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

Abstract

The invention discloses a high-resolution radar echo extrapolation prediction method fused with satellite data, and the method specifically comprises the following steps: firstly, inputting historical radar echo sequence preprocessing at a previous T moment, including denoising, normalization processing and data set segmentation, and obtaining cleaned data; then, through a deterministic modeling method (SimVP), a fuzzy prediction sequence of a future T duration is obtained, then a variational auto-encoder (VAE) maps an original radar echo image and the fuzzy prediction sequence to a low-dimensional potential space, and two-stage diffusion modeling is carried out on the basis; in the first stage, a space-time converter (ST-Translator) is used to extract space-time evolution characteristics of radar echoes; in the second stage, satellite data at the corresponding time of the previous T moment is input, preprocessing including normalization processing, feature selection and data set segmentation is carried out, cleaned data is obtained, and the influence of the satellite data is dynamically adjusted in the diffusion process by adopting a multi-source fusion denoising network Fsrform so as to make full use of satellite information; and finally, inversely transforming output results of the two stages into a pixel space to obtain a high-resolution radar echo extrapolation prediction result of the future T duration. According to the invention, computing resource consumption can be effectively reduced, and the precision and detail fidelity of short temporary rainfall prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological forecasting, and in particular relates to a multi-source data fusion short-term precipitation prediction method that fuses satellite data. Background Art

[0002] As a core technology in weather forecasting, short-term precipitation prediction is crucial for agricultural production, urban safety, and disaster warning. According to statistics, extreme precipitation-induced disasters in my country cause hundreds of billions of yuan in economic losses each year, making accurate predictions of precipitation within the next 0-2 hours a pressing need.

[0003] Currently, radar echo extrapolation-based forecasting methods are the mainstream approach, but they face two major technical bottlenecks: Most current radar echo extrapolation methods use single radar echo data, which has limited ability to depict the generation 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 understanding of precipitation occurrence and development. Secondly, high-resolution radar echo extrapolation is difficult to achieve using diffusion models. Due to the diffusion model's multi-step denoising and denoising mechanism, the computational overhead is high, making it difficult for the model to run efficiently under high-resolution conditions. The consumption of computing resources is particularly significant in inference tasks involving longer time series, thus limiting the model's application in real-world business scenarios.

[0004] To address these issues, this paper proposes a multi-source data fusion prediction method (SrCast) based on a combination of a variational autoencoder and a diffusion model to achieve efficient latent diffusion modeling. Through two-stage latent space learning and a 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 the present invention is to provide a short-term precipitation prediction method based on multi-source data fusion of variational autoencoder (VAE) and diffusion model, to solve the computational efficiency problem of high-resolution radar echo extrapolation, to break through the prediction limitations of a single data source, and to improve the prediction accuracy of convective inception and heavy precipitation events.

[0006] To achieve the above-mentioned object, 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 previous T historical radar echo sequence for denoising, normalization, and data set segmentation to obtain cleaned data;

[0008] S2: Deterministic modeling method fuzzy prediction and variational autoencoder learning, a deterministic modeling method (SimVP) to obtain a fuzzy prediction sequence of the future time length T. The radar echo image and the fuzzy prediction image are encoded using a variational autoencoder and mapped into a low-dimensional latent space;

[0009] S3: Spatiotemporal transformer learning in latent space, 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 of the corresponding time T moments before and perform preprocessing, including normalization, feature selection, and data set segmentation, to obtain cleaned data;

[0011] S5: Multi-source fusion diffusion learning in latent space, using the multi-source fusion denoising network Fsrformer, dynamically adjusts the influence of satellite data during the diffusion process, enabling the model to fully utilize the guiding role of satellite data in precipitation forecasting;

[0012] S6: Inverse transformation and output of the results. The prediction results of the latent space are inversely transformed to the pixel space using the variational self-decoder to obtain the high-resolution radar echo extrapolation prediction results for the future time period T.

