A radar echo space-time extrapolation method based on asymmetric space modeling

Abstract

The application discloses a radar echo space-time extrapolation method based on asymmetric space modeling, belongs to the cross field of meteorological radar data processing and artificial intelligence technology, and first constructs an asymmetric input space sample containing a target region and a surrounding extended region by preprocessing original meteorological radar observation data, builds an asymmetric space modeling prediction network based on a 3D U-Net architecture, adopts a regional mask mean square error loss function to complete model training, and finally realizes accurate space-time extrapolation of the target region radar echo. The application solves the technical problems of information fragmentation and significant long-time prediction error accumulation of the traditional symmetric modeling method, improves the precision and stability of radar echo extrapolation under complex weather systems, and is suitable for 0-1 hour nowcasting and meteorological disaster accurate early warning scenes.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of meteorological radar data processing and artificial intelligence technology, and in particular relates to a radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling. Background Technology

[0002] Radar echo spatiotemporal extrapolation is a technique that uses real-time radar scan data to predict future radar echo scenarios. It can be viewed as an estimation and prediction of the changing trends of continuous time-series images, i.e., using existing radar echo images over a certain period to predict radar echo images over a future period. In the field of radar echo spatiotemporal extrapolation, traditional methods (such as optical flow and linear extrapolation) directly predict the future state of a target area by analyzing the motion features in the input data. Deep learning extrapolation methods, such as PredRNN and 3D U-Net, use models to directly learn and map the spatiotemporal features within the input area, achieving accurate prediction of the target area. However, both traditional and deep learning extrapolation methods currently employ symmetric spatial modeling. Symmetric spatial modeling refers to a forecasting model that follows the design paradigm of "input and label size consistency." Its core logic is that the model's input data and predicted output (label) are strictly matched in spatial range; that is, only historical radar echo data of the target area are used as input to directly predict the future echo evolution state of the same area. Symmetric modeling methods have the following shortcomings:

[0003] 1. Severing the Regional Correlation of Severe Convective Systems: Severe convective weather systems exhibit significant spatiotemporal expansion and migration. Their formation, development, and intensity changes depend not only on the internal characteristics of the target region but also on key factors such as the surrounding regional circulation background and water vapor transport. Symmetrical spatial modeling only inputs data from the target region, severing the intrinsic connection between severe convective systems and the surrounding environment, causing the model to fail to capture this crucial background information.

[0004] 2. Significant accumulation of errors in long-term forecasts: The information limitations of symmetric spatial modeling can be partially compensated for by local motion features in short-term forecasts, but in long-term forecasts, due to the lack of supporting information from the surrounding environment, the model cannot accurately predict the path shift, intensity increase or decrease, and morphological evolution of strong convective systems, resulting in the accumulation of errors over the forecast period and a significant decrease in accuracy.

[0005] 3. Limitations on the model's utilization of multi-scale spatiotemporal features: Symmetrical spatial modeling has a fixed input range and is limited to the target region, meaning the model can only learn the local spatiotemporal features of that region and cannot access the interaction patterns between severe convective systems and the surrounding environment at different scales. This results in poor model adaptability to different types of severe convection (such as localized thunderstorms and regional rainstorms) and insufficient generalization ability.

[0006] Currently, there is no asymmetric spatial modeling method for spatiotemporal extrapolation of radar echoes, and the joint features of the target area and its surrounding extended areas have not yet been integrated into the deep learning model, resulting in insufficient accuracy in predicting the cross-regional evolution of complex weather systems.

[0007] Therefore, how to change the modeling method and improve the prediction accuracy of radar echo spatiotemporal extrapolation under complex weather conditions has become a key problem that urgently needs to be solved. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide a radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling. By expanding the input range to include the target area and surrounding areas, the method significantly enhances the model's ability to capture regional correlation features of strong convective systems.

