A radar echo extrapolation method based on space-time attention mechanism

Abstract

The application discloses a radar echo extrapolation method based on a space-time attention mechanism, first obtains radar echo data of a target area, and carries out pretreatment on the obtained radar echo data; the radar echo data after the pretreatment is subjected to sliding grouping and data augmentation, so that a radar echo sequence data set is obtained, and then the radar echo sequence data set is divided into a training set and a test set; a radar echo prediction network model based on a SimVP architecture is constructed and trained and tested, a multi-target loss function is used to supervise model training; and a future radar echo image is predicted in real time. Through the method, complex weather phenomena such as storms can be more accurately captured in dynamic evolution, the precision and interpretability of short-term nowcasting can be significantly improved, more reliable technical support is provided for timely and accurate early warning of meteorological disasters, and the method has important practical application value.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of meteorological radar data processing and artificial intelligence, specifically to a radar echo extrapolation method based on a spatiotemporal attention mechanism. Background Technology

[0002] Radar echo extrapolation is a technique that uses real-time radar scan data to predict future radar echo scenes. It can be viewed as an estimation and prediction of the changing trends of continuous time-series images, that is, using existing radar echo images over a certain period to predict radar echo images over a future period. Traditional radar echo extrapolation methods, including optical flow methods and 3D convolutional neural networks, have been applied in certain specific scenarios, but still have the following shortcomings:

[0003] 1. Insufficient spatial feature extraction: Existing extrapolation methods often fail to effectively distinguish between strong echo regions and noisy regions, especially in complex weather phenomena such as storms, resulting in insufficient clarity in the extrapolation results. Traditional methods fail to fully extract key spatial features from radar echo images, making it impossible to guarantee accurate predictions of highly important regions such as storm areas.

[0004] 2. Limited spatiotemporal modeling capabilities: Existing methods often rely solely on simple convolution operations to extract spatiotemporal features, which is insufficient for modeling complex echo motion patterns, especially the nonlinear deformation and multi-scale interactions of echo morphology. In real weather, echo motion is not nonlinear and exhibits different dynamic characteristics at different time scales; existing technologies have failed to effectively capture these complex spatiotemporal evolution patterns.

[0005] 3. High model complexity: In order to improve accuracy, some methods often introduce a large number of learnable parameters. Although this improves the model performance to a certain extent, too many parameters not only increase the computational cost, but also easily lead to overfitting, reducing the model's generalization ability and stability. Especially with limited datasets, the robustness of model training is poor. Summary of the Invention

[0006] The purpose of this invention is to provide a radar echo extrapolation method based on a spatiotemporal attention mechanism, which can improve the accuracy, physical consistency and computational efficiency of radar echo extrapolation under complex weather conditions.

[0007] To achieve the above functions, this invention designs a radar echo extrapolation method based on a spatiotemporal attention mechanism. For the target area, the following steps S1-S5 are executed to predict the future radar echo image of the target area:

[0008] Step S1: Collect radar echo data of the target area, preprocess the obtained radar echo data, and construct a radar echo dataset based on the preprocessed radar echo data;

[0009] Step S2: Perform sliding grouping on the radar echo dataset obtained in step S1 to obtain radar echo sequence data, and perform data augmentation to obtain radar echo sequence dataset. Then, divide the radar echo sequence dataset into training set and test set.

[0010] Step S3: Construct a radar echo prediction network model based on the SimVP architecture, including an encoder, a translator, and a decoder connected in sequence; the encoder is also connected to the decoder. The radar echo prediction network model takes radar echo sequence data in the radar echo sequence dataset as input and the predicted future radar echo image of the target area as output.

[0011] Step S4: Use the training set and test set divided in step S2 to train and test the radar echo prediction network model constructed in step S3, and construct a multi-target loss function. Use the multi-target loss function to supervise the training of the radar echo prediction network model to obtain the trained radar echo prediction network model.

[0012] Step S5: Collect radar echo data in real time, preprocess it using the same method as in step S1, and perform sliding grouping using the same method as in step S2 to obtain a real-time radar echo sequence dataset. Input the dataset into the radar echo prediction network model trained in step S4 to obtain the predicted future radar echo image.

[0013] As a preferred technical solution of the present invention: the preprocessing in step S1 includes data slicing, format conversion, data cleaning, coordinate transformation, data resampling and gridding processing, and data normalization for radar echo data.

[0014] As a preferred technical solution of the present invention, the specific method of data slicing is as follows: first, the target latitude and longitude range is set, and then the original radar echo data is sliced ​​according to the target latitude and longitude range to obtain the radar echo data slice required for the target area.

