A short-term radar echo extrapolation method and system

Abstract

The application provides a short-term radar echo extrapolation method and system, and belongs to the technical field of radar monitoring meteorological conditions. The method comprises the following steps: acquiring multi-source data and preprocessing the multi-source data; constructing a neural network with multi-source data input and multi-branch feature extraction structure; training the neural network by using a training method based on an interval optimization strategy; and outputting future radar echo extrapolation data by using the trained neural network according to the preprocessed multi-source data, so as to complete the extrapolation processing of the short-term radar echo. The application solves the problems of long-term prediction result severe blur and lack of end-to-end automatic deployment of the short-term radar echo extrapolation.

Description

Technical Field

[0001] This invention belongs to the field of radar monitoring of meteorological conditions, and in particular relates to a method and system for extrapolating short-term radar echoes. Background Technology

[0002] Short-term forecasting refers to the prediction of weather phenomena on mesoscale and small-scale scales from the current moment until the next 6 hours, and it is an important task of meteorological forecasting. Radar echo data directly reflects cloud patterns and is directly related to meteorological elements such as wind and precipitation, supporting short-term forecasting. Radar echo extrapolation methods utilize time series composed of radar data from multiple previous moments to infer the echo state at a certain time later based on the changing trends, thus forecasting short-term weather conditions. Short-term weather forecasts based on radar echo extrapolation are faster and have significant application value in early warning of short-term severe weather.

[0003] Radar echoes are influenced by complex meteorological and physical processes, and further research is needed to improve the accuracy of echo extrapolation. Traditional numerical extrapolation methods do not fully utilize historical radar echo data and struggle to handle nonlinear motion characteristics. Deep learning is currently a hot topic in extrapolation research; however, it suffers from long-term prediction ambiguity and low prediction indices. This stems from the insufficient ability of existing methods to extract spatiotemporal feature information and the tendency of existing training strategies to overfit. How to further improve the learning ability of spatiotemporal features, refine training strategies, and solve the problem of long-term prediction ambiguity is a research area with both practical value and challenges.

[0004] Currently, radar echo extrapolation methods are divided into traditional methods and learning-based methods. Traditional extrapolation methods calculate the speed and direction of radar echo movement to obtain motion vectors, and then infer the spatial position and echo intensity of the radar at future times. Compared with traditional methods, deep learning methods can learn from a large amount of historical data to obtain nonlinear motion features. Deep learning methods treat radar echo extrapolation as a spatiotemporal sequence prediction problem, using the previous few frames of the radar echo sequence to predict the next few frames. Common solutions to this problem include Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). RNN-based methods are mainly used for modeling time-series data, and their temporal inference is highly interpretable. However, gradients in RNN-based methods are prone to runaway during training, and the error accumulation of image prediction at each time step is significant, resulting in low computational efficiency. CNN-based methods require less GPU memory and have higher computational efficiency. Because they output the temporal extrapolation results end-to-end in one go, they can fully utilize historical data and thus achieve better predictions. However, they suffer from the problem of image details disappearing as the prediction time increases. Generative Adversarial Networks (GANs) improve the resolution of generated radar echo images by using a discriminator to extrapolate images, but they face problems such as training convergence difficulties and model crashes. ViT is a novel attention-based temporal image method that has been widely used in video prediction, but the attention mechanism is computationally intensive. This invention tested some RNN-based and Transformer-based methods and found that these networks require more than 40GB of GPU memory to process a single sample no larger than 480x480 pixels.

[0005] In terms of training strategies, existing deep learning methods generally employ pixel-level loss functions, requiring each pixel's predicted value to perfectly match the true value. This is difficult to achieve and instead leads to overfitting on local datasets. Consequently, the predicted data tends to output an average predicted value in order to minimize the average loss function value, resulting in fuzzy predictions after long-term training. These limitations in training strategies significantly reduce the effectiveness of existing deep learning methods in radar echo prediction. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a short-range radar echo extrapolation method and system, which solves the problems of low accuracy and lack of end-to-end deployment in existing radar echo extrapolation methods.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a short-range radar echo extrapolation method, comprising the following steps:

[0008] Acquire multi-source data and preprocess the multi-source data;

[0009] A neural network with both multi-source data input and multi-branch feature extraction structure is constructed, and the neural network is trained using a training method based on interval optimal strategy.

[0010] Based on the preprocessed multi-source data, the trained neural network outputs future radar echo extrapolation data to complete the extrapolation processing of short-term radar echoes.

[0011] The beneficial effects of this invention are: This invention can input multi-source data, mainly radar data and supplemented by other types of meteorological data; at the same time, based on x86 CPU or GPU hardware, and with a multi-branch spatiotemporal neural network as the core model, it outputs radar echo prediction results for the next 1-6 hours, solving the problem of severe ambiguity and low accuracy in long-term prediction results of short-term radar echo extrapolation.

