radar observation data and X-band radar observation data corresponding to the same rainfall event. The S-band radar observation data is preferably a reflectivity factor product with multiple elevation angles or multiple altitude layers, and the X-band radar observation data is preferably a reflectivity factor product with a spatiotemporal overlap coverage area with the S-band radar.

[0032] In one specific embodiment, a single urban heavy rainfall event was selected as the sample. The S-band radar used reflectivity factor products at elevation angles of 0.5° and 1.5°, while the X-band radar used low-elevation reflectivity factor products from the same observation time. The time matching window was set to ±3 minutes, and the common coverage area was cropped to 64km × 64km, with all samples uniformly resampled to a 250m × 250m grid. After preprocessing, each sample included two layers of S-band reflectivity factor fields, an X-band reflectivity factor field, and an effective observation mask M. valid And the attenuation region mask M, which is jointly identified by the low-value region of the X-band, the missing region, and the difference in reflectivity factor between the S / X-band. att .

[0033] In this embodiment, the S-band radar observation data and X-band radar observation data are preprocessed as follows: 1. Time synchronization processing: Match S-band radar observation data with X-band radar observation data according to the same observation time or preset time window; 2. Spatial registration processing: Transform observation data from different radar coordinate systems to a unified geographic coordinate system and crop to obtain a common coverage area; 3. Grid resampling processing: Unify S-band radar observation data and X-band radar observation data to the same spatial resolution and grid coordinate system, preferably to the target grid resolution of X-band radar; 4. Quality control processing: Remove missing values, outliers, and non-meteorological echoes, retain valid precipitation echo areas, and form standardized input data.

[0034] After the above processing, S-band grid data S and X-band grid data X are obtained on a unified spatiotemporal grid.

[0035] (II) Construction of Composite Physics Guidance Diagram A* This implementation differs from methods that rely solely on S-band adjacent elevation angle differences or single gradient maps to construct guidance maps. Instead, it constructs a composite physical guidance map A* by fusing S-band vertical structure information, X-band suspected attenuation hole spatial morphology information, and cross-band local difference information to characterize the potential attenuation risk level at different spatial locations.

[0036] 1. Construction of the first attenuation characterization quantity Based on the reflectivity factors of multiple elevation angles or multiple altitude layers in the S-band, characteristic quantities reflecting the vertical structure changes of precipitation are extracted.

[0037] For two adjacent elevation angles or adjacent height layers, the S-band reflectivity factor Z s (h1) and Z s (h2), where h1 and h2 represent two adjacent elevation angles or adjacent height layers, and h1 and h2 correspond to the same spatial grid position, and a difference term G1=|Z can be constructed. s (h1) Z s (h2)|, and the vertical gradient term G2=| can be further constructed. Z s |

[0038] in, Z s This represents the gradient change of the S-band reflectivity factor along the vertical direction.

[0039] The aforementioned first attenuation characterization quantity is used to reflect the rapid change characteristics of echoes on the vertical structure under heavy rainfall conditions, and serves as an important a priori basis for potential attenuation risks.

[0040] 2. Construction of the second attenuation characterization quantity In addition to the S-band vertical structure characteristics, this embodiment further utilizes the spatial morphological information related to attenuation and cross-band difference information in the X-band data to construct a second attenuation characterization quantity.

[0041] For suspected attenuation hole regions R in the X-band xConnectivity analysis is performed to extract features such as hole area, boundary morphology, and aspect ratio, constructing a connectivity structure term G3. G3 can be obtained by normalizing and weighting the connected region area, boundary complexity, and aspect ratio of the suspected decaying region, expressed as G3 = η1Area_norm + η2Boundary_norm + η3Ratio_norm, where Area_norm, Boundary_norm, and Ratio_norm are the normalized connected region area, boundary complexity, and aspect ratio, respectively, and η1, η2, and η3 are weighting coefficients. The boundary complexity Boundary can be calculated as Boundary = P² / (4πS), where P is the perimeter of the connected region boundary, and S is the area of ​​the connected region. After normalizing the boundary complexity Boundary, Boundary_norm is obtained, and Boundary_norm is used to construct the connectivity structure term G3.