[0015] The specific method for format conversion is as follows: the collected radar echo data is in binary format, the radar echo data is sliced, decoded, and its data format is converted into a multidimensional array form;

[0016] The specific method for data cleaning is as follows: set negative and invalid values ​​in the radar echo data slices to zero; reset radar echo data exceeding 70dBZ to 70dBZ;

[0017] The specific method for coordinate transformation is to transform radar echo data in a three-dimensional polar coordinate system to a three-dimensional Cartesian rectangular coordinate system.

[0018] The specific method for data resampling and gridding is as follows: First, the radar echo data is resampled using spline interpolation. Then, the radar echo data is mapped to the geographic coordinate grid according to its latitude and longitude coordinates to obtain regular gridded radar echo data in the geographic coordinate system.

[0019] The specific method for data normalization is as follows: linear normalization is used to map the gridded radar echo data from 0dBZ to 70dBZ to the range of [0, 1].

[0020] As a preferred embodiment of the present invention, the specific method of sliding grouping in step S2 is as follows:

[0021] A sliding window technique is used to group radar echo data, with 6 minutes of radar echo data forming one frame. T + The continuous sequence of radar echo data is divided into a separate group, serving as a sequence sample in the radar echo sequence dataset; within the sequence sample, T Frame radar echo data was used as input features, while Frame radar echo data are defined as corresponding labels; the radar echo sequence dataset has the following dimensions: ( T + , C , H , W ),in, T + Indicates the number of time steps and the corresponding label; C Indicates the number of channels; H Indicates altitude; W Indicates width;

[0022] The specific method for data augmentation is as follows: Radar echo sequence data augmentation is performed using a rotation data augmentation method with rotation angles of 90°, 180°, and 270°.

[0023] The ratio of the training set to the test set is 9:1.

[0024] As a preferred embodiment of the present invention: the encoder in step S3 includes four sets of convolutional layers, each with a kernel size of 3×3. The hidden expression of the encoder is as follows:

[0025] ;

[0026] in, SiLU represents the nonlinear activation function. This indicates the normalization layer operation. This indicates a convolutional layer operation with a kernel size of 3×3. Represents input to the number i Tensors in each convolutional layer; i When =1, This represents the tensor input to the encoder; the stride of the convolutional layer is 1, and downsampling is performed once between every two convolutional layers with a stride of 2. The encoder performs a total of two downsampling operations.

[0027] As a preferred technical solution of the present invention: the Translator in step S3 includes eight spatiotemporal attention modules, each of which sequentially includes a channel attention module and a spatial attention module; firstly, the channel attention module extracts global information of the input features through global average pooling, and processes the input features using channel convolution and activation functions to generate a channel attention map; the channel attention map is multiplied element-wise with the input features to obtain channel-weighted features; subsequently, the channel-weighted features are input to the spatial attention module to further generate a spatial attention map; the spatial attention map is channel-mixed by the convolution module, and finally, the output features are obtained by multiplying the channel mixing result element-wise with the channel-weighted features; the hidden expression of the spatiotemporal attention module is as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] in, This represents the input features of the spatiotemporal attention module; and These represent the outputs of the channel attention module and the spatial attention module, respectively. This represents deep convolutional layer operations with kernel sizes of 1×3 and 3×1; The operation can represent multiple separable dilated convolutional layers with different kernel sizes; This represents a convolutional layer operation with a kernel size of 1×1. Represents global average pooling; Indicates the Kronecker product; It represents the Hadamardi (or Hadama) stack; This represents the output characteristics of the spatiotemporal attention module.

[0032] As a preferred embodiment of the present invention: the Decoder in step S3 is structurally symmetrical with the Encoder, including four sets of deconvolution layers with a kernel size of 3×3, and a skip connection is added between the first layer of the Encoder and the last layer of the Decoder; the hidden expression of the Decoder is as follows:

[0033] ;

[0034] in, SiLU represents the nonlinear activation function. This indicates the normalization layer operation. This indicates a deconvolutional layer operation with a kernel size of 3×3; Indicates input to the first i Tensors in a deconvolution layer i When =1, This represents the tensor input to the Decoder; the stride size of the deconvolution layer is set to 1, and then an upsampling operation is performed once for every two deconvolution layers using PixelShuffle, for a total of two upsampling operations by the Decoder.

[0035] As a preferred technical solution of the present invention: the multi-objective loss function constructed in step S4 The formula is as follows:

[0036] ;

[0037] ;

[0038] ;

[0039] in, Let the mean squared error loss function be used. For structural similarity loss, The structural similarity loss function; N It is the number of radar echo sequence data samples. It is the actual value. It is a predicted value. , They represent the first i The true and predicted values ​​of a sample of radar echo sequence data; and These are the mean squared error loss functions. and structural similarity loss function Weighting coefficients;

[0040] During training, the multi-objective loss function is calculated. Value, in the form of a multi-objective loss function The objective is to minimize the value, and backpropagation is performed by iterating through a multi-objective loss function. This process continues until the radar echo prediction network model converges. After training, the trained radar echo prediction network model is obtained.

