Radar approaching rainfall prediction method based on frequency domain perception and conditional diffusion

Abstract

The invention discloses a radar approaching rainfall prediction method based on frequency domain perception and conditional diffusion, and belongs to the technical field of image data processing, and the method comprises the steps: obtaining a radar echo data set, and pre-training a variational auto-encoder MVAE; the method comprises the following steps: constructing a convolution enhanced frequency domain Transform; constructing an improved denoising network based on a multi-layer convolution enhanced frequency domain Transform, generating an improved conditional diffusion model based on MVAE, and training to obtain an approaching rainfall prediction model; the method is used for approaching rainfall prediction. Through frequency domain-spatial domain collaborative modeling and efficient conditional diffusion architecture design, the method not only overcomes the core defects of prediction fuzziness and the like caused by unsmooth local and global feature fusion and signal-noise confusion in the prior art, but also enhances the adaptation capability to a complex meteorological scene, and improves the prediction accuracy. Therefore, high-frequency details and edge features of the convection system can be more accurately captured, the prediction precision of a weak echo region is remarkably improved, and the accuracy of short and temporary rainfall prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, and particularly relates to a radar near precipitation prediction method based on frequency domain perception and conditional diffusion. BACKGROUND

[0002] Precipitation nowcasting using radar echo extrapolation, i.e. predicting high-resolution rainfall intensity in a short time (such as 0-2 hours), plays a key role in modern meteorological services and provides indispensable guidance to alleviate the social and economic impact of severe weather phenomena. Current radar echo prediction techniques mainly include traditional numerical model methods and deep learning methods.

[0003] In the traditional method: numerical weather prediction (NWP) is based on fluid mechanics equations modeling, and predicts precipitation probability and intensity by assimilating observation data, but it has high computational complexity, relies on supercomputing resources and is sensitive to initial conditions in short-time and small-scale prediction, making it difficult to capture the dynamic evolution of small-scale systems. The optical flow method and Lagrangian extrapolation method extrapolate echo positions through the motion vector of adjacent radar frames, such as the PySTEPS algorithm combined with a statistical model to generate ensemble predictions, but this method assumes that echo motion is stable and cannot model the birth and death of strong convective weather, and the error accumulates significantly over time, and the prediction accuracy of weak echo areas decreases sharply.

[0004] In the deep learning method: Models based on recurrent structures such as ConvLSTM and TrajGRU have long sequence prediction degradation problems, and face gradient vanishing in long sequence prediction, especially in strong convective processes, where error accumulation leads to a sharp drop in prediction accuracy, and the ability to capture details of weak echoes and strong convective boundaries is limited.

[0005] Existing diffusion models (such as SRNDiff) can generate diverse samples, but have local texture modeling and efficiency imbalance problems, and the modeling of local textures such as weak echo boundaries is ambiguous and computationally expensive, making it difficult to balance efficiency and accuracy, and when diffusion models are combined with CNNs, the limited receptive field of convolutional operations makes it difficult to effectively capture long-range spatial dependencies, limiting the effectiveness and accuracy of nearcasting.

[0006] The visual Transformer (ViT) also has obvious defects in application. Its patch-based tokenization strategy divides the continuous radar echo map into non-overlapping discrete patches, essentially reducing the fine-grained spatial continuity between pixels, resulting in a large loss of fine texture and edge information, and weakening the model's ability to model local correlation. At the same time, in the iterative denoising process, the high-frequency signals that constitute the real weather structure (such as the clear edges of convective cells) are difficult to distinguish from the additive Gaussian noise used in the forward process, forming a signal-noise confusion phenomenon, prompting the denoising network to mistakenly suppress real physical details as noise, resulting in a decrease in structure fidelity.

[0007] In addition, traditional hybrid architectures such as CNN+ViT have defects in feature fusion, treating the two as independent modules with a single fusion strategy and relying on additional fusion blocks to assist integration, resulting in blocked information flow between local details and global dependencies, unable to dynamically coordinate interactions, and thus causing boundary blurring and context fragmentation problems. The above defects collectively cause existing models (whether based on CNN or RNN) to often produce overly smooth and blurred prediction results, underperforming in capturing the inherent high-frequency details of convective weather systems (such as the sharp gradients of storm core intensity and rain bands), ultimately reducing the reliability of predictions in practical applications.

[0008] The pixel values of the radar echo map essentially correspond to the reflectivity factor (Z), with units of dBZ (decibels per Z). The larger the value, the stronger the reflected electromagnetic wave energy, indicating that there are more and larger precipitation particles (raindrops, snowflakes, hail, etc.) in the atmosphere. Therefore, the pixel values in the radar echo map can represent the precipitation at the corresponding location.

[0009] Glossary: Space-time attention: As a key technology in the field of deep learning, space-time attention mechanism significantly improves the processing ability of time series data and space-time modeling by capturing the dependence of data in time and space dimensions. Including separated space-time attention, joint space-time attention, and space-time continuous attention, the "Divided Space-Time Attention" layer used in TimeSformer is a typical representative of separated space-time attention, which adopts a cascaded structure of space first and then time to achieve efficient modeling. In specific implementation, first, the attention calculation in the spatial dimension is performed independently for each time frame, and the multi-head attention mechanism is used to establish the spatial correlation between pixels, focusing on capturing the local details and global structural features of single-frame images. Subsequently, based on the enhancement of spatial features, cross-frame attention correlation is established along the time axis, effectively capturing the motion evolution law in the video or the depth correlation of the volume data. This space-time separation design reduces the computational complexity from O(T²H²W²) to O(T²+H²W²), significantly improving the computational efficiency while maintaining the feature expression ability.