[0009] Technical solution: The present invention provides a method for spatiotemporal extrapolation of radar echoes based on asymmetric spatial modeling, comprising the following steps:

[0010] Step 1: Preprocessing of meteorological radar observation data: Obtain raw meteorological radar observation data, and sequentially perform data reading and decoding, geospatial registration and grid unification, comprehensive spatiotemporal range extraction of asymmetric input space, outlier removal and physical constraint correction, spatial resolution standardization and numerical scale normalization to obtain standardized radar echo data suitable for asymmetric spatial modeling.

[0011] Step 2, Construction of training sample set: Based on the normalized radar echo data obtained in Step 1, time series grouping is performed using a sliding window strategy. Each consecutive time series window is used as an independent training sample to construct the model training sample set, and the training set, validation set and test set are divided according to a preset ratio.

[0012] Step 3: Construction of Asymmetric Spatial Modeling Prediction Network: Build a radar echo prediction network model based on the 3D U-Net architecture and construct an asymmetric spatial modeling structure; the network takes the spatial range including the target study area and the surrounding extended area as input and the spatial range including only the target study area as output, forming an asymmetric constraint where the input range is greater than the output range;

[0013] Step 4, Model Training and Parameter Optimization: Based on the training set and validation set obtained in Step 2, train the radar echo prediction network model constructed in Step 3; construct the regional mask mean square error loss function, perform loss supervision only on the target output region, complete the model parameter optimization, and obtain the trained radar echo prediction network model.

[0014] Step 5, Model Prediction and Comprehensive Evaluation: Input the test set divided in Step 2 into the radar echo prediction network model trained in Step 4, and output the predicted radar echo image; use multi-dimensional evaluation indicators to perform a quantitative and qualitative comprehensive evaluation of the prediction results.

[0015] Furthermore, step 1 specifically includes the following steps:

[0016] Step 1.1, Data Reading and Decoding: The acquired binary raw radar data is parsed and converted into computable multidimensional array structure data;

[0017] Step 1.2, Geospatial Registration and Grid Unification: Based on latitude and longitude information, the multidimensional array structure data obtained in Step 1.1 is mapped to a unified geographic grid to complete spatial location registration;

[0018] Step 1.3, Extraction of the comprehensive spatiotemporal range of the asymmetric input space: Based on the core research area, the boundary of the surrounding environmental area is set outward to construct an asymmetric input space that includes the target area and the surrounding extended area. Combined with the target observation time, the radar echo data within the comprehensive spatiotemporal range is extracted.

[0019] Step 1.4, Outlier Removal and Physical Constraint Correction: Correct invalid and out-of-range values ​​in the radar echo data. Set invalid values ​​to 0dBZ, values ​​less than 0dBZ to 0, and values ​​exceeding 70dBZ to 70dBZ.

[0020] Step 1.5, Spatial Resolution Standardization: The resolution of radar echo data is adjusted using a downsampling strategy to achieve spatial size uniformity;

[0021] Step 1.6, Numerical scale normalization: Map the intensity of radar echo data to the standard interval [0,1], corresponding to the range of 0dBZ to 70dBZ.

[0022] Furthermore, in step 2, the sliding window strategy is specifically as follows: the window length is set to T frames, each time window is treated as a complete set of samples, and the dimension of each data sample is (T, C, H, W), where C represents the number of channels, and H and W represent the spatial height and width of the data, respectively.

[0023] Furthermore, in step 3, the asymmetric spatial modeling structure is specifically as follows: the input spatial range is larger than the output spatial range, the input includes the extended environmental region, and the output only retains the target prediction region; the asymmetric spatial modeling prediction network adopts an encoder-bottleneck layer-decoder structure, and realizes spatiotemporal feature extraction, fusion and upsampling through three-dimensional convolution and deconvolution, and uses skip connections to retain high-resolution detailed features, thereby completing the spatiotemporal feature learning and prediction output of radar echoes in asymmetric space.

