Radar echo extrapolation method and system based on frequency domain enhancement

Abstract

The invention discloses a radar echo extrapolation method and system based on frequency domain enhancement, and the method mainly comprises the following steps: obtaining and preprocessing a historical radar echo grayscale image sequence, generating a sequence sample through a sliding window, and dividing the sequence sample into a training set, a verification set and a test set; the method comprises the following steps: constructing a frequency domain enhanced U-Net network comprising an encoder-decoder structure, introducing a multi-scale deep convolution structure into an encoder and a decoder, and enhancing frequency domain features by using a frequency domain dynamic attention mechanism in jump connection; inputting the training set into the model for training by adopting a composite loss function comprising intensity weighted loss, frequency domain consistency loss and structural similarity loss; and inputting the test set into the trained model, and outputting a radar echo prediction result at a future moment. The method can be effectively applied to the fields of short-term and temporary weather forecast, severe convection monitoring and the like, and provides more accurate and reliable radar echo prediction support for meteorological disaster early warning.

Description

Technical Field

[0001] This invention belongs to the field of short-term meteorological forecasting technology, and in particular relates to a radar echo extrapolation method and system based on frequency domain enhancement. Background Technology

[0002] Short-term precipitation forecasting is a high-precision forecast of precipitation distribution and intensity within the next two hours, and it is an important research area in meteorology. Unlike medium- and long-term weather forecasts, short-term precipitation forecasting focuses on responding to sudden and locally significant precipitation events, especially severe convective weather such as heavy rain and thunderstorms. In recent years, extreme weather events have become more frequent, and the accuracy of short-term precipitation forecasts is of great significance for urban flood control, agricultural irrigation, and traffic management, while also providing core technical support for disaster prevention and mitigation. Early traditional numerical weather prediction (NWP) solved atmospheric motion control equations and used high-performance computers to simulate future weather evolution. However, due to the uncertainty of initial conditions and the high computational resource consumption, the accuracy and timeliness of numerical weather prediction are limited.

[0003] Radar echo extrapolation is a short-term precipitation forecasting method based on real-time radar observation data. This technique analyzes radar echo images from consecutive time intervals, extracts the motion evolution characteristics of the echo field, establishes a spatiotemporal extrapolation model, and predicts the echo distribution within the next 0-2 hours. Based on this, and combined with the ZrR relationship, the predicted echo intensity is quantitatively converted into precipitation intensity, ultimately achieving short-term precipitation forecasting. Commonly used methods in traditional radar echo extrapolation include cross-correlation, centroid tracking, and optical flow. Cross-correlation estimates the translational velocity of the echo by performing correlation matching on local areas of continuous radar images, but it is easily affected by changes in the shape of local heavy precipitation echoes. Centroid tracking extracts and tracks the centroid position of individual echo cells or strong centers, suitable for predicting the movement of isolated cells, but difficult to handle echo systems with multiple cells or complex structures. Optical flow, based on the image brightness preservation assumption, calculates the velocity vector field between continuous images, capable of depicting a more detailed spatial flow field, but prone to large errors in echo generation, dissipation, and non-translational motion. These methods generally rely on the morphological stability of the echo, making it difficult to effectively capture the evolution of echo intensity and morphological changes.

[0004] With the development of deep learning technology, data-driven methods have received widespread attention in the field of radar echo extrapolation and have significantly improved the accuracy and timeliness of nowcasting. Traditional extrapolation methods mostly rely on the translation assumption of the echo field motion, which is difficult to characterize complex nonlinear evolution. Deep neural networks, with their powerful feature extraction and spatiotemporal modeling capabilities, can directly learn the evolution law of echoes from historical radar images. Shi et al. [1] regarded the short-term precipitation forecast problem as a spatiotemporal sequence prediction problem and proposed the ConvLSTM model. ConvLSTM introduces the convolution structure into the recurrent neural network. By preserving spatial structure information in the hidden state, it realizes the joint modeling of echo movement and morphological changes. It has been widely used in precipitation prediction, typhoon path forecasting and other tasks. Kevin et al. [2] proposed the SmaAt-UNet model. Based on the classic U-Net framework, it introduces a deep separable convolution and attention mechanism, which effectively reduces the number of parameters while enhancing the feature expression ability, and realizes a lighter and more efficient echo prediction. However, the inductive bias of these models tends to smooth the output, making it difficult to effectively capture localized sudden precipitation in severe convective weather. This results in blurred edges and loss of details in the predicted echoes, leading to a high rate of underreporting of small-scale extreme weather events.

[0005] [1]Shi X, Chen Z, Wang H, et al. Convolutional LSTM network: Amachine learning approach for precipitation nowcasting[J]. Advances in neuralinformation processing systems, 2015, 28.

[0006] [2]Trebing K, Staǹczyk T, Mehrkanoon S. SmaAt-UNet: Precipitationnowcasting using a small attention-UNet architecture[J]. Pattern RecognitionLetters, 2021, 145: 178-186. Summary of the Invention

[0007] The purpose of this invention is to provide a radar echo extrapolation method and system based on frequency domain enhancement, which aims to solve key problems such as echo ambiguity and distortion and loss of high echo regions in traditional time domain extrapolation techniques. By converting radar echo data to the frequency domain for analysis and modeling, it can effectively capture the multi-scale motion characteristics of meteorological systems while suppressing high-frequency noise interference, thereby improving the shortcomings of traditional time domain extrapolation in modeling nonlinear processes and exhibiting higher stability and detail preservation capabilities in short-term forecasts.

[0008] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0009] In a first aspect, the present invention provides a radar echo extrapolation method based on frequency domain enhancement, which includes the following steps:

[0010] S1: Obtain a sequence of historical radar echo grayscale images composed of radar echo grayscale images, convert the grayscale values ​​in the radar echo grayscale images into radar echo intensities, obtain radar echo images, and form a radar echo sequence. Use the sliding window technique on the preprocessed radar echo sequence to construct radar echo sequence samples.