[0010] Dual-channel space-time channel attention: It is a three-dimensional extension structure based on the Squeeze-and-Excitation (SE) mechanism, which realizes the cooperative feature enhancement of channels and space-time dimensions through a parallel dual-branch architecture. In the channel attention branch, the input features are first compressed by global average pooling in the space-time dimension to generate channel descriptors, then transformed through a bottleneck structure containing two linear layers and GELU activation, and finally output the channel attention weight matrix F a . The space-time attention branch uses 3D convolution operation to model the input features in space-time, and realizes feature compression through single-channel convolution kernel, and generates the space-time attention weight matrix F b . The feature enhancement stage realizes double attention modulation through the element-wise product of F a and F b , and fuses with the original features in the form of a learnable scaling coefficient residual connection, while preserving the basic features and realizing the adaptive enhancement of key space-time channel features. Through the cooperative action of channel re-labeling and space-time position selection, this structure significantly improves the representation ability of three-dimensional features. Summary of the invention

[0011] The purpose of the present application is to provide a radar near precipitation prediction method based on frequency domain perception and conditional diffusion, which can more accurately capture the high-frequency details and edge features of the convection system, significantly improve the prediction accuracy of weak echo area, and maintain stable prediction performance in strong convection scenes such as heavy rain and thunderstorm.

[0012] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion, comprising the following steps; S1, obtain radar echo data set D1. The samples in D1 are time-sequential radar echo images. Any S adjacent samples constitute an image sequence. The pixel value in the sample represents the precipitation at the corresponding location. S2, pre-trained variational autoencoder M VAE , preset shape one and shape two; The variational autoencoder includes a VAE encoder and a VAE decoder, wherein the VAE encoder is used to transform a batch of image sequences X IS Encoding from pixel space to latent space to obtain the latent feature sequence Z IS , the VAE decoder is used to transform Z IS Decode back to pixel space to get the predicted image sequence , where Z IS The shape is [B, S, Cz, Hz, Wz], where B is the batch size, Cz, Hz, and Wz are the number of channels, width, and height of the image in the latent space, respectively. IS Each image can be divided into N×N image blocks by the image block layer and form a feature sequence of length L=N×N; Preset shape 1 [B, S, L, Cp] and shape 2 [B, S, Cp, Hp, Wp], where Cp, Hp, and Wp are the number of feature channels of shape 2 and the height and width of the image respectively; S3, constructs a convolution-enhanced frequency-domain Transformer, which includes a noise-aware frequency-domain gating module, a layer normalization layer, a spatiotemporal attention layer, a local feature aggregation module, and an MLP. The noise-aware frequency domain gating module is used to obtain the input features of shape 1 , adjusted to shape 2, sequentially undergoing DCT transformation, frequency domain selection enhancement and IDCT transformation, and then adjusted back to shape 1 to obtain the output characteristics of the noise-aware frequency domain gating module ; The layer normalization layer and the spatiotemporal attention layer are sequentially Perform layer normalization and spatiotemporal attention operations to obtain attention features X att ,and and X att After skip connection, it is used as the input feature of the local feature aggregation module , The shape is shape one; The local feature aggregation module is used to Transformed to shape 2, after multi-scale feature extraction and fusion, dual-path spatiotemporal channel attention operation, and then transformed back to shape 1, the output features of the local feature aggregation module are obtained ; The MLP is used for input and The output feature of the convolution-enhanced frequency domain Transformer after the skip connection ; S4, based on M VAE And an improved conditional diffusion model is constructed by improving the denoising network, wherein the improved denoising network is obtained by connecting M layers of convolution-enhanced frequency domain Transformers in a Unet structure, M≥13 and is an odd number; S5, training the improved conditional diffusion model to convergence to obtain a nearby precipitation prediction model, which is used to predict the radar echo sequence of the second time period according to the radar echo sequence of the first time period, and the interval between the endpoints of the two time periods is less than 2 hours; S6, obtaining the conditional radar echo sequence for prediction , outputting the predicted noise through the precipitation prediction model, and obtaining the predicted target radar echo sequence based on the predicted noise.

[0013] As preferred, the noise-aware frequency domain gating module is configured to obtain The method is as follows: Sa1, adjusting the shape of to X1 according to the following formula; , In the formula, Reshape(·) is a Reshape function, is shape one, and X1 is shape two; Sa2, generating a DCT transformation matrix D of the image sequence composed of latent space images in length, height and width according to the following formula S , Hp , Wp ; , In the formula, is a dimension index, , D S ∈R S×S , Hp ∈R Hp×Hp , Wp ∈R Wp×Wp , is the value of the ith row and jth column, k is the frequency domain index in the row direction, i is the spatial position index of D in the column direction, 0≤k,i≤τ-1, and orthogonal to its transpose matrix ; Sa3, DCT transform, frequency domain selection enhancement and IDCT transform are performed according to the following formula to obtain the frequency domain feature X dct , enhanced feature X scaled and IDCT feature X idct ; , , , In the formula, , , respectively, are matrix multiplications in the 3rd, 4th and 5th dimensions, σ (scaled) is a channel-level gating weight generated by a Sigmoid function on a learnable weight parameter scaled, and T is a transpose operation; Sa4, adjusts X idct to shape one to obtain .

