A weather radar missing frame reconstruction method fusing space-time context information

Abstract

This invention discloses a method for reconstructing missing frames from meteorological radar by fusing spatiotemporal context information. The method includes: preprocessing raw meteorological radar observation data to obtain standardized radar echo sequence data; employing a random masking strategy to simulate missing frames in the standardized radar echo sequence data, constructing a training sample set; introducing an efficient upsampling module and a multi-scale progressive fusion mechanism to construct a meteorological radar missing frame reconstruction model that integrates spatiotemporal context information and training the model; preserving high-level semantics and low-level details through residual fusion; and aligning cross-level features spatially through upsampling with embedded coordinate information; inputting the actual missing radar echo sequence into the trained missing frame reconstruction model, and outputting a completed radar echo sequence. This method significantly improves the integrity and usability of radar data, providing reliable data support for downstream operations such as accurate monitoring and timely early warning of meteorological disasters.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of meteorological data quality control, remote sensing information processing, and next-generation electronic information, and particularly to a method for reconstructing missing frames of meteorological radar by integrating spatiotemporal context information. Background Technology

[0002] Compared to ground-based weather stations and meteorological satellites, radar offers higher spatial resolution and shorter scanning intervals, enabling faster acquisition of weather conditions within the observation area. Derivative products of weather radar, such as combined reflectivity and vertical liquid water data, play a crucial role in precipitation nowcasting. These products effectively reflect the intensity and evolution of severe convective weather systems, providing a data foundation for precipitation nowcasting. However, due to factors such as weather radar hardware failures and network transmission problems, incomplete radar echo data is often not collected or transmitted, resulting in data gaps, limiting data quality, and hindering the provision of more accurate and reliable data support for downstream research and applications (such as precipitation nowcasting).

[0003] Traditional interpolation methods, including linear interpolation, spline interpolation, and Kriging interpolation, have been applied in certain specific scenarios, but they still have some shortcomings: Traditional methods often assume that the changes between data are smooth and linear, but the evolution of clouds in real weather systems has significant nonlinear characteristics, and their morphology and structure often change drastically in a short period of time, such as splitting, merging, strengthening, and dissipating. Cloud movement often exhibits non-uniform and non-linear dynamic behaviors such as acceleration and turning. Traditional interpolation methods, based on the assumption of linear or smooth transitions, struggle to capture such complex spatiotemporal evolution processes and cannot effectively reflect the nonlinear evolution of cloud morphology and the acceleration of motion between adjacent frames, resulting in significant deviations in the spatiotemporal continuity and physical plausibility of the interpolation results.

[0004] Therefore, improving the accuracy of reconstructing missing frames from weather radar data has become a critical issue that urgently needs to be addressed. Summary of the Invention

[0005] To address the problems of traditional interpolation methods being unable to handle nonlinear cloud evolution and low reconstruction accuracy, this invention proposes a meteorological radar missing frame reconstruction method that integrates spatiotemporal context information. By simulating data missing situations in real-world scenarios, an efficient deep learning reconstruction method is designed to achieve high-precision completion of missing meteorological radar data.

[0006] This invention adopts the following technical solution: a method for reconstructing missing frames of a weather radar by fusing spatiotemporal context information, comprising the following steps:

[0007] Step 1: Preprocessing of raw weather radar observation data: Obtain raw weather radar observation data and preprocess it. After data decoding, spatial alignment, spatiotemporal slicing, data cleaning, spatial downsampling, normalization and sequence grouping, normalized radar echo sequence data is obtained.

[0008] Step 2, Construction of the training sample set: Using a self-supervised reconstruction paradigm, a missing data simulation is performed on the normalized radar echo sequence data through a random masking strategy to construct the training sample set;

[0009] Step 3: Construction and Training of Missing Frame Reconstruction Model: Based on the encoder-decoder structure, and considering the characteristics of strong continuity and complex spatial structure in the temporal dimension of radar echo sequence data, a meteorological radar missing frame reconstruction model integrating spatiotemporal context information is constructed. An efficient upsampling module and a multi-scale progressive fusion mechanism are introduced. High-level semantics and low-level details are preserved through residual fusion, and cross-level features are aligned in spatial position through upsampling with embedded coordinate information. A hybrid reconstruction loss function is then constructed for model training.