[0011] S2: Two improvements are made to the U-Net network to obtain a frequency-domain enhanced U-Net network with an encoder-decoder structure. The first improvement is to replace the convolutional layers in the encoder and decoder of the U-Net network with multi-scale deep convolutional modules. The second improvement is to embed a frequency-domain dynamic attention module at each skip connection of the U-Net network to enhance the features of the encoder output.

[0012] S3: Train the frequency domain augmented U-Net network on radar echo sequence samples. During training, a composite loss function including intensity weighted loss, frequency domain consistency loss and structural similarity loss is used. The parameters of the frequency domain augmented U-Net network are updated based on minimizing the composite loss function.

[0013] S4: Input the radar echo sequence to be predicted into the trained frequency domain augmented U-Net network, output the radar echo prediction results for multiple future time steps, and finally visualize the radar echo prediction results to complete the entire extrapolation process.

[0014] Based on the above scheme, each step can be implemented in the following preferred manner.

[0015] As a preferred embodiment of the first aspect, the specific process of preprocessing the radar echo sequence and constructing radar echo sequence samples in step S1 is as follows: Radar echo images with a maximum echo intensity lower than a preset effective echo intensity threshold are removed from the radar echo sequence, and the retained radar echo images are normalized to form a preprocessed radar echo sequence; on the preprocessed radar echo sequence, a fixed-length sliding window is used for sample extraction. Each radar echo sequence sample contains input data and label data. The input data consists of N consecutive radar echo images in the preprocessed radar echo sequence, and the label data consists of N consecutive radar echo images following the input data in the preprocessed radar echo sequence.

[0016] As a preferred embodiment of the first aspect, in step S2, the multi-scale deep convolution module adopts a two-stage processing architecture: In the first stage, the input features are first subjected to layer normalization. The normalized features are then expanded by a 1×1 convolutional channel and fed into parallel first, second, and third deep convolutional layers. The features output by the three deep convolutional layers are concatenated. The concatenated features are then compressed by a 1×1 convolutional channel, subjected to an activation function, adaptively weighted by a channel attention module, and restored by a 1×1 convolutional channel to obtain the first feature. The first feature is then residually fused with the input features to form the first-stage output feature. In the second stage, the first-stage output feature is normalized again and fed into a feedforward neural network to obtain the second-stage output feature. Finally, the second-stage output feature and the first-stage output feature are residually connected to output the processing result of the multi-scale deep convolution module.

[0017] As a preferred embodiment of the first aspect, the kernel size of the first deep convolutional layer is 3×3, the kernel size of the second deep convolutional layer is 5×5, and the kernel size of the third deep convolutional layer is 7×7; the feedforward neural network is composed of a 1×1 convolutional layer, a GELU activation function, and a 1×1 convolutional layer cascaded sequentially.

[0018] As a preferred embodiment of the first aspect mentioned above, in step S2, the frequency domain dynamic attention module takes the features output by the multi-scale deep convolution module as input features. First, it performs a two-dimensional fast Fourier transform on the input features to transform them from the spatial domain to the frequency domain, generating shallow frequency domain features and separating their real and imaginary parts. The decomposed real and imaginary parts are then concatenated along the channel dimension to form an enhanced feature representation containing amplitude and phase information. The enhanced feature representation undergoes a 3×3 deep convolution to form deep frequency domain features, which are then fed into the frequency dynamic convolution module to generate frequency domain features after dynamic frequency modeling. An inverse Fourier transform is then applied to these features to return them to the spatial domain, resulting in frequency-enhanced spatial domain features. These enhanced spatial domain features are then fed into the multi-frequency attention module for processing, forming fused spatial domain features. Finally, the frequency-enhanced spatial domain features are residually connected with the fused spatial domain features to obtain the output features of the frequency domain dynamic attention module.

[0019] As a preferred embodiment of the first aspect mentioned above, the specific processing procedure in the frequency dynamic convolution module is as follows: K depth convolution kernels are set. The deep frequency domain features are first processed by global average pooling, 1×1 convolution, and softmax function to generate the weights of each depth convolution kernel. Then, the deep frequency domain features are fed into each depth convolution kernel for convolution operation. The outputs of each depth convolution kernel are weighted and summed according to their corresponding weights to obtain the frequency domain features after frequency dynamic modeling.

[0020] As a preferred embodiment of the first aspect mentioned above, the specific processing procedure in the multi-frequency attention module is as follows: The spatial domain features enhanced in the frequency domain are first obtained by two-dimensional fast Fourier transform to obtain a frequency domain tensor and regarded as a two-dimensional spectrogram. The two-dimensional spectrogram is divided into frequency bands according to the frequency level using a ring mask rule based on polar coordinate radius to extract the complex spectral components of each frequency band. Then, the complex spectral components of each frequency band are restored to the spatial domain by inverse Fourier transform to obtain the corresponding spatial domain feature map. Each spatial domain feature map is modeled by a channel attention module. Then, the weighted feature maps generated by all channel attention modules are concatenated in the channel dimension and compressed and fused by the fusion convolution module to generate the fused spatial domain features. The fusion convolution module consists of a cascaded 3×3 convolution, a ReLU activation function, and a 1×1 convolution.

[0021] As a preferred embodiment of the first aspect mentioned above, in step S3, the composite loss function used is a weighted sum of intensity-weighted loss, frequency domain consistency loss, and structural similarity loss: the intensity-weighted loss calculates the sum of the weighted mean square error and the mean absolute error between the predicted image and the real image, and the weights used in the weighting are proportional to the radar echo intensity; the frequency domain consistency loss is the mean absolute error of the frequency domain amplitude calculated after performing Fourier transforms on the predicted image and the real image respectively; the structural similarity loss constrains the similarity of the local structures of the predicted image and the real image through the SSIM exponent; wherein, the predicted image is the radar echo map output by the frequency domain enhanced U-Net network predicting the input data, and the real image is the label data corresponding to the input data.

