Radar nowcasting method based on frequency domain perception and conditional diffusion

Abstract

The application discloses a radar near precipitation prediction method based on frequency domain perception and conditional diffusion, and belongs to the technical field of image data processing. VAE The method comprises the following steps: obtaining a radar echo data set, pre-training a variational autoencoder M VAE ; constructing a convolution enhanced frequency domain Transformer; constructing an improved denoising network and M VAE based on the multi-layer convolution enhanced frequency domain Transformer, generating an improved conditional diffusion model and training a near precipitation prediction model; and being used for near precipitation prediction. The application overcomes the core defects of poor fusion of local and global features, prediction ambiguity caused by signal-noise confusion and the like in the prior art through frequency domain-space collaborative modeling and efficient conditional diffusion architecture design, enhances the adaptation capability to complex weather scenes, can more accurately capture high-frequency details and edge features of a convection system, significantly improves the prediction accuracy of a weak echo area, and improves the accuracy of short-term and near-term precipitation prediction.

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 mesoscale 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, with significant error accumulation over time and a sharp decline in prediction accuracy in weak echo areas.

[0004] In the deep learning method:

[0005] 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 decline in prediction accuracy, and the ability to capture details of weak echoes and strong convective boundaries is limited.

[0006] Existing diffusion models (such as SRNDiff) can generate diverse samples, but have local texture modeling and efficiency imbalance problems, with fuzzy local texture modeling of weak echo boundaries and high computational cost, 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.

[0007] 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.

[0008] 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.

[0009] 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.

[0010] Glossary:

[0011] 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, space-time continuous attention, etc. The "Divided Space-Time Attention" 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 space dimension attention calculation 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 structure 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.

[0012] Dual-channel space-time channel attention: It is a three-dimensional extension structure based on Squeeze-and-Excitation (SE) mechanism, which realizes the cooperative feature enhancement of channel and space-time dimensions through 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 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

[0013] 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.

[0014] In order to achieve the above object, the technical scheme adopted by the present application is as follows: a radar near precipitation prediction method based on frequency domain perception and conditional diffusion, comprising the following steps:

[0015] S1, obtaining a radar echo data set D1, the samples in D1 are time series arranged radar echo maps, any adjacent S samples constitute an image sequence, and the pixel value in the sample represents the precipitation at the corresponding position;

[0016] S2, pre-training a variational autoencoder M VAE , presetting a shape one and a shape two;

[0017] 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 constitutes a feature sequence with a length L=N*N;

[0018] The 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;

[0019] 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 which are sequentially arranged;

[0020] The noise perception frequency domain gating module is used for obtaining the input feature of the shape one , adjusting it to the shape two, sequentially performing DCT transformation, frequency domain selection enhancement and IDCT transformation, and then adjusting it back to the shape one to obtain the output feature of the noise perception frequency domain gating module ;

[0021] 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 the shape one;

[0022] The local feature aggregation module is configured to aggregate transform to shape two, multi-scale feature extraction and fusion, double path space-time channel attention operation, and then transform back to shape one to obtain output features of the local feature aggregation module ;

[0023] The MLP is configured to input and features after the skip connection, and output output features of the convolution-enhanced frequency domain Transformer ;

[0024] S4, based on M VAE and an improved de-noising network, an improved conditional diffusion model is constructed, wherein the improved de-noising network is obtained by connecting M layers of convolution-enhanced frequency domain Transformers in a Unet structure, M≥13 and is an odd number;

[0025] S5, training the improved conditional diffusion model to convergence to obtain a nearby precipitation prediction model, which is configured to predict a radar echo sequence of a second time period based on a radar echo sequence of a first time period, and the interval between the end of the two time periods is less than 2 hours;

[0026] S6, obtaining a conditional radar echo sequence for prediction , outputting a predicted noise through the precipitation prediction model, and obtaining a predicted target radar echo sequence based on the predicted noise.