[0041] As a preferred technical solution of the present invention: in step S5, the future radar echo image output by the radar echo prediction network model is analyzed to determine the weather conditions by analyzing the spatial characteristics of the radar echo image, and the temporal evolution characteristics of the radar echo image are analyzed to determine the weather change trend and to issue a corresponding warning.

[0042] As a preferred technical solution of the present invention: the spatial features of radar echo images include radar echo intensity distribution, radar echo morphology and structural features;

[0043] The temporal evolution characteristics of radar echo images include the direction and speed of radar echo movement, the trend of radar echo intensity change, the evolution of radar echo morphology, and the characteristics of its generation and disappearance.

[0044] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0045] By designing a spatiotemporal attention fusion mechanism in the encoder and spatiotemporal converter sections, the model's ability to model the complex spatiotemporal evolution of radar echoes is significantly enhanced.

[0046] In the spatiotemporal transformer, a lightweight spatiotemporal transformer structure (combining one-dimensional horizontal and vertical depthwise convolution (DW Conv) to capture local receptive fields, multi-scale one-dimensional horizontal and vertical depthwise separable dilated convolution (DW-D Conv) to establish long-range connections, and the fusion of spatial and channel attention modules) effectively models the nonlinear motion, morphological evolution, and multi-scale interactions of echoes, overcoming the weakness of traditional methods in spatiotemporal modeling. While significantly improving feature representation capabilities, it maintains the model's lightweight nature, avoiding the computational overhead and overfitting risk caused by excessively increasing learnable parameters. Combined with a carefully designed radar echo prediction network model, the radar echo prediction network model also exhibits good training robustness and generalization ability even with limited datasets.

[0047] Employing a multi-objective loss function (Joint mean squared error loss function) and structural similarity loss function At the same time, it optimizes the numerical accuracy (accuracy of reflectivity values) and structural consistency (echo morphology, brightness, contrast) of the prediction results, effectively supervising the generation of future radar echo images with more reasonable physical meaning and richer details.

[0048] The combined effect of these technological advantages enables the model to more accurately capture the dynamic evolution of complex weather phenomena such as storm systems (including spatial structure, direction of movement, and intensity trends). Ultimately, this significantly improves the accuracy and interpretability of short-term nowcasting (0-1 hour), providing more reliable technical support for timely and accurate early warning of meteorological disasters (heavy rain, thunderstorms, strong winds, hail, etc.), and has significant practical application value. Attached Figure Description

[0049] Figure 1 This is a flowchart of a radar echo extrapolation method based on a spatiotemporal attention mechanism according to an embodiment of the present invention;

[0050] Figure 2 This is an example diagram of radar echo data provided according to an embodiment of the present invention;

[0051] Figure 3 This is an overview diagram of the radar echo prediction network model provided in an embodiment of the present invention;

[0052] Figure 4 This is an internal structural diagram of the spatiotemporal attention module provided in an embodiment of the present invention;

[0053] Figure 5 This is a comparison chart of prediction results of different prediction models provided in the embodiments of the present invention on the radar combined reflectivity CREF dataset. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0055] This invention provides a radar echo extrapolation method based on a spatiotemporal attention mechanism, which, for a target area, refers to... Figure 1 Perform the following steps S1-S5 to complete the prediction of future radar echo images of the target area:

[0056] Step S1: Collect radar echo data of the target area, preprocess the obtained radar echo data, and construct a radar echo dataset based on the preprocessed radar echo data;

[0057] One embodiment of the present invention involves conducting a radar echo sequence prediction experiment on a radar combined reflectivity (CREF) dataset. The target region is selected as southern China, specifically within the latitude and longitude range of 109.64°E to 114.76°E and 20.34°N to 25.46°N, mainly covering Guangdong Province, Guangxi Zhuang Autonomous Region, Hong Kong, and Macau. The radar combined reflectivity (CREF) dataset for the target region consists of two years' worth of CREF data from 2023 to 2024, with a temporal resolution of 6 minutes, a spatial resolution of 0.01°, and a size of 512×512 km.

[0058] The preprocessing includes data slicing, format conversion, data cleaning, coordinate transformation, data resampling and gridding, and data normalization for radar echo data.

[0059] The specific method for data slicing is as follows: First, the target latitude and longitude range is set. Then, the original radar echo data is sliced ​​according to the target latitude and longitude range to obtain the radar echo data slices required for the target area. There are usually no strict restrictions on the shape of the slices. For the convenience of subsequent analysis, they are generally cut into rectangles or squares. They can also be cut into irregular shapes such as polygons according to the terrain of the target area. However, it is necessary to ensure the integrity and consistency of the slices so as to cover key information and not affect the model input. In terms of resolution, it can be selected according to research needs and data accuracy. High resolution is beneficial for detailed analysis but the data volume is large, while low resolution is suitable for large-scale research and preliminary analysis.