[0022] Furthermore, a composite loss function of the following form is adopted. It is composed of intensity-weighted loss Frequency domain consistency loss and structural similarity loss The weighted sum is obtained by:

[0023]

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, Weights for loss items; This indicates the number of predicted radar echo maps output. and These represent the height and width of the radar echo image, respectively. Representing radar echo images in two-dimensional coordinates The weight of each pixel; and Representing the real image and the predicted image in two-dimensional coordinates respectively. Pixel value at; Representing radar echo images in two-dimensional coordinates Radar echo intensity at the location; and Represent the two-dimensional coordinates of the real image and the predicted image in the frequency domain, respectively. The amplitude value at that point; For real images and predicted images Structural similarity index between them; and Represents two non-zero constants; and These are the mean and variance of the real image, respectively; and These are the mean and variance of the predicted image, respectively; This represents the covariance between the real image and the predicted image.

[0030] As a preferred embodiment of the first aspect, in step S3, during the training of the frequency domain augmented U-Net network, an initial learning rate and a learning rate decay strategy are set. Radar echo sequence samples are input into the frequency domain augmented U-Net network in batches for forward propagation. The weights of the frequency domain augmented U-Net network are optimized by calculating the composite loss function and performing backpropagation using the Adam optimizer. A validation set monitoring and early stopping strategy are adopted. After iterative training, the parameters of the frequency domain augmented U-Net network with the best performance on the validation set are finally saved to obtain the trained frequency domain augmented U-Net network.

[0031] Secondly, the present invention provides a radar echo extrapolation system based on frequency domain enhancement, comprising:

[0032] The data acquisition module is used to acquire the radar echo sequence to be predicted;

[0033] The prediction module is used to predict the radar echo sequence to be predicted according to the frequency domain enhancement-based radar echo extrapolation method described in the first aspect above, obtain the radar echo prediction results for multiple future time steps, and display them visually.

[0034] The present invention has the following advantages and effects compared with the prior art:

[0035] 1. This invention introduces multi-scale depth convolution and designs a multi-scale depth convolution module. By performing three different scales of depth convolution in parallel, it captures small-scale detail changes and large-scale echo structure features in radar images. Compared with traditional single-scale convolution, this module can improve the model's ability to express complex spatial patterns without significantly increasing parameter overhead, thereby improving the prediction accuracy and stability of radar echo images.

[0036] 2. This invention analyzes and models frequency domain information, and designs a frequency domain dynamic attention module. It maps radar echo images from the spatial domain to the frequency domain using Fourier transform, introduces dynamic convolution operations in the frequency domain, and adaptively selects different convolution kernels based on the input frequency domain features to achieve fine modeling and structural enhancement of different frequency components. Furthermore, the frequency domain features are divided into multiple frequency bands, and an independent attention network is used to highlight frequency bands containing key structural information, such as edge variations in high frequencies or main body contours in low frequencies. This overcomes problems such as edge blurring and detail loss, improving the stability and detail preservation capability of radar echo image prediction.

[0037] 3. This invention optimizes the loss function by designing a composite loss function that integrates intensity-weighted loss, frequency domain consistency loss, and structural similarity loss to comprehensively improve the prediction quality of radar echo images. Specifically, the frequency domain consistency loss, by aligning the predicted and ground truth images in terms of frequency domain amplitude, enhances the modeling ability of the global structure and periodic features of the image, improving the ability to reconstruct key areas such as heavy precipitation edges and echo morphology. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of the method of the present invention;

[0039] Figure 2 This is a schematic diagram of the frequency-domain enhanced U-Net network of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of the multi-scale depth convolution module of the present invention;

[0041] Figure 4 This is a schematic diagram of the frequency domain dynamic attention module of the present invention;

[0042] Figure 5 This is a schematic diagram illustrating the prediction effect in an embodiment of the present invention;

[0043] Figure 6 This is a system block diagram of the present invention. Detailed Implementation

[0044] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0045] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0046] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned radar echo extrapolation method based on frequency domain enhancement includes the following steps S1 to S4. The specific implementation process of each step will be described in detail below.

[0047] S1: Obtain a historical radar echo grayscale image sequence composed of radar echo grayscale images, convert the grayscale values ​​in the radar echo grayscale images into radar echo intensities, obtain radar echo images, and form a radar echo sequence. Use the sliding window technique on the preprocessed radar echo sequence to construct radar echo sequence samples.

[0048] It should be noted that the specific process of preprocessing the radar echo sequence and constructing radar echo sequence samples in step S1 is as follows: Radar echo images with a maximum echo intensity lower than a preset effective echo intensity threshold are removed from the radar echo sequence, and the retained radar echo images are normalized to form a preprocessed radar echo sequence. On the preprocessed radar echo sequence, a fixed-length sliding window is used for sample extraction. Each radar echo sequence sample contains input data and label data. The input data consists of N consecutive radar echo images in the preprocessed radar echo sequence, and the label data consists of N consecutive radar echo images following the corresponding input data in the preprocessed radar echo sequence.

[0049] In this embodiment, step S1 specifically includes the following sub-steps:

[0050] 1) Obtain the sequence of grayscale radar echoes detected by the Shanghai Doppler radar from October 2015 to July 2018 (i.e., the historical radar echo grayscale image sequence). Each radar echo grayscale image has a resolution of 501×501. Use bicubic interpolation to scale the radar echo grayscale images to 128×128.

[0051] 2) The processed radar echo grayscale image is converted into radar echo intensity using the formula E = I ÷ 255.0 × R, thus obtaining a radar echo image and forming a radar echo sequence. Here, E is the radar echo intensity, I is the pixel value of the processed radar echo grayscale image, and R is the maximum value of the radar echo intensity. In this embodiment, the maximum detectable radar echo intensity is 70 dBZ.

[0052] 3) When the radar echo intensity is less than the effective echo intensity threshold of 15dBZ, it usually indicates no precipitation. Therefore, 15dBZ is set as the effective echo intensity threshold, and invalid radar echo image frames with the maximum echo intensity lower than this threshold are removed from the above sequence.