[0014] As a preferred, the local feature aggregation module comprises a first deformation layer, a multi-scale depth separable convolution layer, a layer normalization layer, a first linear layer, a two-way spatio-temporal channel attention layer, a second linear layer, a layer scaling layer, a random depth regularization layer and a second deformation layer connected in sequence; The first deformation layer is configured to adjust the input feature X from shape one to shape two to obtain the feature X ; The multi-scale depth separable convolution layer is configured to perform 3D depth separable convolution on X respectively with 1x1, 3x3 and 5x5 convolution kernels to obtain convolution features X , , , the channel dimensions of which are spliced and the channel number is adjusted to Cp to obtain the fused feature X ; The layer normalization layer and the first linear layer are configured to perform layer normalization operation on X in sequence, and then adjust the output channel number from Cp to 4Cp to obtain the expanded feature X ; The two-way spatio-temporal channel attention layer is configured to perform attention in the channel dimension and the spatio-temporal dimension on X to generate a channel attention weight matrix F a and a spatio-temporal attention weight matrix F b , and generate a two-way attention feature X , according to the formula, wherein γ is a learnable scaling factor; The second linear layer is configured to adjust X The number of channels is adjusted from 4Cp to Cp, and a feature is obtained ; The layer scaling layer and the random depth regularization layer are sequentially subjected to layer scaling and random depth regularization operations, and a feature is obtained ; ; The second deformation layer is used to adjust from shape two to shape one, and a feature is obtained .

[0015] Preferably, the improved conditional diffusion model comprises a VAE encoder, a noise adding unit, an image block layer, a time encoder, a splicing unit, an improved denoising network, an image block reorganization layer, a denoising module, and a VAE decoder. The VAE encoder is used to encode the conditional radar echo sequence X cond and the target radar echo sequence X0 into a latent space, respectively, to obtain the corresponding latent conditional sequence Z cond and the latent target sequence Z0. The noise adding unit is used to add Gaussian noise to Z0 step by step in a forward diffusion process to obtain a noise sequence Z T ; The image block layer is used to convert Z cond , Z T to shape one, respectively, to obtain the corresponding feature sequence , ; The time encoder is used to encode the time step into a feature space to obtain a time sequence T s ; The splicing unit is used to splice T s , , along the channel dimension and then map to shape one to obtain a joint noise sequence Z mix ; The improved denoising network is obtained by connecting M layers of convolution enhanced frequency domain Transformers in a Unet structure, and is used to output a predicted noise sequence according to Z mix ; The image block reorganization layer is used to convert the predicted noise sequence to [B, S, Cz, Hz, Wz] to obtain a predicted noise ; The denoising module is used to remove T from Z to obtain a predicted latent target sequence ; The VAE decoder is used to output a predicted target radar echo sequence according to .

[0016] As preferred, in S5, one round of training comprises steps S51-S53; S51, obtain a batch of image sequences of the first time period from D1 to form a conditional radar echo sequence X cond a batch of image sequences of the second time period to form a target radar echo sequence X0; S52, encode X cond and X0 into latent space respectively to obtain corresponding latent conditional sequence Z cond and latent target sequence Z0, add Gaussian noise to Z0 at each time step by a forward diffusion process to obtain noise sequence Z T S53, convert Z cond and Z T to shape one respectively to obtain corresponding feature sequence , S54, encode time steps into shape one to obtain time sequence T s S55, concatenate T s , , along the channel dimension and map to shape one to obtain joint noise sequence Z mix ; S56, input Z mix into the improved denoising network to output a predicted noise sequence, and convert to [B, S, Cz, Hz, Wz] to obtain predicted noise S57, calculate the L2 loss of predicted noise and , and adjust the network parameters of the improved denoising network to minimize the L2 loss.

[0017] As preferred, S6 specifically comprises S61-S63; S61, obtain a conditional radar echo sequence for prediction input into the nowcasting precipitation prediction model, encode and convert to shape one to obtain corresponding feature sequence S62, randomly sample a Gaussian noise sequence T convert to shape one to obtain corresponding feature sequence ; S63, construct joint noise sequence s , , from T , predict a noise sequence by the improved denoising network and convert to [B, S, Cz, Hz, Wz] to obtain predicted noise T from which predicted noise is removed to obtain predicted latent target sequence ; S64, input the predicted latent target sequence into the nowcasting precipitation prediction model to obtain a predicted nowcasting precipitation sequence The predicted target radar echo sequence is generated by the VAE decoder.