[0012] Furthermore, the acquisition and preprocessing of multi-source data specifically includes:

[0013] The radar data format is read, which includes binary files and image formats. When reading radar data, each element must be in dBz. If it is echo signal intensity, it needs to be converted to dBz. If terrain data is available, the grid data of the corresponding area in the radar data is read. D ,make d The number of layers in the radar data is used to combine the radar data to obtain fused radar grid data.

[0014] Read satellite data formats, including satellite-specific storage files and image formats;

[0015] Read meteorological data, where each meteorological data point includes longitude, latitude, and time information;

[0016] Time-aligned processing is performed on satellite data, meteorological data, and radar data;

[0017] Spatial alignment processing is performed on satellite data, meteorological data, and radar data;

[0018] A training set is constructed by offline processing of historical offline data that has been aligned spatially and temporally, where satellite data and meteorological data are optional.

[0019] The beneficial effects of the above-mentioned further solutions are as follows: This invention provides a multi-source data preprocessing scheme, including a data preprocessing scheme based on radar as the main source and satellite and other types of meteorological data such as temperature, humidity, pressure and topography as the auxiliary sources. It can read meteorological data formats such as grb and nc or general image formats such as tif (grid data can be stored as images), and perform dbz data conversion, temporal resolution interpolation, spatial data alignment, multi-height layer combination, signal filtering, etc., to form a spatiotemporally registered multi-source dataset.

[0020] Furthermore, the expression for the fused radar grid data is as follows:

[0021] ;

[0022] in, This represents the fused radar grid data. This represents the normalization constant after weighted summation of grid points within the neighborhood. i and j These represent the discretized latitude and longitude coordinates of the grid points within the neighborhood, respectively. , , and Both represent the neighborhood range. x and y These represent the planar spatial coordinates of the current grid point after discretization. Indicates the height layer of the radar data. This indicates the total number of height layers in the radar data. Representation and height layer d The relevant weighting coefficients, Represents pixel value, This represents the weight set based on the current pixel value range, used to filter out outliers. This indicates the weighting of the geographical area.

[0023] The beneficial effect of the above-mentioned further scheme is that different altitude layers have different impacts on the region's future weather conditions, such as precipitation and snowfall. By weighting and combining different altitude layers according to the local geographical, topographical and climatic characteristics, radar signals that have a significant impact on future weather conditions can be highlighted.

[0024] Furthermore, the time alignment process specifically includes:

[0025] Let the time of the current radar data be... t The signal values ​​of satellite data and meteorological data are I ; and will be with t The time series of the two most recent meteorological data is set to... t 1 and t 2;

[0026] like Using the following formula, in time t Interpolation is performed at the point:

[0027]

[0028] in, Indicates time t The discrete coordinates of latitude and longitude are i, j The height layer discretized coordinates are d signal strength, Indicates time t Signal strength at point 1 Indicates time t Signal strength at point 2;

[0029] For those unable to obtain t 2, then use directly. t 1 or according to the t 2. Make predictions and complete time alignment processing.

[0030] The beneficial effect of the above-mentioned further scheme is that there are differences in the sampling time points of radar data and other meteorological elements such as satellites and observation stations (temperature, humidity, wind pressure). Aligning the sampling time points of all elements with the sampling time points of radar data can reduce the interference caused by time inconsistency.

[0031] Furthermore, the spatial alignment process specifically involves:

[0032] Based on the resolution and size of the radar grid data, the actual coverage area of ​​the radar grid data is obtained;

[0033] Based on the resolution of satellite data and meteorological data, the offset and scale of satellite grid data and meteorological data within the actual coverage area of ​​radar grid data are determined respectively.

[0034] Based on the actual coverage area, offset, and scale of the radar grid data, cropping is performed; if the resolution of the satellite grid data and the resolution of the radar grid data are not multiples of 2 or fractions of each other, the satellite grid data is sampled to make the resolution of the satellite grid data a multiple of 2 or fractions of the resolution of the radar grid data.

[0035] The beneficial effects of the above-mentioned further scheme are: maintaining spatial consistency between radar data and gridded data from other meteorological elements such as satellites and observation stations (temperature, humidity, wind pressure), and adjusting their resolutions to [the required values]. The multiple (which can be 2, 4, 8 times, or 1 / 2, 1 / 4, 1 / 8 times, etc.) makes it convenient for the network model to obtain feature data of known size at low cost after extracting and sampling features from different data.

[0036] Furthermore, the neural network includes:

[0037] The backbone encoder is used to extract primary features of radar spatiotemporal sequences.

[0038] A dual-branch intermediate network is used to learn the spatiotemporal evolution of primary features of radar spatiotemporal sequences and extract spatiotemporal features.

