Meteorological radar image sequence structured noise detection and space-time repair method and system

Abstract

The application relates to the technical field of radar image denoising and repairing, in particular to a weather radar image sequence structured noise detection and space-time repairing method and system. The method comprises the following steps: performing feature extraction on a raw radar image sequence with a mask to obtain gated coding features; performing optical flow field estimation on the gated coding features; performing cross-frame information transmission and local time sequence completion on the estimation result of the motion field and the gated coding features; performing global space-time completion on the repaired features after the local time sequence completion to obtain final repaired features; performing sub-pixel decoding processing on the final repaired features to obtain a prediction result; synthesizing the prediction result and a raw radar image frame with a mask to obtain a deep learning repairing result; and performing real echo texture reconstruction on the deep learning repairing result to obtain a repaired radar image sequence. The application can fill in a large area of missing data with high fidelity and restore the space-time continuity of cloud movement.

Description

Technical Field

[0001] This application relates to the field of radar image denoising and restoration technology, and in particular to a method and system for structured noise detection and spatiotemporal restoration of meteorological radar image sequences. Background Technology

[0002] Meteorological radar image sequences are crucial observational data for acquiring information on the evolution of precipitation clouds and severe convective weather systems. They continuously record the spatial structure and temporal changes of targets, playing a vital role in weather monitoring, disaster early warning, and meteorological analysis. With the widespread application of artificial intelligence technology in meteorology, radar image sequences have been extensively used for tasks such as target identification, echo extrapolation, and spatiotemporal sequence analysis. However, these applications are highly dependent on the accuracy, completeness, and spatiotemporal continuity of radar observation data. Therefore, effectively improving radar image quality has become a significant research issue in the field of meteorological data processing.

[0003] In actual radar observations, structured noise is commonly present in raw radar images due to factors such as terrain obstruction, reflections from fixed buildings, radar equipment malfunctions, and electromagnetic interference. This type of noise mainly manifests as anomalous echoes of ground features, radial fan-shaped artifacts, gray bands in the central blind zone, and striped interference. It typically possesses strong spatial structural characteristics, with relatively fixed positions over long time series, but its influence range and echo intensity fluctuate over time. Unlike random noise, structured noise often covers a large area and spatially overlaps with real meteorological echoes, severely affecting the accuracy and usability of radar data.

[0004] Existing methods typically employ threshold filtering, morphological processing, manual rules, and deep learning to detect and suppress structured noise. In recent years, various processing methods based on variational optimization models and deep learning networks have been proposed for image noise removal. However, these methods still have limitations when dealing with structured noise in weather radar images. Firstly, both structured noise and strong convective clouds can appear as high-reflectivity areas, making it difficult to effectively distinguish between them using only the spatial features of a single frame, potentially leading to the accidental deletion of real meteorological targets or residual noise. Secondly, most existing methods process images at a single time point, failing to fully utilize the spatiotemporal differences between the relatively stable position of structured noise and the continuous movement and change of real meteorological targets, thus hindering accurate identification of structured noise regions in complex scenes. Furthermore, the aforementioned methods primarily focus on the noise suppression process, lacking effective solutions for the large-scale missing regions and spatiotemporal continuity restoration issues resulting from structured noise removal. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a method and system for structured noise detection and spatiotemporal restoration of meteorological radar image sequences. It aims to solve problems such as the difficulty in accurately identifying structured noise in meteorological radar image sequences, the inability to maintain spatiotemporal continuity in large missing areas, and the tendency for restoration results to exhibit over-smoothing and loss of detail. By detecting, restoring, and post-processing structured noise, the authenticity, integrity, and spatiotemporal consistency of the radar image sequence are restored, providing high-quality input data for subsequent meteorological analysis and spatiotemporal sequence prediction. The essence of this invention lies in proposing a three-stage collaborative technique for the restoration and reconstruction of structured noise in meteorological radar images: classical temporal statistical prior, deep learning motion-gated restoration, and classical morphological high-fidelity reconstruction.

[0006] On the one hand, a method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences is provided, including: The weather radar image sequence to be processed is obtained, and structured noise detection is performed on the weather radar image sequence to be processed to obtain the original radar image sequence with masking. The masked original radar image sequence is input into a trained image sequence restoration model to obtain the deep learning restoration result. Specifically, the trained image sequence restoration model extracts features from the masked original radar image sequence to obtain gated coded features; optical flow field estimation is performed on the gated coded features to obtain motion field estimation results; cross-frame information transfer and local temporal completion are performed on the motion field estimation results and gated coded features to obtain locally temporally completed restoration features; global spatiotemporal completion is performed on the locally temporally completed restoration features to obtain the final restoration features; sub-pixel decoding processing is performed on the final restoration features to obtain the prediction result; and the prediction result is synthesized with the masked original radar image frames to obtain the deep learning restoration result. The deep learning restoration results are used to reconstruct the real echo texture, the reconstructed image is mapped to the polar coordinate system, radar artifacts are filtered out, and the restored radar image sequence is obtained.

[0007] On the other hand, a structured noise detection and spatiotemporal restoration system for weather radar image sequences is provided, including: The preprocessing module is configured to: acquire the weather radar image sequence to be processed, perform structured noise detection on the weather radar image sequence to be processed, and obtain the original radar image sequence with mask; The deep learning module is configured to: input a masked original radar image sequence into a trained image sequence restoration model to obtain a deep learning restoration result; wherein, the trained image sequence restoration model is used to extract features from the masked original radar image sequence to obtain gated coded features; perform optical flow field estimation on the gated coded features to obtain motion field estimation results; perform cross-frame information transfer and local temporal completion on the motion field estimation results and gated coded features to obtain locally temporally completed restoration features; perform global spatiotemporal completion on the locally temporally completed restoration features to obtain the final restoration features; perform sub-pixel decoding processing on the final restoration features to obtain prediction results; and synthesize the prediction results with the masked original radar image frames to obtain the deep learning restoration result. The reconstruction module is configured to: reconstruct the real echo texture from the deep learning restoration results, map the reconstructed image to the polar coordinate system, filter out radar artifacts, and obtain the restored radar image sequence.

[0008] The above technical solution has the following advantages or beneficial effects: (1) Overcoming the technical bottleneck of the difficulty in distinguishing between “strong cloud echo” and “fixed structured noise”: This invention innovatively introduces statistical features of the time dimension, utilizes the difference between the spatiotemporal dynamics of the real echo and the spatiotemporal statics of the structured noise, captures the distribution of potential structured noise in the dataset in an unsupervised and adaptive manner and generates a high-precision spatial mask, breaking through the discrimination bottleneck of traditional single-frame denoising.