[0042] For cross-band difference information, construct a local difference term G4=max(Z s -Z x ,0), Among them, Z s and Z x These represent the S-band and X-band reflectivity factors at the corresponding locations, respectively. If the S-band echo is significantly stronger than the X-band echo, then that location is more likely to correspond to the X-band attenuation region.

[0043] 3. Construction of the composite attenuation characterization quantity G* By weighting and combining the above multiple characterization terms, we obtain the composite attenuation characterization quantity G*: G*=αG1+βG2+γG3+δG4, Among them, α, β, γ, and δ are the corresponding weight coefficients, which can be determined based on sample statistical results, empirical parameters, or adaptive adjustment methods during the training phase.

[0044] In the above specific embodiments, α, β, γ, and δ can be set to 0.25, 0.25, 0.30, and 0.20 respectively as initial weights; they can also be updated as learnable parameters during the training phase. The composite physics guidance graph A* is obtained by normalizing the composite decay characterization G*, and its expression is A*=min(max((G* T min ) / (T max T min ),0),1), where T min and T maxTo preset the normalization threshold, the 5th and 95th quantiles of the composite attenuation characterization quantity G* in the training samples can be used, or the minimum and maximum effective values ​​of G* can be used based on the sample statistical results. After normalizing G* to [0,1], two thresholds of 0.65 and 0.35 can be set: positions where A* ≥ 0.65 are identified as high attenuation risk areas, positions where 0.35 < A* < 0.65 are identified as transition fusion areas, and positions where A* ≤ 0.35 are identified as low attenuation reliable areas. The risk partitioning thresholds can be adjusted according to the distribution of A* in the training samples, for example, by taking the high-risk quantile threshold and the low-risk quantile threshold of the A* distribution. The A* obtained through this step shows a high value in the downstream region of the suspected strong attenuation rain core, an intermediate value at the effective observation edge of the X-band, and a low value in the region without obvious attenuation.

[0045] 4. Mapping of the composite physics-guided diagram A* The composite attenuation characterization quantity G* is segmented and normalized or partitioned by threshold mapping to obtain the composite physical guidance map A*. A* is normalized to the [0,1] interval, and its value increases with the degree of potential attenuation risk. Multiple risk level thresholds are set to partition A* for attenuation risk in different value intervals, so as to perform differentiated feature fusion and training constraints in the subsequent process. The composite physical guidance map A* constructed through the above steps can more comprehensively characterize the attenuation risk of X-band heavy rainfall compared to a single difference map or a single gradient map, and provides physical priors for subsequent partitioning fusion.

[0046] (III) Construction of Deep Learning Correction Model like Figure 3 As shown, the deep learning correction model takes S-band input data, X-band input data, and a composite physical guidance map A* as inputs; the S-band input data is processed by a semantic structure feature extraction module to obtain S-band semantic structure features Fs, and the X-band input data is processed by a detail feature extraction module to obtain X-band local detail features F. x In the physical guidance feature fusion module, the two are combined with A* for partitioned adaptive fusion to obtain the fused feature F. fuse The high-resolution reconstruction output module then generates the correction result X0.

[0047] The deep learning correction model in this embodiment includes at least a semantic structure feature extraction module, a detail feature extraction module, a physically guided feature fusion module, and a high-resolution reconstruction output module.

[0048] 1. Semantic structural feature extraction module like Figure 4As shown, the semantic structure feature extraction module includes an initial feature extraction unit, multiple scale semantic feature extraction units, and a downsampling unit. When using an S-band multi-layer reflectivity field as input, the module extracts semantic features at the first, second, and third scales, denoted as Fi, respectively. s 1 F s 2 and F s 3 It is used to characterize the overall direction, rain core location, and boundary structure of heavy rainfall echoes.

[0049] The semantic structure feature extraction module is used to extract multi-scale semantic structure features from S-band gridded data S. This module employs a multi-scale coding structure to encode the heavy rainfall organization morphology, rain core distribution trend, and macroscopic structure in the S-band echo, obtaining semantic structure features F at different scales. s l .

[0050] Where l represents the number of feature scale layers.

[0051] 2. Detail Feature Extraction Module The detail feature extraction module is used to extract multi-scale local detail features from X-band grid data. This module preferably employs a convolutional feature extraction structure to encode the high-resolution textures, local edges, and spatial details still preserved in the X-band, obtaining detail features F at different scales. x l .