[0039] The multi-data processing branch is used to extract initial features from satellite or meteorological grid data using the branch encoder to obtain primary features of satellite / meteorological data, and to obtain deep spatiotemporal latent features based on the primary features of satellite / meteorological data. The output size of the deep spatiotemporal latent features is adjusted by sampling, and the adjusted deep spatiotemporal latent features are merged with the spatiotemporal features extracted by the dual-branch intermediate network. The multi-data processing branch includes a branch encoder and a branch spatiotemporal feature extraction network.

[0040] The backbone decoder is used to reconstruct and generate prediction sequences based on the spatiotemporal latent features of merged radar, satellite, and meteorological grid data.

[0041] Furthermore, the dual-branch intermediate network includes a backbone spatiotemporal feature branch and an attention enhancement branch;

[0042] The main spatiotemporal feature branch is used to learn features based on the primary features of the radar spatiotemporal sequence output by the main encoder, using the spatiotemporal latent feature space to learn the spatiotemporal change trend information of the primary features of the radar spatiotemporal sequence in the latent feature space, and to extract spatiotemporal features.

[0043] The attention enhancement branch is used to extract and learn detailed information about the primary feature data of radar spatiotemporal sequence in spatiotemporal changes, wherein a meta-transformer framework is used for feature learning.

[0044] The beneficial effect of the above-mentioned further scheme is that the neural network model enhances its spatiotemporal feature learning ability by fusing the spatiotemporal feature extraction branch and the attention learning branch. The main spatiotemporal feature branch adopts the UNet framework, using Inception network layers or 2D / 3D convolutional operators as the base network layers to extract multi-scale spatiotemporal information as the basis for prediction. Both the UNet framework and the Inception network layer have been proven effective in extracting overall and local detail information of images. The attention enhancement branch adopts the Metaformer framework, similar to Transformer, and includes an attention extraction module and a convolution-based feedforward network (FFN). This structure is used to improve the accuracy of the predicted location and fuses the information from the spatiotemporal feature network to obtain high-quality intermediate spatiotemporal features, thereby obtaining better output results.

[0045] Furthermore, the expression for the loss function of the neural network is as follows:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] Among them, threshold τ have K Each value corresponds to P Threshold of the function τ k , i , j and d This represents the discretized X, Y, and Z axis coordinates on the latitude and longitude plane of the prediction range. t Indicates time, The loss function of a neural network. This represents the mean squared error loss function. k Indicating signal strengthk Each level corresponds to a threshold. τ k , K This indicates the total number of signal strength levels. Indicates different thresholds τ Weighted loss function This represents the intensity threshold used to assess prediction accuracy. This represents a multi-threshold loss function. T Indicates the length of the predicted time series. D The height layer discretized coordinates represent the prediction range. This represents the discretized Y-axis length on the prediction range plane. This represents the length of the discretized X-axis on the prediction range plane. Indicates the weighting coefficient. Indicates the predicted value. Indicates the gold standard value. This indicates that the predicted value is greater than the threshold. τ k The probability, β Indicates hyperparameters, This indicates that the gold standard value is greater than the threshold. τ k The probability is 1. This indicates that when the threshold is set to τ k The critical accuracy of the prediction area. This indicates that both the prediction region and the gold standard region are greater than the threshold. τ k The overlap probability, Indicates control variables, β and eps Used to control the range of the region with the lowest loss function value. This indicates that the predicted region is smaller than the threshold. τ k However, the corresponding gold standard area is greater than the threshold. τ k The overlap probability, This indicates that the predicted region is larger than the threshold. τ k However, the corresponding gold standard area is less than the threshold. τ k The probability of overlap.

[0056] The beneficial effects of the above-mentioned further solution are as follows: The new loss function constructed in this invention consists of a multi-threshold loss function and a mean squared error (MSE) loss function. The multi-threshold loss function is the core, directly optimizing the Critical Success Index (CSI) metric. The core idea is to optimize the network's learning process, i.e., how the loss function decreases, from a process of continuously approaching the minimum point to a process of approaching the minimum interval (e.g., ...). Figure 3 As shown in the figure, this improves the overfitting problem in network training. This training method is also effective on other existing deep learning networks.

[0057] Furthermore, the extrapolation of short-range radar echoes is completed by using the trained neural network to output future radar echo extrapolation data based on the preprocessed multi-source data. Specifically, this process involves:

[0058] The system is initialized by loading the trained neural network and acquiring multi-source input data. Among the multi-source data, radar data is required, while satellite data and meteorological data are optional.

[0059] The multi-source data is preprocessed to obtain aligned multi-source data, and then the aligned multi-source data is input into the neural network for inference to obtain future radar echo extrapolation data.