[0009] (2) Restoring spatiotemporal coherence and avoiding negative migration: The present invention adopts an integrated motion-gated radar image sequence repair model. The motion field estimation-completion network directly regresses the bidirectional optical flow that has been filled in the missing area. The bidirectional gated cyclic propagation is combined with a confidence gate to align adjacent frames according to the motion field and adaptively switch between the propagation content and the observation data. Then, the global spatiotemporal context is aggregated with an axial memory Transformer, thereby filling in large-area missing areas with high fidelity and restoring the spatiotemporal continuity of cloud movement.

[0010] (3) High-fidelity texture and lossless protection of real targets: This invention designs a high-fidelity echo texture reconstruction mechanism and a multi-dimensional feature artifact filtering mechanism, which cleverly combines the coherence of the macroscopic structure of deep learning with the authenticity of the original microscopic texture. It also innovatively introduces polar coordinate dimensionality reduction and "morphological shield" mechanism to completely remove fan-shaped artifacts and isolated noise while achieving lossless protection of real meteorological clouds, fundamentally blocking the spatiotemporal propagation of noise and error amplification in the downstream prediction model. Attached Figure Description

[0011] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0012] Figure 1 This is a flowchart of the method in Example 1. Detailed Implementation

[0013] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0014] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of this application. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] To avoid noise affecting subsequent predictions, some methods choose to directly zero out or remove noisy regions. However, when the structured noise coverage is large, this strategy creates large areas of missing information in radar images, causing discontinuities in the real cloud targets within the time series and disrupting the spatiotemporal integrity of the original observation data. Furthermore, existing image inpainting methods often lack effective spatiotemporal constraint mechanisms when filling in large missing regions, making it difficult to guarantee the consistency of cross-frame structure and easily leading to problems such as morphological distortion, motion discontinuities, and false structures.

[0016] Meanwhile, existing deep learning-based image inpainting methods generally suffer from oversmoothing. Although they can restore the overall structure of the missing area, they often struggle to reconstruct fine-grained texture information and local high-frequency features in radar echoes, resulting in missing details and blurred boundaries in the inpainted results. This, in turn, affects the ability of subsequent prediction models to identify small-scale convective systems and severe weather features.

[0017] Therefore, to address the problem of structured noise in meteorological radar image sequences, there is an urgent need for a processing method that can fully utilize spatiotemporal sequence features to achieve accurate detection of structured noise, spatiotemporal consistency restoration of large-scale missing areas, and reconstruction of high-frequency texture details. This would restore the authenticity, integrity, and spatiotemporal continuity of radar data, providing high-quality input data for subsequent meteorological analysis, target identification, data assimilation, and spatiotemporal sequence prediction.

[0018] Currently, the following technical challenges remain in addressing the aforementioned issues: (1) Structured noise is difficult to detect. Structured noise and strong meteorological echoes have similar high reflectivity characteristics in a single frame image. Existing methods are difficult to distinguish between the two accurately, which easily leads to false detection and false negative detection.

[0019] (2) It is difficult to restore the spatiotemporal consistency of large-scale missing areas. The large-area void areas formed after the removal of structured noise lack effective observation information. Existing restoration methods cannot make full use of motion correlation information across time frames, resulting in spatiotemporal discontinuities and structural distortion in the restoration results.

[0020] (3) Difficulty in preserving details during the restoration process. Existing deep learning restoration models are prone to over-smoothing, making it difficult to recover high-frequency textures and local details in real radar echoes, thus affecting the performance of subsequent meteorological target recognition and spatiotemporal sequence prediction.

[0021] Example 1 This embodiment provides a method for structured noise detection and spatiotemporal restoration of weather radar image sequences; like Figure 1 As shown, the method for structured noise detection and spatiotemporal restoration of weather radar image sequences includes: S101: Obtain the weather radar image sequence to be processed, perform structured noise detection on the weather radar image sequence to be processed, and obtain the original radar image sequence with mask; S102: The masked original radar image sequence is input into the trained image sequence restoration model to obtain the deep learning restoration result. The trained image sequence restoration model is used to extract features from the masked original radar image sequence to obtain gated coded features; optical flow field estimation is performed on the gated coded features to obtain the motion field estimation result; cross-frame information transfer and local temporal completion are performed on the motion field estimation result and the gated coded features to obtain the locally temporally completed restoration feature; global spatiotemporal completion is performed on the locally temporally completed restoration feature to obtain the final restoration feature; sub-pixel decoding processing is performed on the final restoration feature to obtain the prediction result; the prediction result is synthesized with the masked original radar image frame to obtain the deep learning restoration result. S103: Reconstruct the real echo texture from the deep learning restoration results, map the reconstructed image to the polar coordinate system, filter out radar artifacts, and obtain the restored radar image sequence.

[0022] Further, in step S101: acquire the weather radar image sequence to be processed, wherein the weather radar image sequence includes several frames of weather radar images; the weather radar image is a visualization map that intuitively displays the distribution intensity and movement of precipitation after the radar emits electromagnetic waves to detect precipitation particles such as clouds, rain, and hail in the air and processes the reflected echoes.

[0023] Further, step S101: performing structured noise detection on the weather radar image sequence to be processed to obtain a masked original radar image sequence, specifically including: S101-1: Perform pixel-level temporal mean aggregation on radar image sequences within a preset time window, and extract static or semi-static noise basis by utilizing the time-varying nature of real meteorological echoes and the difference in spatiotemporal stability of structured noise. S101-2: Convert the noise substrate into a grayscale image and perform global binarization to obtain a structured noise binary mask with the same size as the original radar image, which is used to indicate the location of the area to be repaired.

[0024] Further, S101-1: performing pixel-level temporal mean aggregation on the radar image sequence within a preset time window, specifically including: Take within the time window Calculating the temporal mean of each pixel in a frame-by-pixel cleansing image : ; in, Represents the spatial coordinates of pixels in the image along the horizontal direction; Represents the spatial coordinates of pixels in the vertical direction in an image; Indicates the first time within the time window Frame radar image in pixel coordinates The echo intensity value at that location; This represents the time index, used to indicate the temporal position of each radar image frame within a time window; This indicates the number of radar image frames included in the time-domain mean calculation within the time window.

[0025] Furthermore, S101-1: Utilizing the time-varying nature of real meteorological echoes and the difference in spatiotemporal stability between structured noise, a static or semi-static noise basis is extracted, specifically including: The radar images at each moment within the time window are processed according to the formula. Construct a noise basis map, where, Represents pixel coordinates The noise floor intensity value at that location, This represents the temporal mean of the corresponding pixel within the time window.