[0010] Step 4: Complete the missing radar echo sequence: Input the actual missing radar echo sequence into the trained missing frame reconstruction model, output the completed radar echo sequence, perform inverse normalization on the output complete radar echo sequence to restore the original intensity scale of the radar echo data, and obtain complete radar data for downstream meteorological operations.

[0011] As a preferred embodiment, the preprocessing of the raw weather radar observation data in step 1 includes the following sub-steps:

[0012] 1.1 Data Decoding: Decode the acquired raw binary radar data and convert it into a multidimensional array format that can be processed by a computer.

[0013] 1.2 Spatial Alignment: Based on the latitude and longitude coordinates of the data, map them to a unified geographic coordinate grid to complete spatial alignment and generate standardized grid data with geographic reference consistency;

[0014] 1.3 Spatiotemporal slicing: Based on the spatiotemporal range of actual business needs, the decoded radar data is sliced ​​to extract radar echo data within the target area and target time period, and irrelevant areas and redundant time data are removed.

[0015] 1.4 Data Cleaning: For the sliced ​​radar echo data, outlier handling is performed. Invalid values ​​in the radar echo data are set to zero, data exceeding 70dBZ are reset to 70dBZ, and data less than 0dBZ are reset to 0dBZ to ensure the physical rationality of the data.

[0016] 1.5 Spatial Downsampling: A local region averaging downsampling strategy is used to spatially downsample the radar echo data. This strategy calculates the average value of adjacent pixel regions as the downsampled data;

[0017] 1.6 Normalization Processing: The max-min normalization method is used to map the gridded radar echo data to [0, 1], corresponding to the range of 0 dBZ to 70 dBZ. The normalization formula is:

[0018] ;

[0019] in, It is the intensity value of the radar echo data. and These are the minimum and maximum values ​​in the radar echo data, which are 0 and 70, respectively. Normalized radar echo data values;

[0020] 1.7 Sequence Grouping: A sliding window strategy is used to group the radar echo sequences, with the window length set to [value missing]. This ensures that the data within the window is continuous in time; each window serves as a complete label sample, with a sliding step of 6 minutes, consistent with the regular radar scan interval, meaning that each window moves backward by one frame to generate a new sample. The final radar echo sequence data has the following dimension ( (C, H, W);

[0021] in, The time step is represented by C, the number of channels is represented by H, the height is represented by W, and the width is represented by W.

[0022] As a preferred approach, the construction of the training sample set in step 2 adopts a self-supervised reconstruction paradigm, firstly by extracting a continuous segment without missing data. The radar sequence is used as a complete label; then a manual full-frame mask is applied to the sequence to generate model input with missing data; the model learns to reconstruct the complete sequence from partial observations, thereby gaining the ability to repair the real missing data.

[0023] Specifically, it includes the following sub-steps:

[0024] 2.1 Construction of the mask sequence: Let It is a length of The original radar echo sequence ( = ),in, Represents the first in the sequence Radar echo characteristic map at each observation time;

[0025] Construction and Binary mask sequences of the same length This is used to indicate whether each position is masked:

[0026] ;

[0027] in, Represents the first in the sequence One observation value, = ; Represents the first in the sequence A binary mask.

[0028] 2.2 Simulation of missing data processing: Element-wise multiplication is used to process the radar echo sequence. After masking, the radar echo sequence with simulated missing 0s is obtained. The calculation formula is:

[0029] ;

[0030] in, This is an element-wise multiplication operation.

[0031] By adjusting the mask sequence middle The ratio can simulate data scenarios with different degrees of missing data, making the constructed training sample set more consistent with the randomness and diversity of missing data in actual applications, and improving the generalization ability of the model.

[0032] The missing images may be at the beginning, end, or middle, and may be a single image or multiple images in succession. The missing rate is set to 10%-70%.

[0033] 2.3. Simulate the missing radar echo sequence As input to the model, the corresponding original radar echo sequence serves as a supervision label, forming training sample pairs.

[0034] As a preferred embodiment, the missing frame reconstruction model construction and training in step 3 includes the following sub-steps:

[0035] 3.1 Model Architecture Design: Based on the encoder-decoder structure, this model constructs a missing frame reconstruction model to address the characteristics of meteorological radar echo sequences, such as strong continuity in the time dimension and complex spatial structure.

