A radar intelligent echo extrapolation method based on global-local aggregation model
The radar intelligent echo extrapolation method based on a global-local aggregation model solves the problem of underestimating the coverage and intensity of heavy rainfall in existing technologies, and achieves accurate prediction of meteorological radar echoes and improved precipitation skill scores.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2022-09-30
- Publication Date
- 2026-06-23
AI Technical Summary
Existing radar echo extrapolation models suffer from systematic underestimation of the coverage and intensity of heavy rainfall, and cannot effectively capture the global and local spatiotemporal interactions of echo sequences.
A radar intelligent echo extrapolation method based on a global-local aggregation model is adopted. By constructing a global-local aggregation model and combining it with an encoding-prediction architecture, echo motion representation guided by optical flow information, global-local aggregation branch based on attention mechanism and channel attention feature fusion, the accurate prediction of meteorological radar echoes can be achieved.
It improved the ability to predict the coverage area and intensity of heavy rainfall, enhanced the precipitation skill score, and significantly improved the prediction performance, especially in long-term series forecasting.
Smart Images
Figure CN115598611B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of weather radar technology, and in particular to a radar intelligent echo extrapolation method based on a global-local aggregation model. Background Technology
[0002] Short-term precipitation forecasting is of great significance for meteorological disaster prevention and mitigation. Weather radar has the advantages of wide coverage and rapid data updates, enabling timely acquisition of large-area precipitation data and providing high-quality observational information for short-term forecasting. Weather radar echo extrapolation is the most commonly used technique for short-term forecasting. Current mainstream echo extrapolation methods include traditional linear extrapolation based on estimating radar echo motion vectors, and intelligent extrapolation methods based on deep learning models.
[0003] Traditional linear extrapolation methods based on estimating radar echo motion vectors, including centroid tracking and optical flow methods, offer advantages such as simplicity and high predictive clarity. However, these methods result in significant distortion and warping in the predicted radar images, failing to capture the complex nonlinear motion of real-world precipitation echoes and exhibiting poor precipitation skill scores. Furthermore, traditional methods cannot fully utilize the vast amounts of historical weather radar echo data.
[0004] Intelligent extrapolation methods model the meteorological radar echo extrapolation problem as a deep spatiotemporal sequence prediction problem, relying on pure convolutional networks or convolutional recurrent neural networks to construct the echo extrapolation model. While demonstrating potential superiority over traditional methods in precipitation skill scoring, they still suffer from a systematic underestimation of the coverage and intensity of heavy rainfall. This is because, due to the limited receptive field of the model and the scarcity of natural heavy rainfall data, existing echo extrapolation models inherently struggle to simultaneously capture the global and local spatiotemporal interactions of the echo sequence, thus limiting their extrapolation performance. Summary of the Invention
[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a radar intelligent echo extrapolation method based on a global-local aggregation model. The main technical problem solved is that the existing extrapolation model systematically underestimates the coverage area and intensity of heavy rainfall.
[0006] The technical solution of this invention is:
[0007] A radar intelligent echo extrapolation method based on a global-local aggregation model, the steps of which include:
[0008] A. Obtain the meteorological radar precipitation echo image sequence dataset stored in grayscale image format, and obtain the optical flow sequence dataset corresponding to the meteorological radar precipitation echo image sequence based on the obtained meteorological radar precipitation echo image sequence dataset.
[0009] B. Construct a global-local aggregation model;
[0010] The constructed global-local aggregation model includes a prediction branch based on an encoding-prediction architecture and a global-local aggregation branch based on an attention mechanism;
[0011] The attention-based global-local aggregation branch includes an echo motion representation submodule guided by optical flow information, an attention-based global-local aggregation submodule, and a feature fusion submodule based on channel attention.
[0012] C. Use the meteorological radar precipitation echo image sequence dataset and meteorological radar precipitation echo optical flow sequence dataset obtained in step A to train the global-local aggregation model constructed in step B, and obtain the trained global-local aggregation model.
[0013] D. Input the test set sample data into the global-local aggregation model trained in step C to obtain the echo extrapolation prediction image sequence of the test set sample data, and complete the radar intelligent echo extrapolation based on the global-local aggregation model.
[0014] In step A, the pixel grayscale value in the meteorological radar precipitation echo image sequence dataset stored in grayscale image form represents the true echo reflectivity factor, and the default reflectivity value is set to a pixel value of 255 in the grayscale image.