[0060] The specific method for format conversion is as follows: the collected radar echo data is in binary format. The radar echo data is sliced, decoded, and its data format is converted into a multidimensional array. The NumPy library is usually used for this operation. Finally, the processed array data is saved as a .npy file for efficient data reading and processing in the future.

[0061] The specific data cleaning method is as follows: Negative and invalid values ​​in the radar echo data slices are set to zero to exclude invalid or unreliable data and avoid interference with subsequent analysis; radar echo data exceeding 70 dBZ is reset to 70 dBZ; and the echo intensity range of the radar echo data slices is ensured to be 0-70 dBZ. Simultaneously, abnormal data in the radar echo data is removed; examples of abnormal radar echo data are shown in the image below. Figure 2 ;

[0062] The specific method for coordinate transformation is to transform radar echo data in a three-dimensional polar coordinate system to a three-dimensional Cartesian rectangular coordinate system.

[0063] The specific method for data resampling and gridding is as follows: First, the radar echo data is resampled using spline interpolation to ensure that the radar echo data has a uniform spatial resolution; then, the radar echo data is mapped to the geographic coordinate grid according to its latitude and longitude coordinates to obtain regular gridded radar echo data under the geographic coordinate system, so as to form a standardized input format.

[0064] The specific method for data normalization is as follows: linear normalization is used to map the gridded radar echo data from 0dBZ to 70dBZ to the range [0, 1]. The normalization formula is as follows:

[0065] ;

[0066] in, It is the intensity value of the radar echo data. and These are the minimum (0dBZ) and maximum (70dBZ) values ​​in the radar echo data, respectively. Normalized radar echo data values;

[0067] After the acquired radar echo data is preprocessed as described above, a single frame of radar echo data is obtained.

[0068] Step S2: Perform sliding grouping on the radar echo dataset obtained in step S1 to obtain radar echo sequence data, and perform data augmentation to obtain radar echo sequence dataset. Then, divide the radar echo sequence dataset into training set and test set.

[0069] The specific method for sliding grouping in step S2 is as follows:

[0070] A sliding window technique is used to group radar echo data, with 6 minutes of radar echo data forming one frame. T + The continuous sequence of radar echo data is divided into a separate group, serving as a sequence sample in the radar echo sequence dataset; within the sequence sample, T Frame radar echo data was used as input features, while Frame radar echo data are defined as corresponding labels; the radar echo sequence dataset has the following dimensions: ( T + , C , H , W ),in, T + Indicates the number of time steps and the corresponding label; C Indicates the number of channels; H Indicates altitude; W Indicates width;

[0071] The specific method for data augmentation is as follows: Radar echo sequence data augmentation is performed using a rotation data augmentation method with rotation angles of 90°, 180°, and 270°.

[0072] The ratio of the training set to the test set is 9:1.

[0073] In this embodiment, a radar echo sequence dataset for the experiment is constructed based on the radar echo dataset obtained in step S1. 90% of the radar echo sequence dataset is used as the training set, and the remaining 10% is used as the test set. Each radar echo sequence data includes 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 labels for the radar echo prediction network model. The model parameters are optimized using a loss function along with the model output. That is, the radar echo sequence of the past hour is used to predict the radar echo sequence of the next hour.

[0074] Step S3: Construct a radar echo prediction network model based on the SimVP architecture, referring to... Figure 3 It includes an encoder, a translator, and a decoder connected in sequence; the encoder is also connected to the decoder. The radar echo prediction network model takes radar echo sequence data in the radar echo sequence dataset as input and the predicted future radar echo image of the target area as output.

[0075] To encode the input high-dimensional radar echo past frames into a low-dimensional latent space, the encoder uses four sets of convolutional layers, each with a kernel size of 3×3. The latent expression of the encoder is as follows:

[0076]

[0077] in, SiLU represents the nonlinear activation function. This indicates the Groupnorm operation for the normalization layer. This indicates a convolutional layer operation with a kernel size of 3×3. Represents input to the number i Tensors in each convolutional layer; i When =1, This represents the tensor input to the encoder; the stride of the convolutional layer is 1, and downsampling is performed once between every two convolutional layers with a stride of 2. The encoder performs a total of two downsampling operations.