[0053] 4) For the retained valid radar echo image, divide the echo intensity value of each pixel by the maximum detection intensity of the radar system (i.e., the maximum value of radar echo intensity, which is 70 dBZ in this embodiment) to normalize the radar echo image and linearly map the radar echo intensity to the [0,1] interval.

[0054] 5) Sample extraction is performed using a fixed-length sliding window with a length of 20. The interval between two adjacent radar echo images is 6 minutes. A radar echo sequence sample is formed by 10 consecutive normalized radar echo images, which are then used as input data for the model. The subsequent 10 normalized radar echo images are used as predicted target labels, with a coverage time of 2 hours.

[0055] 6) Divide the processed data into training set, validation set and test set in a ratio of 8:1:1. The training set has 63,072 samples, the validation set has 7,884 samples, and the test set has 7,884 samples, which will be used for subsequent model training.

[0056] S2: Two improvements are made to the U-Net network to obtain a frequency-domain enhanced U-Net network with an encoder-decoder structure. The first improvement is to replace the convolutional layers in both the encoder and decoder of the U-Net network with multi-scale deep convolutional modules. The second improvement is to embed a frequency-domain dynamic attention module at each skip connection of the U-Net network to enhance the features of the encoder output.

[0057] It should be noted that this invention uses the U-Net architecture as the basic model, which consists of an encoder and a decoder. In this embodiment, the input is a multi-channel radar echo image sequence with a size of 128×128×10, where the spatial resolution of each radar echo image is 128×128 and the channel dimension is 10, representing radar echo images at 10 consecutive time steps. First, the input image is passed through an initial convolutional layer to increase the number of input channels from 10 to 64, and the output feature size is 128×128×64. Subsequently, the model enters the encoder part, which contains a four-level downsampling module. Each downsampling module consists of a convolutional layer and a pooling layer. Through layer-by-layer downsampling, the spatial resolution is halved at each level, and the number of channels is doubled at each level. The first-level downsampling module outputs a size of 64×64×128, the second-level downsampling module outputs 32×32×256, the third-level downsampling module outputs 16×16×512, and the fourth-level downsampling module outputs 8×8×1024. The encoder output enters the bottleneck layer, which continues to extract deep semantic features through convolutional operations, maintaining the feature dimension at 8×8×1024. Next, the model enters the decoder section. Each upsampling module first doubles the spatial size through transposed convolution, and then concatenates the upsampled feature map with the corresponding encoder stage feature map along the channel dimension, thereby fusing shallow details and deep semantic information. The concatenated feature map further fuses information through convolutional layers, with the channel dimension decreasing layer by layer. The first-level upsampling module has an output size of 16×16×512, the second-level upsampling module has an output size of 32×32×256, the third-level upsampling module has an output size of 64×64×128, and the fourth-level upsampling module has an output size of 128×128×64. Finally, the output layer uses a 1×1 convolution to map the number of channels from 64 to the target number of channels, resulting in a prediction result with a size of 128×128×10.

[0058] It should be noted that, as Figure 2As shown, this invention designs a multi-scale deep convolutional module (MDConv Block) and a frequency domain dynamic attention module (FreqDyAtt Block), and introduces these two modules to improve the U-Net architecture, thereby constructing a frequency domain enhanced U-Net network. The frequency domain enhanced U-Net network adopts a standard encoder-decoder structure. The encoder consists of four downsampling modules, each containing a pooling layer and a multi-scale deep convolutional module; the decoder consists of four upsampling modules, each containing a transposed convolutional layer and a multi-scale deep convolutional module. Resolution is restored through transposed convolution, and the upsampling module is concatenated with the feature maps of the corresponding layers in the encoder via skip connections to achieve feature fusion. A frequency domain dynamic attention module is embedded at each skip connection. Specifically, the multi-scale deep convolutional module is located during the convolution process of the encoder and decoder, and the frequency domain dynamic attention module is deployed after the multi-scale deep convolutional module of each downsampling module, concatenated to the input of the corresponding layer in the decoder via skip connections to achieve frequency domain feature fusion.

[0059] To facilitate understanding by those skilled in the art, the specific processing flow of the multi-scale depth convolution module and the frequency domain dynamic attention module designed in this invention will be described in detail below.

[0060] It should be noted that in this invention, the multi-scale deep convolution module adopts a two-stage processing architecture: The first stage first performs layer normalization on the input features, and then expands the normalized features through a 1×1 convolution channel before feeding them into the parallel first, second, and third deep convolution layers. The features output by the three deep convolution layers are concatenated, and the concatenated features are sequentially compressed through a 1×1 convolution channel, activated by an activation function, adaptively weighted by a channel attention module, and restored through a 1×1 convolution channel to obtain the first feature. The first feature is then residually fused with the input features to form the first-stage output feature. The second stage performs layer normalization on the first-stage output feature again before feeding it into a feedforward neural network to obtain the second-stage output feature. Finally, the second-stage output feature and the first-stage output feature are residually concatenated to output the processing result of the multi-scale deep convolution module.

[0061] Furthermore, the kernel size of the first deep convolutional layer is 3×3, the kernel size of the second deep convolutional layer is 5×5, and the kernel size of the third deep convolutional layer is 7×7.

[0062] Furthermore, the feedforward neural network is composed of cascaded 1×1 convolutional layers, GELU activation functions, and 1×1 convolutional layers.

[0063] In this embodiment, the specific processing flow in the multi-scale depthwise convolution module is as follows: Figure 3As shown, the multi-scale deep convolution module integrates multiple design concepts such as multi-scale deep convolution, channel attention, residual connections, and feedforward networks, aiming to significantly enhance its feature representation capabilities while maintaining the model's lightweight nature. This module consists of two stages: multi-scale feature modeling and feedforward feature enhancement. The specific steps are as follows:

[0064] In the first stage, input features First, the feature distribution is stabilized through layer normalization (LayerNorm2d), and then a 1×1 convolution is used for channel expansion to improve the feature representation capability.