[0052] 3. Physical guidance feature fusion module The physical guidance feature fusion module is the core module of this invention. It is used to partition the spatial location into attenuation risk zones based on the composite physical guidance map A*, and to perform differentiated fusion of S-band semantic structural features and X-band local detail features at different scales and spatial locations. The physical guidance feature fusion module includes: a scale matching unit; a weight generation unit; a weighted fusion unit; and a cross-modal attention enhancement unit.

[0053] (1) Scale matching unit Due to the composite physics guiding diagram A* and feature diagram F s l F x l The spatial scales may be different, so A* is first mapped to a scale consistent with the current scale features through a scale matching unit to obtain a scale-aligned guide map.

[0054] (2) Weight generation unit The scale-matched guide graph is input into the weight generation unit to generate the gating weight W at the current scale. lW l This is used to characterize the preference of each spatial location at the l-th scale for S-band structural information and X-band detail information.

[0055] (3) Weighted fusion unit like Figure 5 As shown, the physical guidance feature fusion module first performs scale matching on the composite physical guidance graph A*, and then generates the current scale gating weight W. l And the semantic structure features of the S-band at scale l are F s l and X-band local detail features F x l Input to the weighted fusion unit; the weighted fusion result then enters the cross-modal attention enhancement unit to obtain the l-th scale fusion feature F. fuse l .

[0056] At the l-th scale, according to the gating weight W l For F s l and F x l Weighted fusion is performed to obtain preliminary fusion characteristics: F mix l =W l ⊙F s l +(1 W l )⊙F x l , Here, ⊙ represents element-wise multiplication.

[0057] The physical significance of this fusion method lies in the fact that when a certain spatial location corresponds to a high risk of attenuation, W l Larger bands prioritize inheriting S-band semantic structure information; when a spatial location corresponds to a low-attenuation reliable region, W... l Smaller sizes prioritize preserving local details in the X-band; when a spatial location corresponds to a transition fusion region, W... l By taking the median value, the S and X features are synergistically fused.

[0058] (4) Cross-modal attention enhancement unit To further enhance the effective interaction between different modal features, the preliminary fusion feature F can be modified. mix l Cross-modal attention enhancement is performed to obtain the final fused feature F. fuse l Using F x l As a query, with F sl As Key, with F mix l As a cross-modal attention mechanism for Value, it enables synergistic enhancement between detailed information and structural information.

[0059] 4. High-resolution reconstruction output module The high-resolution reconstruction output module is used to progressively reconstruct and scale-restore the fused features at various scales, outputting the corrected X-band reflectivity factor map X0. X0 represents the correction result directly output by the model.

[0060] (iv) Adaptive fusion strategy based on attenuation risk partitioning In this implementation, instead of using a uniform fusion rule for different spatial locations, adaptive fusion is performed based on the risk level represented by the composite physical guidance map A*.

[0061] The spatial location is divided into: high attenuation risk zone; transition fusion zone; and low attenuation reliable zone.

[0062] For high-attenuation-risk areas, S-band semantic structure features are used first to enhance the recovery capability of the structure of hollow areas, severely attenuated areas and core areas of heavy rainfall; for low-attenuation reliable areas, X-band local detail features are preserved first to minimize interference with the original valid observations; for transitional fusion areas, S / X feature collaborative fusion is adopted to ensure the continuity of transition between different areas.

[0063] Compared with conventional unified feature splicing or unified gating fusion methods, this implementation method achieves spatial partitioning control based on radar attenuation mechanism through composite physical guidance map A*, making the fusion process more consistent with the physical characteristics of different areas under heavy rainfall scenarios.

[0064] (V) Training Strategy and Loss Function Design like Figure 6 As shown, Figure 6 As shown, during the training phase, the model output X0, the reference supervision X, and the non-decaying region mask M are used. non Input different loss calculation units to calculate content loss, structural constraint loss, and region adaptive structure preservation loss respectively, and then obtain the total loss L through the total loss calculation unit. total This training process enables the model to prioritize structure recovery in regions of attenuation risk and prioritize maintaining consistency with the original observations in reliable, non-attenuated regions.