[0026] The effect of the above scheme is that the real echo is greatly diluted under long-term average due to motion generation and cancellation, while the fixed structured noise is retained at a high intensity due to long-term static state, thereby adaptively separating the static and semi-static noise substrates.

[0027] Further, step S101-2: converting the noise substrate into a grayscale image and performing global binarization to obtain a structured noise binary mask with the same size as the original radar image, used to indicate the location of the area to be repaired, specifically including: After converting the noise floor to grayscale, apply a threshold. Binarization is performed, with pixels having a grayscale value not lower than the threshold marked as 1, and others marked as 0. The output is a binary mask of the same size as the original image. This provides a priori location information for subsequent targeted repair.

[0028] Further, S102: the trained image sequence restoration model is used to extract features from the masked original radar image sequence to obtain gated coded features, specifically including: First, standardize the image sequence and the corresponding binary mask By concatenating the data along the channel dimension, we obtain the input tensor of the gated coding network: ; in, , This indicates a channel-level concatenation operation. Indicates batch size, Indicates the number of frames in the image sequence. and These represent the image height and width, respectively.

[0029] The input tensor Feature extraction is performed on a gated coding network with shared parameters input in time frames to obtain gated motion coding features for motion field estimation. For the input of frame t... (in The processing procedure of a gated coding network is represented as follows:

[0030]

[0031]

[0032] in, This represents the first gated 2D convolutional layer, used to downsample the input features to... Resolution; This indicates a second gated 2D convolutional layer, used to further downsample the features to... Resolution; and These represent the first residual block and the second residual block, respectively. and They represent the first The intermediate features of the frame after first- and second-level gated coding are used to... After performing shared parameter encoding operations on each frame, the overall gated coding features are obtained. .

[0033] At the same time, the mask sequence is downsampled to the same spatial resolution using nearest neighbor interpolation.

[0034] Furthermore, the gated coding network includes: a first gated two-dimensional convolutional layer, a first residual block, a second gated two-dimensional convolutional layer, and a second residual block connected in sequence; The operation of the first gated 2D convolutional layer and the second gated 2D convolutional layer is the same. The specific operation of the first gated 2D convolutional layer is as follows: ; ; ; in, This represents the feature of the first gated 2D convolutional layer input in frame t; This represents the content features output by the linear feature extraction branch; This represents the gating weights that generate the branch outputs from the gating coefficients; This represents the output features of the first gated 2D convolutional layer; and These represent the kernel weights and biases of the linear feature extraction branch, respectively. and These represent the kernel weights and biases of the gating coefficient generation branches, respectively; Represents a two-dimensional convolution operation; This represents the Sigmoid activation function, used to constrain the gate weights to between 0 and 1; This represents the LeakyReLU activation function; This indicates element-wise multiplication. Through the above gating mechanism, the network can adaptively adjust the feature responses of missing or unreliable regions.

[0035] Furthermore, the working processes of the first residual block and the second residual block are consistent. The working process of the first residual block is as follows: ; ; ; in, This represents the input features of the first residual block, which is also the output features of the first gated 2D convolutional layer; This represents the output characteristics of the first residual block; This represents the intermediate features of the first convolutional layer and the nonlinear activation in the residual block; This represents the residual mapping feature output by the second convolutional layer in the residual block; , These represent the kernel weights and biases of the first convolutional layer in the residual block, respectively. , These represent the kernel weights and biases of the second convolutional layer in the residual block, respectively. Represents a two-dimensional convolution operation; This represents the LeakyReLU activation function. It involves mapping residual features... Input features By adding them together, the residual blocks enhance the feature representation capabilities while retaining the original information in the input features.

[0036] Further, step S102: performing optical flow field estimation on the gated coding features to obtain the motion field estimation result, specifically including: To obtain motion information of the missing region, for any adjacent time... and The gating motion coding features, denoted as the first Frame features are , No. Frame features are The corresponding downsampling masks are respectively and ,in, ， .

[0037] By concatenating the features of adjacent frames with the corresponding downsampling mask along the channel dimension, we obtain the forward optical flow estimation input and the backward optical flow estimation input:

[0038]

[0039] in, Indicates the use of estimating the first Frame to the Input characteristics of frame forward optical flow; Indicates the use of estimating the first Frame to the Input characteristics of frame backward optical flow; This indicates a channel dimension splicing operation.

[0040] Specifically, the forward low-resolution optical flow and the backward low-resolution optical flow are represented as follows: ; ; in, This represents the optical flow decoding network; Indicates from the first Frame points to the first The forward low-resolution optical flow of a frame, i.e. , used to describe the In the next time step, the pixel or feature point in the frame Motion displacement within a frame; Indicates from the first Frame points to the first The backward low-resolution optical flow of the frame, i.e. This describes the motion displacement of a pixel or feature point in frame i+1 when it returns to the previous frame i. The same optical flow decoding network estimates the forward and backward optical flow by exchanging the order of the input features, thus achieving parameter sharing.

[0041] Subsequently, the forward low-resolution optical flow With backward low-resolution optical flow The motion field estimation results are obtained by upsampling to the original image resolution and simultaneously performing displacement amplitude scaling.

[0042] Furthermore, the optical flow decoding network is used to map the features of the stitched adjacent frames into a two-dimensional motion displacement field. The optical flow decoding network processing is represented as follows:

[0043]

[0044]

[0045]

[0046] in, This represents the splicing characteristics of the input optical flow decoding network when estimating the forward optical flow. When estimating backward optical flow ; and These represent the third gated two-dimensional convolutional layer and the fourth gated two-dimensional convolutional layer, respectively. Indicates the third residual block; This indicates an optical flow output convolutional layer used to output a 2-channel displacement field. , These represent the kernel weights and biases of the optical flow output convolutional layer, respectively. This represents a two-dimensional convolution operation; , , These represent intermediate features in the optical flow decoding network; This represents the low-resolution optical flow field output, with its two channels representing horizontal and vertical displacements, respectively.

[0047] Furthermore, taking the aforementioned forward low-resolution optical flow as an example, the upsampling calculation process is characterized by the following formula:

[0048]

[0049] in, This represents the spatial height of the forward low-resolution optical flow. Space width ; Indicates a spatial upsampling operator (preferably using bilinear interpolation). This indicates the forward transition optical flow that has been spatially upsampled but has not yet undergone displacement amplitude scaling. and The horizontal displacement component matrix and the vertical displacement component matrix obtained by decomposing the forward transition optical flow in the channel dimension are respectively characterized; This represents the forward full-resolution target optical flow after synchronous scaling of displacement amplitude; and These represent the height and width of the original input image, respectively.

[0050] The backward low-resolution optical flow Using the exact same upsampling and amplitude scaling methods, the corresponding back-side full-resolution target light was calculated. Finally, the complete motion field estimation result is obtained by combining the forward and backward full-resolution target optical flows corresponding to all adjacent frames in the sequence.