[0024] Furthermore, the encoder comprises four resolution levels, each consisting of two sets of 3D convolutional layers with a kernel size of 3×3×3, a batch normalization layer, and so on. The activation function consists of a 3D max-pooling layer with a kernel size of 2×2×2 and a stride of 2 after each layer for downsampling, achieving layer-by-layer extraction and dimensionality reduction of image features; the encoder expression is shown below:

[0025]

[0026]

[0027]

[0028] in, This represents the three-dimensional radar echo spatiotemporal tensor input into the encoder convolutional layer; This is the input for the first resolution level; For encoder number Convolutional features at each level; This is a dimensionality reduction operation following downsampling; This is a 3D convolution operation. This is a 3D max pooling operation, and BN is a batch normalization layer. It is a non-linear activation function; the number of channels in the encoder doubles layer by layer with downsampling, realizing the expansion of feature dimensions, and maintaining the three-dimensional spatial structure features of the tensor throughout the process.

[0029] Furthermore, the bottleneck layer employs two consecutive 3×3×3 three-dimensional convolutions, batch normalization, and... The activation function is composed of a structure; the bottleneck layer receives the deepest feature map after encoder downsampling, performs nonlinear transformation and depth enhancement on global spatiotemporal features, and extracts the core features of large-scale motion trends of radar echoes and strong convective system correlation.

[0030] Furthermore, the decoder and encoder have a symmetrical structure, containing resolution levels corresponding to the encoder. Each level first undergoes upsampling through a 3D deconvolution layer with a kernel size of 2×2×2 and a stride of 2. The upsampled features are then concatenated and fused with features of the same resolution from the encoder. Next, the decoder sequentially passes through two sets of 3D convolutional layers with kernel sizes of 3×3×3, a batch normalization layer, and so on. The non-linear activation function refines the features; the decoder's hidden representation is shown below:

[0031]

[0032]

[0033]

[0034]

[0035] in, This represents the encoder deep feature tensor input to the decoder; This refers to the downsampled output features of the encoder's last layer. For decoder number The upsampling features of the first level are the decoder's first... Upsampled features at each level, upsampling features Downsampling at the corresponding level of the encoder Features after splicing and fusion The refined features are obtained by performing two 3D convolutions, batch normalization, and ReLU activation to fuse the features; This represents the downsampled features of the encoder's 4-i layer; This is a feature concatenation operation that fuses high-resolution features from the encoder with upsampled features from the decoder. This is a 3D deconvolution upsampling operation; the last layer of the decoder maps the number of feature channels to the number of output channels through a 1×1×1 3D convolutional layer to obtain the radar echo prediction result of the target area.

[0036] Furthermore, in step 4, the mean square error loss function of the region mask... The formula is as follows:

[0037]

[0038] Where M is the total number of elements with a value of 1 in the mask, that is, the number of elements in the target region; Let b be the tensor of the model predictions, c be the batch index, p be the height index, and q be the width index. For the real label value tensor; This is a binary mask tensor, with a value of 1 for the region of interest and a value of 0 for the region of masking.

[0039] During training, the mean square error loss function of the region mask is calculated. The mean square error loss function of the region mask The objective is to minimize the mean squared error loss function of the region mask, and backpropagation is performed through multiple iterations. This process continues until the model converges. After training, the final radar echo prediction network model is obtained.

[0040] Furthermore, in step 5, the model prediction and comprehensive evaluation employ multi-dimensional evaluation indicators to comprehensively evaluate the prediction results, including image quality evaluation indicators and meteorological operational evaluation indicators. Specifically, the image quality evaluation indicators include mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM).

[0041]

[0042]

[0043]

[0044] in, This represents the total number of pixels in the radar echo predicted image and the actual image. Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; It represents the maximum possible value of a pixel in the image; and These represent the pixel values ​​of the predicted image and the real image, respectively; and These are the pixel mean values ​​of the predicted image and the real image, respectively; and These are the pixel standard deviations of the predicted image and the real image, respectively; To determine the covariance between the predicted image and the real image; and It is a constant;

[0045] The meteorological operational evaluation indicators include hit rate (POD), false alarm rate (FAR), critical success index (CSI), and accuracy, specifically:

[0046]

[0047]

[0048]

[0049]

[0050] in, This represents the number of grid points that actually occurred and were correctly predicted. This represents the number of grid points that actually occurred but were not predicted. This represents the number of grid points that were predicted to occur but did not actually occur. This represents the number of grid points that neither actually occurred nor were predicted.