[0065] In this embodiment, the deep learning correction model is trained using training samples. The training loss function includes at least content loss, structural constraint loss, and region adaptive structure preservation loss.

[0066] In the specific embodiments described above, the training samples can be divided into training, validation, and test sets according to the rainfall process. The model input channels include S-band multilayer reflectivity factor, X-band reflectivity factor, and A*, and the output channel is the corrected X-band reflectivity factor map X0. During training, a batch input method is used, and the weights λ1, λ2, and λ3 for content loss, structural constraint loss, and region adaptive structure preservation loss can be set to 1.0, 0.2, and 0.5, respectively. After training, in the X-band strong attenuation region, the output results can fill in local holes caused by low values ​​or missing measurements in the original X-band; in the low attenuation reliable region, the output results are consistent with the original X-band observations, avoiding excessive smoothing of effective details.

[0067] 1. Content loss Content loss is used to constrain the numerical deviation between the model output and the reference supervision, employing L1 loss. Within the low-attenuation reliable region, a non-attenuation region mask M is introduced. non Content loss L content Apply constraints: L content =||M non ⊙(X0 X)||1, Where X represents the reference supervision, which can be the effective observation value of the non-attenuation region of the X-band within the low attenuation reliable region, and X0 represents the corrected X-band reflectivity factor map directly output by the deep learning correction model.

[0068] 2. Structural constraint loss Structural constraint loss is used to constrain the consistency between the model output and the reference supervision in terms of edge, gradient, or structural similarity, thereby enhancing the spatial structure preservation capability of the output echo field. A gradient difference loss L based on the Sobel operator is employed. edge :L edge =||M non ⊙( X0 X)||1.

[0069] in, The gradient operator is used to calculate the spatial variation of the reflectivity factor map in the horizontal and vertical directions. It can be implemented by the Sobel operator or the first-order difference operator.

[0070] 3. Region-adaptive structure preservation loss To embody the partitioned adaptive supervision concept of this invention, a region-adaptive structure preservation loss L is introduced in addition to the content loss and structural constraint loss. region : L region =λ h ·mean(Mh ⊙L s_struct )+λ m ·mean(M m ⊙L smooth )+λ l ·mean(M l ⊙L x_cons ), where L region M represents the region-adaptive structure preservation loss; h M m M l These are masks for high attenuation risk regions, transition fusion regions, and low attenuation reliable regions, respectively; L s_struct L smooth and L x_cons These are the grid-by-grid loss matrices with the same size as the output results; L s_struct The S-band macrostructure consistency term can be calculated from the gradient difference or structural similarity difference between the output and the S-band reference structure; L smooth The term representing the spatial continuity of the transition region can be calculated from the difference in reflectivity factors between adjacent grids; L x_cons The X-band original observation consistency term can be calculated from the absolute error between the output result and the original X-band observation; mean represents the average value of the grid loss over the selected area of ​​the mask; ⊙ represents element-wise multiplication; λ h , λ m , λ l To maintain the weight coefficients of the terms within the loss for the region adaptive structure.

[0071] The regional adaptive structure preservation loss is used to apply differentiated supervision rules to different regions based on the different attenuation risk regions represented by the composite physical guidance map A*. Specifically, it includes: increasing the weight of the S-band macrostructure consistency term in the high attenuation risk region; increasing the weight of the X-band original observation consistency term in the low attenuation reliable region; and increasing the weight of the spatial continuity term in the transition fusion region to suppress false edges and unnatural transitions at the boundary between the attenuation region and the non-attenuation region.

[0072] 4. Mask structure Construct an effective observation mask M for the X-band valid and attenuation region mask M att In one specific embodiment, a preset low-value threshold of 15 dBZ and a preset strong echo threshold of 40 dBZ are set. When X-band data is missing at a certain grid location, or when the X-band reflectivity factor at that location is lower than 15 dBZ and the corresponding S-band reflectivity factor is higher than 40 dBZ, the grid is marked as a suspected attenuation region and used to construct an attenuation region mask M. attThe above thresholds can also be adjusted according to different radar models, observation areas, historical precipitation sample statistical quantiles, or operational quality control standards, and the non-attenuation region mask can be obtained by the following expression: M non =M valid ⊙(1 M att ), M non During the training phase, it can be used to ensure that the output results in the low-attenuation reliable region remain consistent with the original X-band observations.