[0018] Compared with the prior art, the present application has the advantages that: (1) A noise-aware frequency gating module is designed, which decomposes the noise-added feature map into the frequency domain through discrete cosine transform (DCT), and realizes frequency selection enhancement through learnable gating weights. This way can well couple the physical characteristics of radar echo signals (meteorological signals) and the diffusion generation logic. For example, in the early strong noise stage, the high-frequency region saturated by pure noise is actively shielded through the gating weight, forcing the subsequent module to focus on the key region (low-frequency band) of signal-noise mixture. In the later low-noise stage, the high frequency is enhanced and the low frequency is weakened to fine-grainedly restore the storm edge details. This design essentially fits the generation logic of the diffusion model "first main body and then details", realizing the deep coupling of meteorological physical laws and deep learning principles. Based on this design, the noise-aware frequency gating module can differentially suppress the noise-dominant frequency components through the frequency domain dynamic filtering mechanism of noise level perception, while retaining and enhancing the high-frequency signals corresponding to the real physical structure, thereby significantly reducing the confusion between signal and noise, improving the signal reconstruction quality and the convergence speed of the model. This design makes the improved conditional diffusion model more accurate in predicting noise and reconstructing real meteorological features.

[0019] (2) A local feature aggregation module is designed to aggregate From shape one to shape two, shape one [B, S, L, Cp], shape two [B, S, Cp, Hp, Wp], two shapes excluding batch B and image quantity S, shape one is a one-dimensional feature of LxCp, and shape two is a two-dimensional image feature of Cp x Hp x Wp. The local feature aggregation module first converts a one-dimensional feature into a two-dimensional image feature to adapt to convolution operation, captures context features of different scales through different multi-scale feature extraction, and significantly improves the spatiotemporal perception ability through multi-branch feature fusion. The fused features are stably distributed through layer normalization, the channel dimension is expanded to 4 times the original size through the first linear layer to enhance the representation ability, then the double-path spatiotemporal channel attention mechanism is introduced: attention is performed in the channel and spatiotemporal dimensions respectively, and the results are weighted and fused to form "channel-spatiotemporal" collaborative enhancement. The enhanced features are restored to the original channel dimension through the second linear layer, and the feature intensity distribution is fine-tuned through the learnable layer scaling parameter. Finally, random depth regularization (path dropout probability 0.1) is applied to suppress overfitting. The processed spatial feature is mapped back to shape one to realize the closed-loop interaction of global semantics and local features. This design enhances the representation ability of complex visual patterns, and is especially suitable for tasks that require multi-granularity feature understanding.

[0020] (3) A convolution-enhanced frequency domain Transformer is designed, which integrates a noise perception frequency domain gating module and a local feature aggregation module. Without increasing unnecessary calculations, the local feature aggregation module is seamlessly integrated into the frequency domain prior transformer architecture, solving the complementary defects of CNN and ViT. The local feature extraction capability of CNN is combined with the global context understanding of ViT to enhance the fine expression capability of radar echo spatio-temporal features.

[0021] (4) An improved denoising network is constructed based on the convolution-enhanced frequency domain Transformer, which replaces the original denoising network of the diffusion model in the reverse diffusion process. Finally, an end-to-end improved conditional diffusion model is obtained. The model can model the multi-scale spatio-temporal evolution law of the meteorological system by using its heterogeneous fusion characteristics. Experiments show that the architecture can effectively coordinate the frequency domain-spatial domain features and significantly improve the prediction accuracy of complex weather patterns.

[0022] In summary, the model can more accurately capture the high-frequency details and edge features of the convective system by using the frequency domain-spatial domain collaborative modeling and efficient conditional diffusion architecture design, significantly improving the prediction accuracy of weak echo areas and maintaining stable prediction performance in heavy rain, thunderstorm and other strong convective scenarios. This not only effectively overcomes the core defects of poor fusion of local and global features, signal-noise confusion and prediction ambiguity in the prior art, but also enhances the adaptation ability to complex weather scenarios through architecture optimization, ultimately providing more reliable technical support for 0-2 hour short-term precipitation prediction, and helping to improve the timeliness and accuracy of extreme weather warning. BRIEF DESCRIPTION OF DRAWINGS

[0023] Fig. 1 is a structure diagram of a convolution-enhanced frequency domain Transformer; Fig. 2 is a structure diagram of a local feature aggregation module; Fig. 3 is a schematic diagram of an improved conditional diffusion model training process; Fig. 4 is a schematic diagram of an improved conditional diffusion model inference process; Fig. 5 is a comparison diagram of a weak radar echo sequence generated by different prediction models; Fig. 6 is a comparison diagram of a strong radar echo sequence generated by different prediction models. DETAILED DESCRIPTION

[0024] The application will be further described below in conjunction with the embodiments and the drawings.