[0013] In step S1, the radar echo data is first de-noised. Outliers are filtered by setting thresholds of 15 dBZ and 70 dBZ. Non-precipitation clutter below 15 dBZ and super-refractive outliers above 70 dBZ are uniformly set to 0 to retain valid information about the actual precipitation area. The processed radar echo intensity is then mapped to a pixel range of 0-255 through linear transformation, and the pixel values ​​are further normalized to the range of [0, 1] using the minimum-maximum normalization method to provide standardized data for subsequent model input. At the same time, 12 consecutive frames of radar echo data are intercepted at 10-minute intervals to form a sample sequence, of which the first 6 frames are used as input and the last 6 frames are used as labels. The sample sequence is divided into training, validation, and test sets in a ratio of 7:2:1.

[0014] In the above step S2, the input radar echo sequence is first pre-extrapolated using a deterministic modeling method (using SimVP based on the CNN model) to obtain a fuzzy prediction in a pixel space. The frame-by-frame variational autoencoder will encode and compress the input and fuzzy prediction of the pixel space respectively, and obtain the input radar features and fuzzy prediction features of the latent space for subsequent learning of the latent space. The encoder and decoder in the VAE adopt an up- and down-sampling structure based on U-Net. The optimization objectives of VAE training are divided into three parts, namely KL divergence loss, reconstruction loss, and GAN loss. For the encoder, its goal is to learn the mapping from the input data x to the latent variable z, and ensure that its distribution is close to the standard normal distribution N(0,I), so as to maintain the continuity of the latent space. This process assumes that the latent variable z obeys a Gaussian distribution, that is, z~N(μ(x),σ(x)), where μ and σ are output by the encoder respectively. In order to make E φ (z|x) approximates the standard normal distribution p(z)=N(0,I), so the KL divergence loss can be calculated as follows:

[0015]

[0016] Among them L KL Represents the KL divergence loss value, D KL is the symbol for the KL divergence, which indicates the degree of difference between two probability distributions. It represents the approximate posterior distribution, E represents the mathematical expectation, z is the latent variable, x represents the original input data, and σ represents the model parameters. is the prior probability distribution with respect to the latent variable z, i is the index of the summation, indicating the dimension to be traversed, and d is the dimension of the latent variable. represents the mean of the approximate posterior distribution in the i-th dimension. It indicates the central position of the variable in 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 generated space, allowing VAE to perform smooth interpolation.

[0017] For the decoder, its goal is to generate the radar echo image x′ 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 represents the reconstruction loss, E represents the mathematical expectation, and D θ The distribution defined by the 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 deceive the discriminator C so that the generated samples look like real data, that is, maximize the score of the discriminator C on the generated samples:

[0020] L GAN = -E[logC(x′)]

[0021] L GAN Represents the loss function of the generator in the generative adversarial network, E represents the mathematical expectation, C represents the discriminator, and x′ represents the generated data. At the same time, for the discriminator, its task is to distinguish between real data and generated data, and its loss needs to be optimized independently. The binary cross entropy loss is as follows:

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

[0023] L C Denotes the loss function of the discriminator, C represents the discriminator, x1 represents the real data, that is, the sample from the real distribution. x′ represents the generated data.

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

[0025] Where λ1, λ2, and λ3 are the weight coefficients of the three types of losses for the total loss. Compared with the existing diffusion model prediction method, step 2 maps the high-resolution radar data into a low-dimensional space, reducing the computational cost.

[0026] In step S3 above, a spatiotemporal transformer (ST-Translator) is designed to model the deterministic features of precipitation evolution in latent space. ST-Translator adopts a multi-level ST-Inception structure, combining local spatiotemporal feature extraction with multi-scale information aggregation to improve the learning ability of precipitation evolution patterns. ST-Translator mainly consists of multiple ST-Inception modules. The latent input radar feature z after VAE encoding i and fuzzy prediction feature z pAfter splicing, it is used as the input of ST-Translator, and local spatiotemporal features are extracted by gradually stacking multiple layers of ST-Inception to capture the spatial structure and temporal evolution pattern of the precipitation system. The deterministic potential representation information z after processing by ST-Translator o This is the final output of the first stage and is used for further modeling of the diffusion model in the second stage. Compared to existing diffusion model methods, in step three, the ST-Translator module is designed to calculate the MSE with real data to narrow the gap between the prediction results and the real data, thereby improving the stability of the diffusion model input data and thus improving the final forecast accuracy.