[0051] Furthermore, the optical flow decoding network includes: a third gated two-dimensional convolution, a third residual block, and a fourth gated two-dimensional convolution connected in sequence; The working process of the third and fourth gated two-dimensional convolutions is consistent with that of the first gated two-dimensional convolution.

[0052] The working process of the third residual block is the same as that of the first residual block.

[0053] To avoid dependence on real optical flow annotations, this invention employs a self-supervised training strategy based on photometric consistency to optimize the motion field estimation network. The optical flow-guided image reconstruction error is used as a supervision signal to achieve the learning of the motion field in the missing region.

[0054] The masked radar image sequence is input into the motion field network. Based on the shared features of adjacent frames and the mask of the current frame, the bidirectional optical flow field of the missing region is estimated and upsampled to the original resolution as the basis for subsequent cross-frame alignment.

[0055] Further, step S102: performs cross-frame information transfer and local temporal completion on the motion field estimation results and gated coding features to obtain the repaired features after local temporal completion, specifically including: The estimation results of the sports field and the gated coding features are input into the bidirectional gated loop propagation module. The number of inputs and hidden state channels of the bidirectional gated loop propagation module are both [missing information]. During unidirectional propagation, the bidirectional gated cyclic propagation module performs the following alignment and fusion operations at each time step: S102-1: First, based on the optical flow in the corresponding direction, perform a deformation alignment operation on the hidden states propagated from adjacent time points to obtain aligned hidden states. .

[0056] The deformation alignment operation refers to resampling the latent state features from the previous propagation moment in spatial coordinates based on the optical flow displacement relationship between adjacent frames. This latent state feature is then mapped to the coordinate system of the current frame, thereby aligning the propagated latent state with the features of the current frame in spatial position. This operation can be implemented using optical flow-based reverse mesh sampling, i.e., resampling the latent state features at the current frame position. The corresponding sampling position is found in the source hidden state according to the displacement given by the optical flow, and the hidden state feature value at that position is obtained by interpolation.

[0057] Let the current time be The hidden state at the previous propagation moment is The optical flow field used for the current alignment direction is The deformation alignment process is represented as:

[0058] in, Indicates alignment to the current time. The hidden state; This represents the hidden state at the previous propagation moment; This represents a deformation alignment operator based on optical flow; This indicates the optical flow field used in the current alignment step; and They represent In position The displacement components along the horizontal and vertical coordinate directions; and These represent the horizontal and vertical coordinates in the feature map, respectively.

[0059] For forward propagation of an image sequence, let Indicates by the first Frame points to the first The backward optical flow in the frame direction is used to... The hidden state sampled from the frame propagation is taken to the first frame. In frame coordinates, forward alignment is represented as: ; For the reverse propagation of the image sequence, let Indicates by the first Frame points to the first Forward optical flow in the frame direction is used to... The hidden state sampled from the frame propagation is taken to the first frame. In frame coordinates, reverse alignment is represented as: ; in, This indicates alignment to the first position during forward propagation. The hidden state of a frame; Indicates alignment to the first during backpropagation. The hidden state of a frame; Indicating the first step in positive propagation Hidden state after frame update; Indicating the first step in backpropagation Hidden state after frame update; Indicates by the first Frame points to the first Backward optical flow in the frame direction; Indicates by the first Frame points to the first Forward optical flow in the frame direction. If there are no adjacent hidden states to be propagated in the first frame of forward propagation and the last frame of reverse propagation, the features of the corresponding current frame can be initialized to the hidden state at that moment.

[0060] S102-2: Then, combined with a confidence gate network Current frame gated motion coding features And the optical flow used in the current step Calculate confidence level gating weights together .

[0061] The confidence gate network is a learned convolutional subnetwork used to adaptively determine whether to use the propagated hidden state or the current frame features more at each spatial location and in each channel, based on the aligned propagation features, the current frame observation features, and optical flow displacement information. Specifically, the confidence gate network is represented as follows:

[0062]

[0063] in, Indicates the direction of propagation. At that time, it was a positive transmission. This is a reverse propagation; This represents the confidence gate network corresponding to the propagation direction. Forward and backward propagation can use confidence gate networks with the same structure but independent parameters. Indicates the kernel size as The number of input channels is The number of output channels is Convolutional layers; This indicates that the kernel size is 3. 3. The number of input and output channels is [number missing]. Convolutional layers; This represents the LeakyReLU activation function; This indicates splicing along the channel dimension; Represents the Sigmoid function; Indicates the number of feature channels; In From and Channel splicing, additional Each channel originates from the horizontal and vertical displacement components of the optical flow; This represents the confidence level gating weight.

[0064] Furthermore, confidence-gated weights are used to perform gated fusion of the aligned hidden state and the current frame features to obtain the fused hidden state. :

[0065] in, Indicates the first Frame in the direction of propagation On the gated fusion hidden state; Indicates the confidence level gating weight; This represents the propagation hidden state after optical flow deformation alignment; Indicates the first Gated motion coding features of frames; Indicates and A uniformly shaped full-1 tensor; This indicates element-wise multiplication. When... When the value is greater than the set threshold, it indicates that the current position has more confidence in the propagation hidden state after optical flow alignment; when When the value is less than the set threshold, it means that the current position trusts the features of the current frame more.

[0066] S102-3: Subsequently, the current frame's gated motion coding features are... With the fused hidden state The inputs are fed into the Convolutional Gated Recurrent Unit (ConvGRU) to update the hidden state at the current time step. .

[0067] In this implementation, ConvGRU uses the features of the current frame as input. ,by As input to the previous hidden state The update process is represented as follows:

[0068]

[0069]

[0070]

[0071]

[0072] in, Indicates the direction of propagation. At that time, it was a positive transmission. This is a reverse propagation; Indicates that ConvGRU is in the... The input features of the frame are used to extract the gated motion coding features of the current frame in this step. ; This represents the previous hidden state of the input ConvGRU, and the gated fusion hidden state is taken in this step. ; This represents the update gate, used to control the proportion of candidate hidden states written to the current hidden state; This represents the reset gate, which controls the proportion of information from the previous hidden state retained when calculating candidate hidden states. Indicates a candidate hidden state; This represents the hidden state at the current time after the ConvGRU update; This indicates the function used to generate both the update gate and the reset gate simultaneously. A convolutional layer has 100 input channels. The number of output channels is And divided into channels and ; This represents the state used to generate candidate hidden states. A convolutional layer has 100 input channels. The number of output channels is ; Represents the hyperbolic tangent activation function; Indicates channel splicing; This represents element-wise multiplication. Therefore, ConvGRU, through recursive computation of update gates, reset gates, and candidate hidden states, introduces current frame observation features while preserving temporal propagation information, thus achieving local temporal information completion.

