High-resolution radar echo extrapolation prediction method based on fused satellite data

Abstract

The application discloses a kind of high-resolution radar echo extrapolation prediction methods of fusion satellite data, specifically as follows, first, input history radar echo sequence pretreatment of previous T time, including denoising, normalization processing, data set segmentation, obtain cleaned data;Then, by deterministic modeling method (SimVP), obtain the fuzzy prediction sequence of future T length, then variational autoencoder (VAE) respectively original radar echo image and fuzzy prediction sequence are mapped to low-dimensional latent space, and two-stage diffusion modeling is carried out on this basis;For the first stage, utilize space-time converter (ST-Translator) to extract the space-time evolution characteristics of radar echo;Second stage first input corresponding time satellite data of previous T time, pretreatment is carried out, including normalization processing, feature selection, data set segmentation, obtain cleaned data, adopt multi-source fusion denoising network Fsrformer, dynamically adjust the influence of satellite data in diffusion process, to make full use of satellite information;Finally, the output result of two stages is inversely transformed to pixel space, and the high-resolution radar echo extrapolation prediction result of future T length is obtained.The application can effectively reduce the consumption of computing resources, improve the precision and detail fidelity of short-term precipitation prediction.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a multi-source data fusion method for short-term precipitation forecasting that integrates satellite data. Background Technology

[0002] Short-term precipitation forecasting, as a core technology of meteorological forecasting, is of crucial significance for agricultural production, urban safety, and disaster early warning. Statistics show that extreme precipitation-related disasters cause hundreds of billions of yuan in economic losses annually in my country, making accurate prediction of precipitation patterns within the next 0-2 hours an urgent need.

[0003] Currently, radar echo extrapolation is the mainstream prediction method, but it faces two major technical bottlenecks: First, most current radar echo extrapolation methods use single radar echo data, limiting their ability to characterize the formation and dissipation mechanisms of precipitation systems. In contrast, satellite remote sensing data can provide more comprehensive atmospheric information, such as cloud top temperature, cloud water content, and water vapor distribution, which helps improve the model's ability to understand precipitation occurrence and development. Second, diffusion models struggle to achieve high-resolution radar echo extrapolation. Due to the multi-step denoising and noise-adding mechanisms of diffusion models, computational costs are high, making it difficult for the model to run efficiently at high resolutions. This is particularly evident in long-term inference tasks, where computational resource consumption is significant, thus limiting the model's application in real-world business scenarios.

[0004] To address the aforementioned issues, this invention proposes a multi-source data fusion prediction method (SrCast) based on a combination of variational autoencoders and diffusion models to achieve efficient potential diffusion modeling. Through two-stage potential spatial learning and the proposed satellite data fusion strategy, the accuracy and computational efficiency of high-resolution precipitation forecasts are improved. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-source data fusion method for short-term precipitation prediction based on variational autoencoder (VAE) and diffusion model, which solves the computational efficiency problem of high-resolution radar echo extrapolation, breaks through the prediction limitations of a single data source, and improves the prediction accuracy of convective initiation and heavy precipitation events.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a high-resolution radar echo extrapolation prediction method based on fused satellite data, comprising the following steps:

[0007] S1: Radar data input and preprocessing. Input the historical radar echo sequence from the previous T, perform denoising, normalization, and dataset segmentation to obtain cleaned data.

[0008] S2: Deterministic modeling method for fuzzy prediction and variational autoencoder learning. The deterministic modeling method (SimVP) obtains a fuzzy prediction sequence for the future time period T. A variational autoencoder is used to encode the radar echo image and the fuzzy prediction image respectively, mapping them to a low-dimensional latent space;

[0009] S3: Latent space spatiotemporal transformer learning, extracting spatiotemporal evolution features from the latent representation through the designed spatiotemporal transformer module ST-Translator, providing high-quality input to the next stage;

[0010] S4: Satellite data input and preprocessing. Input the satellite data for the corresponding time T before the input, and perform preprocessing, including normalization, feature selection, and dataset segmentation, to obtain cleaned data.

[0011] S5: Multi-source fusion diffusion learning in the latent space adopts the multi-source fusion denoising network Fsrformer to dynamically adjust the influence of satellite data during the diffusion process, so that the model can make full use of the guiding role of satellite data in precipitation forecasting.

[0012] S6: Inverse transformation and output of results. The variational autodecoder is used to inversely transform the prediction results in the latent space to the pixel space to obtain the high-resolution radar echo extrapolation prediction results for the future time T.

[0013] In step S1 above, denoising processing is first performed on the radar echo data. Outliers are filtered by setting thresholds of 15dBZ and 70dBZ. Non-precipitation clutter below 15dBZ and super-refractive outliers above 70dBZ are uniformly set to 0 to retain effective information of the real precipitation area. Then, the processed radar echo intensity is mapped to the pixel range of 0-255 through linear transformation, and the pixel values ​​are further normalized to the [0,1] range using the min-max normalization method to provide standardized data for subsequent model input. At the same time, 12 consecutive frames of radar echo data are extracted at 10-minute intervals to form a sample sequence, with the first 6 frames as input and the last 6 frames as labels, and divided into training set, validation set and test set in a ratio of 7:2:1.