[0073] 5. Total Loss Function Total loss function L total Represented as: L total =λ1·L content +λ2·L edge +λ3·L region , Where λ1, λ2, and λ3 are the weighting coefficients of the corresponding loss terms.

[0074] 6. Acquisition of monitoring tags The reference supervision X corresponding to the training samples preferably consists of at least one of the following data: effective reflectance factor grid data in the non-attenuation region of the X-band; high-quality X-band reflectance factor grid data; an X-band reflectance factor reference field formed after manual quality control; or an X-band reflectance factor reference field generated based on S-band mapping. The X-band reflectance factor reference field formed after manual quality control is a two-dimensional reflectance factor numerical field located in the same spatiotemporal grid as the model output X0, used as reference supervision X during the training phase. For attenuation hole regions, missing measurements are not directly used as supervision; instead, indirect constraints are applied by combining region adaptive structure preservation loss.

[0075] (vi) Correcting output and ensuring consistency in splicing like Figure 7 As shown in the example of the correction effect, after the radar data to be corrected is input into the correction model after training, the output correction result has higher spatial resolution and more continuous echo structure. Figure 7 This is used to illustrate the effect of the present invention on the restoration of the main structure, edge continuity and local strong echo details of rain clusters in the scenario of heavy rainfall echo attenuation.

[0076] During the inference phase, the S-band radar observation data and X-band radar observation data to be corrected are input into the trained deep learning correction model, and the model correction result X0 is output. In one implementation, X0 can be directly output as the correction result.

[0077] In another implementation, to further improve the reliability of the results, a non-attenuation region mask M can be combined.non The final output result is obtained by consistent stitching with the original X-band observations: X out =M non ⊙X+(1 M non )⊙X0, Within the non-attenuation reliable region, original X-band observations are preferentially retained; within the attenuation risk region, model correction results are used, X... out This represents the final output result after consistent stitching of the non-attenuated regions. This consistent stitching mechanism avoids unnecessary modifications to the originally reliable observation regions and improves the reliability of the final output result in business applications.

[0078] The specific embodiments described above address the practical technical problem of X-band radar under heavy rainfall conditions, where propagation path attenuation often leads to significantly lower or even missing downstream echoes from the rain core. Traditional single-band interpolation or unified deep learning repair methods are prone to structural discontinuities, false edges, or excessive modification of reliable observation areas. This invention, through the joint constraints of stable macroscopic structure in the S-band, high-resolution details in the X-band, and A* partitioning guidance, enables the correction results to have a more complete rainfall echo structure in high-attenuation-risk areas, smoother boundary continuity in transition areas, and maintain original observation details in low-attenuation reliable areas. Therefore, compared to existing unified fusion strategies, it demonstrates better physical consistency and operational reliability.

[0079] (vii) Description of the technical effects of this embodiment Compared with existing schemes that employ a single difference guide graph, a unified fusion strategy, or a globally unified loss function, this implementation method has at least the following advantages: 1. By constructing a composite physical guidance map A* that combines S-band vertical structure characteristics, X-band cavity morphology characteristics, and cross-band difference information, potential attenuation risks can be characterized more accurately. 2. By using A*-based partitioned adaptive feature fusion, the structural completion capability of attenuation risk regions and the detail preservation capability of low attenuation regions are improved. 3. By using a region-adaptive structure-preserving loss and consistency splicing mechanism, the regional continuity, structural rationality, and output reliability of the results are improved.