[0051] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0052] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0053] This invention proposes a radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling. By expanding the input range to include the target area and surrounding areas, it significantly enhances the model's ability to capture regional correlation features of strong convective systems. Compared with traditional symmetric spatial modeling methods, it effectively reduces error accumulation in forecast lead times of 30 to 60 minutes, and significantly improves image quality indicators (PSNR, SSIM, etc.) and core meteorological indicators (POD, CSI, etc.).

[0054] This invention employs a region mask mean square error loss function, calculating only the mean square error loss of a specified region of the mask, thus accurately optimizing the prediction effect of the target region and effectively supervising the model to generate more detailed and accurate predicted echo images.

[0055] The combined effect of these technological advantages enables the model to more accurately capture the dynamic changes of complex weather phenomena such as storm systems. Ultimately, in radar echo extrapolation with a forecast lead time of 30 to 60 minutes, the image quality index PSNR is improved by ≥1 dB, SSIM by ≥0.03, and the meteorological operational index CSI by ≥0.04, effectively reducing error accumulation and providing more reliable technical support for timely and accurate early warning of meteorological disasters, thus possessing significant practical application value. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to the present invention. Figure 2 This is a flowchart of the radar echo prediction network model of the 3D U-Net architecture of the present invention; Figure 3 The graph shows the comparison results of various indicators on the model used under different scale input data; where (a) represents the change of MSE index, and the vertical axis represents the mean squared error (unit: (dBZ)). 2 (a) PSNR change, with peak signal-to-noise ratio (dB) on the ordinate; (b) SSIM change, with structural similarity on the ordinate; (c) POD change, with hit rate on the ordinate; (d) FAR change, with false alarm rate on the ordinate; (e) CSI change, with critical success index on the ordinate. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] Example

[0059] This application example demonstrates radar echo sequence prediction experiments using the Radar Combined Reflectance (CREF) dataset. The target area is Jinhua City, Zhejiang Province, specifically ranging from 119.1°E to 120.38°E and 28.22°N to 29.50°N. The surrounding areas include a region extending approximately 40 kilometers outward from Jinhua City and the entire Zhejiang Province. The region extending approximately 40 kilometers outward from Jinhua City has a specific latitude and longitude range of 118.3°E to 121.18°E and 28.14°N to 29.58°N, while the entire Zhejiang Province has a range of 118.02°E to 122.98°E and 27.0°N to 31.0°N. The CREF dataset for the target area consists of two years' worth of CREF data from 2023 to 2024, with a temporal resolution of 6 minutes and a spatial resolution of 0.01°.

[0060] This invention is used to perform radar echo sequence prediction tasks on the radar product data composite reflectivity CREF dataset, such as... Figure 1 As shown, the specific implementation steps are as follows:

[0061] Step S1: Preprocess the raw meteorological radar observation data to obtain standardized radar echo data. In this example, radar combined reflectivity (CREF) data for Jinhua City (119.1°E to 120.38°E, 28.22°N to 29.50°E), radar combined reflectivity (CREF) data for an area extending approximately 40 kilometers outward from Jinhua City (118.3°E to 121.18°E, 28.14°N to 29.58°N), and radar combined reflectivity (CREF) data for the entire Zhejiang Province (118.02°E to 122.98°E, 27.0°N to 31.0°N) are obtained. The acquired radar echo data is then preprocessed to obtain standardized radar echo data.

[0062] Step S2: Constructing the radar echo sequence dataset. In this example, based on the radar echo data obtained in Step 1, a sliding window strategy is used to group the data temporally, using each consecutive temporal window as an independent training sample to construct the radar echo sequence dataset. The sliding window length T=20, corresponding to a time of 120 minutes, with a temporal resolution of 6 minutes and a spatial resolution of 0.01°. 70% of the resulting radar echo sequence dataset is used as the training set, 15% as the test set, and 15% as the validation set. Each radar echo sequence consists of 20 consecutive frames of radar echo data. The first 10 frames are used as input to the radar echo prediction network model; the last 10 frames are used as the label for the radar echo prediction network. The model output is optimized using a loss function. In other words, the radar echo sequence of the past hour is used to predict the radar echo sequence of the next hour.