[0060] The future radar echo extrapolation data will be saved in a gridded data storage format;

[0061] Based on the saved radar echo extrapolation data, a base map of latitude and longitude range is obtained;

[0062] The latest predicted data is retrieved based on the timestamp, and the plotting library is called in conjunction with the retrieved base map to draw graphics, thus completing the extrapolation processing of short-term radar echoes.

[0063] The beneficial effects of the above-mentioned further solutions are as follows: This invention provides a multi-branch deep learning method and software for radar echo extrapolation. It can be used with a personal computer or workstation with CPU or GPU hardware. It can acquire radar data as the main input, integrate other meteorological grid data such as satellite input, and use an innovatively constructed multi-branch spatiotemporal prediction network to output radar echo prediction results for the next 1-6 hours. The results are in grid data format such as grb or tif, and can be visualized using Python software such as cartopy.

[0064] The present invention also provides a short-range radar echo extrapolation system, comprising:

[0065] The first processing module is used to acquire multi-source data and preprocess the multi-source data;

[0066] The second processing module is used to construct a neural network that combines multi-source data input and multi-branch feature extraction structure, and to train the neural network using a training method based on interval optimal strategy.

[0067] The third processing module is used to output future radar echo extrapolation data based on the preprocessed multi-source data and the trained neural network, thereby completing the extrapolation processing of short-term radar echoes.

[0068] The beneficial effects of this invention are:

[0069] This invention constructs a multi-branch spatiotemporal fusion deep learning framework to achieve faster speed, lower resource consumption, and higher spatiotemporal prediction accuracy. The invention extracts features by combining spatial, temporal, and channel dimensions in different ways. The main spatiotemporal feature branch adopts the UNet framework, using Inception network layers or 2D / 3D convolutional operators as the base network layers to construct the spatiotemporal feature extraction branch. Encoding branches from other sources, such as satellite data, are constructed and integrated into the spatiotemporal feature extraction branch of radar data as supplementary prior information. An attention enhancement branch is designed to enhance prediction details. This invention constructs a multi-branch CNN-based network framework to enhance spatiotemporal features, effectively improving radar echo prediction accuracy while maintaining faster speed and lower resource consumption than RNN-based and Transformer-based methods. For example, on a 480×480 km area dataset with a 1km spatial resolution, a 2-hour short-term extrapolation requires only about half the GPU memory of a Transformer-based method with the same number of network layers, and the execution speed on an Nvidia RTX 4090 graphics card is only 1-2 seconds.

[0070] This invention effectively suppresses the long-term prediction ambiguity problem caused by overfitting by constructing a novel training strategy. The invention innovatively designs a multi-threshold loss function and a weighted mean square error loss function for network training, allowing predicted pixel values ​​to approach optimality as long as they fall within the gold standard range, rather than forcibly requiring every pixel value to be completely identical. This effectively suppresses and reduces the long-term prediction ambiguity problem caused by overfitting. Attached Figure Description

[0071] Figure 1 This is a flowchart of the method of the present invention.

[0072] Figure 2 This is a schematic diagram of the neural network structure of the present invention.

[0073] Figure 3 This is a comparison chart of mean loss function and multi-threshold loss function.

[0074] Figure 4This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0076] Example 1

[0077] like Figure 1 As shown, this invention provides a method for extrapolating short-range radar echoes, the implementation of which is as follows:

[0078] S1. Acquire multi-source data and preprocess the multi-source data, specifically as follows:

[0079] The radar data format is read, which includes binary files and image formats. When reading radar data, each element must be in dBz. If it is echo signal intensity, it needs to be converted to dBz. If terrain data is available, the grid data of the corresponding area in the radar data is read. D ,make d To determine the number of layers in the radar data, the radar data is combined to obtain fused radar grid data; satellite data formats are read, including satellite-specific storage files and image formats; meteorological data is read, with each meteorological data point including longitude, latitude, and time information; the satellite data, meteorological data, and radar data are time-aligned; the satellite data, meteorological data, and radar data are spatially aligned; and offline processing is performed using spatially and temporally aligned historical offline data to construct a training set, where satellite data and meteorological data are optional. The spatial alignment process specifically involves:

[0080] Based on the resolution and size of the radar grid data, the actual coverage area of ​​the radar grid data is obtained; according to the resolution of the satellite data and meteorological data, the offset and scale of the satellite grid data and meteorological data within the actual coverage area of ​​the radar grid data are determined respectively; based on the actual coverage area of ​​the radar grid data and the offset and scale, cropping is performed; if the resolution of the satellite grid data and the resolution of the radar grid data are not multiples of 2 or fractions, the satellite grid data is sampled to make the resolution of the satellite grid data a multiple of 2 or fractions of the resolution of the radar grid data.