[0027] As a preferred, the noise-aware frequency domain gating module is configured to obtain by The method is as follows:

[0028] Sa1, reshape to X1 according to the following formula:

[0029] ,

[0030] In the formula, Reshape(·) is a Reshape function, is shape one, and X1 is shape two;

[0031] Sa2, generate a DCT transformation matrix D S , D Hp , D Wp in length, height and width on an image sequence composed of latent space images according to the following formula:

[0032] ,

[0033] In the formula, is a dimension index, , DS ∈R S×S 、D Hp ∈R Hp×Hp 、D Wp ∈R Wp×Wp , for 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 ; is orthogonal to its transpose matrix ;

[0034] 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 ;

[0035] ,

[0036] ,

[0037] ,

[0038] wherein, , , are respectively matrix multiplication of shape two on the 3rd, 4th and 5th dimensions, σ (scaled) is a learnable weight parameter scaled, a channel-level gating weight generated by a Sigmoid function, T is a transpose operation;

[0039] Sa4, X idct is adjusted to shape one to obtain .

[0040] 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;

[0041] The first deformation layer is configured to adjust the input feature X from shape one to shape two by a convolution operation to obtain the feature X ;

[0042] The multi-scale depth separable convolution layer is configured to perform 3D depth separable convolution on X respectively with 1×1, 3×3 and 5×5 convolution kernels to obtain convolution features X , , , the three channel dimensions are spliced and the number of channels is adjusted to Cp to obtain fused features ;

[0043] The layer normalization layer and the first linear layer perform layer normalization operation on in turn, and then the output channel number is adjusted from Cp to 4Cp to obtain expanded features ;

[0044] The dual-path spatio-temporal channel attention layer is configured to perform attention on in the channel dimension and the spatio-temporal dimension respectively, to generate a channel attention weight matrix F a and a spatio-temporal attention weight matrix F b , and generate dual-path attention features according to the formula , where γ is a learnable scaling factor;

[0045] The second linear layer is configured to adjust the channel number from 4Cp to Cp to obtain features ;

[0046] The layer scaling layer and the random depth regularization layer perform layer scaling and random depth regularization operation on in turn to obtain features ;

[0047] The second deformation layer is configured to adjust from shape two to shape one to obtain .

[0048] Preferably, the improved conditional diffusion model comprises a VAE encoder, a noise adding unit, an image blocking layer, a time encoder, a splicing unit, an improved denoising network, an image block recombination layer, a denoising module and a VAE decoder;

[0049] The VAE encoder is configured to encode the conditional radar echo sequence X cond and the target radar echo sequence X0 into latent spaces respectively to obtain corresponding latent conditional sequences Z cond and latent target sequences Z0;

[0050] The noise adding unit is configured to add Gaussian noise to Z0 step by step in time through a forward diffusion process to obtain a noise sequence Z ; T ;

[0051] The image blocking layer is configured to convert Z cond , Z T to shape one respectively to obtain corresponding feature sequences , ;

[0052] The time encoder is used to encode time steps into the feature space to obtain a time sequence T. s ;

[0053] The splicing unit is used to combine T s , , After concatenation along the channel dimension and mapping to shape one, the joint noise sequence Z is obtained. mix ;

[0054] The improved denoising network is obtained by connecting M layers of convolutional enhanced frequency domain Transformers according to a Unet structure, and is used to determine the denoising network based on Z. mix Output the predicted noise sequence;

[0055] The image patch reconstruction layer is used to transform the predicted noise sequence to [B,S,Cz,Hz,Wz] to obtain the predicted noise. ;

[0056] The noise reduction module is used to reduce noise from Z T Remove from middle To obtain the predicted potential target sequence ;

[0057] The VAE decoder is used to... Output the predicted target radar echo sequence.

[0058] Preferably, in S5, one round of training includes steps S51 to S53;

[0059] S51, Obtain a batch of image sequences from D1, constituting the radar echo sequence X. cond A batch of images from the second time period constitutes the target radar echo sequence X0;

[0060] S52, X cond X0 and X1 are encoded into the latent space to obtain the corresponding latent conditional sequences Z. cond The potential target sequence Z0 is to have Gaussian noise added to it step-by-step during the forward diffusion process. Obtain the noise sequence Z T , will Z cond Z T Transform them to shape one respectively to obtain the corresponding feature sequences. , Encode the time steps into shape 1 to obtain the time series T. s Then T s , , After concatenation along the channel dimension and mapping to shape one, the joint noise sequence Z is obtained. mix ;

[0061] S53, obtaining Z mix inputting the improved denoising network, outputting a predicted noise sequence, and converting to [B, S, Cz, Hz, Wz] to obtain a predicted noise , calculating an L2 loss of the predicted noise and , and adjusting the network parameters of the improved denoising network to minimize the L2 loss.