[0015] In step B, the encoder-prediction architecture consists of two parts: an encoder and a predictor. The encoder alternately sets up convolutional downsampling layers and convolutional long short-term memory network units, and the predictor alternately sets up convolutional upsampling layers and convolutional long short-term memory network units. Each layer of convolutional long short-term memory network units in the encoder corresponds one-to-one with each layer of convolutional long short-term memory network units in the predictor. In the prediction branch based on the encoder-prediction architecture, the encoder input is the currently observed echo sequence, and the output is the echo sequence to be predicted.
[0016] In step B, the echo motion characterization submodule guided by optical flow information has an encoder-decoder structure. The encoder input is the currently observed echo sequence, and the output is the optical flow sequence corresponding to the echo sequence. The echo motion characterization submodule guided by optical flow information consists of one residual connection layer, three downsampling residual connection layers, three upsampling layers, a 1×1×1 convolutional layer, a residual connection layer, and one 1×1×1 convolutional layer. The meteorological radar precipitation echo image first passes through the one residual connection layer, and then through the three downsampling layers. The image is processed by a residual connection layer, followed by three convolutional layers consisting of an upsampling layer, a 1×1×1 convolutional layer, and a residual connection layer stacked sequentially. Finally, a 1×1×1 convolutional layer is used to adjust the number of channels to obtain the optical flow sequence corresponding to the input meteorological radar precipitation echo image. Skip-layer connections are added between the downsampling layer and the upsampling layer. After the third downsampling layer of the encoder, a motion semantic segmenter is connected. The input of the motion semantic segmenter is the motion semantic features output from the third layer of the encoder, and the output is vectorized abstract motion information.
[0017] In step B, the input to the global-local aggregation submodule based on the attention mechanism is vectorized abstract motion information. The core components of the global-local aggregation submodule are the self-attention mechanism and the cross-attention mechanism. Formally, it adopts a Transformer decoder stacked structure and is equipped with a global information memory pool. First, the vectorized abstract motion information is linearly mapped to obtain query, key, and value vectors, and then the self-attention mechanism is performed. Subsequently, after residual connection and layer normalization, the output feature is used as the query vector of the next layer cross-attention mechanism. When calculating the cross-attention mechanism, the key and value vector pairs are provided by the memory pool. After the cross-attention mechanism is performed, it is then processed through two layers of residual connection and layer normalization (with a feedforward neural network layer in between the two layers) to obtain the output feature of the global-local aggregation submodule.
[0018] In step B, the feature fusion submodule based on channel attention performs a weighted fusion of the cellular state and hidden state features of the deepest convolutional long short-term memory network units in the prediction branch with the output features of the global-local aggregation submodule, specifically as follows:
[0019] First, the cellular features and the output features of the global-local aggregation submodule are concatenated in the channel dimension. Then, the concatenated features are fed into a max pooling layer and an average pooling layer respectively for further spatial dimension aggregation, and then passed through a shared feedforward neural network layer. The two features are then added together and fed into an activation function to obtain channel attention weights. The channel attention weights are multiplied by the output features of the global-local aggregation submodule to achieve channel attention weighting. The weighted features are concatenated in the channel dimension with the latent features and passed through a 1×1 convolutional layer to adjust the number of channels, resulting in refined latent features. Finally, the refined latent features are passed layer by layer through upsampling layers and shallow convolutional long short-term memory network units to iteratively obtain the output echo image.
[0020] In step C, the meteorological radar precipitation echo image sequence dataset obtained in step A is normalized and then divided into training set, validation set and test set according to a set ratio.
[0021] The optical flow sequence dataset corresponding to the meteorological radar precipitation echo image sequence obtained in step A is normalized, and then divided into training set, validation set and test set according to a set ratio.
[0022] In step C, when training the global-local aggregation model, a two-stage training strategy is set up. The first training stage is the memory pool update stage. The input of this stage is the long-time radar echo sequence. The length of the long-time radar echo sequence is less than or equal to the total length of each radar echo sequence sample in the training set, but greater than the input sequence length specified in the extrapolation algorithm test stage. During the training stage, the weights of the global information memory pool are continuously updated. This stage can be regarded as the long-time echo sequence motion feature storage stage. The second training stage is the memory pool fixing stage. The input of this stage is the short-time radar echo sequence, i.e., the input sequence length specified in the extrapolation algorithm test stage. In this stage, the weights and gradients of the memory pool are locked and no longer updated. This stage can be regarded as the matching stage between the motion features of the short-time echo sequence and the stored motion features of the long-time echo sequence.