[0014] In step S2 above, the input radar echo sequence is first pre-extrapolated using a deterministic modeling method (using SimVP based on a CNN model) to obtain a pixel-space fuzzy prediction. The frame-by-frame variational autoencoder sequentially encodes and compresses the pixel-space input and fuzzy prediction respectively, obtaining the input radar features and fuzzy prediction features in the latent space, which are used for subsequent latent space learning. The encoder and decoder in the VAE adopt a U-Net-based upsampling and downsampling structure. The optimization objective of VAE training is divided into three parts: KL divergence loss, reconstruction loss, and GAN loss. For the encoder, the goal is to learn the mapping from input data x to latent variable z, ensuring its distribution approximates a standard normal distribution N(0,I), thus maintaining the continuity of the latent space. This process assumes that the latent variable z follows a Gaussian distribution, i.e., z ~ N(μ(x), σ(x)), where μ and σ are output by the encoder. To make E φ (z|x) approximates the standard normal distribution p(z)=N(0,I), therefore the KL divergence loss can be calculated as follows:

[0015]

[0016] Where L KL D represents the KL divergence loss value. KL `` is the symbol for KL divergence, used to represent the degree of difference between two probability distributions. `` represents the approximate posterior distribution, `E` represents the mathematical expectation, `z` is the latent variable, `x` represents the original input data, and `σ` are the model parameters. `` is the prior probability distribution of the latent variable `z`, `i` is the index of the summation indicating traversal of the dimensions, and `d` is the dimension of the latent variable. This represents the mean of the approximate posterior distribution along the i-th dimension. It indicates the central location of the variable along that dimension. KL divergence loss makes the distribution of latent variables as close as possible to the standard normal distribution to ensure the continuity of the generator space, enabling VAE to perform smooth interpolation.

[0017] For the decoder, the goal is to make the radar echo image x′ generated from the latent variable z as close as possible to the original input data x, so the mean square error (MSE) is used as the reconstruction loss:

[0018]

[0019] L rec Let E represent the reconstruction loss, and D represent the mathematical expectation. θ The distribution defined by parameter θ, where D θ(z|x) is the conditional distribution of the latent variable z given the original input data x. x′ represents the generated data. To improve the realism of the generated images, the VAE in SrCast introduces a GAN structure. That is, the decoder needs to minimize the reconstruction error of the VAE and also deceive the discriminator C to make the generated samples look like real data, i.e., maximize the discriminator C's score on the generated samples.

[0020] L GAN = -E[logC(x′)]

[0021] L GAN Let E represent the loss function of the generator in the generative adversarial network, C represent the discriminator, and x′ represent the generated data. The discriminator's task is to distinguish between real and generated data, requiring independent optimization of its loss. A binary classification cross-entropy loss is used as follows:

[0022] L C =-E[logC(x1)]-E[log(1-C(x′))]

[0023] L C Let x' represent the loss function of the discriminator, C represent the discriminator, x1 represent the real data, i.e., samples from the true distribution, and x′ represent the generated data.

[0024] L VAE =λ1L KL +λ2L rec +λ3L GAN

[0025] Where λ1, λ2, and λ3 are the weighting coefficients of the three types of loss with respect to the total loss, respectively. Step two, compared to existing diffusion model prediction methods, maps high-resolution radar data to a low-dimensional space, reducing computational costs.

[0026] In step S3 above, a spatiotemporal converter (ST-Translator) was designed to model the deterministic features of precipitation evolution in the potential space. The ST-Translator employs a multi-level ST-Inception structure, combining local spatiotemporal feature extraction with multi-scale information aggregation to enhance the learning ability of the precipitation evolution model. The ST-Translator mainly consists of multiple ST-Inception modules. The VAE-encoded potential input radar features z i and fuzzy prediction features z pAfter being stitched together, the data serves as input to the ST-Translator. Through the progressive stacking of multiple ST-Inception layers, local spatiotemporal features are extracted, thereby capturing the spatial structure and temporal evolution patterns of the precipitation system. The deterministic latent representation information z after processing by the ST-Translator is then obtained. o As the final output of the first stage, it is used for further modeling of the diffusion model in the second stage. Compared with existing diffusion model methods, step three improves the stability of the diffusion model input data by designing the ST-Translator module to calculate the MSE with real data, thereby reducing the difference between the prediction results and the real data and improving the final forecast accuracy.

[0027] In step S4 above, satellite data is processed. Based on meteorological principles and correlation analysis, four infrared channels sensitive to precipitation system evolution—the water vapor channel (WV), mid-infrared channel (MI), infrared channel 1 (IR1), and infrared channel 2 (IR2)—are selected, while redundant visible light channels are removed. The satellite data is resampled from its original resolution to a resolution consistent with radar data using an algorithm, and timestamps are strictly aligned to ensure spatiotemporal synchronization. A channel-specific minimum-maximum normalization method is used to process the data from each satellite channel, eliminating dimensional differences and then pairing it with radar data chronologically to form a multi-source fusion training sample set. This provides spatiotemporally aligned and physically meaningful input data for subsequent potential spatial modeling. Compared to existing single radar echo extrapolation methods, step four incorporates satellite data and supplements atmospheric features such as cloud cover, further improving the accuracy of precipitation forecasts.