[0065]

[0066] in, Representation layer normalization; This represents a 1×1 convolution used for channel expansion; This represents the features after being expanded by a 1×1 convolution channel.

[0067] Subsequently, multi-scale deep convolution processing is performed, consisting of three parallel deep convolution paths: 3×3, 5×5, and 7×7, each corresponding to a different receptive field size. This captures precipitation structure features at different scales in radar echo images, thereby obtaining local and global features. The feature maps output from the three convolution paths are concatenated along the channel dimension and then compressed and fused using a 1×1 convolution.

[0068]

[0069] in, , as well as This represents the 3×3, 5×5, and 7×7 depth convolutional layers mentioned above; Indicates feature concatenation operation; This represents a 1×1 convolution used for channel compression; This represents the multi-scale features after compression of a 1×1 convolution channel.

[0070] Multiscale features After processing by the activation function, the signal is fed into the channel attention module, which includes global average pooling, 1×1 convolution, and sigmoid activation to generate dynamic weight vectors. The above multi-scale features are dynamically weighted to automatically enhance the response to key features such as strong echoes:

[0071]

[0072] in, This indicates point-by-point multiplication.

[0073] Features enhanced by the channel attention module Then, a 1×1 convolution is used to restore the number of channels back to the original dimensions, while setting the first learnable scaling factor. Control the residual strength and fuse it with the input features in a weighted manner:

[0074]

[0075] in, This represents a 1×1 convolution used for channel restoration; This represents the output features of the first stage.

[0076] In the second stage, the output features from the first stage are first normalized again, and then fed into a feedforward neural network consisting of two 1×1 convolutions. The first 1×1 convolution is responsible for increasing the channel dimension to expand the feature subspace, and then introducing a non-linear expression through the GELU activation function. The second 1×1 convolution restores the features back to the original number of channels.

[0077]

[0078] in, This represents the first 1×1 convolution in the feedforward neural network; This represents the second 1×1 convolution in the feedforward neural network; This represents the GELU activation function described above; This represents the characteristics of the feedforward neural network output, i.e., the output characteristics of the second stage.

[0079] Finally, set a second learnable scaling factor. The output of the feedforward neural network is residually fused with the output features of the first stage to form a dual-path residual structure, thereby obtaining the processing results of the multi-scale deep convolution module. :

[0080]

[0081] It should be noted that in this invention, the frequency domain dynamic attention module takes the features output by the multi-scale deep convolution module as input features. First, it performs a two-dimensional fast Fourier transform on the input features, thereby transforming the input features from the spatial domain to the frequency domain, generating shallow frequency domain features and separating the real and imaginary parts. The decomposed real and imaginary parts are concatenated along the channel dimension to form an enhanced feature representation containing amplitude and phase information. The enhanced feature representation is then subjected to a 3×3 deep convolution to form deep frequency domain features, which are then fed into the frequency dynamic convolution module to generate frequency domain features after frequency dynamic modeling. An inverse Fourier transform is then applied to these features to return them to the spatial domain, resulting in frequency domain enhanced spatial domain features. These are then fed into the multi-frequency attention module for processing, forming fused spatial domain features. Finally, the frequency domain enhanced spatial domain features and the fused spatial domain features are residually connected to obtain the output features of the frequency domain dynamic attention module.

[0082] Furthermore, the specific processing steps in the frequency dynamic convolution module are as follows: K depth convolution kernels are set. The deep frequency domain features are first processed by global average pooling, 1×1 convolution, and softmax function to generate the weights of each depth convolution kernel. Then, the deep frequency domain features are fed into each depth convolution kernel for convolution operation. The outputs of each depth convolution kernel are weighted and summed according to their corresponding weights to obtain the frequency domain features after frequency dynamic modeling.

[0083] Furthermore, the specific processing steps in the multi-frequency attention module are as follows: The spatial domain features enhanced in the frequency domain are first obtained by two-dimensional fast Fourier transform to obtain a frequency domain tensor and regarded as a two-dimensional spectrogram. The two-dimensional spectrogram is divided into frequency bands according to the frequency level using a ring mask rule based on polar coordinate radius to extract the complex spectral components of each frequency band. Then, the complex spectral components of each frequency band are restored to the spatial domain by inverse Fourier transform to obtain the corresponding spatial domain feature map. Each spatial domain feature map is modeled by a channel attention module. Then, the weighted feature maps generated by all channel attention modules are concatenated in the channel dimension and compressed and fused by the fusion convolution module to generate the fused spatial domain features.

[0084] Furthermore, the fusion convolution module is composed of a cascaded layer of 3×3 convolution, a ReLU activation function, and a 1×1 convolution.

[0085] In this embodiment, as Figure 4 As shown, the frequency domain dynamic attention module, with frequency domain transformation as its core, combines frequency domain dynamic convolution and multi-frequency attention mechanisms to enhance the model's ability to perceive periodic and structural features in echo images, thereby improving the spatial and temporal consistency of precipitation evolution prediction. The specific processing flow in this module is as follows:

[0086] First, the features output by the multi-scale deep convolution module are mapped from the spatial domain to the frequency domain using a two-dimensional fast Fourier transform. Then, the complex signal (the aforementioned shallow frequency domain features) is decomposed into real and imaginary parts and concatenated to form an enhanced feature representation containing amplitude and phase information.

[0087]

[0088] in, The above-mentioned enhanced features are represented in two-dimensional coordinates. The value at; This represents the features output by the multi-scale deep convolution module in two-dimensional coordinates. The value at; and These represent the height and width of the output features from the multi-scale depth convolution module, respectively. Represents the imaginary unit; This represents the e-index.