[0023] When training a global-local aggregation model, the memory pool update phase and the memory pool fixation phase are performed alternately, and the model with the best performance on the validation set is saved for testing.
[0024] The precipitation rate is estimated by extrapolating the reflectance factor in the image sequence based on the echo of the test set sample data and applied to short-term precipitation forecasting.
[0025] The beneficial effects of this invention are:
[0026] (1) According to the radar intelligent echo extrapolation method based on global-local aggregation model provided by the present invention, a global-local aggregation model is built and trained by constructing a dataset of grayscale image sequences of meteorological radar echoes and their corresponding optical flow sequence datasets, so as to achieve accurate prediction of future meteorological radar echo image sequences.
[0027] (2) This scheme introduces a novel motion information-guided global-local aggregation model, which uses the optical flow information of the echo sequence as motion guidance information and effectively integrates the spatiotemporal information of the echo sequence at different time scales with the help of the attention mechanism. It emphasizes the global-local spatiotemporal aggregation urgently needed in the echo extrapolation task and improves the ability to predict the coverage area and intensity of heavy rainfall in long-term series. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the radar intelligent echo extrapolation method based on the global-local aggregation model of the present invention;
[0029] Figure 2 This is a schematic diagram of the global-local aggregation model of the present invention;
[0030] Figure 3 This is a schematic diagram of the echo motion characterization submodule of the present invention;
[0031] Figure 4 This is a schematic diagram of the channel attention module of the present invention. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0033] Example
[0034] This embodiment provides a radar intelligent echo extrapolation method based on a global-local aggregation model. Please refer to [link to relevant documentation]. Figure 1-4 The main steps include the following:
[0035] S101. Obtain the weather radar echo grayscale image sequence and its corresponding optical flow sequence dataset.
[0036] S102. Build a global-local aggregation model;
[0037] S103. Train the global-local aggregation model. After the model training is complete, input the test set sample data into the trained model to obtain the echo extrapolation predicted image sequence.
[0038] Obtain the dataset of meteorological radar precipitation echo image sequences stored in grayscale image format, and the corresponding optical flow sequence dataset. The meteorological radar precipitation echo images are stored in grayscale format, where the pixel grayscale value represents the true echo reflectivity. The default reflectivity value in the grayscale image is set to a pixel value of 255. Calculate the optical flow field sequence data of the echo image sequences using EpicFlow; normalize the radar image data and divide it into training, validation, and test sets in a 4:1:1 ratio. The optical flow field sequence data are also correspondingly divided into training, validation, and test sets.
[0039] Given a historical radar echo sequence As input to the extrapolation model, S represents 1:t The k-th frame echo image. The echo image is stored as a grayscale image, therefore X k The number of channels C = 1. Extrapolation model The aim is to make the predicted sequence As close as possible to the actual observation S t+1:T In this embodiment, a dataset is constructed using weather radar image data updated every 6 minutes. The input sequence length t is set to 10, and the total sequence length T is set to 20. That is, based on the weather radar observations of the past hour, the weather radar echo sequence for the next hour is predicted. In this embodiment, the spatial scale of the weather radar image is H, W = 384, 384.
[0040] Build a global-local aggregation model. For example... Figure 2 As shown in (a) and (b), the overall model structure includes a prediction branch based on an encoder-prediction architecture and a global-local aggregation branch based on an attention mechanism. The encoder-prediction architecture consists of an encoder and a predictor. The encoder alternates between four convolutional downsampling layers and four convolutional long short-term memory (LSM) network units, while the predictor alternates between four convolutional upsampling layers and four convolutional LSM network units. Each LSM network unit in the encoder corresponds one-to-one with each LSM network unit in the predictor. The number of hidden state channels in the four LSM network units is set to 16, 64, 128, and 128 from shallowest to deepest, respectively.