[0028] In step S5 above, the output of the first stage is accepted as input for iterative noise addition and denoising training. To further integrate additional observation information from satellites, Fsrformer, a multi-source fusion denoising network based on multi-layer SR-DiT stacking, was invented. Its inputs include the noise-added latent radar features (output of the first stage), time-step encoding, and latent satellite features. The latent satellite features are extracted from the original satellite data using a CNN spatial encoder similar to a variational autoencoder, thereby achieving spatial consistency between satellites and radar to facilitate information fusion in the latent space. The latent radar features and latent satellite features are first processed by Patchify, converting radar and satellite data into fixed-size token representations, then mapped to a high-dimensional latent space through a linear embedding layer, and subsequently fed into the SR-DiT module for cross-modal feature fusion and spatiotemporal relationship modeling.

[0029] The SR-DiT module employs a time-step-dependent scaling-shift transformation (SST) to fine-tune the input, adaptively adjusting the feature distribution at different time steps t. In the self-attention stage, the input radar token x first undergoes layer normalization (LN), followed by the introduction of time-step-dependent scaling and shift transformations to generate initial transformed features.

[0030] x′=γ1(t)·LN(x)+β1(t)

[0031] Where γ1(t) and β1(t) are learnable parameters dependent on time step t, which can adaptively adjust the distribution of input features. Then, spatiotemporal relationship modeling is performed through a multi-head self-attention (MHSA) module, and its calculation formula is as follows:

[0032]

[0033] Where Q, K, V represent linear projections onto x′, d1 represents the feature dimensions of Q and K, the output of the self-attention layer is adjusted by a scaling factor α1(t), and the first-stage output is formed through residual connections:

[0034] h1=α1(t)·MHSA(x′)+x′

[0035] h1 represents the output hidden state, α1 represents the weight coefficient, and t represents the time step. In the multi-source data fusion stage, to fully utilize the supplementary role of satellite data to radar information, the input satellite Tokenx... s Similarly, after layer normalization, feature adjustment is performed using another set of time-step-dependent scaling-offset transformations:

[0036] x s =γ2(t)·LN(x) s )+β2(t).

[0037] Next, x s After transformation, the output variable γ2 is the second scaling factor, LN represents layer normalization, t represents the time step, and h1 is used as the query Q. r The satellite feature x adjusted by scaling-offset transformation s ′ respectively as bond K s Sum V s The calculation process for the input satellite radar cross-attention module (SRCA) is as follows:

[0038] Q r =W q h1,K s =W k x s ′,Vs =W v x′

[0039] h2=α2(t)·Attention(Q r ,K s V s )+h1

[0040] Among them, W q W k W v These are the linear projection weights, and the scaling factor α2(t) is also dynamically adjusted according to time step t. In the feedforward network (FFN) stage, the fused features h2 are normalized again and preprocessed using another set of time-step related scaling-offset transformations:

[0041] h2′=γ3(t)·LN(h2)+β3(t)

[0042] Where h2′ represents the preprocessed features, γ3 represents the third scaling factor, LN is the layer normalization, h2 represents the input hidden state after layer normalization, β3 represents the offset factor, and t represents the time step. A point-based feedforward network is used for nonlinear mapping, and after adjustment with the scaling factor α3(t), a residual connection is made with the original features. The final output is:

[0043] h3=α3(t)·FFN(h2′)+h2

[0044] Where h3 represents the final output feature, α3 represents the weighting coefficient, and h2 represents the layer-normalized input hidden state. Step five, compared to the general diffusion model prediction method with a single radar echo data input, adds a satellite data input channel and designs a satellite weighting strategy that adjusts with time steps, thereby improving precipitation forecast accuracy.

[0045] In step S6 above, the sampled feature map is used as the decoder D. θ The input (x|z) is converted back to the original data space to obtain the reconstructed radar echo image. The upsampling part corresponds to the VAE decoder, which improves the image resolution by gradually upsampling from low-dimensional features to restore the image resolution and complete the reconstruction.

[0046] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned high-resolution radar echo extrapolation prediction method based on fused satellite data.

[0047] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the aforementioned high-resolution radar echo extrapolation prediction method based on fused satellite data.

[0048] Compared with the prior art, the advantages of the present invention are as follows:

[0049] 1) This invention addresses the computational burden problem in the multiple iterations of denoising in diffusion models. By introducing a VAE, this invention moves the diffusion process from the pixel space to the latent space, performs diffusion generation in a low-dimensional representation, reduces computational costs, and enhances the learning ability of key structures.

[0050] 2) This invention makes full use of the prior information of satellites to supplement the limitations of single radar extrapolation, and improves both prediction accuracy and image detail, especially in the prediction of heavy precipitation events. At the same time, it can effectively predict potential convective initiation.

[0051] 3) The present invention uses a diffusion model to solve the problems that existing deep learning models generally have, such as overly smooth predicted images, lack of detail, low accuracy, and a significant decrease in prediction timeliness as the prediction time increases, thereby improving prediction accuracy and detail fidelity.