[0078] The Translator spatiotemporal converter includes eight spatiotemporal attention modules, referencing... Figure 4Each spatiotemporal attention module sequentially comprises a channel attention module and a spatial attention module. First, the channel attention module extracts global information from the input features using global average pooling and processes the input features with channel convolution and activation functions to generate a channel attention map. The channel attention map is then element-wise multiplied with the input features to obtain channel-weighted features. Subsequently, the channel-weighted features are input to the spatial attention module to further generate a spatial attention map. The spatial attention map undergoes channel blending via a convolution module to enhance the integration ability between features. Finally, the output features are obtained by element-wise multiplying the channel blending result with the channel-weighted features. The hidden expression of the spatiotemporal attention module is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] in, This represents the input features of the spatiotemporal attention module. B represents the batch size, T represents the time step, C represents the number of channels, H represents the height, and W represents the width; that is, the output of the 4th convolutional layer of the encoder. and These represent the outputs of the channel attention module and the spatial attention module, respectively. , ; These represent deep convolutional layer operations with kernel sizes of 1×3 and 3×1, used to capture the local receptive field within a single channel; The operation of depth-separable dilated convolutional layers, representing multiple convolutional kernels of different sizes, is used to establish connections between distant receptive fields. This represents a 1×1 convolutional layer operation, used to implement interaction between channels; Represents global average pooling; Indicates the Kronecker product; It represents the Hadamardi (or Hadama) stack; This represents the output characteristics of the spatiotemporal attention module.

[0083] The decoder has a symmetrical structure to the encoder, including four deconvolutional layers with a kernel size of 3×3. Skip connections are added between the first layer of the encoder and the last layer of the decoder to preserve spatial features. The hidden expression of the decoder is as follows:

[0084] ;

[0085] in, SiLU represents the nonlinear activation function. This indicates the Groupnorm operation for the normalization layer. This indicates a deconvolutional layer operation with a kernel size of 3×3; Indicates input to the first i Tensors in a deconvolution layer i When =1, This represents the tensor input to the Decoder; the stride size of the deconvolution layer is set to 1, and then an upsampling operation is performed once for every two deconvolution layers using PixelShuffle, which is an upsampling method. The Decoder performs a total of two upsampling operations.

[0086] In step S3, assume that the spatiotemporal tensor of the input radar echo sequence data is... Here, B represents the batch size, T represents the number of time steps, C represents the number of channels, H represents the height, and W represents the width. In the encoder and decoder, the tensor is first reshaped into (B×T, C, H, W) to treat the spatial features of each time step as independent samples, considering only spatial correlations. In the translator, the tensor is reshaped into (B, T×C, H, W), capturing spatial and temporal correlations by treating time as an additional channel dimension and learning the evolution of temporal features.

[0087] Step S4: Train the radar echo prediction network model constructed in Step S3 using the training set divided in Step S2. Adjust parameters such as training batch, learning rate, optimizer, and early stopping, and construct a multi-objective loss function. Use the multi-objective loss function to supervise the training of the radar echo prediction network model. Use the divided radar echo sequence test set to test the radar echo prediction network model to obtain the predicted future radar echo image of the target area. Use meteorological evaluation indicators to evaluate the prediction performance of the radar echo prediction network model.

[0088] Obtain a trained radar echo prediction network model;

[0089] Constructed multi-objective loss function The formula is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] in, The mean squared error loss function directly minimizes the pixel-by-pixel difference between the predicted and true values, ensuring the accuracy of the overall reflectance values. For structural similarity loss, The structural similarity loss function focuses on the structural similarity of the echo signals, considering not only numerical differences but also factors such as brightness, contrast, and structure; N is the number of radar echo sequence data samples. It is the actual value. It is a predicted value. , They represent the first i The true and predicted values ​​of a sample of radar echo sequence data;

[0094] Through weighted combination and Multi-objective loss function optimization improves numerical accuracy and structural consistency; and These are the mean squared error loss functions. and structural similarity loss function In one embodiment, the weighting coefficients are taken as 0.64 and 0.36, respectively;

[0095] The radar echo sequence tensor is input into the radar echo prediction network model as follows: Where B represents batch size, T represents time step number, C represents channel number, H represents height, and W represents width, a dynamic radar echo sequence data is formed based on the changes over T time steps. Future radar echo sequences are then predicted based on the given T frames of radar echo sequence data. Frame. Using a multi-target loss function. The radar echo prediction neural network model is trained; during training, a multi-target loss function is calculated. Value, in the form of a multi-objective loss function The objective is to minimize the value, and backpropagation is performed by iterating through a multi-objective loss function. This process continues until the radar echo prediction network model converges. After training, the trained radar echo prediction network model is obtained.

[0096] The prediction results of the radar echo prediction network model are evaluated using meteorological evaluation indicators such as hit rate (POD), false alarm rate (FAR), critical success index (CSI), and accuracy, as detailed below:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] 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.