[0073] By traversing the image sequence in both the forward and reverse directions as described above, the forward propagation hidden state sequence can be obtained. and backpropagation of hidden state sequence :

[0074]

[0075] in, Indicates the number of frames in the image sequence; This indicates organization or splicing along the time dimension; and All shapes are , Indicates batch size. and These represent the height and width of the feature map, respectively.

[0076] S102-4: Finally, the hidden states obtained from forward and backward propagation are concatenated along the channel dimension, fused through convolution, and then combined with the gated motion coding feature tensor. The residuals are summed to obtain the repaired features after local temporal completion. , is represented as:

[0077]

[0078] in, Represents the forward propagation hidden state sequence Backpropagation of hidden state sequence The feature sequence concatenated along the channel dimension has the following shape: ; This indicates that the number of channels will be changed from... Mapped to of Convolutional layer; This indicates that the number of channels is maintained. of Convolutional layer; This represents the gated motion coding feature of the input bidirectional gated cyclic propagation module, and its shape is as follows: ; This represents the repaired feature after local temporal completion, and its shape is also... + indicates that the element-wise residuals are added together. This indicates a channel-level concatenation operation; This represents the LeakyReLU activation function; Indicates batch size. and These represent the height and width of the feature map, respectively.

[0079] Through this residual connection, the module can supplement bidirectional temporal propagation information while retaining the original intra-frame features extracted by the encoder.

[0080] Along both forward and backward time directions, the hidden states of adjacent frames are aligned using the completed optical flow, and the propagation information and the current observation information are adaptively fused through the confidence gate mechanism and the convolution gated recurrent unit.

[0081] After gating propagation, the features have obtained temporal consistency information within a local time range. To further compensate for missing content over long time spans and large spatial regions, an axial memory spatiotemporal completion module is introduced to model the global spatiotemporal context.

[0082] Further, step S102: Performing global spatiotemporal completion on the repaired features after local temporal completion to obtain the final repaired features, specifically including: The repaired features after local temporal completion are input into the axial memory spatiotemporal completion network. Axial memory spacetime completion network Output the reconstructed feature tensor, sum the reconstructed feature tensor with the repaired features after local temporal completion, and obtain the final repaired features.

[0083] For example, the reconstructed feature tensor is residually fused with the features of the input axial memory spatiotemporal completion network to obtain the final repaired features:

[0084] in, These are the repair features after local temporal completion; This represents the final repaired feature after global spatiotemporal completion. This represents an axial memory spatiotemporal completion network. The axial memory spatiotemporal completion network includes a cascaded process of feature tokenization, conditional position encoding fusion, multiple axial memory Transformer modules, and token recombining operations.

[0085] First, the input features are processed by convolution with stride. Tokenization is performed to obtain spatiotemporal feature tokens: ,in, Indicates batch size, Indicates the time dimension. Indicates the number of space tokens. Indicates the feature dimension.

[0086] Among them, strided convolution refers to a convolution operation in which the convolution kernel slides on the feature map according to a set strid. In this embodiment, the convolution kernel size can be [size missing]. Step size is The convolutional layer maps each local spatial region to a... This allows for spatial downsampling and feature dimension mapping by using dimensional feature vectors.

[0087] Specifically, no. The tokenization process of a frame is represented as follows:

[0088]

[0089] in, This represents the spatiotemporal feature token tensor obtained through tokenization. This indicates a strided convolutional layer used for tokenization; Indicates the first The restoration features after local temporal completion corresponding to the frame; Indicates the first Embedded features obtained after a frame undergoes convolution with stride; and These represent the height and width of the tokenized space, respectively. This indicates the number of space tokens corresponding to each frame; Represents the feature dimension of each token; Indicates batch size; Indicates the number of frames in the image sequence; and These represent the height and width of the input feature map, respectively. This involves unfolding the convolutional spatial features and adjusting their dimensionality order to form a spatiotemporal token tensor suitable for Transformer processing. Tokenization transforms the two-dimensional spatial feature map into a token sequence composed of multiple feature vectors, enabling subsequent axial memory Transformers to perform attention modeling in both the temporal and spatial dimensions.

[0090] Subsequently, the spatiotemporal feature token tensor is encoded using depthwise separable convolution to generate conditional position codes, which are then superimposed onto the spatiotemporal feature token tensor to inject spatial position information.

[0091] in, This represents the feature token tensor after fusion conditional position encoding, which serves as the input to the subsequent cascaded axial memory Transformer module; This represents the spatiotemporal feature token obtained from the previous tokenization step; This indicates a conditional position encoding module; Indicates the kernel size as Channel-wise convolution.

[0092] Feature token tensor after fusion of conditional position encoding The data is then input into multiple cascaded axis memory Transformer modules for sequential processing.

[0093] For the The processing of the axial memory Transformer module is as follows: , , , , , in, , They represent the first , The output tensor of the axial memory Transformer module; , They represent the first , The global memory state corresponding to each axis memory Transformer module; This represents a temporal attention operation used to capture cross-temporal relationships; This represents a spatial attention operation used to capture spatial structural dependencies; This represents a feedforward neural network; This indicates a memory write operation, which aggregates the current spatiotemporal features into a global memory unit; This indicates a memory readout operation, which uses global memory to supplement the contextual information missing from the current feature.

[0094] Furthermore, time-attention operations specifically refer to: For the first The output tensor of the axial memory Transformer module Reorganize according to the spatial token position to obtain This makes each spatial token position correspond to a string of length [length missing]. Time series, and in the time dimension Execute long self-attention:

[0095]

[0096] in, This means rearranging the features into a form that performs attention calculations along the time dimension; This indicates a bullish attention-based operation. , , All of them originate from the same normalized features, therefore they belong to self-attention. This represents the output result of multi-head self-attention; This indicates that the output result of self-attention will be restored to... shape; Indicates batch size; Indicates the number of frames in the image sequence; This indicates the number of space tokens corresponding to each frame; This represents the feature dimension of each token.

[0097] Furthermore, spatial attention operations specifically refer to: The output of the time attention operation at this layer Reassembled according to time frames, we get This ensures that each time frame position corresponds to a line of length [length missing]. The spatial token sequence, and in the spatial token dimension Execute long self-attention:

[0098]

[0099] in, This means rearranging the features into a form where attention is computed along the spatial token dimension; This indicates a bullish attention-based operation. , , All of them originate from the same normalized features, therefore they belong to self-attention. This represents the output result of multi-head self-attention; This indicates that the output result of self-attention will be restored to... shape; Indicates batch size; Indicates the number of frames in the image sequence; This indicates the number of space tokens corresponding to each frame; This represents the feature dimension of each token.