[0089] Secondly, the enhanced feature representation formed by splicing is passed through a 3×3 depthwise convolution module to enhance local sensitivity in the frequency domain and preserve the spatial positional relationship between high-frequency and low-frequency components. The features output from the 3×3 depthwise convolution then enter a frequency dynamic convolution module. In this module, this embodiment sets K 3×3 depthwise convolution kernels and an adaptive attention weight generation module. The adaptive attention weight generation module generates the weights of each convolution kernel through global average pooling and 1×1 convolution, and normalizes them in the branch dimension using the softmax function to ensure that the sum of all convolution kernel weights is 1. The features output from the 3×3 depthwise convolution are fed into each depthwise convolution kernel for convolution operations, and the outputs of each branch are weighted and summed according to their corresponding weights. After frequency dynamic modeling, the frequency domain features generated by the frequency dynamic convolution module are decomposed back into real and imaginary parts, reassembled into a complex tensor, and then returned to the spatial domain through an inverse Fourier transform.

[0090]

[0091] Next, the enhanced spatial domain features are fed into the multi-frequency attention module for further processing. First, a frequency domain tensor is obtained through a two-dimensional fast Fourier transform. This frequency domain tensor is treated as a two-dimensional spectrum, and a ring masking rule based on polar coordinate radii is used for frequency band division. This ring masking rule is existing technology; it is based on a polar coordinate system and is used to divide the frequency domain into concentric rings (ring masks), each ring corresponding to a frequency band. This embodiment briefly describes it. Specifically, the frequency band division operation involves calculating the Euclidean distance (i.e., the radius in polar coordinates) from the center of the two-dimensional spectrum to the origin, using the center of the two-dimensional spectrum as the origin. The maximum radius is the maximum distance from the center of the two-dimensional spectrogram to its boundary. The two-dimensional spectrum is divided into four frequency bands based on different radius intervals, each corresponding to a different frequency range. and Two-dimensional spectrum diagrams The height and width. Each frequency band corresponds to a mask, and each mask is a binary matrix of the same size as the two-dimensional spectrogram, where the values ​​are 1 within the corresponding radius interval and 0 elsewhere. The first mask... Corresponding to the low frequency band, Second mask Corresponding to the mid-low frequency band, Third mask Corresponding to the mid-to-high frequency band, Fourth mask Corresponding to high frequency bands, Finally, through bitwise multiplication ( Extract the complex spectral components of each frequency band:

[0092]

[0093] in, Indicates the first Complex spectral components of each frequency band.

[0094] Then, channel attention weights are applied to each frequency band using a channel attention module. The implementation of the channel attention module is existing technology and will be described in this embodiment. In this module, adaptive average pooling is used to obtain the global description vector for each channel. This vector is then input into a two-layer multilayer perceptron structure for nonlinear transformation. The first layer compresses the number of channels to 1 / 16 of the original number using a 1×1 convolution. After ReLU activation, the second layer uses a 1×1 convolution to restore the original number of channels, and a Sigmoid activation function is used to generate channel attention weights. The spatial feature map corresponding to each frequency band is multiplied point-by-point along the channel dimension with its corresponding attention weight to form a weighted feature map (i.e., the output of each channel attention module).

[0095] Finally, the weighted feature maps generated by all channel attention modules are concatenated along the channel dimension and compressed and fused through a two-layer fusion convolutional module. The first 3×3 convolution of the fusion convolutional module expands the number of channels by a factor of 2, and the second 1×1 convolution compresses the number of channels back to the original number, with a ReLU activation function inserted in between. The fusion convolutional module generates fused spatial domain features. To preserve the correlation between the original information and the enhanced features, a residual connection mechanism is introduced, adding the fused spatial domain features element-wise to the frequency-enhanced spatial domain features to obtain the final output features.

[0096] S3: Train the frequency domain augmented U-Net network on radar echo sequence samples. During training, a composite loss function including intensity weighted loss, frequency domain consistency loss and structural similarity loss is used. The parameters of the frequency domain augmented U-Net network are updated based on minimizing the composite loss function.

[0097] It should be noted that in step S3 of this invention, the composite loss function used is a weighted sum of intensity-weighted loss, frequency domain consistency loss, and structural similarity loss: the intensity-weighted loss calculates the sum of the weighted mean square error and the mean absolute error between the predicted image and the real image, and the weights used in the weighting are proportional to the radar echo intensity; the frequency domain consistency loss is the mean absolute error of the frequency domain amplitude calculated after performing Fourier transform on the predicted image and the real image respectively; the structural similarity loss constrains the similarity of the local structures of the predicted image and the real image through the SSIM exponent; wherein, the predicted image is the radar echo map output by the frequency domain enhanced U-Net network predicting the input data, and the real image is the label data corresponding to the input data.

[0098] It should be noted that in step S3 of this invention, during the training of the frequency domain augmented U-Net network, an initial learning rate and a learning rate decay strategy are set. Radar echo sequence samples are input into the frequency domain augmented U-Net network in batches for forward propagation. The weights of the frequency domain augmented U-Net network are optimized by calculating the composite loss function and performing backpropagation using the Adam optimizer. A validation set monitoring and early stopping strategy are adopted. After iterative training, the parameters of the frequency domain augmented U-Net network with the best performance on the validation set are finally saved to obtain the trained frequency domain augmented U-Net network.

[0099] In step S3 of this embodiment, specifically, based on the characteristics of the radar echo image prediction task, a composite loss function of the following form is adopted. It is composed of intensity-weighted loss Frequency domain consistency loss and structural similarity loss The weighted sum is obtained by:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] in, Weights for loss items; This indicates the number of predicted radar echo maps output. and These represent the height and width of the radar echo image, respectively. Representing radar echo images in two-dimensional coordinates The weight of each pixel; and Representing the real image and the predicted image in two-dimensional coordinates respectively. Pixel value at; Representing radar echo images in two-dimensional coordinates Radar echo intensity at the location; and Represent the two-dimensional coordinates of the real image and the predicted image in the frequency domain, respectively. The amplitude value at that point; For real images and predicted images Structural similarity index between them; and Represents two non-zero constants; and These are the mean and variance of the real image, respectively; and These are the mean and variance of the predicted image, respectively; This represents the covariance between the real image and the predicted image.