[0025] Example 1: see Fig. 1~Fig. 4A radar near-rainfall prediction method based on frequency domain perception and conditional diffusion, comprising the following steps: S1, obtaining a radar echo data set D1, the samples in D1 are time-series arranged radar echo maps, any adjacent S samples form an image sequence, and the pixel value in the sample represents the rainfall at the corresponding position; S2, pre-training a variational autoencoder M VAE , presetting shape one and shape two; The variational autoencoder comprises a VAE encoder and a VAE decoder, the VAE encoder is used for encoding a batch of image sequences X IS from a pixel space to a latent space to obtain a latent feature sequence Z IS , and the VAE decoder is used for decoding Z IS back to the pixel space to obtain a predicted image sequence , wherein Z IS is of shape [B, S, Cz, Hz, Wz], B is the batch size, Cz, Hz and Wz are respectively the channel number, width and height of the image in the latent space, and Z IS each image in the latent space can be divided into N*N image blocks by an image blocking layer and form a feature sequence with a length L=N*N; The preset shape one is [B, S, L, Cp], and the shape two is [B, S, Cp, Hp, Wp], Cp, Hp and Wp are respectively the feature channel number, height and width of the image in the shape two; S3, constructing a convolution enhanced frequency domain Transformer, comprising a noise perception frequency domain gating module, a layer normalization layer, a space-time attention layer, a local feature aggregation module and an MLP arranged in sequence; The noise perception frequency domain gating module is used for obtaining the input feature of shape one , adjusting to shape two, sequentially performing DCT transformation, frequency domain selection enhancement and IDCT transformation, and adjusting back to shape one to obtain the output feature of the noise perception frequency domain gating module ; The layer normalization layer and the space-time attention layer sequentially perform layer normalization and space-time attention operation on to obtain an attention feature X att , and and X att are connected in series as the input feature of the local feature aggregation module , The shape is shape one; The local feature aggregation module is used for transforming to shape two, performing multi-scale feature extraction and fusion and double-path space-time channel attention operation, and transforming back to shape one to obtain the output feature of the local feature aggregation module ; The MLP is used as input and Features after skip connection, output features of convolution-enhanced frequency domain Transformer ; S4, based on M VAE An improved conditional diffusion model is constructed with the improved denoising network, wherein the improved denoising network is obtained by connecting M layers of convolution-enhanced frequency-domain Transformers according to a Unet structure, where M is an odd number and is ≥ 13; S5, training the improved conditional diffusion model until convergence, and obtaining an approaching precipitation prediction model, which is used to predict the radar echo sequence of the second period based on the radar echo sequence of the first period, and the interval between the end points of the two periods is less than 2 hours; S6, obtain the conditional radar echo sequence for prediction , the precipitation prediction model outputs the predicted noise, and the predicted target radar echo sequence is obtained based on the predicted noise.

[0026] The noise-aware frequency domain gating module is based on get The method is: Sa1, according to the following formula The shape is adjusted to X1; , Where Reshape(∙) is the Reshape function, is shape one, X1 is shape two; Sa2, generates the DCT transformation matrix D in length, height and width of the image sequence composed of the latent space image according to the following formula S 、D Hp 、D Wp ; , Where, is the dimension index, , D S ∈R S×S 、D Hp ∈R Hp×Hp 、D Wp ∈R Wp×Wp , for The value of row i and column j, k is The frequency domain index in the row direction, i is the spatial position index of D in the column direction, 0≤k,i≤τ-1, and With its transposed matrix orthogonal; Sa3, according to the following formula, DCT transformation, frequency domain selection enhancement and IDCT transformation are performed to obtain the frequency domain feature Xdct , enhanced feature X scaled , and IDCT feature X idct ; , , , wherein, , , are matrix multiplications in the shape two on the 3rd, 4th, 5th dimensions respectively, and σ (scaled) is a channel-level gating weight generated by a Sigmoid function on a learnable weight parameter scaled, and T is a transpose operation; Sa4, X idct is adjusted to shape one, to obtain .

[0027] The local feature aggregation module comprises a first deformation layer, a multi-scale depth separable convolution layer, a layer normalization layer, a first linear layer, a two-way spatio-temporal channel attention layer, a second linear layer, a layer scaling layer, a random depth regularization layer and a second deformation layer connected in sequence; the first deformation layer is configured to adjust an input feature from shape one to shape two to obtain a feature ; the multi-scale depth separable convolution layer is configured to perform 3D depth separable convolution on respectively with a 1x1, 3x3 and 5x5 convolution kernel, to obtain convolution features , , , which are spliced in the channel dimension and adjusted in the channel number to Cp to obtain a fused feature ; the layer normalization layer and the first linear layer are configured to perform a layer normalization operation on in sequence, and then adjust the output channel number from Cp to 4Cp to obtain an expanded feature ; the two-way spatio-temporal channel attention layer is configured to perform attention in the channel dimension and the spatio-temporal dimension on respectively, to generate a channel attention weight matrix F a and a spatio-temporal attention weight matrix F b , and generate a two-way attention feature according to the formula , wherein γ is a learnable scaling factor; the second linear layer is configured to adjust the channel number of from 4Cp to Cp to obtain a feature ; the layer scaling layer and the random depth regularization layer are configured to perform a layer scaling operation and a random depth regularization operation on in sequence to obtain a feature ; the second deformation layer is configured to adjust from shape two to shape one, obtaining .

[0028] The improved conditional diffusion model comprises a VAE encoder, a noise adding unit, an image block layer, a time encoder, a splicing unit, an improved denoising network, an image block reorganization layer, a denoising module and a VAE decoder; the VAE encoder is used for encoding a conditional radar echo sequence X cond and a target radar echo sequence X0 into a latent space respectively, to obtain corresponding latent condition sequences Z cond and latent target sequences Z0; the noise adding unit is used for adding Gaussian noise to Z0 step by step in time through a forward diffusion process to obtain a noise sequence Z T ; the image block layer is used for converting Z cond , Z T to shape one respectively, to obtain corresponding feature sequences , ; the time encoder is used for encoding time steps into a feature space to obtain a time sequence T s ; the splicing unit is used for splicing T s , , along a channel dimension and then mapping to shape one to obtain a joint noise sequence Z mix ; the improved denoising network is obtained by connecting M layers of convolution enhanced frequency domain Transformers in a Unet structure, and is used for outputting a predicted noise sequence according to Z mix ; the image block reorganization layer is used for converting the predicted noise sequence to [B, S, Cz, Hz, Wz] to obtain a predicted noise ; the denoising module is used for removing from Z T to obtain a predicted latent target sequence ; and the VAE decoder is used for outputting a predicted target radar echo sequence according to .