[0100] Specifically, memory write and read operations refer to: The output of the spatial attention operation at this layer By reorganizing the spatiotemporal dimensions, we obtain Then focus attention:

[0101]

[0102] in, The tensor representing the output of the spatial attention operation at this layer; Represents the recombined tensor; , They represent the first , The global memory state corresponding to each axis memory Transformer module; This indicates a bullish attention-based strategy; This indicates that the multi-head attention output features are restored to... This step enables each spacetime token to read long-range context information from global memory to supplement the spacetime content that is still missing after local propagation.

[0103] After iterative processing by multiple axial memory Transformer modules, a token reassembly operation is performed on the output feature tensor, that is, the spatial resolution is reduced from 1×1 convolution to 1×1. Restore to Where h = H / 4, w = W / 4; Indicates batch size; Indicates the number of frames in the image sequence; This indicates the number of space tokens corresponding to each frame; This represents the feature dimension of each token; C represents the number of channels per frame.

[0104] The beneficial effects of the above technical solution are as follows: Through the aforementioned axial decomposition and memory enhancement mechanism, traditional 3D spatiotemporal attention is decomposed into two independent modeling processes: temporal and spatial dimensions. This reduces computational complexity while achieving effective aggregation and transmission of global spatiotemporal context information, thereby completing the long-range spatiotemporal details not yet recovered during the gating propagation stage. The propagated features are then input into the axial memory Transformer, and through temporal and spatial modeling, as well as the global memory read / write mechanism, the long-range spatiotemporal context information still missing after propagation is completed.

[0105] Further, step S102: performing sub-pixel decoding processing on the final repaired features to obtain the prediction result, specifically including: After completing global spatiotemporal completion, the final features will be repaired. The input to the subpixel upsampling decoder is upsampled fourfold to restore the original image resolution, yielding the prediction result. .

[0106] Specifically, the subpixel upsampling decoder workflow is as follows:

[0107]

[0108]

[0109]

[0110] in, This represents the final repaired feature after global spatiotemporal completion. 、 、 This represents the intermediate processing result; This represents the predicted repair outcome; Indicates the kernel size as The number of input channels is The number of output channels is Two-dimensional convolution; Indicates the multiplier. The subpixel rearrangement operation is used to rearrange the channels in the channel dimension. Sub-pixel features are rearranged into spatial dimensions, thus enlarging both the height and width of the feature map. times; twice Cascading achieves Double upsampling; Indicates the number of intermediate channels in the decoder; This represents the LeakyReLU activation function; This represents the hyperbolic tangent activation function, used to constrain the output pixel value to the interval [-1, 1].

[0111] The completed features are input into the subpixel upsampling decoder to restore the full-resolution predicted image. The image is then synthesized according to the strategy of "using the predicted result in the masked area and retaining the original image in the known area" to eliminate the spatiotemporal discontinuity caused by structured noise.

[0112] Further, step S102: synthesizing the prediction result with the masked original radar image frame to obtain the deep learning restoration result, specifically including: To preserve the true information of known regions in each frame of the original image sequence, a mask is used to synthesize the prediction stream with the original frames:

[0113] in, This represents the normalized image sequence of the initial input; Represents the binary mask sequence of the initial input; Represents the sequence of predicted repair results; This indicates element-wise multiplication; This is a result of deep learning repair.

[0114] The process fills in missing data and eliminates spatiotemporal discontinuities caused by structured noise. Through the above fusion process, only the missing regions are reconstructed, while the known regions retain their original observations, thus obtaining a spatiotemporally continuous and structurally consistent restoration result.

[0115] Further, in step S102: the original radar image sequence with mask is input into the trained image sequence restoration model to obtain the deep learning restoration result, specifically including: The training process of the trained image sequence restoration model includes: Construct a training set, which is a sequence of masked original radar images with known deep learning restoration results; The training set is input into the image sequence restoration model to train it. Training is stopped when the model's loss function no longer decreases or the number of iterations exceeds a set number, resulting in the trained image sequence restoration model.

[0116] Further, step S103: reconstructing the real echo texture from the deep learning repair result, specifically includes: Repairing deep learning results Grayscale and will Based on effective echo threshold Binarization yields an echo structure mask that represents only "where valid cloud clusters exist". :

[0117] in, This indicates the deep learning repair result at the pixel location. The grayscale value at that location; and These represent the horizontal and vertical coordinates in the image, respectively. This indicates the set effective echo threshold; Indicates pixel position The echo structure mask value at that location, This indicates that there is a valid cloud echo at that pixel location. This indicates that there is no valid cloud echo at this pixel location.

[0118] echo structure mask After being expanded to three channels, the original radar image corresponding to the timestamp is multiplied pixel by pixel:

[0119] in, This represents a radar image after texture reconstruction. This represents the original radar image with the same timestamp as the current deep learning restoration result; Indicates single-channel echo structure mask expansion; This indicates element-wise multiplication. The fixed structured noise in the original image corresponding to that location is completely zeroed out. The original high-frequency textures that have not been smoothed by deep learning are directly preserved, thereby improving the realism of local textures while ensuring structural coherence and alleviating the problem of over-smoothing in deep learning repair.

[0120] The deep learning restoration results are used to generate a spatial guidance matrix, which is then fused pixel by pixel with the original radar image at the corresponding time. This process preserves the overall structural consistency of the restoration results while restoring high-frequency texture details from the original observation as much as possible and reducing over-smoothing.

[0121] Further, step S103: mapping the reconstructed image to a polar coordinate system, filtering out radar artifacts, and obtaining a repaired radar image sequence specifically includes: Mapping the image to a polar coordinate system reduces the dimensionality of the radially diverging fan-shaped ray artifacts to a horizontal linear structure. By morphologically closing the discontinuous line segments of the horizontal linear structure and then performing an opening operation to preserve the long lines, ray candidates are obtained. ; To avoid accidentally damaging valid data, the effective echo mask in polar coordinates is used. With angle width threshold The vertical opening operation is used to expand and generate a "morphological shield". The package contains real cloud formations.

[0122] Subsequently, the ray candidate mask was... With morphological shield Taking the intersection of the inverse masks yields the noise mask in polar coordinates:

[0123] in, This represents the ray candidate mask obtained by horizontal morphological closing and opening operations in polar coordinates. Indicates by The morphological shield obtained through vertical opening and dilation is used to mark the real cloud area that should be preserved; express Invert the mask; This represents the mask intersection operation; This represents the polar coordinate noise mask that is ultimately determined to be artifact noise. The formula indicates that only candidates belonging to the ray will be included. And it does not belong to the actual cloud cluster protection area. The pixel position is identified as noise, thereby reducing the false damage to valid meteorological echoes while stripping away radial artifacts.