[0107] In step S3 of this embodiment, specifically, radar echo sequence samples are input as training data into the frequency domain enhanced U-Net network, and the Adam optimizer is used for parameter optimization. The initial learning rate is set to 0.001. When the loss function value does not decrease within 5 consecutive epochs, the learning rate decay mechanism is triggered, and the current learning rate is automatically multiplied by a decay factor, which is set to 0.1. During training, the loss function value of the validation set is continuously monitored. When the loss function value on the validation set does not show a significant improvement within 10 consecutive epochs, the training process will be terminated early to avoid invalid training and overfitting. Finally, the optimal model parameters on the validation set are saved.

[0108] S4: Input the radar echo sequence to be predicted into the trained frequency domain augmented U-Net network, output the radar echo prediction results for multiple future time steps, and finally visualize the radar echo prediction results to complete the entire extrapolation process.

[0109] In step S4 of this embodiment, the test set obtained in step S1 (6) is input into the trained frequency-domain augmented U-Net network to obtain 10 radar echo images for the next hour. Then, based on the reflectivity intensity range, the radar echo grayscale image is converted into a pseudo-color image through color mark mapping. The final partial radar image effect is as follows. Figure 5 As shown.

[0110] Furthermore, to quantify the prediction results, this embodiment uses CSI (Critical Success Index), POD (Point of Detection), FAR (False Alarm Rate), and HSS (Heidke Skill Score) for evaluation. CSI, POD, FAR, and HSS are calculated at thresholds of 20 dBZ, 30 dBZ, and 40 dBZ, respectively, and the results are shown in Table 1. The calculated confusion matrix results are shown in Table 2.

[0111] Among the above indicators, CSI measures the accuracy of predicting rainfall; a higher value indicates a more accurate prediction of the rainfall area. POD measures the ability to capture actual rainfall areas; a higher value indicates fewer missed reports. FAR represents the proportion of incorrectly reported rainfall in the forecast; a lower value indicates a more robust model. HSS evaluates the improvement of the model compared to random predictions, taking into account hits, missed reports, and false alarms; a value closer to 1 indicates better performance. The specific calculation methods for each of the above evaluation indicators are as follows:

[0112]

[0113]

[0114]

[0115]

[0116] TP indicates that the predicted intensity is greater than the threshold and the actual intensity is greater than the threshold; FP indicates that the predicted intensity is greater than the threshold but the actual intensity is less than the threshold; FN indicates that the predicted intensity is greater than the threshold but the actual intensity is less than the threshold; TN indicates that the predicted intensity is less than the threshold and the actual intensity is less than the threshold.

[0117] Table 1. Evaluation Indicators

[0118] Table 2. Confusion Matrix

[0119] In addition, to verify the effectiveness of the proposed multi-scale convolution module (MDConv) and frequency domain dynamic attention module (FreqDyAtt) in radar echo extrapolation, a series of ablation experiments were designed to analyze the changes in model performance as the aforementioned key modules were gradually introduced.

[0120] The specific configuration includes the following four model structures: U-Net model: containing only the basic U-Net structure without introducing any enhancement modules; U-Net+MDConv: adding a multi-scale convolution module on the basis of U-Net; U-Net+FreqDyAtt: adding a frequency domain dynamic attention module on the basis of U-Net; This invention: a complete model that includes both MDConv and FreqDyAtt modules.

[0121] The experiment used the same dataset, optimizer configuration and loss function combination. The evaluation metrics included step CSI, POD, FAR and HSS metrics. The thresholds were set to 20dBz, 30dBz and 40dBz. The results are shown in Table 3.

[0122] Table 3. Ablation Experiment

[0123] Experimental results show that introducing the multi-scale convolution module and the frequency domain dynamic attention module respectively significantly improves the CSI, POD, FAR and HSS metrics compared to the traditional U-Net model, proving the effectiveness of the multi-scale convolution module and the frequency domain dynamic attention module. Furthermore, when used simultaneously, the model performs best in CSI, FAR and HSS metrics at all thresholds, indicating that the two modules have complementary advantages in spatial and frequency domain feature enhancement.

[0124] It should also be noted that the frequency-domain enhanced radar echo extrapolation method in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a frequency-domain enhanced radar echo extrapolation system corresponding to the frequency-domain enhanced radar echo extrapolation method provided in the above embodiments, such as... Figure 6 As shown, it includes:

[0125] The data acquisition module is used to acquire the radar echo sequence to be predicted;

[0126] The prediction module is used to predict the radar echo sequence to be predicted according to the frequency domain enhancement-based radar echo extrapolation method described in the above embodiments, obtain the radar echo prediction results for multiple future time steps, and display them visually.

[0127] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0128] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A radar echo extrapolation method based on frequency domain enhancement, characterized in that, Includes the following steps: S1: Obtain a sequence of historical radar echo grayscale images composed of radar echo grayscale images, convert the grayscale values ​​in the radar echo grayscale images into radar echo intensities, obtain radar echo images, and form a radar echo sequence. Use the sliding window technique on the preprocessed radar echo sequence to construct radar echo sequence samples. S2: Two improvements are made to the U-Net network to obtain a frequency-domain enhanced U-Net network with an encoder-decoder structure. The first improvement is to replace the convolutional layers in the encoder and decoder of the U-Net network with multi-scale deep convolutional modules. The second improvement is to embed a frequency-domain dynamic attention module at each skip connection of the U-Net network to enhance the features of the encoder output. S3: Train the frequency domain augmented U-Net network on radar echo sequence samples. During training, a composite loss function including intensity weighted loss, frequency domain consistency loss and structural similarity loss is used. The parameters of the frequency domain augmented U-Net network are updated based on minimizing the composite loss function. S4: Input the radar echo sequence to be predicted into the trained frequency domain augmented U-Net network, output the radar echo prediction results for multiple future time steps, and finally visualize the radar echo prediction results to complete the entire extrapolation process.