[0027] In step S4, the satellite data is processed. Based on meteorological principles and correlation analysis, four infrared channels sensitive to precipitation system evolution are selected: the water vapor channel (WV), the mid-infrared channel (MI), the infrared channel 1 (IR1), and the infrared channel 2 (IR2). Redundant channels such as visible light are eliminated. An algorithm is used to resample the satellite data from its original resolution to a resolution consistent with the radar data, and the timestamps are strictly aligned to ensure spatiotemporal synchronization. The channel-by-channel minimum-maximum normalization method is used to process each satellite channel data, eliminating dimensional differences and then pairing them with the radar data in chronological order to form a multi-source fusion training sample set. This provides input data with spatiotemporal alignment and clear physical meaning for subsequent latent space modeling. Compared to existing single-radar echo extrapolation methods, step 4 incorporates satellite data and supplements atmospheric features such as clouds, further improving precipitation forecast accuracy.

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

[0029] The SR-DiT module designs a time-step-dependent scale-shift transformation (SST) to fine-grainedly control the input. This adaptively adjusts the feature distribution at different time steps t. In the self-attention stage, the input radar token x first undergoes layer normalization (LN), and then introduces time-step-dependent scale and shift transformations to generate the initial transformed features:

[0030] x′=γ1(t)·LN(x)+β1(t)

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

[0032]

[0033] Among them, Q, K, V represent the linear projection of x′, d1 represents the feature dimension of Q and K, and the output of the self-attention layer is adjusted by the scaling factor α1(t) and formed through the residual connection to form the first stage output:

[0034] h1=α1(t)·MHSA(x′)+x′

[0035] h1 represents the hidden state of the output, α1 represents the weight coefficient, and t represents the time step. In the multi-source data fusion stage, in order to make full use of the complementary role of satellite data to radar information, the input satellite Tokenx s After layer normalization, feature conditioning is performed through another set of time-step-dependent scaling-offset transformations:

[0036] x s ′=γ2(t)·LN(x s )+β2(t).

[0037] Next, x s ′ is the transformed 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 , and the satellite characteristics x adjusted by the scale-offset transformation s ' respectively as key K s Sum V s Enter the Satellite Radar Cross Attention Module (SRCA), whose calculation process is:

[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 are the linear projection weights, and 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 again layer-normalized and preprocessed using another set of time-step-dependent 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 layer normalization, h2 represents the layer-normalized input hidden state, β3 represents the offset factor, and t represents the time step. Nonlinear mapping is performed through a point-wise feedforward network, and the scaling factor α3(t) is adjusted and then residually connected with the original features to produce the final output:

[0043] h3=α3(t)·FFN(h2′)+h2

[0044] Where h3 represents the final output feature, α3 represents the weight coefficient, and h2 represents the layer-normalized input hidden state. Compared to the general diffusion model prediction method with a single radar echo data input, step 5 adds a satellite data input channel and designs a satellite weighting strategy that adjusts over time, thereby improving the precipitation forecast accuracy.

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

[0046] An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for extrapolating and predicting high-resolution radar echoes based on fused satellite data is implemented.

[0047] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement a 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) The present invention addresses the computational burden problem in the multiple-iteration denoising process of the diffusion model. By introducing VAE, the present invention migrates 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) The present invention makes full use of satellite prior information to supplement the limitations of single radar extrapolation, improving both prediction accuracy and image detail, especially in the prediction of heavy rainfall events. It can also effectively predict the potential initiation of convection.

[0051] 3) The present invention adopts the diffusion model method to solve the common problems of existing deep learning models, such as overly smooth predicted images, missing details, low accuracy, and a significant decrease in forecast timeliness as the forecast time increases, thereby improving the prediction accuracy and detail fidelity.

[0052] 4) This paper introduces a multi-source fusion denoising network, Fsrformer, and dynamically adjusts the influence of satellite data through the SR-DiT module, enabling the model to fully utilize satellite information during the initial stages of convection. This significantly enhances the model's sensitivity to convection inception and its early warning capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0054] Figure 2 Schematic diagram of radar data preprocessing according to 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 of the present invention.