[0081] In this embodiment, radar data is read in a format typically with the suffix ".bin" or ".tif," where ".bin" indicates a binary file and ".tif" is an image format. The radar dataset includes time, echo grid data, latitude and longitude, and layer height, usually located at the beginning of the ".bin" file. Reading the data requires each element to be in dBz. If it is echo signal strength, it needs to be converted to dBz. A reference conversion formula is as follows:

[0082] ;

[0083] in, This represents the radar reflectivity factor value (dBZ) of the current image. This indicates the strength of the original radar echo signal.

[0084] Multi-layer fusion of radar data: If terrain data is available, grid data of the corresponding area from the radar data is read. d To determine the number of layers in the radar data, the radar data is combined as follows:

[0085] ;

[0086] in, This represents the fused radar grid data. This represents the normalization constant after weighted summation of grid points within the neighborhood. i and j These represent the discretized latitude and longitude coordinates of the grid points within the neighborhood, respectively. , , and Both represent the neighborhood range. x and y These represent the planar spatial coordinates of the current grid point after discretization. Indicates the height layer of the radar data. This indicates the total number of height layers in the radar data. Representation and height layer d The relevant weighting coefficients, Represents pixel value, This represents the weight set based on the current pixel value range, used to filter out outliers. This indicates the weighting of the geographical area. and It is used to process different terrains and landforms such as plateaus, basins, and plains. , , and When the value is always 1, the above results are the average of the radar data; if for the same Different height levels of the location d , Only If the maximum value is 1, and the value is 0 otherwise, the result is the projection of the radar data along the maximum value of the height layer.

[0087] The system can read satellite data in various formats, typically with the extensions .awx or .tif. .awx is a satellite-specific storage file, while .tif is an image format. One method is to use the Python meteva library to read .awx files. For data such as cloud top brightness and temperature, phase inversion operations are required, such as calculating the maximum and minimum values ​​of the entire dataset. and For each pixel value Update value for .

[0088] Optional meteorological data such as temperature, humidity, and pressure can be read, typically in formats with suffixes like grb, grb2, or nc. Taking the grb series as an example, the ecCodes package in Python is used to read GRIB data files. ecCodes accesses the content of GRIB messages using a key / value pair and supports reading data from GRIB versions 1 and 2. Each meteorological dataset includes longitude, latitude, and time information. The `eccodes.codes_grib_new_from_file` function reads a message from the GRIB file, and the `eccodes.codes_get` function retrieves the message name, typically shortName or name. The name is used to determine if the data is intended for reading.

[0089] Time alignment of satellite or other data. This data needs to be time-aligned with radar data. Let the time of the current radar data be... t , I The signal values ​​representing satellite data and meteorological data, and t The time series of the two most recent other meteorological data is as follows: t 1 and t 2, of which Then in time t Interpolation is performed at:

[0090]

[0091] in, Indicates time t The discrete coordinates of latitude and longitude are i, j The height layer discretized coordinates are d signal strength, Indicates time t Signal strength at point 1 Indicates timet Signal strength at point 2.

[0092] During the deployment of inference, it is assumed that the following information cannot be obtained. t At time 2, it can be used directly. t Estimation based on time-1 data or the average trend of meteorological data from previous times. t Data at time 2.

[0093] Spatial dimension alignment of satellite or other data. Based on the actual latitude and longitude range of the radar data, data with the same latitude and longitude range are cropped from satellite and meteorological data. Assume the radar data resolution is... The resolution of the meteorological data is The resolution ratio is adjusted to a multiple of 2 or the same resolution. For example, if the radar's resolution in each dimension is twice that of the meteorological data, then methods such as nearest neighbor or bilinear interpolation are used to directly perform a 2x resampling. If the radar's resolution in each dimension is 3.6 times that of the meteorological data, then the meteorological data is sampled to make the radar resolution 2 or 4 times that of the meteorological data. Specifically:

[0094] First, based on the resolution and size of the radar grid data, calculate the actual coverage area of ​​the radar grid data. Then, based on the resolution of the satellite grid data, calculate the offset and size of the satellite grid data within the corresponding coverage area, and perform cropping. If the satellite grid resolution and the radar grid resolution are not... If the relationship is multiples, image sampling techniques can be used to resample the satellite grid data, such as nearest neighbor and bilinear sampling methods.

[0095] In this embodiment, the training set is constructed offline by retrieving spatially and temporally aligned historical offline data. In actual deployment, the input data is acquired in real time, and the above preprocessing steps are performed. Radar data is required; other data are optional.

[0096] S2. Construct a neural network that combines multi-source data input and multi-branch feature extraction structures, and train the neural network using a training method based on interval optimality strategy; the neural network includes:

[0097] The backbone encoder is used to extract primary features of radar spatiotemporal sequences.