[0062] Preferably, S6 specifically comprises S61-S63.

[0063] S61, obtaining a conditional radar echo sequence for prediction inputting the nowcasting model, encoding and converting shape one to obtain a corresponding feature sequence , randomly sampling a Gaussian noise sequence T , converting shape one to obtain a corresponding feature sequence ;

[0064] S62, obtaining T s , , to form a joint noise sequence , predicting a noise sequence through the improved denoising network and converting to [B, S, Cz, Hz, Wz] to obtain a predicted noise, and T subtracting the predicted noise from to obtain a predicted latent target sequence

[0065] S63, obtaining generating a predicted target radar echo sequence through the VAE decoder.

[0066] Compared with the prior art, the present application has the following advantages:

[0067] (1) A noise-aware frequency gating module is designed. The module decomposes the noisy feature map into the frequency domain through discrete cosine transform (DCT), and uses learnable gating weights to achieve frequency selection and enhancement. This approach can well couple the physical characteristics of radar echo signals (meteorological signals) with 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-tune the storm edge details. This design essentially fits the generation logic of the diffusion model "first main body and then details", achieving deep coupling of meteorological physical laws and deep learning principles. Based on this design, the noise-aware frequency gating module can use the frequency domain dynamic filtering mechanism based on noise level perception to differentially suppress noise-dominant frequency components while preserving and enhancing high-frequency signals corresponding to real physical structures, thereby significantly reducing signal-noise confusion and improving signal reconstruction quality and model convergence speed. This design enables the improved conditional diffusion model to more accurately predict noise and reconstruct real meteorological features.

[0068] (2) A local feature aggregation module is designed to 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 into a two-dimensional image feature to adapt to convolution operations, captures context features of different scales through different multi-scale feature extraction, and significantly improves 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, and then the dual-channel spatiotemporal attention mechanism is introduced: attention is performed in the channel and spatiotemporal dimensions respectively, and the results are weighted and fused to form a "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 learnable layer scaling parameters. Finally, random depth regularization (path dropout probability 0.1) is applied to suppress overfitting, and the processed spatial feature map is mapped back to shape one to realize closed-loop interaction between global semantics and local features. This design enhances the representation ability of complex visual patterns, especially for tasks that require multi-granularity feature understanding.

[0069] (3) A convolution-enhanced frequency domain Transformer is designed, which integrates a noise-aware 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 combination of CNN's local feature extraction capability and ViT's global context understanding enhances the fine expression ability of radar echo spatio-temporal features.

[0070] (4) Based on the convolution-enhanced frequency domain Transformer, an improved denoising network is constructed, which replaces the original denoising network in the diffusion model during the back diffusion process. Finally, an end-to-end improved conditional diffusion model is obtained. This model can model the multi-scale spatio-temporal evolution law of the meteorological system by utilizing its heterogeneous fusion characteristics. Experiments show that this architecture can effectively coordinate the frequency domain-spatial domain features and significantly improve the prediction accuracy of complex weather patterns.

[0071] In summary, through frequency domain-spatial domain collaborative modeling and efficient conditional diffusion architecture design, the model can more accurately capture the high-frequency details and edge features of convective systems, significantly improve the prediction accuracy in weak echo areas, and maintain 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 leading to 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, helping to improve the timeliness and accuracy of extreme weather warning. BRIEF DESCRIPTION OF DRAWINGS

[0072] Fig. 1 is a convolution-enhanced frequency domain Transformer structure diagram;

[0073] Fig. 2 is a local feature aggregation module structure diagram;

[0074] Fig. 3 is an improved conditional diffusion model training process schematic diagram;

[0075] Fig. 4 is an improved conditional diffusion model inference process schematic diagram;

[0076] Fig. 5 is a weak radar echo sequence generated by different prediction models for comparison of predicted target radar echo sequences;

[0077] Fig. 6 is a strong radar echo sequence generated by different prediction models for comparison of predicted target radar echo sequences. DETAILED DESCRIPTION

[0078] The application will be further described below with reference to the embodiments and drawings.