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

[0058] Figure 6 Schematic diagram of the detailed structure of the Fsrformer network in the present invention. DETAILED DESCRIPTION

[0059] In order to deepen the knowledge and understanding of the present invention, the present invention is further described below with reference to the accompanying drawings and specific embodiments.

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

[0061] S1: Radar data input and preprocessing: input the previous T historical radar echo sequence for denoising, normalization, and data set segmentation to obtain cleaned data;

[0062] S2: Deterministic modeling method fuzzy prediction and variational autoencoder learning, a deterministic modeling method (SimVP) to obtain a fuzzy prediction sequence of the future time length T. The radar echo image and the fuzzy prediction image are encoded using a variational autoencoder and mapped into a low-dimensional latent space;

[0063] S3: Spatiotemporal transformer learning in latent space, 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 of the corresponding time T moments before and perform preprocessing, including normalization, feature selection, and data set segmentation, to obtain cleaned data;

[0065] S5: Multi-source fusion diffusion learning in latent space, using the multi-source fusion denoising network Fsrformer, dynamically adjusts the influence of satellite data during the diffusion process, enabling the model to fully utilize the guiding role of satellite data in precipitation forecasting;

[0066] S6: Inverse transformation and output of the results. The prediction results of the latent space are inversely transformed to the pixel space using the variational self-decoder to obtain the high-resolution radar echo extrapolation prediction results for the future time period T.

[0067] In step S1, Figure 2 As shown in the figure, first, denoising is performed on the radar echo data. By setting the thresholds of 15dBZ and 70dBZ, outliers are filtered out, and non-precipitation clutter below 15dBZ and super-refractive outliers above 70dBZ are uniformly set to 0 to retain the 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 minimum-maximum normalization method is further used to normalize the pixel value to the interval of [0,1] to provide standardized data for subsequent model input; at the same time, 12 consecutive frames of radar echo data are intercepted at intervals of 10 minutes to form a sample sequence, of which the first 6 frames are used as input and the last 6 frames are used as labels. The sample sequence is divided into training set, validation set and test set in a ratio of 7:2:1.

[0068] In step S2, Figure 4 As shown, the input radar echo sequence is first pre-extrapolated using a deterministic modeling method (using SimVP based on the CNN model) to obtain a fuzzy prediction in a pixel space. The frame-by-frame variational autoencoder will encode and compress the input and fuzzy prediction in the pixel space respectively, and obtain the input radar features and fuzzy prediction features in the latent space for subsequent learning of the latent space. The encoder and decoder in the VAE adopt an up- and down-sampling structure based on U-Net. The optimization objectives of VAE training are divided into three parts, namely KL divergence loss, reconstruction loss, and GAN loss. For the encoder, its goal is to learn the mapping from input data x to latent variable z, and ensure that its distribution is close to the standard normal distribution N(0,I), so as to maintain the continuity of the latent space. This process assumes that the latent variable z obeys a Gaussian distribution, that is, z~N(μ(x),σ(x)), where μ and σ are output by the encoder respectively. In order to make E φ (z|x) approximates the standard normal distribution p(z)=N(0,I), so the KL divergence loss can be calculated as follows:

[0069]

[0070] Among them L KL Represents the KL divergence loss value, D KL is the symbol for the KL divergence, which indicates the degree of difference between two probability distributions. It represents the approximate posterior distribution, E represents the mathematical expectation, z is the latent variable, x represents the original input data, and σ represents the model parameters. is the prior probability distribution with respect to the latent variable z, i is the index of the summation, indicating the dimension to be traversed, and d is the dimension of the latent variable. represents the mean of the approximate posterior distribution in the i-th dimension. It indicates the central position of the variable in 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 generated space, allowing VAE to perform smooth interpolation.