[0041] The global-local aggregation branch includes a motion representation submodule, an attention-based global-local aggregation submodule, and a feature fusion submodule based on channel attention. For example... Figure 3As shown in (a), the motion representation module is an encoder-decoder structure. The encoder input is the currently observed echo sequence, and the output is the optical flow sequence corresponding to the echo sequence. The input echo sequence first passes through one residual connection layer, then through three downsampled residual connection layers, then through three convolutional layers consisting of an upsampled layer, a 1×1×1 convolutional layer, and a residual connection layer stacked sequentially, and finally through one 1×1×1 convolutional layer to adjust the number of channels, thus obtaining the optical flow sequence corresponding to the input echo sequence. like Figure 3 As shown by the dashed line in (a), a skip connection is added between the downsampling layer and the upsampling layer. Specifically, as... Figure 3 As shown in (b), the residual layer backbone is composed of batch normalization (BN) layers, leaky linear rectified function (LReLU), 3×3×3 convolutional layers with stride of 1, batch normalization (BN) layers, leaky linear rectified function (LReLU), random deactivation layers (Dropout), and 3×3×3 convolutional layers with stride of 1 (Conv 3×3×3, stride=1) stacked together. Figure 3 As shown in (c), the downsampling layer can be considered a special residual layer. The stride of the first convolutional layer in its main branch is set to 2 (Conv 3×3×3, stride = 2). An identity mapping branch adds a 3×3×3 convolution with a stride of 2 and a batch normalization layer. The output channels of the first residual layer are set to 32. Afterward, each time a feature map passes through a downsampling block, its channel dimension doubles. The remaining three upsampling blocks maintain their channel dimensions. The first three 1×1×1 convolutional layers are used to adjust the channel dimensions of skip connections. The last 1×1×1 convolutional layer has a channel dimension of 2 for 2D optical flow vector estimation. After the third downsampling layer of the encoder, an additional motion semantic segmenter is connected. The input to the motion semantic segmenter is the motion semantic features output from the third layer of the encoder. Where D t D hw These are the downsampling factors for the time scale and the spatial scale, respectively, set to 2 and 12 in this embodiment. C1 is the number of channels for feature A, set to 128. Figure 3 As shown in (d), the motion semantic segmenter consists of a 3×3×3 convolutional layer with a stride of 1, a leaky linear rectified function (LReLU) layer, a max pooling layer (MaxPool), an average pooling layer (AvgPool), and a flattening operation layer stacked together. The output is vectorized abstract motion information. Here, L is set to 256.
[0042] Vectorized abstract motion information input is based on a global-local aggregation submodule using an attention mechanism. For example... Figure 2As shown in (b) and (c), the core components of the global-local aggregation submodule are the self-attention mechanism and the cross-attention mechanism. Formally, it adopts a Transformer decoder stacked structure and is equipped with a global information memory pool. First, the vectorized abstract motion information is linearly mapped to obtain the query Q, key K, and value vector V. Then, the self-attention mechanism is used for computation, denoted as follows:
[0043]
[0044]
[0045] in Let N be the query, key, and value vector for each attention head in a multi-head self-attention mechanism. heads The number of heads used is set to 8 here, i = 1, 2, 3, ..., N head ,d k The scaling factor is set to 128 here, h is the output of the multi-head self-attention mechanism, Concat[·] is the concatenation of feature channel dimensions, and W d The first part is a linear mapping matrix; then, after residual connection and layer normalization, the output feature is used as the query vector of the next layer cross-attention mechanism; when calculating the cross-attention mechanism, the calculation process is similar to that of the self-attention mechanism, but here the key and value vector pairs are provided by the memory pool; after the cross-attention mechanism operation, it goes through two layers of residual connection and layer normalization (with a feedforward neural network in between the two layers) to obtain the output feature of the global-local aggregation submodule.
[0046] The feature fusion submodule based on channel attention will predict the cell state of the deepest convolutional long short-term memory network unit in the branch. Latent state characteristics (The superscript represents the 4th layer convolutional long short-term memory network unit) and the output features of the global-local aggregation submodule are weighted and fused. For example... Figure 4 As shown, firstly, the cell-state features and the output features of the global-local aggregation submodule are concatenated along the channel dimension; then, the concatenated features are input into a max-pooling layer F in two separate paths. max With a single average pooling layer F avg The spatial features are further aggregated, then passed through a shared feedforward neural network layer. The two feature paths are then summed and fed into an activation function to obtain channel attention weights. These channel attention weights are then multiplied by the Hadamard product of the output features θ′ from the global-local aggregation submodule to achieve channel attention weighting. The weighted features θ′ are then... att Features of latent states Channel-dimensional splicing is performed, and the number of channels is adjusted by a 1×1 convolutional layer to obtain the refined hidden state features. Finally, the hidden state features were refined. The output echo image is obtained by iteratively passing through upsampling layers and shallow convolutional long short-term memory network units.