[0079] Embodiment 1: see Fig. 1~Fig. 4 A radar near precipitation prediction method based on frequency domain perception and conditional diffusion includes the following steps:

[0080] 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 precipitation at the corresponding position;

[0081] S2, pre-training a variational autoencoder M VAE , presetting shape one and shape two;

[0082] The variational autoencoder includes a VAE encoder and a VAE decoder, the VAE encoder is used for encoding a batch of image sequences X IS from the pixel space to the 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 each image in Z IS can be divided into N×N image blocks by an image blocking layer and form a feature sequence with a length of L=N×N;

[0083] 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 shape two;

[0084] S3, constructing a convolution enhanced frequency domain Transformer, including 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;

[0085] 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 ;

[0086] 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 X attThe input feature of the local feature aggregation module after the skip connection , The shape is shape one.

[0087] The local feature aggregation module is configured to aggregate transformed to shape two, multi-scale feature extraction and fusion, double-path space-time channel attention operation, and then transformed back to shape one to obtain output features of the local feature aggregation module ;

[0088] The MLP is configured to input and the features after the skip connection, and output the output features of the convolution-enhanced frequency domain Transformer ;

[0089] S4, based on M VAE and an improved denoising network, an improved conditional diffusion model is constructed, 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;

[0090] S5, training the improved conditional diffusion model to convergence to obtain a nearby precipitation prediction model, which is configured to predict a radar echo sequence of a second time period based on a radar echo sequence of a first time period, and the interval between the end of the two time periods is less than 2 hours;

[0091] S6, obtaining a conditional radar echo sequence for prediction , outputting a predicted noise through the precipitation prediction model, and obtaining a predicted target radar echo sequence based on the predicted noise.

[0092] The noise-aware frequency domain gating module is configured to obtain the method for is as follows:

[0093] Sa1, adjusting the shape of to X1 according to the following formula:

[0094] ,

[0095] In the formula, Reshape(·) is a Reshape function, shape one and X1 is shape two.

[0096] Sa2, generating a DCT transformation matrix D S , D Hp , and D Wp in length, height, and width on an image sequence composed of latent space images according to the following formula:

[0097] ,

[0098] wherein, is a dimension index, , D S ∈R S×S , D Hp ∈R Hp×Hp , D 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 is orthogonal to its transpose matrix ;

[0099] Sa3, DCT transforms, frequency domain selection enhancement and IDCT transform are performed according to the following formula to obtain the frequency domain feature X dct , the enhanced feature X scaled and the IDCT feature X idct ;

[0100] ,

[0101] ,

[0102] ,

[0103] wherein, , , are respectively matrix multiplication in the 3rd, 4th and 5th dimensions, σ (scaled) is a learnable weight parameter scaled, a channel-level gating weight generated by a Sigmoid function, T is a transpose operation;

[0104] Sa4, X idct is adjusted to shape one to obtain .

[0105] 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 double-path space-time 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 from shape one to shape two to obtain the feature ; the multi-scale depth separable convolution layer is configured to perform 3D depth separable convolution on respectively with a 1×1, 3×3 and 5×5 convolution kernel to obtain convolution features , , The three channels are concatenated and the number of channels is adjusted to Cp to obtain the fused feature. The layer normalization layer and the first linear layer are applied sequentially to... Perform layer normalization, then adjust the number of output channels from Cp to 4Cp to obtain the extended features. The dual-channel spatiotemporal attention layer is used to respectively... Attention is performed in both the channel and spatiotemporal dimensions, generating 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 was adjusted from 4Cp to Cp to obtain the features. The scaling layer and random depth regularization layer are applied sequentially to... Perform layer scaling and random depth regularization operations to obtain features. The second deformation layer is used to... Adjusting shape two to shape one, we get .