[0098] The dual-branch intermediate network is used to learn the spatiotemporal evolution of primary features of radar spatiotemporal sequences and extract spatiotemporal features; the dual-branch intermediate network includes a backbone spatiotemporal feature branch and an attention enhancement branch;

[0099] The main spatiotemporal feature branch is used to learn the primary features of the radar spatiotemporal sequence based on the output of the main encoder. It utilizes the spatiotemporal latent feature space to learn the spatiotemporal variation trend information of the primary features of the radar spatiotemporal sequence in the latent feature space, and uses a U-shaped network (UNet) based on Inception network layers or stacked 2D / 3D convolution operators for feature extraction.

[0100] The attention enhancement branch is used to extract and learn detailed information about the spatiotemporal variations of primary feature data of radar spatiotemporal sequences. The feature learning is performed using a metaformer framework.

[0101] The multi-data processing branch is used to extract initial features from satellite or meteorological grid data using the branch encoder to obtain primary features of satellite / meteorological data, and to obtain deep spatiotemporal latent features based on the primary features of satellite / meteorological data. The output size of the deep spatiotemporal latent features is adjusted by sampling, and the adjusted deep spatiotemporal latent features are merged with the spatiotemporal features extracted by the dual-branch intermediate network. The multi-data processing branch includes a branch encoder and a branch spatiotemporal feature extraction network.

[0102] The backbone decoder is used to reconstruct and generate prediction sequences based on the spatiotemporal latent features of merged radar, satellite, and meteorological grid data.

[0103] In this embodiment, the multiple data processing branches are equivalent to other input data, each configured with its own independent encoder network and spatiotemporal feature extraction network, operating at the same level as the backbone encoder network and the dual-branch intermediate network. This avoids interference with the processing of the main data, i.e., radar features. The input data for the multiple data processing branches is matrix (grid) data, and the corresponding branch encoders and branch spatiotemporal feature extraction networks have essentially the same structure as the backbone network, only with different parameters. Finally, they are merged into the decoder to assist in correcting the results of the radar backbone network.

[0104] In this embodiment, as Figure 2 As shown, the main body of the neural network includes a backbone encoder, a dual-branch intermediate network, multiple data processing branches, and a backbone decoder.

[0105] The backbone encoder extracts primary features from the radar spatiotemporal sequence and reduces the number of parameters, providing them to the dual-branch intermediate network for spatiotemporal evolution learning to extract spatiotemporal features. The backbone decoder is responsible for restoring the spatiotemporal latent features of the merged radar, satellite, and meteorological grid data to generate the prediction sequence. This model uses radar echo data from the previous N (1-6) hours as the primary input, with optional inputs such as meteorological satellite data and temperature and humidity data, to predict the radar echo sequence for the next M (1-6) hours. The encoder and decoder are CNN network layers. The main network components of the encoder and decoder are 2D or 3D convolutional layers, consisting of 2D / 3D convolutional layers, regularization layers (such as Group Normalization), and activation layers (such as Sigmoid Linear Units, SiLU; Leaky ReLU, etc.). In the backbone encoder, feature downsampling is achieved by setting the stride parameter (the step size the convolutional kernel moves on the input feature map) of the convolutional layers or by using pooling layers. In the backbone decoder, upsampled features can be obtained through methods such as sub-pixel shuffle or transposed convolution. The last layer of the backbone decoder converts the data size to B×T×C×H×W before outputting the prediction result. Throughout the framework, the image is downsampled and upsampled N times. Considering that gridded data such as radar, meteorological, and satellite data contain four dimensions (latitude, longitude, altitude, and time), three-dimensional convolution operators can be used for the convolutional layers. If the altitude dimension is considered as the channel dimension, the time and altitude dimensions can be merged, and two-dimensional convolution operators can be used for the convolutional layers. If 2D convolution is used, the input data needs to be transformed through dimensionality transformation. or After determining the dimensions, input the encoder. Among them, B Indicates batch, T Indicates the time dimension. C Indicates channel dimension, H and W This represents the length and width dimensions of a plane.

[0106] In this embodiment, the backbone encoder is used to extract primary features of the radar spatiotemporal sequence. Input data is first fed into the backbone encoder, and the results are fed into a dual-branch intermediate network composed of the backbone spatiotemporal feature branch and the attention enhancement branch. The backbone encoder mainly consists of stacked convolutional layers, with kernel sizes typically set to 3×3 or 5×5 to acquire feature information from multiple channels. In the backbone encoder, the stride parameter in the convolution operator can be set to downsample the length and width of the feature map; if downsampling is not performed, the stride parameter is set to 1. The number of convolutional layers and whether each layer performs downsampling are manually set based on metadata such as the size and resolution of the input data.