[0036] 3.1.1 The encoder employs multiple downsampling stages, each stage including: 3D convolution, batch normalization, ReLU activation function, and max pooling operation, to progressively extract multi-scale spatiotemporal features;

[0037] The decoder restores spatial resolution through upsampling and introduces an efficient upsampling module and a multi-scale progressive fusion mechanism. The former preserves high-level semantics and low-level details through residual fusion, while the latter ensures strict spatial alignment of cross-level features through upsampling with embedded coordinate information. The two work together to achieve spatial alignment and semantic enhancement of cross-level features.

[0038] 3.1.2 Multi-scale progressive fusion mechanism: A multi-scale progressive fusion mechanism is embedded in the decoder path to efficiently integrate multi-scale spatiotemporal features from different levels of encoders and decoders, achieving hierarchical and progressive feature fusion. The specific processing flow is as follows:

[0039] Feature concatenation: combining the input feature maps of the current decoding layer. The feature map output by the high-efficiency upsampling module of the previous layer decoder By performing splicing processing along the channel dimension, a fused feature containing multi-scale information is obtained;

[0040] Nonlinear transformation: A nonlinear transformation module consisting of three serial residual blocks is applied to the spliced ​​fused features. Each residual block consists of two parts: a 3D convolution, an activation function, and a batch normalization. Residual connections are used to avoid the gradient vanishing problem in deep network training and enhance the model's ability to extract complex features.

[0041] Feature output: The nonlinearly transformed features are compared with the original input feature map of the current decoding layer. Element-wise addition is performed in a specific ratio to obtain the output feature map of the current decoding layer. The mathematical expression is:

[0042] ;

[0043] in, This indicates a splicing operation at the channel dimension.

[0044] This mechanism progressively fuses multi-scale features, enabling the model to simultaneously capture local details and global trends in radar echo data, thereby enhancing its ability to model nonlinear evolution features.

[0045] 3.1.3 High-efficiency upsampling module: In order to solve the problem of inconsistent size of different feature maps during the skip connection process, this invention designs a high-efficiency channel upsampling module to unify the spatial resolution of feature maps at different levels.

[0046] The specific structure of this module is as follows:

[0047] Trilinear interpolation upsampling: Input feature map Spatial upsampling is performed using a trilinear interpolation layer to increase the feature map resolution to match that of the corresponding decoding layer features. Figure 1 Therefore, this operation can initially recover the spatial details of the feature map without introducing additional parameters;

[0048] CoordConv layer processing: The feature map after trilinear interpolation is input into the CoordConv layer. During the convolution process, the spatial coordinate information of the feature map is embedded. The CoordConv layer first generates a normalized x-coordinate matrix and y-coordinate matrix with the same spatial size as the input feature map, which represent the relative position of each pixel in the width and height directions, respectively.

[0049] Specifically, if the size of the input feature map is H×W, then it is defined as:

[0050] ;

[0051] ;

[0052] in, These represent the row index and column index of the pixel, respectively. and It is a normalized coordinate tensor with a value range of [0,1].

[0053] Subsequently, these two coordinate matrices are concatenated with the original feature map along the channel dimension as additional channels to form enhanced features:

[0054] ;

[0055] in, The feature map after trilinear interpolation. This is the enhanced feature map.

[0056] Finally, a convolution operation is applied to this enhanced feature:

[0057] ;

[0058] in, This indicates a convolution operation.

[0059] This allows the network to explicitly perceive absolute spatial location during feature extraction, enabling the model to more accurately perceive the spatial positional relationships of radar echo data and reduce the impact of spatial deviations caused by terrain and site location on the reconstruction results.

[0060] Depthwise separable convolution processing: The feature map output from the CoordConv layer is input into the depthwise separable convolution layer, which consists of Pointwise Convolution (PWConv) and Depthwise Convolution (DWConv).

[0061] PWConv adjusts the number of channels in the feature map using a 1×1×1 convolution kernel to achieve cross-channel information interaction; DWConv performs spatial convolution independently in each channel using a 3×3×3 convolution kernel, which significantly reduces the number of model parameters and computational load while ensuring feature extraction capabilities.

[0062] Output of high-efficiency upsampling module The calculation formula is:

[0063] ;

[0064] in, This represents trilinear interpolation. This represents a coordinate convolutional layer, used to embed spatial coordinate information into feature maps to enhance the model's ability to perceive spatial location. This represents pointwise convolution. This represents depthwise convolution.