[0047] Training the global-local aggregation model. A two-stage training strategy is used to train the global-local aggregation model. The first training stage is the memory pool update stage, and the input for this stage is a long-term radar echo sequence. The length of the long-term radar echo sequence is less than or equal to the total length of each radar echo sequence sample in the training set, but greater than the input sequence length specified in the extrapolation algorithm's testing phase. Furthermore, during the training phase, the weights of the global information memory pool are continuously updated via backpropagation. This phase can be considered the long-term echo sequence motion feature storage phase. The second training phase is the memory pool fixing phase, where the input is the short-term radar echo sequence S. short =S 1:t This refers to the input sequence length specified during the testing phase of the extrapolation algorithm. During this phase, the weights and gradients in the memory pool are locked and no longer updated. This phase can be viewed as a matching phase between the motion features of the short-term echo sequence and the stored motion features of the long-term echo sequence. After model training is complete, the test set sample data is input into the trained global-local aggregation model to obtain the echo extrapolation predicted image sequence.
[0048] In this embodiment, the proposed global-local aggregation model was tested on the SRAD2018 dataset. The Heidegger Skill Score (HSS) and Critical Success Index (CSI) at a 40 dBZ reflectivity threshold reached 0.5907 and 0.4540, respectively, while the HSS and CSI of the traditional convolutional long short-term memory network were 0.5802 and 0.4423, respectively. It is evident that the radar intelligent echo extrapolation method based on the global-local aggregation model designed in this invention, by introducing optical flow information as motion guidance information and using a global-local aggregation module based on an attention mechanism, enhances the model's interactive perception ability of short-term local and long-term global echo spatiotemporal features, significantly improving the precipitation skill score for heavy rainfall forecasting.
[0049] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a computer storage medium (ROM / RAM, magnetic disk, optical disk) for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, the present invention is not limited to any particular hardware and software combination.
[0050] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A radar intelligent echo extrapolation method based on a global-local aggregation model, characterized in that... The steps of this method include: A. Obtain the meteorological radar precipitation echo image sequence dataset stored in grayscale image format, and obtain the optical flow sequence dataset corresponding to the meteorological radar precipitation echo image sequence based on the obtained meteorological radar precipitation echo image sequence dataset. B. Construct a global-local aggregation model; C. Use the meteorological radar precipitation echo image sequence dataset and meteorological radar precipitation echo optical flow sequence dataset obtained in step A to train the global-local aggregation model constructed in step B, and obtain the trained global-local aggregation model. D. Input the test set sample data into the global-local aggregation model trained in step C to obtain the echo extrapolation prediction image sequence of the test set sample data, and complete the radar intelligent echo extrapolation based on the global-local aggregation model. In step A, the pixel grayscale value in the meteorological radar precipitation echo image sequence dataset stored in grayscale image form represents the true echo reflectivity factor, and the default reflectivity value is set to a pixel value of 255 in the grayscale image. In step B, the constructed global-local aggregation model includes a prediction branch based on an encoding-prediction architecture and a global-local aggregation branch based on an attention mechanism. The attention-based global-local aggregation branch includes an echo motion representation submodule guided by optical flow information, an attention-based global-local aggregation submodule, and a feature fusion submodule based on channel attention. In step B, the encoder-prediction architecture consists of two parts: an encoder and a predictor. The encoder alternately sets up convolutional downsampling layers and convolutional long short-term memory network units, and the predictor alternately sets up convolutional upsampling layers and convolutional long short-term memory network units. Each layer of convolutional long short-term memory network units in the encoder corresponds one-to-one with each layer of convolutional long short-term memory network units in the predictor. In the prediction branch based on the encoder-prediction architecture, the encoder input is the currently observed echo sequence, and the output is the echo sequence to be predicted. In step B, the echo motion characterization submodule guided by optical flow information has an encoder-decoder structure. The encoder input is the currently observed echo sequence, and the output is the optical flow sequence corresponding to the echo sequence. The echo motion characterization submodule guided by optical flow information consists of one residual connection layer, three downsampling residual connection layers, three convolutional layers stacked sequentially with an upsampling layer, a 1×1×1 convolutional layer, and a residual connection layer, and one 1×1×1 convolutional layer. The meteorological radar precipitation echo image