[0052] 4) This invention introduces the multi-source fusion denoising network Fsrformer, which dynamically adjusts the influence of satellite data through the SR-DiT module, enabling the model to fully utilize satellite information in the initial stage of convection. This significantly enhances the model's sensitivity to and early warning capability for the initial stage of convection. Attached Figure Description

[0053] Figure 1 This is a flowchart of the high-resolution radar echo extrapolation method based on fused satellite data according to the present invention.

[0054] Figure 2 This is a schematic diagram of the radar data preprocessing of the present invention.

[0055] Figure 3 This is a schematic diagram of satellite data preprocessing according to the present invention.

[0056] Figure 4 This is a schematic diagram of the high-resolution radar echo extrapolation method based on fused satellite data according to the present invention.

[0057] Figure 5 This is a detailed structural diagram of the potential space ST-Translator in this invention.

[0058] Figure 6 This is a schematic diagram showing the detailed structure of the Fsrformer network in this invention. Detailed Implementation

[0059] To enhance understanding of the present invention, further description of the invention is provided below in conjunction with the accompanying drawings and specific embodiments.

[0060] Example 1: This specific implementation discloses a high-resolution radar echo extrapolation method based on fused satellite data, such as... Figures 1-5 As shown, it includes the following steps:

[0061] S1: Radar data input and preprocessing. Input the historical radar echo sequence from the previous T, perform denoising, normalization, and dataset segmentation to obtain cleaned data.

[0062] S2: Deterministic modeling method for fuzzy prediction and variational autoencoder learning. The deterministic modeling method (SimVP) obtains a fuzzy prediction sequence for the future time period T. A variational autoencoder is used to encode the radar echo image and the fuzzy prediction image respectively, mapping them to a low-dimensional latent space;

[0063] S3: Latent space spatiotemporal transformer learning, extracting spatiotemporal evolution features from the latent representation through the designed spatiotemporal transformer module ST-Translator, providing high-quality input to the next stage;

[0064] S4: Satellite data input and preprocessing. Input the satellite data for the corresponding time T before the input, and perform preprocessing, including normalization, feature selection, and dataset segmentation, to obtain cleaned data.

[0065] S5: Multi-source fusion diffusion learning in the latent space adopts the multi-source fusion denoising network Fsrformer to dynamically adjust the influence of satellite data during the diffusion process, so that the model can make full use of the guiding role of satellite data in precipitation forecasting.

[0066] S6: Inverse transformation and output of results. The variational autodecoder is used to inversely transform the prediction results in the latent space to the pixel space to obtain the high-resolution radar echo extrapolation prediction results for the future time T.

[0067] In step S1, such as Figure 2 As shown, denoising processing is first performed on the radar echo data. Outliers are filtered by setting thresholds of 15dBZ and 70dBZ. Non-precipitation clutter below 15dBZ and superrefractive outliers above 70dBZ are uniformly set to 0 to retain effective information of the real precipitation area. Then, the processed radar echo intensity is mapped to the pixel range of 0-255 through linear transformation, and the pixel values ​​are further normalized to the [0,1] range using the min-max normalization method to provide standardized data for subsequent model input. At the same time, 12 consecutive frames of radar echo data are extracted at 10-minute intervals to form a sample sequence, with the first 6 frames as input and the last 6 frames as labels, divided into training set, validation set and test set in a ratio of 7:2:1.

[0068] In step S2, such as Figure 4 As shown, the input radar echo sequence is first pre-extrapolated using a deterministic modeling method (using SimVP based on a CNN model) to obtain a pixel-space fuzzy prediction. The frame-by-frame variational autoencoder sequentially encodes and compresses the pixel-space input and fuzzy prediction respectively, obtaining the input radar features and fuzzy prediction features in the latent space, which are used for subsequent latent space learning. The encoder and decoder in the VAE adopt a U-Net-based upsampling and downsampling structure. The optimization objective of VAE training is divided into three parts: KL divergence loss, reconstruction loss, and GAN loss. For the encoder, the goal is to learn the mapping from input data x to latent variable z, ensuring its distribution approximates the standard normal distribution N(0,I), thus maintaining the continuity of the latent space. This process assumes that the latent variable z follows a Gaussian distribution, i.e., z~N(μ(x),σ(x)), where μ and σ are output by the encoder. To make E φ (z|x) approximates the standard normal distribution p(z)=N(0,I), therefore the KL divergence loss can be calculated as follows:

[0069]

[0070] Where L KL D represents the KL divergence loss value. KL `` is the symbol for KL divergence, used to represent the degree of difference between two probability distributions. `` represents the approximate posterior distribution, `E` represents the mathematical expectation, `z` is the latent variable, `x` represents the original input data, and `σ` are the model parameters. `` is the prior probability distribution of the latent variable `z`, `i` is the index of the summation indicating traversal of the dimensions, and `d` is the dimension of the latent variable. This represents the mean of the approximate posterior distribution along the i-th dimension. It indicates the central location of the variable along that dimension. KL divergence loss makes the distribution of latent variables as close as possible to the standard normal distribution to ensure the continuity of the generator space, enabling VAE to perform smooth interpolation.