[0106] The improved conditional diffusion model includes a VAE encoder, a noise addition unit, an image block segmentation layer, a time encoder, a stitching unit, an improved denoising network, an image block reconstruction layer, a denoising module, and a VAE decoder; the VAE encoder is used for conditional radar echo sequences X. cond The target radar echo sequence X0 and the target radar echo sequence X0 are encoded into the latent space to obtain the corresponding latent condition sequence Z. cond and the potential target sequence Z0; the noise-adding unit is used to add Gaussian noise to Z0 step by step via a forward diffusion process. Obtain the noise sequence Z T The image block layer is used to divide Z into blocks respectively. cond Z T Transform to shape one to obtain the corresponding feature sequence. , The time encoder is used to encode time steps into the feature space to obtain a time series T. s The splicing unit is used to connect T s , , After concatenation along the channel dimension and mapping to shape one, the joint noise sequence Z is obtained. mix The improved denoising network is obtained by connecting M layers of convolutional enhanced frequency domain Transformers according to a Unet structure, and is used to determine the denoising network based on Z. mix The output is a predicted noise sequence; the image patch reconstruction layer is used to transform the predicted noise sequence to [B, S, Cz, Hz, Wz] to obtain the predicted noise. The noise reduction module is used to reduce noise from Z... T Remove from middle To obtain the predicted potential target sequence The VAE decoder is used to... Output the predicted target radar echo sequence.

[0107] In S5, one round of training includes steps S51 to S53;

[0108] S51, Obtain a batch of image sequences from D1, constituting the radar echo sequence X. cond A batch of images from the second time period constitutes the target radar echo sequence X0;

[0109] S52, X cond X0 and X1 are encoded into the latent space to obtain the corresponding latent conditional sequences Z. cond The potential target sequence Z0 is to have Gaussian noise added to it step-by-step during the forward diffusion process. Obtain the noise sequence Z T , will Z cond Z T Transform them to shape one respectively to obtain the corresponding feature sequences. , Encode the time steps into shape 1 to obtain the time series T. s Then T s , , After concatenation along the channel dimension and mapping to shape one, the joint noise sequence Z is obtained. mix ;

[0110] S53, Z mix The input is an improved denoising network, the output is a predicted noise sequence, which is then transformed to [B,S,Cz,Hz,Wz] to obtain the predicted noise. Calculate the predicted noise and The L2 loss is calculated, and the network parameters of the improved denoising network are adjusted to minimize the L2 loss.

[0111] S6 specifically includes S61~S63;

[0112] S61, Obtain the conditional radar echo sequence for prediction. Input the near-term precipitation prediction model, encode it, and convert it to shape 1 to obtain the corresponding feature sequence. Randomly sample a Gaussian noise sequence T Transform shape one to obtain the corresponding feature sequence. ;

[0113] S62, Ts 、 、 constitute a joint noise sequence , the improved denoising network predicts the noise sequence and converts to [B, S, Cz, Hz, Wz], to obtain the predicted noise, from T remove the predicted noise from the middle to obtain the predicted potential target sequence ;

[0114] S63, the predicted target radar echo sequence is generated by the VAE decoder.

[0115] 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.

[0116] The improved denoising network includes multiple layers of convolution enhanced frequency domain Transformers, and each layer of convolution enhanced frequency domain Transformer includes noise perception frequency domain gating module, layer normalization layer, space-time attention layer, local feature aggregation module and MLP arranged in sequence; the part of deleting noise perception frequency domain gating module and local feature aggregation module in convolution enhanced frequency domain Transformer is marked as Transformer.

[0117] 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 including the complete improved denoising network is taken as model 3. Ablation experiments are performed on the CIKM 2017 short-term quantitative precipitation prediction dataset, and the core indicators are shown in Table 1:

[0118] Table 1. Comparison of core indicators of ablation experiment

[0119] 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

[0120] From Table 1, the results are as follows: when model 1 only adopts 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 achieves a significant leap in the performance of model 3, compared with 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, from model 1 to model 2 to model 3, the two error indicators show a consistent 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, providing direct experimental support for the effectiveness of the convolution enhanced frequency domain Transformer.