[0124] Finally, noise points with areas smaller than the connected region area threshold are removed according to the set threshold.

[0125] Reassemble each frame according to its original timestamp, and the resulting spatiotemporal tensor sequence is the final denoising result.

[0126] The image is mapped to a polar coordinate system, and the radially diverging fan-shaped artifacts are converted into regular morphological structures. The artifact candidate regions are extracted by combining morphological opening and closing operations. Furthermore, a protection region is constructed by constraining the angle and width to reduce false alarms to real meteorological cloud clusters. Finally, the residual isolated noise points are removed by combining connected component analysis to obtain a high-purity radar image sequence.

[0127] Example 2 This embodiment provides a structured noise detection and spatiotemporal restoration system for weather radar image sequences, including: The preprocessing module is configured to: acquire the weather radar image sequence to be processed, perform structured noise detection on the weather radar image sequence to be processed, and obtain the original radar image sequence with mask; The deep learning module is configured to: input a masked original radar image sequence into a trained image sequence restoration model to obtain a deep learning restoration result; wherein, the trained image sequence restoration model is used to extract features from the masked original radar image sequence to obtain gated coded features; perform optical flow field estimation on the gated coded features to obtain motion field estimation results; perform cross-frame information transfer and local temporal completion on the motion field estimation results and gated coded features to obtain locally temporally completed restoration features; perform global spatiotemporal completion on the locally temporally completed restoration features to obtain the final restoration features; perform sub-pixel decoding processing on the final restoration features to obtain prediction results; and synthesize the prediction results with the masked original radar image frames to obtain the deep learning restoration result. The reconstruction module is configured to: reconstruct the real echo texture from the deep learning restoration results, map the reconstructed image to the polar coordinate system, filter out radar artifacts, and obtain the restored radar image sequence.

[0128] It should be noted that the preprocessing module, deep learning module, and reconstruction module mentioned above correspond to steps S101 to S103 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0129] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences, characterized by: include: The weather radar image sequence to be processed is obtained, and structured noise detection is performed on the weather radar image sequence to be processed to obtain the original radar image sequence with masking. The masked original radar image sequence is input into a trained image sequence restoration model to obtain the deep learning restoration result. Specifically, the trained image sequence restoration model extracts features from the masked original radar image sequence to obtain gated coded features; optical flow field estimation is performed on the gated coded features to obtain motion field estimation results; cross-frame information transfer and local temporal completion are performed on the motion field estimation results and gated coded features to obtain locally temporally completed restoration features; global spatiotemporal completion is performed on the locally temporally completed restoration features to obtain the final restoration features; sub-pixel decoding processing is performed on the final restoration features to obtain the prediction result; and the prediction result is synthesized with the masked original radar image frames to obtain the deep learning restoration result. The deep learning restoration results are used to reconstruct realistic echo textures. The reconstructed images are mapped to polar coordinates, and radar artifacts are filtered out to obtain the restored radar image sequence, which specifically includes: The image is mapped to polar coordinates, and the radially diverging fan-shaped ray artifacts are reduced to a horizontal linear structure. The horizontal linear structure is then bridged by morphological closing operations to connect discontinuous line segments, and long lines are preserved by opening operations to obtain ray candidates. Effective echo mask in polar coordinates With angle width threshold The vertical opening operation is used to expand and generate a "morphological shield". Enveloping real cloud formations; Subsequently, the ray candidate mask was... With morphological shield Taking the intersection of the inverse masks yields the noise mask in polar coordinates: in, This represents the ray candidate mask obtained by horizontal morphological closing and opening operations in polar coordinates. Indicates by The morphological shield obtained through vertical opening and dilation is used to mark the real cloud area that should be preserved; express Invert the mask; This represents the mask intersection operation; This represents the polar coordinate noise mask that is ultimately determined to be artifact noise; Finally, noise points with areas smaller than the connected component area threshold are removed according to a set threshold; each frame is reassembled according to its original timestamp, and the spatiotemporal tensor sequence is output as the final denoising result.

2. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, Structured noise detection is performed on the weather radar image sequence to be processed to obtain a masked original radar image sequence, specifically including: Pixel-level temporal mean aggregation is performed on the radar image sequence within a preset time window. Static or semi-static noise base is extracted by utilizing the time-varying nature of real meteorological echoes and the difference in spatiotemporal stability of structured noise. The noise base is converted into a grayscale image and subjected to global binarization to obtain a structured noise binary mask with the same size as the original radar image, which is used to indicate the location of the area to be repaired.

3. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, The trained image sequence inpainting model is used to extract features from the masked original radar image sequence to obtain gated coded features, specifically including: First, standardize the image sequence and the corresponding binary mask By concatenating the data along the channel dimension, we obtain the input tensor of the gated coding network: ； in, , This indicates a channel-level concatenation operation. Indicates batch size, Indicates the number of frames in the image sequence. and These represent the image height and width, respectively. The input tensor Feature extraction is performed on a gated coding network with shared parameters input in time frames to obtain gated motion coding features for motion field estimation. For the input of frame t , The processing procedure of the gated coding network is represented as follows: in, This represents the first gated 2D convolutional layer, used to downsample the input features to... Resolution; This indicates a second gated 2D convolutional layer, used to further downsample the features to... Resolution; and These represent the first residual block and the second residual block, respectively. and They represent the first The intermediate features of the frame after first- and second-level gated coding are used to... After performing shared parameter encoding operations on each frame, the overall gated coding features are obtained. .

4. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, Optical flow field estimation is performed on the gated coding features to obtain the motion field estimation results, specifically including: To obtain motion information of the missing region, for any adjacent time... and The gating motion coding features, denoted as the first Frame features are , No. Frame features are The corresponding downsampling masks are respectively and ,in, ， ; By concatenating the features of adjacent frames with the corresponding downsampling mask along the channel dimension, we obtain the forward optical flow estimation input and the backward optical flow estimation input: in, Indicates the use of estimating the first Frame to the Input characteristics of frame forward optical flow; Indicates the use of estimating the first Frame to the Input characteristics of frame backward optical flow; This indicates a channel-level concatenation operation; Forward low-resolution optical flow and backward low-resolution optical flow are respectively represented as: ； ； in, This represents the optical flow decoding network; Indicates from the first Frame points to the first The forward low-resolution optical flow of the frame is used to describe the first frame. In the next time step, the pixel or feature point in the frame Motion displacement within a frame; Indicates from the first Frame points to the first The backward low-resolution optical flow of a frame is used to describe the motion displacement of a pixel or feature point in the (i+1)th frame as it returns to the previous frame i. Subsequently, the forward low-resolution optical flow With backward low-resolution optical flow The motion field estimation results are obtained by upsampling to the original image resolution and simultaneously performing displacement amplitude scaling.

5. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, The estimation results of the motion field and the gated coding features are subjected to cross-frame information transfer and local temporal completion to obtain the repaired features after local temporal completion, specifically including: The estimation results of the sports field and the gated coding features are input into the bidirectional gated loop propagation module. The number of inputs and hidden state channels of the bidirectional gated loop propagation module are both [missing information]. During unidirectional propagation, the bidirectional gated cyclic propagation module performs the following alignment and fusion operations at each time step: First, based on the optical flow in the corresponding direction, a deformation alignment operation is performed on the hidden states propagated from adjacent time points to obtain aligned hidden states. ; Then, by combining through a confidence gate network Current frame gated motion coding features and the optical flow used in the current step Calculate confidence gating weights together The aligned hidden state and the current frame features are gated and fused using confidence-gated weights to obtain the fused hidden state. ; Subsequently, the current frame's gated motion coding features are... With the fused hidden state The inputs are fed into the Convolutional Gated Recurrent Unit (ConvGRU) to update the hidden state at the current time step. ; Finally, the hidden states obtained from forward and backward propagation are concatenated along the channel dimension, fused through convolution, and then combined with the gated motion coding feature tensor. The residuals are summed to obtain the repaired features after local temporal completion. .

6. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, Global spatiotemporal completion is performed on the restored features obtained after local temporal completion to obtain the final restored features, which specifically include: The repaired features after local temporal completion are input into the axial memory spatiotemporal completion network. Axial memory spacetime completion network The recombined feature tensor is output, and the recombined feature tensor is summed with the repaired features after local temporal completion to obtain the final repaired features; the axial memory spatiotemporal completion network includes feature tokenization, conditional position encoding fusion, multiple axial memory Transformer modules, and the cascaded processing of token recombining operations.

7. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, The final repaired features are subjected to sub-pixel decoding to obtain the prediction results, which include: After completing global spatiotemporal completion, the final features will be repaired. The input is upsampled four times by the subpixel upsampling decoder to restore the original image resolution, thus obtaining the prediction result. ; Specifically, the subpixel upsampling decoder workflow is as follows: in, This represents the final repaired feature after global spatiotemporal completion. 、 、 This represents the intermediate processing result; This represents the predicted repair outcome; Indicates the kernel size as The number of input channels is The number of output channels is Two-dimensional convolution; Indicates the multiplier. The subpixel rearrangement operation is used to rearrange the channels in the channel dimension. Sub-pixel features are rearranged into spatial dimensions, thus enlarging both the height and width of the feature map. times; twice Cascading achieves Double upsampling; Indicates the number of intermediate channels in the decoder; This represents the LeakyReLU activation function; This represents the hyperbolic tangent activation function, used to constrain the output pixel value to the interval [-1, 1]. The predicted results are synthesized with the masked original radar image frames to obtain the deep learning restoration results, which specifically include: To preserve the true information of known regions in each frame of the original image sequence, a mask is used to synthesize the prediction stream with the original frames: in, This represents the normalized image sequence of the initial input; Represents the binary mask sequence of the initial input; Represents the sequence of predicted repair results; This indicates element-wise multiplication; This is a result of deep learning repair.

8. The method for structured noise detection and spatiotemporal restoration of meteorological radar image sequences as described in claim 1, characterized in that, The deep learning restoration results are then used to reconstruct realistic echo textures, specifically including: Repairing deep learning results Grayscale and will Based on effective echo threshold Binarization yields the echo structure mask. : in, This indicates the deep learning repair result at the pixel location. The grayscale value at that location; and These represent the horizontal and vertical coordinates in the image, respectively. This indicates the set effective echo threshold; Indicates pixel position The echo structure mask value at that location, This indicates that there is a valid cloud echo at that pixel location. This indicates that there is no valid cloud echo at this pixel location; echo structure mask After being expanded to three channels, the original radar image corresponding to the timestamp is multiplied pixel by pixel: in, This represents a radar image after texture reconstruction. This represents the original radar image with the same timestamp as the current deep learning restoration result; Indicates single-channel echo structure mask expansion; This indicates element-wise multiplication; The fixed structured noise in the original image corresponding to that location is completely zeroed out. The original high-frequency texture, which has not been smoothed by deep learning, is directly preserved.

9. A structured noise detection and spatiotemporal restoration system for weather radar image sequences, characterized in that... include: The preprocessing module is configured to: acquire the weather radar image sequence to be processed, perform structured noise detection on the weather radar image sequence to be processed, and obtain the original radar image sequence with mask; The deep learning module is configured to: input a masked original radar image sequence into a trained image sequence restoration model to obtain a deep learning restoration result; wherein, the trained image sequence restoration model is used to extract features from the masked original radar image sequence to obtain gated coded features; perform optical flow field estimation on the gated coded features to obtain motion field estimation results; perform cross-frame information transfer and local temporal completion on the motion field estimation results and gated coded features to obtain locally temporally completed restoration features; perform global spatiotemporal completion on the locally temporally completed restoration features to obtain the final restoration features; perform sub-pixel decoding processing on the final restoration features to obtain prediction results; and synthesize the prediction results with the masked original radar image frames to obtain the deep learning restoration result. The reconstruction module is configured to: reconstruct realistic echo textures from the deep learning restoration results, map the reconstructed image to a polar coordinate system, filter out radar artifacts, and obtain a restored radar image sequence, specifically including: The image is mapped to polar coordinates, and the radially diverging fan-shaped ray artifacts are reduced to a horizontal linear structure. The horizontal linear structure is then bridged by morphological closing operations to connect discontinuous line segments, and long lines are preserved by opening operations to obtain ray candidates. Effective echo mask in polar coordinates With angle width threshold The vertical opening operation is used to expand and generate a "morphological shield". Enveloping real cloud formations; Subsequently, the ray candidate mask was... With morphological shield Taking the intersection of the inverse masks yields the noise mask in polar coordinates: in, This represents the ray candidate mask obtained by horizontal morphological closing and opening operations in polar coordinates. Indicates by The morphological shield obtained through vertical opening and dilation is used to mark the real cloud area that should be preserved; express Invert the mask; This represents the mask intersection operation; This represents the polar coordinate noise mask that is ultimately determined to be artifact noise; Finally, noise points with areas smaller than the connected component area threshold are removed according to a set threshold; each frame is reassembled according to its original timestamp, and the spatiotemporal tensor sequence is output as the final denoising result.

Images