first passes through the one residual connection layer, and then... The optical flow sequence corresponding to the input meteorological radar precipitation echo image is obtained by passing through three downsampling residual connection layers, followed by three convolutional layers consisting of an upsampling layer, a 1×1×1 convolutional layer, and a residual connection layer stacked sequentially, and finally through one 1×1×1 convolutional layer to adjust the number of channels. Skip-layer connections are added between the corresponding downsampling and upsampling layers. After the third downsampling layer of the encoder, a motion semantic segmenter is connected. The input of the motion semantic segmenter is the motion semantic features output from the third layer of the encoder, and the output is vectorized abstract motion information. In step B, the input to the global-local aggregation submodule based on the attention mechanism is vectorized abstract motion information. The core components of the global-local aggregation submodule are the self-attention mechanism and the cross-attention mechanism. Formally, it adopts a Transformer decoder stacked structure and is equipped with a global information memory pool. First, the vectorized abstract motion information is linearly mapped to obtain query, key, and value vectors, and then the self-attention mechanism is operated. Subsequently, after residual connection and layer normalization, the output feature is used as the query vector of the next layer cross-attention mechanism. When calculating the cross-attention mechanism, the key and value vector pairs are provided by the memory pool. After the cross-attention mechanism is operated, it is then processed through two layers of residual connection and layer normalization. A feedforward neural network is provided between the two layers to obtain the output feature of the global-local aggregation submodule. In step B, the feature fusion submodule based on channel attention performs a weighted fusion of the cellular state and hidden state features of the deepest convolutional long short-term memory network units in the prediction branch with the output features of the global-local aggregation submodule, specifically as follows: First, the cellular features and the output features of the global-local aggregation submodule are concatenated along the channel dimension. Then, the concatenated features are fed into a max-pooling layer and an average-pooling layer respectively for further spatial feature aggregation, followed by a shared feedforward neural network layer. The two sets of features are then summed and fed into an activation function to obtain channel attention weights. These channel attention weights are then multiplied by the Hadamard product of the output features of the global-local aggregation submodule to achieve channel attention weighting. The weighted features are then concatenated along the channel dimension with the latent features, and the number of channels is adjusted using a 1×1 convolutional layer to obtain refined latent features. Finally, the refined latent features are iteratively passed through upsampling layers and shallow convolutional long short-term memory network units to obtain the output echo image.
2. The radar intelligent echo extrapolation method based on a global-local aggregation model according to claim 1, characterized in that: In step C, the meteorological radar precipitation echo image sequence dataset obtained in step A is normalized and then divided into training set, validation set and test set according to a set ratio. The optical flow sequence dataset corresponding to the meteorological radar precipitation echo image sequence obtained in step A is normalized and then divided into corresponding training set, validation set and test set according to a set ratio.
3. The radar intelligent echo extrapolation method based on a global-local aggregation model according to claim 2, characterized in that: In step C, when training the global-local aggregation model, a two-stage training strategy is set to train the global-local aggregation model. The first training stage is the memory pool update stage. The input of this stage is the long-time radar echo sequence. The length of the long-time radar echo sequence is less than or equal to the total length of each radar echo sequence sample in the training set, but greater than the input sequence length specified in the extrapolation algorithm test stage. In addition, the weight of the global information memory pool is continuously updated during the training stage. This stage can be regarded as the long-time echo sequence motion feature storage stage. The second training phase is the memory pool fixing phase. The input for this phase is the short-time radar echo sequence, which is the input sequence length specified in the extrapolation algorithm testing phase. In this phase, the weights and gradients of the memory pool are locked and no longer updated. This phase can be regarded as the matching phase between the motion features of the short-time echo sequence and the motion features of the stored long-time echo sequence. When training a global-local aggregation model, the memory pool update phase and the memory pool fixation phase are performed alternately, and the model with the best performance on the validation set is saved for testing.
4. An application of the radar intelligent echo extrapolation method based on a global-local aggregation model as described in any one of claims 1-3, characterized in that: The precipitation rate is estimated by extrapolating the reflectance factor in the image sequence based on the echo of the test set sample data and applied to short-term precipitation forecasting.