[0071] For the decoder, its goal is to generate a radar echo image x′ from the latent variable z that is 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 represents the reconstruction loss, E represents the mathematical expectation, and D θ The distribution defined by the 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 deceive the discriminator C so that the generated samples look like real data, that is, maximize the score of the discriminator C on the generated samples:

[0074] L GAN = -E[logC(x′)]

[0075] L GAN Represents the loss function of the generator in the generative adversarial network, E represents the mathematical expectation, C represents the discriminator, and x′ represents the generated data. At the same time, for the discriminator, its task is to distinguish between real data and generated data, and its loss needs to be optimized independently. The binary cross entropy loss is as follows:

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

[0077] L C Denotes the loss function of the discriminator, C represents the discriminator, x1 represents the real data, that is, the sample from the real distribution. x′ represents the generated data.

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

[0079] Where λ1, λ2, and λ3 are the weight coefficients of the three types of losses relative to the total loss. Compared with existing diffusion model prediction methods, step 2 innovatively maps high-resolution radar data into a low-dimensional space, reducing computational costs.

[0080] In step S3, Figure 5 As shown in Figure 2, a spatiotemporal transformer (ST-Translator) is designed to model the deterministic features of precipitation evolution in latent space. ST-Translator adopts a multi-level ST-Inception structure, combining local spatiotemporal feature extraction with multi-scale information aggregation to improve the learning ability of precipitation evolution patterns. ST-Translator mainly consists of multiple ST-Inception modules. The latent input radar feature z after VAE encoding i and fuzzy prediction feature z pAfter splicing, it is used as the input of ST-Translator, and local spatiotemporal features are extracted by gradually stacking multiple layers of ST-Inception to capture the spatial structure and temporal evolution pattern of the precipitation system. The deterministic potential representation information z after processing by ST-Translator o It is the final output of the first stage and used for further modeling of the diffusion model in the second stage.

[0081] In step S4, Figure 3 As shown in the figure, the satellite data are processed. Based on meteorological principles and correlation analysis, four infrared channels sensitive to the evolution of precipitation systems, namely water vapor channel (WV), mid-infrared channel (MI), infrared channel 1 (IR1) and infrared channel 2 (IR2), are selected, and redundant visible light channels are eliminated. The satellite data are resampled from the original resolution to the resolution consistent with the radar data through an algorithm, and the timestamps are strictly aligned to ensure spatiotemporal synchronization. The channel-by-channel minimum-maximum normalization method is used to process the data of each satellite channel. After eliminating the dimensional difference, the data are paired one by one with the radar data in chronological order to form a multi-source fusion training sample set, which provides spatiotemporally aligned input data with clear physical meaning for subsequent latent space modeling.

[0082] In step S5, Figure 6 As shown, the output of the first stage is used as input for iterative denoising and denoising training. To further integrate additional observation information from satellites, Fsrformer, a multi-source fusion denoising network based on a multi-layer SR-DiT stack, was invented. Its input includes the denoised latent radar features (output of the first stage), time step encoding, and latent satellite features. Among them, 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 radars to facilitate information fusion in the latent space. The latent radar features and latent satellite features are first processed by Patchify to convert the radar and satellite data into fixed-size token representations. They are then mapped to a high-dimensional latent space through a linear embedding layer and then enter the SR-DiT module for cross-modal feature fusion and spatiotemporal relationship modeling.

[0083] The SR-DiT module designs a time-step-dependent scale-shift transformation (SST) to fine-grainedly control the input. This adaptively adjusts the feature distribution at different time steps t. In the self-attention stage, the input radar token x first undergoes layer normalization (LN), and then introduces time-step-dependent scale and shift transformations to generate the initial transformed features:

[0084] x′=γ1(t)·LN(x)+β1(t)

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

[0086]

[0087] Among them, Q, K, V represent the linear projection of x′, d1 represents the feature dimension of Q and K, and the output of the self-attention layer is adjusted by the scaling factor α1(t) and formed through the residual connection to form the first stage output:

[0088] h1=α1(t)·MHSA(x′)+x′

[0089] h1 represents the hidden state of the output, α1 represents the weight coefficient, and t represents the time step. In the multi-source data fusion stage, in order to make full use of the complementary role of satellite data to radar information, the input satellite Tokenx s After layer normalization, feature conditioning is performed through another set of time-step-dependent scaling-offset transformations:

[0090] x s ′=γ2(t)·LN(x s )+β2(t).