[0080] This invention provides a method for correcting multi-band radar heavy rainfall data based on deep learning and physical guidance feature fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for correcting multi-band radar heavy rainfall data based on deep learning and physical guidance feature fusion, characterized in that, Includes the following steps: Step 1: Acquire S-band radar observation data and X-band radar observation data for the same precipitation process used in training. The S-band radar observation data includes reflectivity factors for two elevation or height layers. Perform time synchronization and spatial registration on the S-band radar observation data and X-band radar observation data used in training, and resample them to the same grid to obtain S-band grid data and X-band grid data used in training. Step 2: Construct a training composite physical guidance map A* based on the training S-band grid data and the training X-band grid data. The training composite physical guidance map A* is used to characterize the potential attenuation risk level of each spatial location in the training samples. Step 3: Construct a deep learning correction model, which includes a semantic structure feature extraction module, a detail feature extraction module, a physically guided feature fusion module, and a high-resolution reconstruction output module. The semantic structure feature extraction module is used to extract multi-scale semantic structure features F from S-band grid data. s ; The detail feature extraction module is used to extract multi-scale local detail features F from X-band grid data. x ; The physical guidance feature fusion module is used to partition the spatial location into attenuation risk zones based on the composite physical guidance map A*, and to apply differentiated cross-band fusion strategies to different zones to obtain the fused feature F. fuse ; The high-resolution reconstruction output module is used to determine the fusion feature F. fuse Output the corrected X-band reflectivity factor map X0; Step 4: Use the training samples consisting of training S-band grid data, training X-band grid data, and training composite physical guidance map A* to train the deep learning correction model. The training loss function includes content loss, structural constraint loss, and region adaptive structure preservation loss. The region adaptive structure preservation loss is used to apply differentiated supervision constraints to the high attenuation risk region, the transition fusion region, and the low attenuation reliable region based on the different attenuation risk regions represented by the composite physical guidance map A*. Step 5: Obtain the S-band radar observation data and X-band radar observation data to be corrected, perform preprocessing and construct the composite physical guidance map A* to be corrected. Input the S-band grid data to be corrected, the X-band grid data to be corrected and the composite physical guidance map A* to be corrected into the trained deep learning correction model to obtain the corrected X-band reflectivity factor map X0.

2. The method according to claim 1, characterized in that, In step 1, the time synchronization involves matching S-band radar observation data and X-band radar observation data according to the same observation time or a preset time window; the spatial registration involves converting data from different radar coordinate systems to a unified geographic coordinate system and cropping it to a common coverage area; and the resampling involves unifying S-band radar observation data and X-band radar observation data to grid data with the same spatial resolution.

3. The method according to claim 2, characterized in that, Step 2 includes: performing differential or gradient calculations on the reflectivity factors of adjacent elevation angles or height layers in the S-band to obtain the first attenuation characterization quantity; The second attenuation characterization quantity is obtained based on the spatial connectivity, morphological distribution, or inconsistency with the S-band high echo region in the low-value region or missing region of X-band. The first and second attenuation characteristics are weighted and combined, and then segmented and mapped to obtain the composite physical guidance diagram A*. The X-band low-value region is a connected region where the X-band reflectivity factor is lower than a preset low-value threshold; the missing region is a region in the X-band radar data where there is no effective reflectivity factor value; the S-band strong echo region is a region where the S-band reflectivity factor is higher than a preset strong echo threshold.

4. The method according to claim 3, characterized in that, In step 2, the first attenuation characterization quantity includes at least one of the following: the difference in reflectivity factor between adjacent elevation angles in the S-band; the difference in reflectivity factor between adjacent height layers in the S-band; and the local vertical gradient in the S-band. The rate of change of reflectivity factor of adjacent grids or adjacent height layers in the S-band strong echo region; The second attenuation characterization quantity includes at least one of the following: the area of ​​the connected domain, boundary morphology, aspect ratio, or hole distribution characteristics of the low-value region or missing region in the X-band; Spatial overlap difference between low-value or missing X-band regions and strong echo S-band regions; reflectivity factor difference between S-band and X-band at corresponding grid locations.

5. The method according to claim 4, characterized in that, In step 2, the expression for the weighted combination of the first attenuation characteristic and the second attenuation characteristic is: G*=αG1+βG2+γG3+δG4, Wherein, G* is the composite attenuation characterization quantity, G1 is the difference in reflectivity factor between adjacent elevation angles or adjacent height layers in the S-band, G2 is the change rate of local vertical gradient or strong echo in the S-band, G3 is the spatial connectivity structure term of low-value area or missing area in the X-band, G4 is the difference in reflectivity factor between the S-band and X-band at corresponding grid positions, and α, β, γ, and δ are weighting coefficients. The composite attenuation characterization quantity G* is obtained by performing piecewise normalization mapping or partition threshold mapping to obtain the composite physical guidance graph A*.