[0065] 3.2 Model Training:

[0066] 3.2.1 Loss Function Selection: A hybrid reconstruction loss function is adopted. For loss function, The function combines the advantages of mean squared error (MSE) and structural similarity index (SSIM), ensuring both the numerical accuracy of the reconstructed data and the structural similarity of the radar echo images. The specific expression is as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] in, Indicates the number of samples. Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; and These represent the average brightness of the predicted image and the actual image, respectively. This represents the covariance between the predicted image and the true image. and Let these represent the variances of the two variables, respectively. and It is the smallest constant, used to avoid the case where the denominator is zero; As a trade-off coefficient, a value of 0.5 is preferred to achieve a balance between mean squared error and structural similarity; and These represent the mean squared error loss and the structural similarity index loss, respectively.

[0071] 3.2.2 During training: The training sample set constructed in step 2 is divided into a training set and a test set in an 8:2 ratio; the Adam optimizer is selected, and the learning rate is set to 1e-4 to minimize the mixed loss function. To achieve this, the reconstruction model is iteratively trained. During the training process, an early stopping strategy is adopted. When the test set loss does not decrease for 20 consecutive epochs, the model training is stopped, the current optimal model parameters are saved, and the trained missing frame reconstruction model is obtained.

[0072] As a preferred embodiment, step 4, which involves completing the missing data, includes the following sub-steps:

[0073] The missing radar echo sequences in practical applications are processed through data decoding, spatiotemporal slicing, spatial downsampling, spatial alignment, and normalization. These processes are then input into the trained missing frame reconstruction model. The model predicts the spatiotemporal characteristics of the missing locations and outputs the completed radar echo sequence. The output complete sequence is then denormalized to restore the original intensity scale of the radar echo data, resulting in complete radar data that can be used by downstream meteorological operations.

[0074] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0075] 1. This invention proposes a method for reconstructing missing frames from meteorological radar by fusing spatiotemporal context information. It combines a multi-scale progressive fusion mechanism with an efficient upsampling module, significantly improving the model's ability to capture the spatiotemporal evolution characteristics of radar echo data. Compared to traditional interpolation methods and existing deep learning methods, this invention does not rely on assumptions of data smoothing or linear changes, and can better simulate complex dynamic processes such as cloud splitting and merging.

[0076] 2. The multi-scale progressive fusion mechanism proposed in this invention effectively captures the nonlinear morphological evolution and cross-scale structural correlation of radar echo data by utilizing spatiotemporal features at different levels, overcoming the limitations of traditional methods that can only handle linear changes and cannot capture complex spatiotemporal correlations. The efficient upsampling module integrates trilinear interpolation, CoordConv, and depthwise separable convolution. This design can accurately recover spatial location information, ensure the consistency of feature scales, and significantly reduce the number of model parameters and computational burden, avoiding the risk of overfitting due to complex architecture, enabling the model to achieve efficient inference even with limited hardware resources. In addition, the use of a hybrid reconstruction loss function optimizes the numerical accuracy and image structure consistency of the reconstruction results, effectively suppressing the numerical bias and morphological distortion problems common in traditional methods, ensuring that the completion results not only meet physical strength constraints but also maintain the true spatial distribution and spatiotemporal continuity.

[0077] 3. The method of this invention can efficiently repair radar data missing problems caused by various reasons such as hardware failure and network interruption, generate high-quality meteorological radar data, improve the integrity of meteorological radar data, and provide reliable data support for downstream operations such as accurate monitoring and timely early warning of meteorological disasters. It is especially suitable for radar data repair in scenarios with high missing rates and complex missing data, and has good practical application value and promotion prospects. Attached Figure Description

[0078] Figure 1 This is an overall flowchart of the meteorological radar missing frame reconstruction method that integrates spatiotemporal context information according to the present invention;

[0079] Figure 2 This is a schematic diagram illustrating the results of simulating missing data in real-world scenarios using a random masking strategy, as presented in this invention.