[0029] In S5, one round of training comprises steps S51-S53; S51, a batch of image sequences of a first time period from D1 constitute a conditional radar echo sequence X cond , and a batch of image sequences of a second time period constitute a target radar echo sequence X0; S52, X cond and X0 are encoded into a latent space respectively to obtain corresponding latent condition sequences Z cond and latent target sequences Z0, Gaussian noise is added to Z0 step by step in time through a forward diffusion process to obtain a noise sequence Z T , Z cond, Z T Respectively convert to shape one, get the corresponding feature sequence 、 Encode the time step to shape one to get the time sequence T s , and then T s 、 、 After splicing along the channel dimension, map to shape one to get the joint noise sequence Z mix ; S53, input Z mix into the improved denoising network, output the predicted noise sequence, and then convert to [B, S, Cz, Hz, Wz] to get the predicted noise , calculate the L2 loss of the predicted noise and , and adjust the network parameters of the improved denoising network to minimize the L2 loss.

[0030] S6 specifically includes S61-S63; S61, obtain the conditional radar echo sequence for prediction Input into the near precipitation prediction model, encode and convert to shape one to get the corresponding feature sequence , randomly sample a Gaussian noise sequence T , convert to shape one to get the corresponding feature sequence ; S62, construct T s 、 、 into a joint noise sequence , predict the noise sequence through the improved denoising network and convert to [B, S, Cz, Hz, Wz] to get the predicted noise, and then T Subtract the predicted noise from to get the predicted latent target sequence ; S63, generate the predicted target radar echo sequence through the VAE decoder.

[0031] Embodiment 2: see Fig. 1~Fig. 4 To verify the synergistic effect of each module in the convolution enhanced frequency domain Transformer, we perform an ablation experiment on the improved denoising network obtained in Embodiment 1.

[0032] The improved denoising network comprises a multi-layer convolution enhanced frequency domain Transformer, and each layer of the convolution enhanced frequency domain Transformer comprises, in sequence, a noise perception frequency domain gating module, a layer normalization layer, a space-time attention layer, a local feature aggregation module and an MLP; the noise perception frequency domain gating module and the local feature aggregation module in the convolution enhanced frequency domain Transformer are marked as Transformers.

[0033] For the improved conditional diffusion model, only the Transformer part is included in the improved denoising network to obtain model 1, only the Transformer part and the noise perception frequency domain gating module part are included in the improved denoising network to obtain model 2, and the improved conditional diffusion model comprising the complete improved denoising network is taken as model 3. Ablation experiments are respectively performed on the CIKM 2017 short-time quantitative precipitation prediction dataset, and the core indicators are shown in Table 1 as follows: Table 1. Comparison of core indicators of ablation experiments Model architecture Mean squared error Mean absolute error Structural similarity index Model 1 326.14 12.53 65.42 Model 2 275.94 11.12 68.73 Model 3 237.88 9.37 71.96 From Table 1, the following results are obtained: when the model 1 adopts only the Transformer architecture and is trained from scratch, the mean square error is 326.14 and the mean absolute error is 11.53; after introducing the noise perception frequency domain gating module to form the model 2 architecture, the mean square error is reduced to 275.94, with a reduction of about 15.6%, and the mean absolute error is reduced to 11.12, with a reduction of about 11.3%, which reflects the effectiveness of DCT frequency domain modeling in reducing prediction error; further fusion of the local feature aggregation module to obtain the performance of the model 3 achieves a significant leap, compared with the model 2, the mean square error is reduced to 237.88, with a reduction of about 13.7%, the mean absolute error is reduced to 9.37, with a reduction of about 15.7%, and the structural similarity index is increased to 71.96, with an increase of about 4.7%. Overall, in the architecture evolution from model 1 to model 2 to model 3, the two error indicators show a continuous downward trend. In summary, the ablation experiment results prove that the noise perception frequency domain gating module, the Transformer global modeling and the multi-scale inverse reconstruction module for local feature extraction and fusion can continuously optimize the model performance, and the synergistic effect of the three is the key to improving the radar echo prediction accuracy, which provides direct experimental support for the effectiveness of the convolution enhanced frequency domain Transformer.

[0034] Embodiment 3: To evaluate the performance of the present application in the radar echo prediction task, a comparative experiment is carried out to verify its effectiveness. The following several classical prediction models are selected for comparison with the present application, and the prediction performance is evaluated under multiple indicators, and the results are shown in Table 2 as follows: Table 2. Comparison of performance indicators of comparative experiments Model SSIM PSNR CSI HSS POD PhyDNet 0.6301 23.12 0.2995 0.1263 0.2148 SmaAt-UNet 0.5904 22.14 0.1599 0.1047 0.2026 ConvLSTM 0.6342 22.94 0.3006 0.3075 0.3536 Rainformer 0.6043 22.07 0.1644 0.1122 0.2065 PredRNN 0.6572 23.47 0.3243 0.3233 0.3772 The invention 0.7128 24.72 0.3537 0.3579 0.4257 In Table 2, PhyDNet, SmaAt-UNet, ConvLSTM, Rainformer, and PredRNN are all existing network architectures. SSIM is the structural similarity index, PSNR is the peak signal-to-noise ratio, CSI is the critical success index, HSS is the Heidke skill score, and POD is the hit rate. CSI, HSS, and POD are calculated at thresholds of 10, 20, 30, and 40 dbz, respectively, and then averaged.