[0063] This experiment sets up a control group and an experimental group. The control group uses a single area of ​​Jinhua City as input and Jinhua City as output, which is the traditional symmetrical modeling control group. The experimental group uses a 40-kilometer area of ​​Jinhua City as input / the entire Zhejiang Province as input and Jinhua City as output, which is the asymmetrical modeling experimental group of this invention. The technical advantages of the method of this invention are verified by comparing multiple dimensions of indicators.

[0064] Step S3, construct the radar echo prediction network model:

[0065] Construct a radar echo prediction network model based on the 3D U-Net "encoder-bottleneck layer-decoder" architecture, such as... Figure 2 As shown, the encoder contains four resolution levels. Each level consists of two sets of 3D convolutional layers with a kernel size of 3×3×3, a batch normalization layer, and a ReLU non-linear activation function. After each set of levels, a 3D max pooling layer with a kernel size of 2×2×2 and a stride of 2 is set for downsampling. The decoder also contains four resolution levels. Each level first performs upsampling through a 3D deconvolutional layer with a kernel size of 2×2×2 and a stride of 2. The upsampled features are then concatenated and fused with the features of the same resolution as the encoder. The features are then refined by passing through two sets of 3D convolutional layers with a kernel size of 3×3×3, a batch normalization layer, and a ReLU non-linear activation function. Input convolutional layers and output convolutional layers are set in the model for dimension adaptation.

[0066] Step S4: Model training and prediction performance evaluation.

[0067] The model is trained based on the training set obtained in step 2. The learning rate is set to 1e-4, the Adam optimizer is selected, and the batch size for training all prediction methods is set to 1. To prevent the model from overfitting, the early stopping patience value is set to 30 epochs, that is, if no improvement in validation loss is observed in 30 consecutive epochs, training will stop.

[0068] The acquired radar echo sequences from the test set were then input into the trained radar echo prediction network to obtain the predicted radar echo sequences. The prediction results were then evaluated using corresponding image quality metrics and meteorological evaluation metrics. The image quality metrics and meteorological evaluation metrics for prediction results on the 3D UNet model for regions of different input sizes are shown in Tables 1 and 2. Jinhua City serves as the traditional symmetric modeling control group; Jinhua City with an extension of 40 km and Zhejiang Province serve as the asymmetric modeling experimental group of this invention.

[0069] Table 1

[0070]

[0071] Table 2

[0072]

[0073] As shown in Table 1, Table 2 and Figure 3 As shown, the asymmetric spatial modeling method, centered on Jinhua City and extending 40 kilometers to encompass the entire Zhejiang province as input, with Jinhua City as the label, outperforms the symmetric spatial modeling method, which uses Jinhua City as input and is labeled Jinhua City, in all image quality and meteorological evaluation indicators. The improvement is even more pronounced in the 30-60 minute forecast lead time. When inputting 10 frames of Zhejiang radar echo sequences, 10 actual radar echo sequences, and 10-frame radar echo sequences predicted by the same model under different scales of input data, the method using Jinhua City as input, extending 40 kilometers to encompass the entire Zhejiang province, shows better performance across all time periods compared to using Jinhua City as input, especially in the 30-60 minute period (the last 6 frames), where the improvement is more significant.

[0074] The test results of the above embodiments demonstrate the effectiveness and superiority of the radar echo prediction method of the present invention. This application example uses Jinhua City, Zhejiang Province as the research object and is only used to verify the effectiveness of the technical solution of the present invention. The method of the present invention can be applied to the spatiotemporal extrapolation of radar echoes in any region, including meteorological monitoring scenarios at different spatial scales such as local, regional, and provincial.