[0091] Next, x s ′ is the transformed 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 , and the satellite characteristics x adjusted by the scale-offset transformation s ' respectively as key K s Sum V s Enter the Satellite Radar Cross Attention Module (SRCA), whose calculation process is:

[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 vare the linear projection weights, and 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 again layer-normalized and preprocessed using another set of time-step-dependent 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 layer normalization, h2 represents the layer-normalized input hidden state, β3 represents the offset factor, and t represents the time step. Nonlinear mapping is performed through a point-wise feedforward network, and the scaling factor α3(t) is adjusted and then residually connected with the original features to produce the final output:

[0097] h3=α3(t)·FFN(h2′)+h2

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

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

[0100] Example 2:

[0101] The radar echo dataset used for validation of this method comes from the Nanjing Institute of Meteorological Science and Technology Innovation (NJIAS). The dataset covers the summer months of June, July, and August each year, from 2019 to 2021. The dataset covers a 300 km × 300 km area in Jiangsu Province, with a spatial resolution of 1 km. Radar scans are performed at a fixed interval of 10 minutes, resulting in 144 radar echo frames per day. Meteorological satellite data for the corresponding time periods were also used. The Himawari-8 (H8) geostationary meteorological satellite, operated by the Japan Meteorological Agency, covers the entire Eastern Hemisphere, including Asia, Australia, and parts of the Pacific and Indian Oceans. H8's observation frequency is every 10 minutes, resulting in six data acquisitions per hour, which matches the temporal resolution of the NJIAS radar echo data. H8 is equipped with 16 spectral channels, including six visible channels (including three near-infrared channels) and 10 infrared channels. Among them, the data from June and July 2019 were used as validation sets and test sets respectively, and the rest were training sets.

[0102] The present invention was experimentally compared with two RNN-based precipitation prediction methods, PredRNNV2 and MAU; two diffusion model-based precipitation prediction methods, SCRD and CasCast; a CNN-based precipitation prediction method, SimVP; and a Transformer-based precipitation prediction method, Earthformer. As shown in Tables 1, 2, 3, 4, 5, and 6, it achieved the best performance in terms of CSI, HSS, and FAR indicators above a 40dBZ threshold, as well as SSIM, Sharpness, and PSNR indicators.

[0103] Table 1 Precision prediction indicators for radar echo intensity greater than 25dBZ

[0104]

[0105] Table 2 Precision prediction indicators for radar echo intensity greater than 35dBZ

[0106]

[0107]

[0108] Table 3 Radar echo intensity greater than 40dBZ accuracy prediction indicators

[0109]

[0110] Table 4 Precision prediction indicators for radar echo intensity greater than 45dBZ

[0111]

[0112] Table 5 Precision prediction indicators for radar echo intensity greater than 50dBZ

[0113]

[0114]

[0115] Table 6 Comparison of experimental results of 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, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.

Claims

1. A high-resolution radar echo extrapolation prediction method integrating satellite data, characterized by: The method comprises the following steps: S1: Radar data input and preprocessing: input the previous T historical radar echo sequence for denoising, normalization, and data set segmentation to obtain cleaned data; S2: Deterministic modeling method fuzzy prediction and variational autoencoder learning, a deterministic modeling method (SimVP), obtains a fuzzy prediction sequence of the future time length T, uses a variational autoencoder to encode the radar echo image and the fuzzy prediction image respectively, and maps them into a low-dimensional latent space; S3: Spatiotemporal transformer learning in latent space, 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 of the corresponding time T moments before and perform preprocessing, including normalization, feature selection, and data set segmentation, to obtain cleaned data; S5: Multi-source fusion diffusion learning in latent space, using the multi-source fusion denoising network Fsrformer, dynamically adjusts the influence of satellite data during the diffusion process, enabling the model to fully utilize the guiding role of satellite data in precipitation forecasting; S6: Inverse transformation and output of the results. The prediction results of the latent space are inversely transformed to the pixel space using the variational self-decoder to obtain the high-resolution radar echo extrapolation prediction results for the future time period T.