[0121] Example 3: To evaluate the performance of the present application in the radar echo prediction task, comparative experiments are carried out to verify its effectiveness. The following several classical prediction models are selected for comparison experiments with the present application, and the prediction performance is evaluated under multiple indicators, and the results are shown in Table 2 as follows:

[0122] Table 2. Comparison of performance indicators of comparative experiments

[0123] 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

[0124] In Table 2, PhyDNet, SmaAt-UNet, ConvLSTM, Rainformer, and PredRNN are all public 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, POD is the hit rate, and CSI, HSS, and POD are calculated at thresholds of 10, 20, 30, and 40 dbz, respectively, and then averaged.

[0125] From Table 2, it can be seen that the application has achieved the best performance in various key indicators. In terms of reconstruction quality, compared with other models, the algorithm of the present application significantly improves the structural similarity and signal-to-noise ratio of the generated image, indicating that the prediction result is closer to the real 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 of the present application also shows a significant advantage, and the CSI value is significantly higher than all the comparison models, which indicates that it has higher accuracy in identifying and positioning different intensity meteorological targets such as heavy rainfall area. Overall, the experimental data fully verify the superior performance of the algorithm in the radar echo prediction task, not only leading in the image reconstruction quality, but also performing well in the key meteorological target detection capability.

[0126] 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, wherein T1~T6 represent the 1st image~the 6th image in the image sequence. The strong radar echo sequence is an image sequence composed of 6 strong radar echo images, wherein T1~T6 represent the 1st image~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 with the passage of time, all models can capture the overall movement trend of the 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 PhyDNet, SmaAt-UNet, ConvLSTM, Rainformer and PredRNN.

[0127] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement 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 nowcasting method based on frequency domain perception and conditional diffusion, characterized in that, The method comprises the following steps; S1, obtaining a radar echo dataset D1, wherein samples in the dataset are time-series arranged radar echo images, any adjacent S samples form an image sequence, and pixel values in the samples represent precipitation at corresponding positions; S2, pre-training a variational autoencoder M VAE , preset shape one, shape two; The variational autoencoder includes a VAE encoder and a VAE decoder, wherein the VAE encoder is used to process a batch of image sequences X. IS The latent feature sequence Z is obtained by encoding from the pixel space to the latent space. IS The VAE decoder is used to convert Z IS Decoding back to pixel space yields the predicted image sequence. 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, and Z is the number of channels. IS Each image can be divided into N×N image blocks through an image block layer, forming a feature sequence of length L=N×N; a preset shape one [B, S, L, Cp] and a shape two [B, S, Cp, Hp, Wp], wherein Cp, Hp and Wp are respectively a feature channel number of the shape two, a height and a width of an image; 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 which are sequentially arranged; The noise perception frequency domain gating module is used for acquiring an input feature of shape one , adjusting to shape two, sequentially performing DCT transformation, frequency domain selection enhancement and IDCT transformation, adjusting back to shape one, and obtaining an output feature of the noise perception frequency domain gating module ; The layer normalization layer and the space-time attention layer are sequentially connected to obtain an attention feature X att , and X att is input to a local feature aggregation module after a skip connection , the shape is shape one;​ The local feature aggregation module is configured to aggregate transform to shape two, multi-scale feature extraction and fusion, double path space-time channel attention operation, and then transform back to shape one to obtain output features of the local feature aggregation module ; The MLP is used for input and The output feature of the convolution enhanced frequency domain Transformer after the jump connection ; S4, based on M VAE and an improved de-noising network is configured into an improved conditional diffusion model, wherein the improved de-noising network is obtained by connecting M layers of convolution enhanced frequency domain Transformers in a Unet structure, M≥13 and M 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 a radar echo sequence of a second time period according to a radar echo sequence of a first time period, and an interval between the two time periods is less than 2 hours; S6, obtaining a conditional radar echo sequence for prediction , the precipitation prediction model outputs prediction noise, and obtains a predicted target radar echo sequence based on the prediction noise; 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 recombination 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 latent spaces respectively to obtain corresponding latent conditional 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 , , after mapping to shape one along a channel dimension 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 .