[0075] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for spatiotemporal extrapolation of radar echoes based on asymmetric spatial modeling, characterized in that, Includes the following steps: Step 1: Preprocessing of meteorological radar observation data: Obtain raw meteorological radar observation data, and sequentially perform data reading and decoding, geospatial registration and grid unification, comprehensive spatiotemporal range extraction of asymmetric input space, outlier removal and physical constraint correction, spatial resolution standardization and numerical scale normalization to obtain standardized radar echo data suitable for asymmetric spatial modeling. Step 2, Construction of training sample set: Based on the normalized radar echo data obtained in Step 1, time series grouping is performed using a sliding window strategy. Each consecutive time series window is used as an independent training sample to construct the model training sample set, and the training set, validation set and test set are divided according to a preset ratio. Step 3: Construction of Asymmetric Spatial Modeling Prediction Network: Build a radar echo prediction network model based on the 3D U-Net architecture and construct an asymmetric spatial modeling structure; the network takes the spatial range including the target study area and the surrounding extended area as input and the spatial range including only the target study area as output, forming an asymmetric constraint where the input range is greater than the output range; Step 4, Model Training and Parameter Optimization: Based on the training set and validation set obtained in Step 2, train the radar echo prediction network model constructed in Step 3; construct the regional mask mean square error loss function, perform loss supervision only on the target output region, complete the model parameter optimization, and obtain the trained radar echo prediction network model. Step 5, Model Prediction and Comprehensive Evaluation: Input the test set divided in Step 2 into the radar echo prediction network model trained in Step 4, and output the predicted radar echo image; use multi-dimensional evaluation indicators to perform a quantitative and qualitative comprehensive evaluation of the prediction results.

2. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1, Data Reading and Decoding: The acquired binary raw radar data is parsed and converted into computable multidimensional array structure data; Step 1.2, Geospatial Registration and Grid Unification: Based on latitude and longitude information, the multidimensional array structure data obtained in Step 1.1 is mapped to a unified geographic grid to complete spatial location registration; Step 1.3, Extraction of the comprehensive spatiotemporal range of the asymmetric input space: Based on the core research area, the boundary of the surrounding environmental area is set outward to construct an asymmetric input space that includes the target area and the surrounding extended area. Combined with the target observation time, the radar echo data within the comprehensive spatiotemporal range is extracted. Step 1.4, Outlier Removal and Physical Constraint Correction: Correct invalid and out-of-range values ​​in the radar echo data. Set invalid values ​​to 0dBZ, values ​​less than 0dBZ to 0, and values ​​exceeding 70dBZ to 70dBZ. Step 1.5, Spatial Resolution Standardization: The resolution of radar echo data is adjusted using a downsampling strategy to achieve spatial size uniformity; Step 1.6, Numerical scale normalization: Map the intensity of radar echo data to the standard interval [0,1], corresponding to the range of 0dBZ to 70dBZ.

3. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 1, characterized in that, In step 2, the sliding window strategy is as follows: the window length is set to T frames, each time window is treated as a complete sample, and the dimension of each data sample is (T, C, H, W), where C represents the number of channels, and H and W represent the spatial height and width of the data, respectively.

4. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 1, characterized in that, In step 3, the asymmetric spatial modeling structure is specifically as follows: the input spatial range is larger than the output spatial range, the input includes the extended environmental region, and the output only retains the target prediction region; the asymmetric spatial modeling prediction network adopts an encoder-bottleneck layer-decoder structure, and realizes spatiotemporal feature extraction, fusion and upsampling through three-dimensional convolution and deconvolution, and uses skip connections to retain high-resolution detailed features, thereby completing the spatiotemporal feature learning and prediction output of radar echoes in asymmetric space.

5. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 4, characterized in that, The encoder contains four resolution levels, each consisting of two sets of 3D convolutional layers with a kernel size of 3×3×3, a batch normalization layer, and so on. The activation function consists of a 3D max-pooling layer with a kernel size of 2×2×2 and a stride of 2 after each layer for downsampling, achieving layer-by-layer extraction and dimensionality reduction of image features; the encoder expression is shown below: ； ； ； in, This represents the three-dimensional radar echo spatiotemporal tensor input into the encoder convolutional layer; This is the input for the first resolution level; For encoder number Convolutional features at each level; This is a dimensionality reduction operation following downsampling; This is a 3D convolution operation. This is a 3D max pooling operation, and BN is a batch normalization layer. It is a non-linear activation function; the number of channels in the encoder doubles layer by layer with downsampling, realizing the expansion of feature dimensions, and maintaining the three-dimensional spatial structure features of the tensor throughout the process.

6. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 4, characterized in that, The bottleneck layer employs two consecutive 3×3×3 three-dimensional convolutions, batch normalization, and... The activation function is composed of a structure; the bottleneck layer receives the deepest feature map after encoder downsampling, performs nonlinear transformation and depth enhancement on global spatiotemporal features, and extracts the motion trend of radar echoes and features associated with strong convective systems.

7. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 4, characterized in that, The decoder and encoder have a symmetrical structure, containing resolution levels corresponding to the encoder. Each level first performs upsampling through a 3D deconvolution layer with a kernel size of 2×2×2 and a stride of 2. The upsampled features are then concatenated and fused with features of the same resolution from the encoder. Next, the decoder sequentially passes through two sets of 3D convolutional layers with kernel sizes of 3×3×3, a batch normalization layer, and so on. Nonlinear activation functions are used for feature refinement. The decoder expression is as follows: ； ； ； ； in, This represents the encoder deep feature tensor input to the decoder; This refers to the downsampled output features of the encoder's last layer. For the decoder Upsampled features at each level, upsampling features Downsampling at the corresponding level of the encoder Features after splicing and fusion The refined features are obtained by performing two 3D convolutions, batch normalization, and ReLU activation to fuse the features; This represents the downsampled features of the encoder's (4-i)th layer. This is a feature concatenation operation that fuses high-resolution features from the encoder with upsampled features from the decoder. This is a 3D deconvolution upsampling operation; the last layer of the decoder maps the number of feature channels to the number of output channels through a 1×1×1 3D convolutional layer to obtain the radar echo prediction result of the target area.

8. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 1, characterized in that, In step 4, the mean square error loss function of the region mask The formula is as follows: ； Where M is the total number of elements with a value of 1 in the mask, that is, the number of elements in the target region; Let b be the tensor of the model predictions, c be the batch index, p be the height index, and q be the width index. For the real label value tensor; This is a binary mask tensor, with a value of 1 for the region of interest and a value of 0 for the region of masking. During training, the mean square error loss function of the region mask is calculated. The mean square error loss function of the region mask The objective is to minimize the mean squared error loss function of the region mask, and backpropagation is performed through multiple iterations. This process continues until the model converges. After training, the final radar echo prediction network model is obtained.

9. The radar echo spatiotemporal extrapolation method based on asymmetric spatial modeling according to claim 1, characterized in that, In step 5, the model prediction and comprehensive evaluation employs multi-dimensional evaluation indicators to comprehensively evaluate the prediction results, including image quality evaluation indicators and meteorological operational evaluation indicators. Specifically, the image quality evaluation indicators include mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM). ； ； ； in, This represents the total number of pixels in the radar echo predicted image and the actual image. Indicates the first The true value of each sample Indicates the first Predicted values ​​for each sample; It represents the maximum possible value of a pixel in the image; and These represent the pixel values ​​of the predicted image and the real image, respectively; and These are the pixel mean values ​​of the predicted image and the real image, respectively; and These are the pixel standard deviations of the predicted image and the real image, respectively; To determine the covariance between the predicted image and the real image; and It is a constant; The meteorological operational evaluation indicators include hit rate (POD), false alarm rate (FAR), critical success index (CSI), and accuracy, specifically: ； ； ； ； in, This represents the number of grid points that actually occurred and were correctly predicted. This represents the number of grid points that actually occurred but were not predicted. This represents the number of grid points that were predicted to occur but did not actually occur. This represents the number of grid points that neither actually occurred nor were predicted.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images