2. The radar echo extrapolation method based on frequency domain enhancement as described in claim 1, characterized in that, In step S1, the specific process of preprocessing the radar echo sequence and constructing radar echo sequence samples is as follows: Radar echo images with a maximum echo intensity lower than a preset effective echo intensity threshold are removed from the radar echo sequence, and the retained radar echo images are normalized to form a preprocessed radar echo sequence. On the preprocessed radar echo sequence, a fixed-length sliding window is used for sample extraction. Each radar echo sequence sample contains input data and label data. The input data consists of N consecutive radar echo images in the preprocessed radar echo sequence, and the label data consists of N consecutive radar echo images following the corresponding input data in the preprocessed radar echo sequence.

3. The radar echo extrapolation method based on frequency domain enhancement as described in claim 1, characterized in that, In step S2, the multi-scale deep convolution module adopts a two-stage processing architecture: the first stage first performs layer normalization on the input features, expands the layer normalized features through 1×1 convolution channels and sends them into the parallel first, second and third deep convolution layers, concatenates the features output by the three deep convolution layers, and then passes the concatenated features through 1×1 convolution channel compression, activation function, adaptive weighting by the channel attention module, and 1×1 convolution channel restoration to obtain the first feature. The first feature is then residually fused with the input features to form the output feature of the first stage. The second stage normalizes the output features from the first stage again and feeds them into the feedforward neural network to obtain the output features of the second stage. Finally, the output features of the second stage and the output features of the first stage are connected by residuals to output the processing result of the multi-scale deep convolution module.

4. The radar echo extrapolation method based on frequency domain enhancement as described in claim 3, characterized in that, The kernel size of the first deep convolutional layer is 3×3, the kernel size of the second deep convolutional layer is 5×5, and the kernel size of the third deep convolutional layer is 7×7; the feedforward neural network is composed of cascaded 1×1 convolutional layers, GELU activation function and 1×1 convolutional layers.

5. The radar echo extrapolation method based on frequency domain enhancement as described in claim 1, characterized in that, In step S2, the frequency domain dynamic attention module takes the features output by the multi-scale deep convolution module as input features. First, it performs a two-dimensional fast Fourier transform on the input features to transform the input features from the spatial domain to the frequency domain, generating shallow frequency domain features and separating the real and imaginary parts. The real and imaginary parts obtained by decomposition are concatenated along the channel dimension to form an enhanced feature representation containing amplitude and phase information. The enhanced feature representation is processed by a 3×3 depth convolution to form a deep frequency domain feature, which is then fed into the frequency dynamic convolution module to generate frequency domain features after frequency dynamic modeling. An inverse Fourier transform is then applied to these features to return them to the spatial domain, resulting in frequency domain enhanced spatial domain features. These enhanced spatial domain features are then fed into the multi-frequency attention module for processing, forming fused spatial domain features. Finally, the frequency domain enhanced spatial domain features and the fused spatial domain features are joined by a residual connection to obtain the output features of the frequency domain dynamic attention module.

6. The radar echo extrapolation method based on frequency domain enhancement as described in claim 5, characterized in that, The specific processing steps in the frequency dynamic convolution module are as follows: K depth convolution kernels are set. The deep frequency domain features are first processed by global average pooling, 1×1 convolution and softmax function to generate the weights of each depth convolution kernel. Then, the deep frequency domain features are fed into each depth convolution kernel for convolution operation. The outputs of each depth convolution kernel are weighted and summed according to their corresponding weights to obtain the frequency domain features after frequency dynamic modeling.

7. The radar echo extrapolation method based on frequency domain enhancement as described in claim 5, characterized in that, The specific processing steps in the multi-frequency attention module are as follows: The spatial domain features enhanced in the frequency domain are first obtained by two-dimensional fast Fourier transform to obtain a frequency domain tensor and regarded as a two-dimensional spectrogram. The two-dimensional spectrogram is divided into frequency bands according to the frequency level using a ring mask rule based on polar coordinate radius to extract the complex spectral components of each frequency band. Then, the complex spectral components of each frequency band are restored to the spatial domain by inverse Fourier transform to obtain the corresponding spatial domain feature map. Each spatial domain feature map is modeled by a channel attention module. Then, the weighted feature maps generated by all channel attention modules are concatenated in the channel dimension and compressed and fused by the fusion convolution module to generate the fused spatial domain features. The fusion convolution module is composed of a cascaded 3×3 convolution, a ReLU activation function and a 1×1 convolution.

8. The radar echo extrapolation method based on frequency domain enhancement as described in claim 1, characterized in that, In step S3, the composite loss function used is a weighted sum of intensity-weighted loss, frequency domain consistency loss, and structural similarity loss: the intensity-weighted loss calculates the sum of the weighted mean square error and the mean absolute error between the predicted image and the real image, and the weights used in the weighting are proportional to the radar echo intensity; the frequency domain consistency loss is the mean absolute error of the frequency domain amplitude calculated after performing Fourier transforms on the predicted image and the real image respectively; the structural similarity loss constrains the similarity of the local structures of the predicted image and the real image through the SSIM exponent; where the predicted image is the radar echo map output by the frequency domain enhanced U-Net network predicting the input data, and the real image is the label data corresponding to the input data.

9. The radar echo extrapolation method based on frequency domain enhancement as described in claim 1, characterized in that, In step S3, during the training of the frequency domain augmented U-Net network, an initial learning rate and a learning rate decay strategy are set. Radar echo sequence samples are input into the frequency domain augmented U-Net network in batches for forward propagation. The weights of the frequency domain augmented U-Net network are optimized by calculating the composite loss function and performing backpropagation using the Adam optimizer. A validation set monitoring and early stopping strategy are adopted. After iterative training, the parameters of the frequency domain augmented U-Net network with the best performance on the validation set are finally saved to obtain the trained frequency domain augmented U-Net network.

10. A radar echo extrapolation system based on frequency domain enhancement, characterized in that, include: The data acquisition module is used to acquire the radar echo sequence to be predicted; The prediction module is used to predict the radar echo sequence to be predicted according to the frequency domain enhancement-based radar echo extrapolation method according to any one of claims 1 to 9, obtain the radar echo prediction results for multiple future time steps, and display them visually.

Images