[0102] In this embodiment, the training set obtained in step S2 is used to train the model. The OneCycleLR learning rate scheduling strategy with an initial learning rate of 1e-4 is used, the Adam optimizer is used, and the batch size of all training is set to 4. In order to prevent the model from overfitting, the early stopping patience value is set to 20 epochs, that is, if no improvement in the validation loss is observed in 20 consecutive epochs, the training will be terminated.

[0103] The acquired radar echo sequences from the test set are then input into the trained radar echo prediction network to obtain future radar echo images of the target area. The prediction results are then evaluated using appropriate meteorological evaluation metrics. Table 1 shows the meteorological evaluation metrics for the prediction results of different methods on the CREF radar combined reflectivity dataset.

[0104] Table 1. Prediction Results and Meteorological Evaluation Indicators of Different Methods

[0105]

[0106] As shown in Table 1, compared with existing mainstream prediction models (such as SimVP, TAU, and ConvLSTM), the method of this invention exhibits higher prediction accuracy in key meteorological evaluation indicators (Popularity of Detection (POD), False Alarm Rate (FAR), and Critical Success Index (CSI), especially in longer prediction periods (such as 60 minutes). The input consists of 10 frames of radar echo sequence data. The actual 10 frames of radar echo sequence data and the 10 frames of radar echo sequence data predicted by different model methods are shown below. Figure 5 As shown, by Figure 5 It can be seen that the prediction results of each model are quite satisfactory within half an hour (the first six frames), but as time goes on, the prediction results of the method of this invention are significantly better than those of other methods.

[0107] The test results of the above embodiments demonstrate the effectiveness and superiority of the present invention in radar echo sequence prediction method.

[0108] Step S5: Deploy the trained radar echo prediction network model to the computing platform, collect radar echo data in real time, perform preprocessing using the same method as in Step S1, and perform sliding grouping using the same method as in Step S2 to obtain a real-time radar echo sequence dataset. Input this dataset into the radar echo prediction network model trained in Step S4 to obtain the predicted future radar echo image. Analyze the spatial features of the radar echo image to determine weather conditions, and analyze the temporal evolution features of the radar echo image to determine weather change trends. This achieves end-to-end real-time processing and prediction of radar echo data, performs weather condition analysis, outputs the results, and applies them to meteorological disaster early warning.

[0109] The specific method is as follows:

[0110] (1) Real-time data acquisition and preprocessing. Continuously receive real-time radar echo data from the meteorological radar station, and process the received real-time radar echo data strictly in accordance with the preprocessing procedure defined in step S1;

[0111] (2) Model Input and Future Echo Prediction. The preprocessed real-time radar echo sequence data is used as input data and input into the radar echo prediction network model trained and optimized in step S4. This model has been deployed on a computer system with corresponding computing capabilities (such as a server or workstation);

[0112] (3) Analyze the spatial features of radar echo images to determine weather conditions. Perform spatial feature analysis on each frame of future predicted radar echo image output by the radar echo prediction network model to infer the specific weather conditions at the corresponding future time.

[0113] The spatial features analyzed include:

[0114] Echo intensity distribution: Identifying the spatial extent, intensity, and structure of echo areas at different intensity levels (e.g., dBZ values) in radar echo images. Strong echo areas (e.g., >40 dBZ) typically indicate heavy precipitation (heavy rain, hail), moderate echo areas (e.g., >25 dBZ) indicate ordinary rain / snow, and weak echo areas (e.g., >15 dBZ) may indicate weak precipitation or clouds.

[0115] Echo morphology and structural characteristics: Identify echo morphologies with specific weather significance, such as: linear echoes (squall lines), blocky echoes and cluster echoes (thunderstorm cells), bow echoes (downburst risk), comma-shaped cloud echoes (mid-latitude cyclones), and stratiform cloud echoes (stable precipitation). Analyze the echo gradient, texture, and boundary sharpness.

[0116] Key area identification: Identify severe weather areas in radar echo images that may have a significant impact on specific areas (such as cities, transportation hubs, disaster-prone areas).

[0117] Based on the comprehensive analysis of the above spatial characteristics, the system can automatically determine or assist forecasters in determining the dominant weather type (such as clear sky, light rain, moderate rain, heavy rain, rainstorm, thunderstorm, hail, snowfall, etc.) and its potential intensity at a specific time and in a specific region in the future.

[0118] (4) Analyze the temporal evolution characteristics of radar echo images to determine the weather change trend. Take the continuous T-frame future radar echo image sequence output by the radar echo prediction network model as a whole and analyze its dynamic change characteristics in the time dimension to reveal the movement, development and weakening trend of the weather system.

[0119] The temporal evolution characteristics analyzed include:

[0120] Echo movement direction and velocity: Track the centroid movement trajectory of the main echo clusters, calculate their movement direction and velocity, and predict their future influence area and arrival time.