[0035] From Table 2, it can be seen that the present application has achieved the best performance in various key indicators. In terms of reconstruction quality, compared with other models, the algorithm in this paper significantly improves the structural similarity and signal-to-noise ratio of the generated image, indicating that its prediction result is closer to the true echo, the details are clearer, and the noise is less. In terms of prediction accuracy (critical success index CSI), which is more practical and valuable, the algorithm in this paper also shows a significant advantage, with a CSI value significantly higher than all the comparison models, which indicates that it has higher accuracy in identifying and locating different intensity meteorological targets such as heavy rainfall areas. Overall, the experimental data fully verify the superior performance of the algorithm in this paper in the radar echo prediction task, not only leading in image reconstruction quality, but also performing well in key meteorological target detection capability.

[0036] Fig. 5 and Fig. 6 respectively show a prediction case of a weak radar echo sequence and a strong radar echo sequence. The weak radar echo sequence is an image sequence composed of 6 weak radar echo images, where T1-T6 represent the 1st image to the 6th image in the image sequence. The strong radar echo sequence is an image sequence composed of 6 strong radar echo images, where T1-T6 represent the 1st image to the 6th image in the image sequence. The color bar in the figure shows the mapping relationship between the dBZ value of the radar echo and the color. It can be observed that over time, all models can capture the overall trend of radar echo distribution. However, PhyDNet, SmaAt-UNet, ConvLSTM, Rainformer, and PredRNN perform poorly in echo intensity and feature distribution, and the predicted area shows obvious echo dissipation over time. In contrast, the method of the present application can effectively capture subtle echo features. The overall effect remains stable, and the image quality is much higher than that of PhyDNet, SmaAt-UNet, ConvLSTM, Rainformer, and PredRNN.

[0037] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion, characterized in that: The following steps are included: S1, obtain radar echo data set D1. The samples in D1 are time-sequential radar echo images. Any S adjacent samples constitute an image sequence. The pixel value in the sample represents the precipitation at the corresponding location. S2, pre-trained variational autoencoder M VAE , preset shape one and shape two; The variational autoencoder includes a VAE encoder and a VAE decoder, wherein the VAE encoder is used to transform a batch of image sequences X IS Encoding from pixel space to latent space to obtain the latent feature sequence Z IS , the VAE decoder is used to transform Z IS Decode back to pixel space to get the predicted image sequence , where Z IS The shape is [B, S, Cz, Hz, Wz], where B is the batch size, Cz, Hz, and Wz are the number of channels, width, and height of the image in the latent space, respectively. IS Each image can be divided into N×N image blocks by the image block layer and form a feature sequence of length L=N×N; Preset shape 1 [B, S, L, Cp] and shape 2 [B, S, Cp, Hp, Wp], where Cp, Hp, and Wp are the number of feature channels of shape 2 and the height and width of the image respectively; S3, constructs a convolution-enhanced frequency-domain Transformer, which includes a noise-aware frequency-domain gating module, a layer normalization layer, a spatiotemporal attention layer, a local feature aggregation module, and an MLP. The noise-aware frequency domain gating module is used to obtain the input features of shape 1 , adjusted to shape 2, sequentially undergoing DCT transformation, frequency domain selection enhancement and IDCT transformation, and then adjusted back to shape 1 to obtain the output characteristics of the noise-aware frequency domain gating module ; The layer normalization layer and the spatiotemporal attention layer are sequentially Perform layer normalization and spatiotemporal attention operations to obtain attention features X att ,and and X att After skip connection, it is used as the input feature of the local feature aggregation module , The shape is shape one; The local feature aggregation module is used to Transformed to shape 2, after multi-scale feature extraction and fusion, dual-path spatiotemporal channel attention operation, and then transformed back to shape 1, the output features of the local feature aggregation module are obtained ; The MLP is used as input and Features after skip connection, output features of convolution-enhanced frequency domain Transformer ; S4, based on M VAE An improved conditional diffusion model is constructed with the improved denoising network, wherein the improved denoising network is obtained by connecting M layers of convolution-enhanced frequency-domain Transformers according to a Unet structure, where M is an odd number and is ≥ 13; S5, training the improved conditional diffusion model until convergence, and obtaining an approaching precipitation prediction model, which is used to predict the radar echo sequence of the second period based on the radar echo sequence of the first period, and the interval between the end points of the two periods is less than 2 hours; S6, obtain the conditional radar echo sequence for prediction , the precipitation prediction model outputs the predicted noise, and the predicted target radar echo sequence is obtained based on the predicted noise.