[0107] like Figure 2 As shown, similar to the backbone encoder, the backbone decoder is mainly composed of stacked convolutional layers. Upsampling layers are manually selected to be inserted after a convolutional layer, but the number of upsampling operations must be consistent with the number of downsampling operations in the encoder. In the backbone decoder, upsampling layers can obtain upsampled features in both width and height through sub-pixel shuffle networks; upsampling can also be achieved through convolutions with bi / trilinear sampling or transposed convolutions. By performing convolutional layer operations and upsampling operations symmetrical to the downsampling operations of the backbone encoder, the last layer of the backbone decoder converts the output data size to the same size as the gold standard sequence, obtaining the final output result. The backbone encoder and backbone decoder are generally symmetrical fully convolutional networks, with the number of convolutional layers and the number of upsampling operations in each layer consistent with the encoder.

[0108] In this embodiment, the dual-branch intermediate network includes a backbone spatiotemporal feature branch and an attention enhancement branch. If a 2D convolution operator is used, the feature data needs to be transformed through dimensionality transformation. or After determining the dimensions, input the intermediate network, where, The number of new channels output by the encoder.

[0109] The main spatiotemporal feature branch is used to learn features in a deeper spatiotemporal latent feature space from the primary features of the radar spatiotemporal sequence output by the main encoder. Multi-scale grouped convolutional layers are used for feature extraction to learn the spatiotemporal variation trend information of the input data in the latent feature space. The main spatiotemporal feature branch adopts a U-shaped network (UNet) structure. Each network layer uses an initial module framework (Inception) or a 2D / 3D convolutional operator design. Each network layer can be upsampled or downsampled, and the number of samplings can be selected by the user. The number of upsampling and downsampling is consistent to maintain the length and width dimensions of the output features consistent with the input data; alternatively, no upsampling or downsampling can be selected. For the initial module framework (Inception), 1×1 small convolutions are first used for mapping, followed by grouped convolutional layers with kernel sizes of 3×3, 5×5, and 7×7 for feature extraction. Group normalization and layer normalization are used for normalization, and LeakyReLU and SiLU activation layers are then applied. The features of the first half of the main spatiotemporal feature branch are concatenated into the output features of the second half of the main spatiotemporal feature branch, and then concatenated and processed.

[0110] The attention enhancement branch is used to extract and learn more detailed information about the spatiotemporal variations of the radar spatiotemporal sequence data from the primary features output by the backbone encoder. It employs a metaformer framework for feature learning, enhancing the accuracy of the high-frequency components in the predicted image sequence. The output feature size of the attention enhancement branch is consistent with that of the backbone spatiotemporal feature branch. The attention branch mainly consists of stacked attention modules and feedforward network (FFN) modules, such as... Figure 2 As shown, the attention module employs a bypass structure, sequentially passing the input features through convolution operators, depthwise splitting convolutions, or dilated convolutions to obtain the features. Then it is split into the first feature. Second feature Bypass output for:

[0111] ;

[0112] in, This represents the Sigmoid function. This represents element-wise matrix multiplication. The FFN module uses either a ResNet or a fully convolutional network framework. First, it performs a 1×1 mini-convolution on the input features In to expand the number of features by a factor of N. Then, it processes the features using convolution operators or grouping / separating convolution operators. The kernel size, expansion factor N, and number of grouping / separating convolutions are set by the user. Finally, a 1×1 mini-convolution is performed to restore the number of features to the same number as the input features. If a residual network is used, the input features are merged into the output features.

[0113] In this embodiment, a multi-data processing branch is used to extract initial features from satellite or other meteorological grid data using a branch encoder, obtaining primary features of the satellite / meteorological data. These primary features are then input into a satellite / meteorological feature extraction branch (i.e., a branch spatiotemporal feature extraction network) to obtain deep spatiotemporal latent feature data. Finally, the latent features are sampled and adjusted to match the output feature size of the main spatiotemporal feature branch, and then merged into the radar latent features obtained by the dual-branch intermediate network. In the multi-data processing branch, the branch encoder uses convolutional layers identical to the main encoder. The number of layers and downsampling times are manually set based on metadata such as the size and resolution of the input data. After extracting the primary satellite / meteorological features, the branch encoder inputs them into a branch spatiotemporal feature extraction network with the same structure as the main spatiotemporal feature branch to obtain spatiotemporal trend information of the satellite / meteorological data in the deep feature space. This branch uses the same UNet framework, with each layer using the Inception module. The number of UNet framework layers, the number of convolutional kernels in the grouped convolutional layers, the number of groups, and the upsampling and downsampling times are manually set based on metadata such as the size and resolution of the input data. Finally, the multiple data processing branches sample the latent feature sizes to match the output feature sizes of the main spatiotemporal feature branches. For example, the latitude and longitude spatial resolution of satellite data is four times coarser than that of radar. Therefore, the latitude and longitude dimensions of satellite images with the same coverage area are also one-quarter of those of radar. After processing by the satellite feature extraction branch, the deep spatiotemporal latent features need to be upsampled to the same size as the radar latent features obtained by the dual-branch intermediate network, and finally merged into the radar latent features obtained by the dual-branch intermediate network. Considering that radar and satellite / weather signals reflect different physical elements, a channel splicing method is generally used for weak coupling.