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

3. The high-resolution radar echo extrapolation prediction method based on satellite data integration according to claim 1 is characterized in that: The specific steps of step S2 are as follows: S21: Deterministic modeling preprocessing: Before inputting into VAE, the radar echo sequence is pre-extrapolated using the CNN-based SimVP (Simple Video Prediction) to obtain a fuzzy preliminary prediction result. This step aims to provide an initial estimate to help 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, respectively, and learn their low-dimensional potential representations. 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 phase. S23: Latent variable optimization: For the discriminator, its task is to distinguish between real data and generated data. It is necessary to optimize the loss independently, using binary cross entropy loss. Therefore, the complete loss function in VAE is as follows: Among them L KL Represents the KL divergence loss value, D KL It is the symbol of KL divergence, which is used to indicate the degree of difference between two probability distributions, indicating the approximate posterior distribution, E indicates the mathematical expectation, z is the latent variable, x indicates the original input data, σ is the parameter of the model, is the prior probability distribution of the latent variable z, i is the index of the summation indicating the traversal of the dimension, d is the dimension of the latent variable, Indicates that in the i-th dimension, For the decoder, its goal is to generate the radar echo image x′ 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 represents the reconstruction loss, E represents the mathematical expectation, and D θ The distribution defined by the 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 deceive the discriminator C so that the generated samples look like real data, that is, maximize the score of the discriminator C on the generated samples: L GAN =-E[logC(x′)] L GAN Represents the loss function of the generator in the generative adversarial network, E represents the mathematical expectation, C represents the discriminator, and x′ represents the generated data. At the same time, for the discriminator, its task is to distinguish between real data and generated data, and its loss needs to be optimized independently. The binary cross entropy loss is as follows: THE C =-E[logC(x1)]-E[log(1-C(x′))] L C represents the loss function of the discriminator, C represents the discriminator, x1 represents the real data, that is, the sample from the real distribution, x′ represents the generated data, L VAE =λ1L KL +λ2L rec +λ3L GAN Among them, λ1, λ2, and λ3 are the weight coefficients of the three losses for the total loss respectively. S24: Latent space feature generation: After the above steps, the latent variables obtained represent the low-dimensional features of the radar echo image. These features not only contain the temporal evolution information of the radar echo, but also retain the spatial details, providing high-quality initial input for the subsequent diffusion model.

4. The high-resolution radar echo extrapolation prediction method based on satellite data integration according to claim 1, characterized in that: Step S3 mainly includes the following steps: S31: Latent feature concatenation and normalization: concatenates the latent radar features generated by VAE with the fuzzy prediction features pre-extrapolated by SimVP in the channel dimension to form spatiotemporal input features. The concatenated features are then layer-normalized to eliminate scale differences and improve model stability. S32: ST-Inception module feature extraction, extracts multi-scale spatial features through group convolution, expands the receptive field to capture the different scale structures of the precipitation system, introduces the channel attention mechanism, learns channel weights through global pooling and fully connected layers, and strengthens the expression of key meteorological features. S33: Difference dispersion regularization, which calculates the inter-frame change between the predicted and true potential features through inter-frame differences, captures the dynamic evolution of the precipitation system, converts the difference into a probability distribution through Softmax, and uses KL divergence to constrain the predicted distribution to be close to the real one, thereby enhancing the sensitivity to time series changes. S34: Multi-level stacking and output, extracting spatiotemporal features layer by layer through multi-layer ST-Inception modules, combining residual connections to prevent gradient vanishing, gradually refining high-level spatiotemporal representations, and outputting spatiotemporal evolution features after layer normalization as input to subsequent diffusion models, improving the ability to model temporal dependencies.