[0080] Figure 3 This is a schematic diagram of the missing frame reconstruction model architecture of the present invention;

[0081] Figure 4 A comparison chart showing the completion effects of different models and methods on the dataset. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0083] In one embodiment of the present invention, a radar echo sequence prediction experiment is conducted on a radar combined reflectivity dataset. The study area is selected as Jiangsu Province, China, specifically ranging from 116.00°E to 121.12°E and from 30.00°N to 35.12°N. The radar combined reflectivity dataset for the study area consists of radar combined reflectivity CREF data from 2023 to 2024, with a temporal resolution of 6 minutes and a spatial resolution of 0.01°.

[0084] The method of this invention is used to perform radar echo sequence prediction tasks on radar product data combined reflectivity datasets, such as... Figure 1 As shown, the specific implementation steps are as follows:

[0085] Step S1: Preprocess the raw meteorological radar observation data to obtain standardized radar echo sequence data.

[0086] In this embodiment, radar combined transmittance (CREF) data of Jiangsu Province, China, with a latitude and longitude range of 116.00° to 121.12° east longitude and 30.00° to 35.12° north latitude, is obtained, and the obtained radar echo data is preprocessed to obtain a radar echo dataset.

[0087] Among them, window length 15 corresponds to a time of 90 minutes, a time resolution of 6 minutes, a spatial resolution of 0.04°, and a shape of 15×1×128×128, where 15 represents 15 consecutive radar echo data, 1 indicates the number of channels, and 128 represents the width and height.

[0088] Step S2: Use a random masking strategy to simulate missing data in the normalized radar echo sequence data and construct a training sample set.

[0089] In this embodiment, the radar echo sequence dataset used in the experiment is constructed based on the radar echo dataset obtained in step 1. For each sample in the training set, 1-10 time steps are randomly selected, where the positions are random and there are cases where it is the first frame, the last frame, or consecutive frames. These are set to all zeros and used as model input. The original 15 frames are used as labels. 80% of the radar echo sequence dataset is used as the training set, and the remaining 20% ​​is used as the test set.

[0090] Specific details of missing data are as follows: Figure 2 As shown, the first row simulates the missing sequence, and the second row represents the complete sequence.

[0091] Step S3: Construct a meteorological radar missing frame reconstruction model that integrates spatiotemporal context information.

[0092] In this embodiment, a missing frame reconstruction model based on an encoder-decoder architecture is constructed, such as... Figure 3 As shown, the encoder is responsible for extracting multi-scale spatiotemporal features; the decoder receives the low-level features from the encoder through skip connections and achieves feature scale alignment by combining an efficient upsampling module; a multi-scale progressive fusion mechanism is embedded in the decoder path to efficiently integrate multi-scale spatiotemporal features. A hybrid reconstruction loss function is adopted. Simultaneously optimize numerical accuracy and structural consistency.

[0093] Step S4: Input the actual missing radar echo sequence into the trained missing frame reconstruction model, and output the completed radar echo sequence.

[0094] In this embodiment, the model is trained based on the training set obtained in step 2. The Adam optimizer is used with an initial learning rate of 1e-4 and a batch size of 2. The cosine annealing learning rate scheduling strategy is adopted. The maximum number of training epochs is set to 200, and the early stopping is set to 20 epochs. That is, if no improvement in the test set loss is observed in 20 consecutive epochs, the training will be terminated.

[0095] Subsequently, the acquired test set radar echo sequences are input into the trained missing frame reconstruction model to obtain the reconstructed radar echo sequences, and the prediction results are evaluated using evaluation metrics.

[0096] Furthermore, baselines were selected from traditional methods such as linear interpolation, spline interpolation, and Kriging interpolation, as well as deep learning methods including U-Net, U-Net v2, ConvLSTM, 3D U-Net, and UCTransNet. The reconstruction evaluation metrics of different methods on the test set are shown in Table 1.

[0097] Table 1

[0098]

[0099] As shown in Table 1, the core evaluation metrics of the method of this invention (Ours) are superior to traditional interpolation methods (linear interpolation, Kriging interpolation) and existing deep learning methods under different missing rate scenarios. The table includes a 10-frame radar echo sequence containing actual missing data, 10-frame sequences reconstructed by each method, and the actual complete sequence. Figure 4 As shown.

[0100] Figure 4 The results demonstrate the cases where data is missing for seven consecutive frames from frame 5 to frame 11, as well as frames 3, 13, and 15. The missing frame reconstruction model outperforms traditional interpolation methods and other deep learning comparison methods, particularly in reconstructing the details of the missing data.