[0121] Echo intensity variation trend: Analyze the change (intensification, maintenance, weakening) of echo intensity in a specific area (or echo cell) over time to predict the evolution of precipitation intensity (such as increased or decreased rainfall).

[0122] Echo morphology evolution: Observe the merging, splitting, dissipation, and formation of echo morphology to predict the development stage (initial, mature, dissipation) of convective systems (heavy rainfall, thunderstorms, etc.) or the evolution of weather systems (such as the passage of fronts).

[0123] Characteristics of the generation and dissipation of echoes: Identify areas where new echoes are generated (which may indicate a convective outbreak) and areas where echoes dissipate rapidly (the end of precipitation).

[0124] By analyzing the temporal evolution characteristics of the entire forecast sequence, we can infer the continuous changing trend of weather conditions over a future period (such as 0-1 hour).

[0125] (5) Output and application of short-term forecast product results. The analysis results of (3) and (4) (including predicted radar echo images, determined weather conditions, predicted weather change trends, key early warning information, etc.) are integrated. The results can be displayed to forecasters in real time through a visualization interface (such as a meteorological forecast platform, GIS map), or automatically generated into structured forecast text and early warning signals. The real-time generated short-term forecast products can be directly used in application scenarios such as meteorological disaster early warning (heavy rain, thunderstorms, strong winds, hail, etc.), urban operation management, traffic scheduling, major event security, and refined weather forecast services for the public.

[0126] The core of this invention lies in the spatiotemporal attention mechanism within the Translator spatiotemporal converter. This mechanism, achieved collaboratively by a spatial attention module and a channel attention module, enables efficient modeling of complex spatiotemporal dynamics. The spatial attention module utilizes deep convolutional layers (DW Conv) to extract local features and depth-separable dilated convolutional layers (DW-D Conv) to establish long-range dependencies, enhancing the spatial resolution of key weather structures. The channel attention module generates temporal weights through global average pooling and 1×1 convolutions, capturing temporal evolution features. Furthermore, this invention employs a multi-objective loss function optimization supervision mechanism to simultaneously optimize the numerical accuracy and structural consistency of predicted echoes. This invention encompasses an end-to-end workflow from data preprocessing, sequence construction, model training, prediction, and weather analysis, applicable not only to all radar echo sequences but also to spatiotemporal sequence prediction tasks in other similar application scenarios.

[0127] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A radar echo extrapolation method based on a spatiotemporal attention mechanism, characterized in that, For the target area, perform the following steps S1-S5 to predict the future radar echo image of the target area: Step S1: Collect radar echo data of the target area, preprocess the obtained radar echo data, and construct a radar echo dataset based on the preprocessed radar echo data. Step S2: Perform sliding grouping on the radar echo dataset obtained in step S1 to obtain radar echo sequence data, and perform data augmentation to obtain radar echo sequence dataset. Then, divide the radar echo sequence dataset into training set and test set. Step S3: Construct a radar echo prediction network model based on the SimVP architecture, including an encoder, a translator, and a decoder connected in sequence; the encoder is also connected to the decoder. The radar echo prediction network model takes radar echo sequence data in the radar echo sequence dataset as input and the predicted future radar echo image of the target area as output. The encoder consists of four convolutional layers, each with a kernel size of 3×3. The hidden expression of the encoder is as follows: ； in, SiLU represents the nonlinear activation function. This indicates the normalization layer operation. This indicates a convolutional layer operation with a kernel size of 3×3. Represents input to the number i Tensors in each convolutional layer; i When =1, This represents the tensor input to the encoder; the stride of the convolutional layer is 1, and downsampling is performed once between every two convolutional layers with a stride of 2. The encoder performs a total of two downsampling operations. The Translator spatiotemporal converter comprises eight spatiotemporal attention modules, each containing a channel attention module and a spatial attention module. First, the channel attention module extracts global information from the input features using global average pooling and processes these features with channel convolution and activation functions to generate a channel attention map. The channel attention map is then element-wise multiplied with the input features to obtain channel-weighted features. Subsequently, the channel-weighted features are input to the spatial attention module to further generate a spatial attention map. The spatial attention map undergoes channel blending via a convolution module. Finally, the output features are obtained by element-wise multiplying the channel blending result with the channel-weighted features. The hidden expression for the spatiotemporal attention module is as follows: ； ； ； in, This represents the input features of the spatiotemporal attention module; and These represent the outputs of the channel attention module and the spatial attention module, respectively. This represents deep convolutional layer operations with kernel sizes of 1×3 and 3×1; The operation can represent multiple separable dilated convolutional layers with different kernel sizes; This represents a convolutional layer operation with a kernel size of 1×1. Represents global average pooling; Indicates the Kronecker product; It represents the Hadamardi (or Hadama) stack; This represents the output characteristics of the spatiotemporal attention module; The decoder has a symmetrical structure to the encoder, including four deconvolutional layers with a kernel size of 3×3, and a skip connection is added between the first layer of the encoder and the last layer of the decoder; the hidden expression of the decoder is as follows: ； in, SiLU represents the nonlinear activation function. This indicates the normalization layer operation. This indicates a deconvolutional layer operation with a kernel size of 3×3; Indicates input to the first i Tensors in a deconvolution layer i When =1, This represents the tensor input to the Decoder; the stride size of the deconvolution layer is set to 1, and then an upsampling operation is performed once for every two deconvolution layers using PixelShuffle, for a total of two upsampling operations in the Decoder; Step S4: Use the training set and test set divided in step S2 to train and test the radar echo prediction network model constructed in step S3, and construct a multi-target loss function. Use the multi-target loss function to supervise the training of the radar echo prediction network model to obtain the trained radar echo prediction network model. Step S5: Collect radar echo data in real time, preprocess it using the same method as in step S1, and perform sliding grouping using the same method as in step S2 to obtain a real-time radar echo sequence dataset. Input the dataset into the radar echo prediction network model trained in step S4 to obtain the predicted future radar echo image.

2. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 1, characterized in that, The preprocessing described in step S1 includes data slicing, format conversion, data cleaning, coordinate transformation, data resampling and gridding, and data normalization for radar echo data.

3. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 2, characterized in that, The specific method for data slicing is as follows: First, the target latitude and longitude range is set, and then the original radar echo data is sliced ​​according to the target latitude and longitude range to obtain the radar echo data slices required for the target area. The specific method for format conversion is as follows: the collected radar echo data is in binary format, the radar echo data slices are decoded, and their data format is converted into a multidimensional array form; The specific method for data cleaning is as follows: set negative and invalid values ​​in the radar echo data slices to zero; reset radar echo data exceeding 70dBZ to 70dBZ; The specific method for coordinate transformation is to transform radar echo data in a three-dimensional polar coordinate system to a three-dimensional Cartesian rectangular coordinate system. The specific method for data resampling and gridding is as follows: First, the radar echo data is resampled using spline interpolation. Then, the radar echo data is mapped to the geographic coordinate grid according to its latitude and longitude coordinates to obtain regular gridded radar echo data in the geographic coordinate system. The specific method for data normalization is as follows: linear normalization is used to map the gridded radar echo data from 0dBZ to 70dBZ to the range of [0, 1].

4. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 1, characterized in that, The specific method for sliding grouping in step S2 is as follows: A sliding window technique is used to group radar echo data, with 6 minutes of radar echo data forming one frame. T + The continuous sequence of radar echo data is divided into a separate group, serving as a sequence sample in the radar echo sequence dataset; within the sequence sample, T Frame radar echo data was used as input features, while Frame radar echo data are then defined with corresponding tags; The radar echo sequence dataset has the following dimensions: ( T + , C , H , W ),in, T + Indicates the number of time steps and the corresponding label; C Indicates the number of channels; H Indicates altitude; W Indicates width; The specific method for data augmentation is as follows: Radar echo sequence data augmentation is performed using a rotation data augmentation method with rotation angles of 90°, 180°, and 270°. The ratio of the training set to the test set is 9:

1.

5. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 1, characterized in that, The multi-objective loss function constructed in step S4 The formula is as follows: ； ； ； in, Let the mean squared error loss function be used. For structural similarity loss, The structural similarity loss function; N It is the number of radar echo sequence data samples. It is the actual value. It is a predicted value. , They represent the first i The true and predicted values ​​of a sample of radar echo sequence data; and These are the mean squared error loss functions. and structural similarity loss function Weighting coefficients; During training, the multi-objective loss function is calculated. Value, in the form of a multi-objective loss function The objective is to minimize the value, and backpropagation is performed by iterating through a multi-objective loss function. This process continues until the radar echo prediction network model converges. After training, the trained radar echo prediction network model is obtained.

6. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 1, characterized in that, In step S5, based on the future radar echo image output by the radar echo prediction network model, the spatial characteristics of the radar echo image are analyzed to determine the weather conditions, the temporal evolution characteristics of the radar echo image are analyzed to determine the weather change trend, and corresponding early warnings are issued.

7. The radar echo extrapolation method based on spatiotemporal attention mechanism according to claim 6, characterized in that, The spatial characteristics of radar echo images include radar echo intensity distribution, radar echo morphology, and structural features. The temporal evolution characteristics of radar echo images include the direction and speed of radar echo movement, the trend of radar echo intensity change, the evolution of radar echo morphology, and the characteristics of its generation and disappearance.

Images