[0071] For the decoder, the goal is to make the radar echo image x′ generated from the latent variable z as close as possible to the original radar echo image x, so the mean square error (MSE) is used as the reconstruction loss:

[0072]

[0073] L rec Let E represent the reconstruction loss, and D represent the mathematical expectation. θ The distribution defined by parameter θ, where D θ(z|x) is the conditional distribution of the latent variable z given the original input data x. x′ represents the generated data. To improve the realism of the generated images, the VAE in SrCast introduces a GAN structure. That is, the decoder needs to minimize the reconstruction error of the VAE and also deceive the discriminator C to make the generated samples look like real data, i.e., maximize the discriminator C's score on the generated samples.

[0074] L GAN = -E[logC(x′)]

[0075] L GAN Let E represent the loss function of the generator in the generative adversarial network, C represent the discriminator, and x′ represent the generated data. The discriminator's task is to distinguish between real and generated data, requiring independent optimization of its loss. A binary classification cross-entropy loss is used as follows:

[0076] L C =-E[logC(x1)]-E[log(1-C(x′))]

[0077] L C Let x' represent the loss function of the discriminator, C represent the discriminator, x1 represent the real data, i.e., samples from the true distribution, and x′ represent the generated data.

[0078] L VAE =λ1L KL +λ2L rec +λ3L GAN

[0079] Where λ1, λ2, and λ3 are the weighting coefficients of the three types of loss with respect to the total loss, respectively. Step two, compared to existing diffusion model prediction methods, innovatively maps high-resolution radar data to a low-dimensional space, reducing computational costs.

[0080] In step S3, such as Figure 5 As shown, a spatiotemporal transducer (ST-Translator) was designed to model the deterministic features of precipitation evolution in the latent space. The ST-Translator employs a multi-level ST-Inception structure, combining local spatiotemporal feature extraction with multi-scale information aggregation to enhance the learning ability of precipitation evolution models. The ST-Translator mainly consists of multiple ST-Inception modules. The latent input radar features z after VAE encoding are shown. i and fuzzy prediction features z pAfter being stitched together, the data serves as input to the ST-Translator. Through the progressive stacking of multiple ST-Inception layers, local spatiotemporal features are extracted, thereby capturing the spatial structure and temporal evolution patterns of the precipitation system. The deterministic latent representation information z after processing by the ST-Translator is then obtained. o This serves as the final output of the first stage and is used for further modeling of the diffusion model in the second stage.

[0081] In step S4, such as Figure 3 As shown, satellite data was processed, and four infrared channels sensitive to the evolution of precipitation systems—the water vapor channel (WV), mid-infrared channel (MI), infrared channel 1 (IR1), and infrared channel 2 (IR2)—were selected based on meteorological principles and correlation analysis. Redundant visible light channels were removed. The satellite data was resampled from its original resolution to a resolution consistent with radar data using an algorithm, and the timestamps were strictly aligned to ensure spatiotemporal synchronization. The sub-channel min-max normalization method was used to process the data of each satellite channel, eliminating dimensional differences, and then paired with radar data one by one in chronological order to form a multi-source fusion training sample set. This provides spatiotemporally aligned and physically meaningful input data for subsequent potential spatial modeling.

[0082] In step S5, such as Figure 6 As shown, the output of the first stage is accepted as input for iterative training involving noise addition and denoising. To further integrate additional observation information from satellites, Fsrformer, a multi-source fusion denoising network based on multi-layer SR-DiT stacking, was invented. Its inputs include the noisy latent radar features (output of the first stage), time-step encoding, and latent satellite features. The latent satellite features are extracted from the original satellite data using a CNN spatial encoder similar to a variational autoencoder, thereby achieving spatial consistency between satellites and radar to facilitate information fusion in the latent space. The latent radar and latent satellite features are first processed by Patchify, converting radar and satellite data into fixed-size token representations, then mapped to a high-dimensional latent space through a linear embedding layer, and subsequently fed into the SR-DiT module for cross-modal feature fusion and spatiotemporal relationship modeling.

[0083] The SR-DiT module employs a time-step-dependent scaling-shift transformation (SST) to fine-tune the input, adaptively adjusting the feature distribution at different time steps t. In the self-attention stage, the input radar token x first undergoes layer normalization (LN), followed by the introduction of time-step-dependent scaling and shift transformations to generate initial transformed features.

[0084] x′=γ1(t)·LN(x)+β1(t)

[0085] Where γ1(t) and β1(t) are learnable parameters dependent on time step t, which can adaptively adjust the distribution of input features. Then, spatiotemporal relationship modeling is performed through a multi-head self-attention (MHSA) module, and its calculation formula is as follows:

[0086]

[0087] Where Q, K, V represent linear projections onto x′, d1 represents the feature dimensions of Q and K, the output of the self-attention layer is adjusted by a scaling factor α1(t), and the first-stage output is formed through residual connections:

[0088] h1=α1(t)·MHSA(x′)+x′

[0089] h1 represents the output hidden state, α1 represents the weight coefficient, and t represents the time step. In the multi-source data fusion stage, to fully utilize the supplementary role of satellite data to radar information, the input satellite Tokenx... s Similarly, after layer normalization, feature adjustment is performed using another set of time-step-dependent scaling-offset transformations:

[0090] x s =γ2(t)·LN(x) s )+β2(t).