[0101] The test results of the above embodiments fully verify the effectiveness and superiority of the present invention in the task of reconstructing missing frames of meteorological radar data.

[0102] In summary, the method of this invention is applicable to repairing the loss of complete time-step radar echo data caused by radar equipment failure, network transmission interruption, or abnormal data acquisition. By fusing the spatiotemporal context information of effective observations before and after, the missing frames are reconstructed with high precision, providing continuous, complete, and high-quality radar data support for meteorological operations such as short-term weather forecasting, severe convection monitoring, and disaster early warning.

[0103] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for reconstructing missing frames from a weather radar by fusing spatiotemporal context information, characterized in that, Includes the following steps: Step 1: Preprocessing of raw weather radar observation data: Obtain raw weather radar observation data and preprocess it. After data decoding, spatial alignment, spatiotemporal slicing, data cleaning, spatial downsampling, normalization and sequence grouping, normalized radar echo sequence data is obtained. Step 2, Construction of the training sample set: Using a self-supervised reconstruction paradigm, a missing data simulation is performed on the normalized radar echo sequence data through a random masking strategy to construct the training sample set; Step 3, Construction and Training of Missing Frame Reconstruction Model: Based on the encoder-decoder structure, and considering the characteristics of strong continuity and complex spatial structure in the time dimension of radar echo sequence data, a meteorological radar missing frame reconstruction model integrating spatiotemporal context information is constructed. An efficient upsampling module and a multi-scale progressive fusion mechanism are introduced. High-level semantics and low-level details are preserved through residual fusion, and cross-level features are aligned in spatial position through upsampling with embedded coordinate information. A hybrid reconstruction loss function is constructed for model training. The multi-scale progressive fusion mechanism has the following processing flow: The input feature map of the current decoding layer Compared with the feature map output by the previous layer's efficient upsampling module By splicing along the channel dimension, the fused features are obtained; A nonlinear transformation module consisting of three serial residual blocks is applied to the fusion feature. Feature extraction is performed, with each residual block consisting of two 3D convolutional layers, a normalization layer, and an activation function to extract high-level semantic information; Compare the features after nonlinear transformation with the original input feature map. Add them together to obtain the output feature map of the current layer. The mathematical expression is: ； in, This indicates a splicing operation along the channel dimension; The high-efficiency upsampling module comprises a trilinear interpolation layer, a CoordConv layer, and a depthwise separable convolutional layer connected in sequence. The specific processing procedure is as follows: Input feature map Spatial upsampling is performed using trilinear interpolation to increase the feature map resolution to match that of the corresponding decoding layer feature map, thus obtaining a preliminary high-resolution feature map. The initial high-resolution feature map is passed through the CoordConv layer, and spatial coordinate information is embedded during the convolution process to enhance the model's ability to perceive spatial location. The output features of the CoordConv layer are processed through a depthwise separable convolutional layer, which includes a PWConv and a DWConv. The DWConv is used to perform spatial convolution independently in each channel, and the PWConv is used to adjust the number of channels. The output of the high-efficiency upsampling module satisfy: ； in, This represents trilinear interpolation. This represents a coordinate convolutional layer, used to embed spatial coordinate information into feature maps to enhance the model's ability to perceive spatial location. This represents pointwise convolution. Represents depthwise convolution; Step 4: Complete the missing radar echo sequence: Input the actual missing radar echo sequence into the trained missing frame reconstruction model, output the completed radar echo sequence, and restore the original intensity scale of the radar echo data through inverse normalization processing to obtain complete radar data for downstream meteorological operations.

2. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, Step 1, which processes the raw weather radar observation data, includes the following sub-steps: Step 1.1, Data Decoding: Decode the acquired raw binary radar data and convert it into a multidimensional array format; Step 1.2, Spatial Alignment: Based on the latitude and longitude coordinates of the data, the decoded radar data is mapped to a unified geographic coordinate grid to complete spatial alignment; Step 1.3, Spatiotemporal Slicing: Slice the data according to the actual research scope and extract radar echo data within the target area and target time period; Step 1.4, Data Cleaning: Set invalid values ​​in the radar echo data to 0dBZ, reset data exceeding 70dBZ to 70dBZ, and reset data less than 0dBZ to 0dBZ; Step 1.5, Spatial downsampling: Spatial downsampling is performed on the radar echo data, using a local area averaging downsampling strategy; Step 1.6, Normalization Processing: Normalize the gridded radar echo data by using the max-min normalization method to map the radar echo data to [0, 1], corresponding to the range of 0dBZ to 70dBZ. Step 1.7, Sequence Grouping: The radar echo sequence is grouped using a sliding window strategy, with the window length set to [value missing]. Frames ensure that the data within a window is continuous in time, with each window serving as a complete label sample, resulting in radar echo sequence data with dimensions ( (C, H, W); where C represents the number of channels, H represents the height, and W represents the width.

3. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, Step 2 involves simulating missing radar echo sequence data, including the following sub-steps: Step 2.1: Construct the mask sequence, let... It is a length of The original radar echo sequence, constructed and Binary mask sequences of the same length Used to indicate whether each position is masked: ； in, Represents the first in the sequence One observation value, = ; Represents the first in the sequence A binary mask; Step 2.2: Simulate missing data processing by performing element-wise multiplication on the radar echo sequence. After masking, the radar echo sequence with simulated missing data is obtained. : ； in, To perform element-wise multiplication, the mask sequence is adjusted. middle The proportion is used to simulate data scenarios with different degrees of missing data; Step 2.3: Simulate the missing radar echo sequence As input to the model, the corresponding original radar echo sequence serves as a supervision label, forming training sample pairs.

4. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, The reconstruction process of the missing frame reconstruction model in step 3 is as follows: Step 3.1: Input the radar echo sequence containing random missing data into the encoder of the missing frame reconstruction model to generate a preliminary feature representation; Step 3.2: The encoder extracts multi-scale spatiotemporal features step by step by downsampling, and passes the encoded features of each level to the corresponding decoding layer through skip connections; Step 3.3: The decoder restores spatial resolution through upsampling and introduces a multi-scale progressive fusion mechanism in the skip connection path. For the current layer, the features after feature extraction of the next layer skip connection are first processed by the efficient upsampling module, and then the input features are concatenated with the features output by the current layer encoder. After feature extraction, the concatenation is sent to the decoder of the corresponding layer. Step 3.4: Reconstruct a complete and continuous radar echo sequence through the output layer to reconstruct the missing data.

5. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, The specific processing procedure of the CoordConv layer is as follows: The trilinearly interpolated feature map is input into the CoordConv layer. During convolution, the spatial coordinate information of the feature map is embedded to generate normalized x-coordinate and y-coordinate matrices with the same spatial dimensions as the input feature map. These matrices represent the relative positions of each pixel in the width and height directions, respectively, and are defined as follows: ； ； The size of the input feature map is... , These represent the row index and column index of the pixel, respectively. and This is a normalized coordinate tensor with a value range of [0,1]. The two coordinate matrices are concatenated with the original feature map along the channel dimension as additional channels to form enhanced features: ； in, The feature map after trilinear interpolation. The enhanced feature map; Applying convolution operations to the enhanced features allows the network to explicitly perceive absolute spatial location during feature extraction: ； in, This indicates a convolution operation.

6. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, The PWConv adjusts the number of channels in the feature map using a 1×1×1 convolution kernel to achieve cross-channel information interaction; the DWConv performs spatial convolution independently in each channel using a 3×3×3 convolution kernel to reduce the number of model parameters and computational load.

7. The method for reconstructing missing frames from a weather radar according to claim 1, characterized in that, The missing frame reconstruction model employs a hybrid reconstruction loss function. The expression for the hybrid reconstruction loss function is as follows: ； ； ； in, Indicates the number of samples. Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; and Representing images respectively and The average brightness of the predicted image and the real image; and Representing images respectively and The variance; Representing two images and covariance; and It is the smallest constant used to avoid the denominator being zero; As a weighting factor, and These represent the mean squared error loss and the structural similarity index loss, respectively.

8. The method for reconstructing missing frames from a weather radar according to claim 7, characterized in that, When training the missing frame reconstruction model, the hybrid reconstruction loss function is minimized. With the goal of reconstructing missing frames, the model parameters are iteratively optimized using the backpropagation algorithm until the model converges, resulting in a fully trained model for reconstructing missing frames.

Images