2. The radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion according to claim 1 is characterized in that: The noise-aware frequency domain gating module is based on get The method is: Sa1, according to the following formula The shape is adjusted to X1; ， Where Reshape(∙) is the Reshape function, is shape one, X1 is shape two; Sa2, generates the DCT transformation matrix D in length, height and width of the image sequence composed of the latent space image according to the following formula S 、D Hp 、D Wp ; ， Where, is the dimension index, , D S ∈R S×S 、D Hp ∈R Hp×Hp 、D Wp ∈R Wp×Wp , for The value of row i and column j, k is The frequency domain index in the row direction, i is the spatial position index of D in the column direction, 0≤k,i≤τ-1, and With its transposed matrix orthogonal; Sa3, according to the following formula, DCT transformation, frequency domain selection enhancement and IDCT transformation are performed to obtain the frequency domain feature X dct , Enhanced Feature X scaled and IDCT feature X idct ; ， ， ， Where, 、 、 are the matrix multiplications of shape 2 on the 3rd, 4th, and 5th dimensions respectively, σ(scaled) is the channel-level gating weight generated by the learnable weight parameter scaled by the Sigmoid function, and T is the transpose operation; Sa4, X idct Adjust to shape 1, and get .

3. The radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion according to claim 1 is characterized in that: The local feature aggregation module includes a first deformation layer, a multi-scale depth-separable convolution layer, a layer normalization layer, a first linear layer, a two-way spatiotemporal channel attention layer, a second linear layer, a layer scaling layer, a random depth regularization layer and a second deformation layer connected in sequence; The first deformation layer is used to transform the input features through convolution operation Adjust from shape one to shape two and get the feature ; The multi-scale depth-wise separable convolutional layer is used to 3D depth-separable convolution is performed with 1×1, 3×3, and 5×5 convolution kernels respectively, and the corresponding convolution features are obtained 、 、 , the three channel dimensions are spliced ​​and the number of channels is adjusted to Cp to obtain the fusion feature ; The layer normalization layer and the first linear layer are successively Perform layer normalization and adjust the number of output channels from Cp to 4Cp to obtain the extended features. ; The dual-path spatiotemporal channel attention layer is used to Perform attention in the channel dimension and spatiotemporal dimension to generate the channel attention weight matrix F a and the spatiotemporal attention weight matrix F b , and generate dual-path attention features according to the formula , , γ is a learnable scaling factor; The second linear layer is used to The number of channels is adjusted from 4Cp to Cp, and the feature ; The layer scaling layer and the random depth regularization layer are sequentially Perform layer scaling and random depth regularization operations to obtain features ; The second deformation layer is used to Adjusting from shape 2 to shape 1, we get .

4. The radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion according to claim 1 is characterized in that: The improved conditional diffusion model includes a VAE encoder, a denoising unit, an image segmentation layer, a temporal encoder, a splicing unit, an improved denoising network, an image block reorganization layer, a denoising module and a VAE decoder; The VAE encoder is used to condition the radar echo sequence X cond and the target radar echo sequence X0 are encoded into the latent space respectively, and the corresponding latent condition sequence Z is obtained cond and potential target sequence Z0; The noise adding unit is used to add Gaussian noise to Z0 time step by time through the forward diffusion process. Get the noise sequence Z T ; The image block layer is used to respectively cond , Z T Convert to shape one and get the corresponding feature sequence 、 ; The time encoder is used to encode the time step into the feature space to obtain the time series T s ; The splicing unit is used to s 、 、 After concatenation along the channel dimension, it is mapped to shape 1 to obtain the joint noise sequence Z mix ; The improved denoising network is obtained by connecting M layers of convolution-enhanced frequency domain Transformer according to the Unet structure, and is used to mix Output predicted noise sequence; The image block reconstruction layer is used to convert the predicted noise sequence to [B, S, Cz, Hz, Wz] to obtain the predicted noise ; The denoising module is used to T Remove , get the predicted potential target sequence ; The VAE decoder is used according to Output predicted target radar echo sequence.

5. The radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion according to claim 1 is characterized in that: In S5, one round of training includes steps S51 to S53; S51, obtain a batch of image sequences of the first period from D1 to form a conditional radar echo sequence X cond , a batch of image sequences of the second period constitutes the target radar echo sequence X0; S52, X cond , X0 are encoded into the latent space to obtain the corresponding latent conditional sequence Z cond , potential target sequence Z0, Gaussian noise is added to Z0 time step by time by the forward diffusion process Get the noise sequence Z T , Z cond , Z T Convert to shape 1 respectively and get the corresponding feature sequence 、 , encode the time step into shape 1 to obtain the time series T s , and then T s 、 、 After concatenation along the channel dimension, it is mapped to shape 1 to obtain the joint noise sequence Z mix ; S53, Z mix Input the improved denoising network, output the predicted noise sequence, and then convert it to [B, S, Cz, Hz, Wz] to get the predicted noise , calculate the prediction noise and The L2 loss is used to minimize the L2 loss and adjust the network parameters of the improved denoising network.

6. The radar approaching precipitation prediction method based on frequency domain perception and conditional diffusion according to claim 1 is characterized in that: S6 specifically includes S61~S63; S61, obtain the conditional radar echo sequence for prediction Input the nearby precipitation prediction model, encode and transform the shape to obtain the corresponding feature sequence , randomly sample a Gaussian noise sequence T , transform shape 1 and get the corresponding feature sequence ; S62, T s 、 、 Constructing a joint noise sequence , the noise sequence is predicted by the improved denoising network and converted to [B, S, Cz, Hz, Wz] to obtain the predicted noise, from T Remove the prediction noise and get the predicted potential target sequence ; S63, will The VAE decoder generates the predicted target radar echo sequence.

Images