[0091] Next, x s After transformation, the output variable γ2 is the second scaling factor, LN represents layer normalization, t represents the time step, and h1 is used as the query Q. r The satellite feature x adjusted by scaling-offset transformation s ′ respectively as bond K s Sum V s The calculation process for the input satellite radar cross-attention module (SRCA) is as follows:

[0092] Q r =W q h1,K s =W k x s ′,V s =W v x′

[0093] h2=α2(t)·Attention(Q r ,K s V s )+h1

[0094] Among them, W q W k W vThese are the linear projection weights, and the scaling factor α2(t) is also dynamically adjusted according to time step t. In the feedforward network (FFN) stage, the fused features h2 are normalized again and preprocessed using another set of time-step related scaling-offset transformations:

[0095] h2′=γ3(t)·LN(h2)+β3(t)

[0096] Where h2′ represents the preprocessed features, γ3 represents the third scaling factor, LN is the layer normalization, h2 represents the input hidden state after layer normalization, β3 represents the offset factor, and t represents the time step. A point-based feedforward network is used for nonlinear mapping, and after adjustment with the scaling factor α3(t), a residual connection is made with the original features. The final output is:

[0097] h3=α3(t)·FFN(h2′)+h2

[0098] Where h3 represents the final output feature, α3 represents the weight coefficient, and h2 represents the normalized input hidden state of the layer.

[0099] In step S6, such as Figure 4 As shown, the sampled feature map is used as the decoder D. θ The input (x|z) is converted back to the original data space to obtain the reconstructed radar echo image. The upsampling part corresponds to the VAE decoder, which improves the image resolution by gradually upsampling from low-dimensional features to restore the image resolution and complete the reconstruction.

[0100] Example 2:

[0101] The radar echo dataset used for validation in this method comes from the Nanjing Institute of Meteorological Science and Technology Innovation (NJIAS), covering the period from 2019 to 2021. It includes radar echo data from June, July, and August each year, covering parts of Jiangsu Province with an area of ​​300 km × 300 km (spatial resolution of 1 km). The radar scan interval is fixed at 10 minutes, resulting in 144 radar echo images per day. Simultaneously, meteorological satellite data corresponding to the same time period is used, sourced from the Japan Meteorological Agency's geostationary meteorological satellite Himawari-8 (H8). H8's observation range covers the entire Eastern Hemisphere, including parts of Asia, Australia, the Pacific Ocean, and the Indian Ocean. Its observation frequency is once every 10 minutes across the Eastern Hemisphere, acquiring data 6 times per hour, consistent with the temporal resolution of the NJIAS radar echo data. H8 is equipped with 16 spectral channels, including 6 visible light channels (including 3 near-infrared channels) and 10 infrared channels. The data from June and July 2019 were used as the validation set and test set, respectively, while the rest were used as the training set.

[0102] This invention was experimentally compared with two RNN-based precipitation prediction methods, PredRNNV2 and MAU; two diffusion model-based precipitation prediction methods, SCRD and CCascast; one CNN-based precipitation prediction method, SimVP; and one Transformer-based precipitation prediction method, Earthformer. As shown in Tables 1, 2, 3, 4, 5, and 6, the invention achieved optimal performance in CSI, HSS, and FAR indices, and SSIM, Sharpness, and PSNR indices, at thresholds greater than 40 dBZ.

[0103] Table 1 Accuracy Prediction Indicators for Radar Echo Intensity Greater Than 25 dBZ

[0104]

[0105] Table 2 Accuracy Prediction Indicators for Radar Echo Intensity Greater Than 35 dBZ

[0106]

[0107]

[0108] Table 3 Accuracy Prediction Indicators for Radar Echo Intensity Greater Than 40 dBZ

[0109]

[0110] Table 4 Accuracy Prediction Indicators for Radar Echo Intensity Greater Than 45 dBZ

[0111]

[0112] Table 5 Accuracy Prediction Indicators for Radar Echo Intensity Greater Than 50 dBZ

[0113]

[0114]

[0115] Table 6 Comparison of Experimental Results for Image Quality Indicators

[0116]

[0117] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A high-resolution radar echo extrapolation prediction method fused with satellite data, characterized in that: The method includes the following steps: S1: Radar data input and preprocessing. Input the historical radar echo sequence from the previous T, perform denoising, normalization, and dataset segmentation to obtain cleaned data. S2: Deterministic modeling method fuzzy prediction and variational autoencoder learning. The deterministic modeling method (SimVP) obtains a fuzzy prediction sequence for the future time T. The variational autoencoder is used to encode the radar echo image and the fuzzy prediction image respectively, and map them to a low-dimensional latent space. S3: Latent space spatiotemporal transformer learning, extracting spatiotemporal evolution features from the latent representation through the designed spatiotemporal transformer module ST-Translator, providing high-quality input to the next stage; S4: Satellite data input and preprocessing. Input the satellite data for the corresponding time T before the input, and perform preprocessing, including normalization, feature selection, and dataset segmentation, to obtain cleaned data. S5: Multi-source fusion diffusion learning in the latent space adopts the multi-source fusion denoising network Fsrformer to dynamically adjust the influence of satellite data during the diffusion process, so that the model can make full use of the guiding role of satellite data in precipitation forecasting. S6: Inverse transformation and output of results. The variational autodecoder is used to inversely transform the prediction results in the latent space to the pixel space to obtain the high-resolution radar echo extrapolation prediction results for the future time T.

2. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: The data preprocessing method in step S1 is as follows: S11: Input data: Process historical radar echo sequences, including radar echo sequence sequences from the past hour; S12: Denoising: Set the radar echo intensity of radar image points below 15dBZ and above 70dBZ to 0; S13: Normalization processing; For radar echo data, first, the minimum and maximum values ​​in the training set are calculated, and then the normalization formula is applied to map it to the [0,1] interval.

3. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: The specific steps of step S2 are as follows: S21: Deterministic Modeling Preprocessing: Before inputting into the VAE, the radar echo sequence is pre-extrapolated using a CNN-based SimVP (Simple Video Prediction) model to obtain a fuzzy preliminary prediction result. This step aims to provide an initial estimate to help the VAE better capture the temporal evolution characteristics of the radar echo. S22: Variational Autoencoder Learning: A variational autoencoder (VAE) is used to encode the radar echo image and the blurred prediction image frame by frame to learn their low-dimensional latent representation. The VAE consists of an encoder and a decoder. The encoder is responsible for converting the radar echo sequence into a latent variable distribution, while the decoder is used to map the generated latent variables back to the pixel space during the inference stage. S23: Latent Variable Optimization: For the discriminator, its task is to distinguish between real data and generated data, requiring independent optimization of its loss. A binary cross-entropy loss is used; therefore, the complete loss function in a VAE is as follows: Where L KL D represents the KL divergence loss value. KL `` is the symbol for KL divergence, used to represent the degree of difference between two probability distributions; `` represents the approximate posterior distribution; `E` represents the mathematical expectation; `z` is the latent variable; `x` represents the original input data; `σ` are the model parameters; `` is the prior probability distribution of the latent variable `z`; `i` is the index of the summation, indicating traversal of the dimensions; and `d` is the dimension of the latent variable. Indicates that in the i-th dimension, For the decoder, the goal is to make the radar echo image x′ generated from the latent variable z as close as possible to the original input data x, so the mean square error (MSE) is used as the reconstruction loss: L rec Let E represent the reconstruction loss, and D represent the mathematical expectation. θ The distribution defined by parameter θ, where D θ (z|x) is the conditional distribution of the latent variable z given the original input data x, and x′ represents the generated data. To improve the realism of the generated images, the VAE in SrCast introduces a GAN structure. That is, the decoder needs to minimize the reconstruction error of the VAE and also deceive the discriminator C to make the generated samples look like real data, i.e., maximize the discriminator C's score on the generated samples. L GAN =-E[logC(x′)] L GAN Let E represent the loss function of the generator in the generative adversarial network, C represent the discriminator, and x′ represent the generated data. The discriminator's task is to distinguish between real and generated data, requiring independent optimization of its loss. A binary classification cross-entropy loss is used as follows: THE C =-E[logC(x1)]-E[log(1-C(x′))] L C Let C represent the discriminator's loss function, x1 represent the real data (samples from the true distribution), and x′ represent the generated data. L VAE =λ1L KL +λ2L rec +λ3L GAN Where λ1, λ2, and λ3 are the weighting coefficients of the three types of loss with respect to the total loss, respectively. S24: Generation of latent spatial features: After the above steps, the latent variables obtained represent the low-dimensional features of the radar echo image. These features contain both the temporal evolution information of the radar echo and the spatial details, providing high-quality initial input for the subsequent diffusion model.

4. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: Step S3 mainly includes the following steps: S31: Latent feature splicing and normalization. The latent radar features generated by VAE and the fuzzy prediction features pre-extrapolated by SimVP are spliced ​​together in the channel dimension to form spatiotemporal input features. Layer normalization is performed on the spliced ​​features to eliminate scale differences and improve model stability. S32: The ST-Inception module extracts features by using grouped convolutions to extract multi-scale spatial features, expanding the receptive field to capture the different scale structures of precipitation systems. It introduces a channel attention mechanism, learning channel weights through global pooling and fully connected layers to enhance the representation of key meteorological features. S33: Differential dispersion regularization calculates the inter-frame variation of predicted and actual latent features through inter-frame differencing, capturing the dynamic evolution of the precipitation system. Softmax is used to convert the difference into a probability distribution, and KL divergence constraints are applied to ensure the predicted distribution closely approximates reality, enhancing sensitivity to temporal variations. S34: Multi-level stacking and output. Spatiotemporal features are extracted layer by layer through multi-level ST-Inception modules. Residual connections are used to avoid gradient vanishing. High-level spatiotemporal representations are gradually refined. After layer normalization, spatiotemporal evolution features are output as input for subsequent diffusion models, thereby improving the ability to model temporal dependencies.

5. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: Step S4 mainly includes the following steps: S41: Input data: Process historical satellite data, including satellite observation data from the past hour; S42: Normalization Processing: For satellite observation data, first, the minimum and maximum values ​​in the training set are calculated, and then a normalization formula is applied to map it to the [0,1] interval. S43: Feature Selection: Using the Pearson correlation analysis method, the correlation between each satellite channel and the radar echo was evaluated, and channels with significant correlation were selected. Redundant information was removed, and infrared channel 7 (I4), water vapor channels (WV, W2, W3) and infrared channels 11-16 (MI, O3, IR, L2, I2, CO) were grouped. Channels with high correlation were selected for subsequent modeling. Finally, channels 8 (WV), 11 (MI), 13 (IR), and 15 (I2) were selected.

6. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: Step S5 mainly includes the following steps: S51: Multi-source input feature preparation involves adding noise to the radar latent features output by the ST-Translator using a Gaussian diffusion process to generate a noisy latent sequence; independently encoding the preprocessed satellite data into latent features of the same dimension; and encoding the diffusion time step into a sinusoidal position vector. These three elements together constitute the Fsrformer input. The noise addition process follows a preset cosine variance scheduling strategy to ensure that the noise intensity changes reasonably with the time step. S52: Tokenization and Feature Embedding. The noisy radar and satellite latent features are divided into fixed-size patches, mapped to a high-dimensional latent space through a linear layer, and positional encoding is added to form a token sequence containing spatiotemporal location information. This provides a unified format of input features for subsequent Transformer processing, enhancing the model's ability to perceive spatial structure and temporal order. S53: Reverse denoising diffusion generation starts with Gaussian noise and generates potential radar features through 1000 steps of reverse diffusion iteration. Each step uses Fsrformer to predict noise and update features. An improved DDPM method is adopted to support batch processing of multiple time steps to improve efficiency. While gradually removing noise, satellite guidance information is injected to ensure that the generated potential features conform to the spatiotemporal evolution of precipitation systems.

7. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: Step S5 is as follows: The SR-DiT module employs a time-step-dependent scaling-shift transformation (SST) to fine-grainedly control the input, adaptively adjusting the feature distribution at different time steps t. In the self-attention stage, the input radar token x first undergoes layer normalization (LN), followed by the introduction of time-step-dependent scaling and shift transformations to generate the initial transformation. feature: x′=γ1(t)·LN(x)+β1(t) Where γ1(t) and β1(t) are learnable parameters dependent on time step t, which can adaptively adjust the distribution of input features. Then, spatiotemporal relationship modeling is performed through a multi-head self-attention (MHSA) module, and its calculation formula is as follows: Where Q, K, V represent linear projections onto x′, d1 represents the feature dimensions of Q and K, the output of the self-attention layer is adjusted by a scaling factor α1(t), and the first-stage output is formed through residual connections: h1=α1(t)·MHSA(x′)+x′ h1 represents the output hidden state, α1 represents the weighting coefficient, and t represents the time step. In the multi-source data fusion stage, to fully utilize the supplementary role of satellite data to radar information, the input satellite... Similarly, after layer normalization, feature adjustment is performed using another set of time-step-dependent scaling-offset transformations: x′ s =γ2(t)·LN(x s )+β2(t). Next, x s After transformation, the output variable γ2 is the second scaling factor, LN represents layer normalization, t represents the time step, and h1 is used as the query Q. r The satellite feature x adjusted by scaling-offset transformation s ′ respectively as bond K s Sum V s The calculation process for the input satellite radar cross-attention module (SRCA) is as follows: Q r =In q h1,K s =In k x′ s ,V s =In v x′ h2=α2(t)·Attention(Q r ,K s ,V s )+h1 Among them, W q W k W v The linear projection weights are denoted as h2 and h3 respectively. The scaling factor α2(t) is also dynamically adjusted according to the time step t. In the feedforward network (FFN) stage, the fused feature h2 is normalized again and preprocessed using another set of time-step related scaling-offset transformations. h′2=γ3(t)·LN(h2)+β3(t) Where h2′ represents the preprocessed features, γ3 represents the third scaling factor, LN is the layer normalization, h2 represents the input hidden state after layer normalization, β3 represents the offset factor, and t represents the time step. A point-based feedforward network is used for nonlinear mapping, and after adjustment with the scaling factor α3(t), a residual connection is made with the original features. The final output is: h3=α3(t)·FFN(h′2)+h2 Where h3 represents the final output feature, α3 represents the weight coefficient, and h2 represents the normalized input hidden state of the layer.

8. The high-resolution radar echo extrapolation prediction method based on fused satellite data according to claim 1, characterized in that: Step S6 mainly includes the following steps: S61: Latent Feature Decoding and Reconstruction. The 128-dimensional radar latent features generated from the diffusion of future T-frames are input into a VAE decoder symmetrical to the encoder. This gradually restores the low-dimensional features into a 300×300 pixel radar echo image, outputting a T×300×300 dimension, achieving effective reconstruction of spatial resolution and detail information. S62: Prediction results output, generating a 60-minute prediction sequence at 10-minute frame intervals, with timestamps aligned with the input data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements a high-resolution radar echo extrapolation prediction method based on any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that: When executed by the processor, the computer instructions implement a high-resolution radar echo extrapolation prediction method based on any one of claims 1-8.

Images