6. The method according to claim 5, characterized in that, In step 3, the attenuation risk partitioning includes dividing the spatial location into a high attenuation risk zone, a transition fusion zone, and a low attenuation reliable zone according to the composite physical guidance map A*; wherein, in the high attenuation risk zone, the contribution of S-band semantic structure features is enhanced, in the low attenuation reliable zone, the preservation of X-band local detail features is enhanced, and in the transition fusion zone, the collaborative fusion of S-band semantic structure features and X-band local detail features is performed.

7. The method according to claim 6, characterized in that, In step 3, the physical guidance feature fusion module includes a scale matching unit, a weight generation unit, a weighted fusion unit, and a cross-modal attention enhancement unit; The scale matching unit is used to map the composite physics guidance graph A* to a scale consistent with the current scale features; the weight generation unit is used to generate gating weights W based on the mapped composite physics guidance graph. l ; The gating weight W l Composite Physical Guidance Map A after Scale Matching l The formula obtained through normalization mapping is as follows: W l =σ(Conv(A l )) , Where A l For the composite physical guidance map at the l-th scale, Conv represents a convolutional mapping with a kernel size of 1×1, and σ represents the Sigmoid function; the weighted fusion unit is used to fuse the features at the l-th scale in the following manner: F mix l =W l ⊙F s l +(1 W l )⊙F x l , Among them, F s l For the semantic structure features of the S-band at the l-th scale, F x l For the local detail features of the X-band at the l-th scale, F mix l This represents the initial fusion feature at the l-th scale; ⊙ indicates element-wise multiplication; The cross-modal attention enhancement unit is used to enhance the preliminary fused features across modalities to obtain the fused feature F. fuse l .

8. The method according to claim 7, characterized in that, In step 4, the content loss is used to constrain the numerical difference between the correction result and the reference supervision; the structural constraint loss is used to constrain the edge, gradient, or structural similarity between the correction result and the reference supervision; and the region adaptive structure preservation loss is used to apply different supervision rules to different decay risk regions according to the composite physics guidance graph A*. The regional adaptive structure preservation loss includes an S-band macroscopic structure consistency term, an X-band original observation consistency term, and a spatial continuity term. The weight of the S-band macroscopic structure consistency term is increased in high-attenuation-risk areas, the weight of the X-band original observation consistency term is increased in low-attenuation-reliable areas, and the weight of the spatial continuity term is increased in transition fusion areas.

9. The method according to claim 8, characterized in that, Step 4 also includes: constructing an effective X-band observation mask M valid and attenuation region mask M att , of which M valid Used to mark grids with effective reflectivity factor values ​​in X-band radar data; M att The grid is used to mark grids that meet preset attenuation criteria. These preset attenuation criteria include: missing X-band data; or, an X-band reflectivity factor lower than a preset low threshold and a corresponding S-band reflectivity factor higher than a preset strong echo threshold. The non-attenuation region mask M is obtained by the following expression. non : M non =M valid ⊙(1 M att ), The non-attenuation region mask M non Used to constrain the consistency of correction results with the original X-band observations within the low-attenuation reliability zone during the training phase.

10. The method according to claim 9, characterized in that, In step 4, the reference supervision X corresponding to the training sample consists of at least one of the following data: effective reflectivity factor grid data of the non-attenuation region of the X-band; high-quality X-band reflectivity factor grid data; X-band reflectivity factor reference field formed after manual quality control; X-band reflectivity factor reference field generated based on S-band mapping; The X-band reflectivity factor reference field formed after manual quality control is a two-dimensional reflectivity factor numerical field located in the same spatiotemporal grid as the corrected X-band reflectivity factor map X0 output by the deep learning correction model, which is used as a reference supervision X during the training phase. In this regard, missing measurements are not directly used as supervision for attenuation hole regions; instead, indirect constraints are applied by combining the region adaptive structure preservation loss. The attenuation hole region refers to the region where X-band data is missing, or where the X-band reflectivity factor is lower than a preset low threshold and the corresponding S-band reflectivity factor is higher than a preset strong echo threshold.

Images