5. The high-resolution radar echo extrapolation prediction method based on satellite data integration 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 within the past hour; S42: Normalization processing: For satellite observation data, first count its minimum and maximum values ​​in the training set, and then apply the normalization formula to map it to the [0,1] interval. S43: Feature selection: The Pearson correlation analysis method was used to evaluate the correlation between each satellite channel and the radar echo, and channels with significant correlation were screened out. Redundant information was removed, and infrared channel 7 (I4), water vapor channels (WV, W2, W3), and infrared channels 11 to 16 (MI, O3, IR, L2, I2, CO) were grouped. Channels with higher 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 satellite data integration according to claim 1, characterized in that: Step S5 mainly includes the following steps: S51: Multi-source input feature preparation: The radar latent features output by ST-Translator are denoised using a Gaussian diffusion process to generate a noisy latent sequence; the preprocessed satellite data are independently encoded into latent features of the same dimension; the diffusion time step is encoded as a sine position vector. The three together constitute the input to Fsrformer. The denoising process follows the preset cosine variance scheduling strategy to ensure that the noise intensity varies reasonably with the time step. S52: Tokenization and feature embedding: The radar and satellite latent features after noise addition are divided into patches of fixed size, mapped to a high-dimensional latent space through a linear layer, and position encoding is added to form a token sequence containing spatiotemporal position information. This provides input features in a unified format for subsequent Transformer processing, enhancing the model's perception of spatial structure and temporal order. S53: Backward denoising diffusion generation, starting with Gaussian noise, generates future radar potential signatures through 1000 steps of back diffusion iteration. At each step, Fsrformer is used to predict noise and update the signature. An improved DDPM method is used to support batch processing of multiple time steps to improve efficiency. The noise is gradually removed while injecting satellite guidance information to ensure that the generated potential signature conforms to the spatiotemporal evolution of the precipitation system.

7. The high-resolution radar echo extrapolation prediction method based on satellite data integration according to claim 1, characterized in that: Step S5 is specifically as follows: The SR-DiT module designs a time-step-dependent scale-shift transformation (SST) to fine-grainedly control the input, that is, adaptively adjust the feature distribution at different time steps t. In the self-attention stage, the input radar token x is first layer-normalized (LN), and then the time-step-dependent scale and shift transformation is introduced to generate the initial transformation feature: x′=γ1(t)·LN(x)+β1(t) Among them, γ1(t) and β1(t) are learnable parameters that depend on the time step t, which can adaptively adjust the distribution of input features. Then, the spatiotemporal relationship modeling is performed through the multi-head self-attention (MHSA) module, and its calculation formula is: Among them, Q, K, V represent the linear projection of x′, d1 represents the feature dimension of Q and K, and the output of the self-attention layer is adjusted by the scaling factor α1(t) and formed through the residual connection to form the first stage output: h1=α1(t)·MHSA(x′)+x′ h1 represents the hidden state of the output, α1 represents the weight coefficient, and t represents the time step. In the multi-source data fusion stage, in order to make full use of the complementary role of satellite data to radar information, the input satellite After layer normalization, feature conditioning is performed through another set of time-step-dependent scaling-offset transformations: x′ s =γ2(t)·LN(x s )+β2(t). Next, x s ′ is the transformed 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 , and the satellite characteristics x adjusted by the scale-offset transformation s ' respectively as key K s Sum V s Enter the Satellite Radar Cross Attention Module (SRCA), whose calculation process is: 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 are the linear projection weights 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 again layer-normalized and preprocessed using another set of time-step-dependent scaling-offset transformations: h′2=γ3(t)·LN(h2)+β3(t) Where h2′ represents the preprocessed feature, γ3 represents the third scaling factor, LN is layer normalization, h2 represents the input hidden state of layer normalization, β3 represents the offset factor, t represents the time step, and nonlinear mapping is performed through the point-wise feedforward network. After adjustment with the scaling factor α3(t), the residual connection is performed with the original feature, and 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 input hidden state of the layer normalization.

8. The high-resolution radar echo extrapolation prediction method based on satellite data integration 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 of the future T frames generated by diffusion are input into the VAE decoder symmetrical with the encoder, and the low-dimensional features are gradually restored to a radar echo image of 300×300 pixels. The output dimension is T×300×300, achieving effective reconstruction of spatial resolution and detail information. S62: Output the prediction results. Generate a 60-minute prediction sequence at a 10-minute frame interval. The timestamp is 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, wherein: When the processor executes the program, the high-resolution radar echo extrapolation prediction method integrating satellite data as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, a high-resolution radar echo extrapolation prediction method integrating satellite data according to any one of claims 1 to 8 is implemented.

Images