2. The radar nowcasting of precipitation method based on frequency domain perception and conditional diffusion according to claim 1, characterized in that, The noise perception frequency domain gating module is configured to obtained The method is as follows: Sai, according to the formula Shape adjustment to X1; ， In the formula, Reshape(·) is a Reshape function, is shape one, and X1 is shape two. Sa2, a DCT transform matrix D of the image sequence composed of the latent space images in length, height, and width is generated according to the following formula S , D Hp , D Wp ; ， wherein is a dimension index, , D S ∈ R S×S , D Hp ∈ R Hp×Hp , D Wp ∈ R Wp×Wp , is a value of the ith row and jth column, k is a frequency domain index in the row direction, i is a spatial position index of D in the column direction, 0≤k,i≤τ-1, and is orthogonal to its transpose matrix . Sa3, DCT transform, frequency domain selective enhancement, and IDCT transform according to the following formula, to obtain the frequency domain feature X dct , enhanced feature X scaled , and IDCT feature X idct ; ， ， ， wherein, , , are shape-wise matrix multiplications in the 3rd, 4th, 5th dimensions, respectively, σ (scaled) is a channel-wise gating weight generated by a Sigmoid function on the learnable weight parameter scaled, and T is a transpose operation. Sa4, X idct adjusting to shape one, resulting in .

3. The method for radar nowcasting precipitation based on frequency domain perception and conditional diffusion according to claim 1, characterized in that, The local feature aggregation module comprises a first deformation layer, a multi-scale deep separable convolution layer, a layer normalization layer, a first linear layer, a double-path space-time channel attention layer, a second linear layer, a layer scaling layer, a random depth regularization layer and a second deformation layer which are sequentially connected; The first deformation layer is configured to apply a convolution operation to the input features from shape one to shape two, to obtain features ; The multi-scale depth separable convolution layer is used for 3D depth separable convolution is performed on the feature image with a 1*1, 3*3 and 5*5 convolution kernel respectively, and convolution features are obtained 、 、 The channel dimensions of the three are spliced, and the channel number is adjusted to Cp to obtain a fusion feature ; The layers are normalized layer, first linear layer in turn Perform layer normalization operation, and then adjust the output channel number from Cp to 4Cp to obtain the expanded feature ; The dual-path space-time channel attention layer is used to respectively generate In the channel dimension and the space-time dimension, attention is performed to generate a channel attention weight matrix F a and a space-time attention weight matrix F b , and a dual-path attention feature is generated according to the formula , , and γ is a learnable scaling factor. The second linear layer is used to transform The number of channels is adjusted from 4Cp to Cp, resulting in the feature ; The layer scaling layer and the random depth regularization layer sequentially perform layer scaling and random depth regularization operations on the feature to obtain a feature ; The second deformation layer is used to change From shape two to shape one, we get .

4. The radar nowcasting of precipitation method based on frequency domain perception and conditional diffusion according to claim 1, characterized in that, In S5, one round of training comprises steps S51-S53; S51, obtain a batch of first time period image sequences from D1 to constitute a condition radar echo sequence X cond a batch of second time period image sequences constitute a target radar echo sequence X0; S52, X cond , X0 are encoded into latent space to get corresponding latent condition sequence Z cond , latent target sequence Z0, by adding Gaussian noise to Z0 at each time step by a forward diffusion process to get noise sequence Z T , Z cond , Z T are converted to shape one respectively to get corresponding feature sequence , , the time step is encoded into shape one to get time sequence T s , T s , , are spliced along the channel dimension and then mapped to shape one to get joint noise sequence Z mix ; S53, Z mix input improved denoising network, output predicted noise sequence, convert to [B, S, Cz, Hz, Wz] again, get 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.

5. The method for radar nowcasting of precipitation based on frequency domain perception and conditional diffusion according to claim 1, characterized in that, S6 specifically comprises S61-S63; S61, obtaining a conditional radar echo sequence for prediction inputting a nowcasting prediction model, encoding and converting shape one to obtain a corresponding feature sequence , randomly sampling a Gaussian noise sequence T , converting shape one to obtain a corresponding feature sequence ; S62, T s 、 、 constitute a joint noise sequence , the improved denoising network predicts the noise sequence and converts to [B, S, Cz, Hz, Wz], to get the predicted noise, from T Subtract the predicted noise from the middle to get the predicted potential target sequence ; S63, to The predicted target radar echo sequence is generated by the VAE decoder.

Images