[0114] In this embodiment, a weighted mean square error loss (MSE loss) and a multi-threshold loss function are combined to train the neural network. The weighted mean square error loss function sets different echo intensity thresholds and corresponding weights to increase the model's attention to strong signals. The weighted mean square error loss function is set as follows:

[0115] ;

[0116] In radar echo signals, weak signals constitute a larger proportion while strong signals constitute a smaller proportion. Strong signals are more strongly associated with severe weather; therefore, the weight of high echo intensity is increased to ensure a higher hit rate. In this experiment, regions with higher signal intensity are generally assigned higher weights. For specific datasets, these weights may need to be adjusted accordingly. The overall training loss function is:

[0117] ;

[0118] In this invention, hyperparameters αk All were set to 1. k This indicates the number of threshold divisions. These hyperparameters can be obtained by evaluating the prediction performance after multiple experiments.

[0119] In this embodiment, the present invention designs a novel multi-threshold loss function, which is combined with the traditional weighted mean square error loss function for training. The multi-threshold loss function is inspired by the Critical Success Index (CSI) function. First, the predicted value of each pixel is... Gold Standard Value In this context, the "gold standard" refers to selecting continuous sequence data from a historical offline dataset as training data. The first two hours of the sequence are used as input data, and the last two hours are used as the gold standard. For example, a 4-hour continuous sequence from a specific moment in a real historical dataset could be selected, with the first two hours as input and the last two hours as the gold standard. The input data is fed into the network model, which outputs a predicted sequence. The predicted sequence and the gold standard have the same dimensions. Loss functions for the predicted sequence and the gold standard can be calculated to optimize the network parameters. Figure 3 As shown, Figure 3 This is a comparison chart of mean loss function and multi-threshold loss function. (A & B) correspond to gold standard values ​​of 20 and 30 dBz, respectively. The multi-threshold loss function... The values ​​are 10, 20, 30, and 40. The minimum of the mean loss function is a single point, while the minimum of the multi-threshold loss function is an interval.

[0120] Gold Standard Value Transform into probabilities:

[0121] ;

[0122] Output predicted value Transform into probabilities:

[0123] ;

[0124] The threshold to be given indicates when both the predicted value and the gold standard are greater than 1. A successful prediction is considered complete when the prediction is successful. The formula for calculating the CSI success rate is then given:

[0125] ;

[0126] in eps It is usually a decimal, generally set to :

[0127] ;

[0128] ;

[0129] ;

[0130] eps and β This collectively controls the range of the region with the lowest loss function value. Typically, the CSI metric requires a combination of multiple thresholds to test performance. Similarly, the Multi-Threshold Loss (MTL) function is defined as follows:

[0131] ;

[0132] The expression for the loss function of a neural network is as follows:

[0133] ;

[0134] Among them, threshold τ have K Each value corresponds to P Threshold of the function τ k , i , j and d This represents the discretized X, Y, and Z (elevation) axis coordinates on the latitude and longitude plane of the prediction range. t Indicates time, The loss function of a neural network. This represents the mean squared error loss function. k Indicating signal strength k Each level corresponds to a threshold. τ k , K This indicates the total number of signal strength levels. Indicates different thresholds τ Weighted loss function This represents the intensity threshold used to assess prediction accuracy. This represents a multi-threshold loss function. T Indicates the length of the predicted time series. D The height layer discretized coordinates represent the prediction range. This represents the discretized Y-axis length on the prediction range plane. This represents the length of the discretized X-axis on the prediction range plane. Indicates the weighting coefficient. Indicates the predicted value. Indicates the gold standard value. This indicates that the predicted value is greater than the threshold. τ k The probability, β Indicates hyperparameters, This indicates that the gold standard value is greater than the threshold. τ kThe probability, This indicates that when the threshold is set to τ k The critical accuracy of the prediction area. This indicates that both the prediction region and the gold standard region are greater than the threshold. τ k The overlap probability, Indicates control variables, β and eps Used to control the range of the region with the lowest loss function value. This indicates that the predicted region is smaller than the threshold. τ k However, the corresponding gold standard area is greater than the threshold. τ k The overlap probability, This indicates that the predicted region is larger than the threshold. τ k However, the corresponding gold standard area is less than the threshold. τ k The probability of overlap.

[0135] This invention employs a cosine training strategy to train the neural network, and optimizers such as Adam and AdamW can be used for network training. A recommended cosine annealing learning rate iteration method is as follows: (Any number of rounds) Corresponding learning rate The expression is as follows:

[0136] ;

[0137] in, This represents the maximum learning rate. This represents the minimum learning rate. That is, the learning rate from Down to The required number of rounds.

[0138] S3. Based on the preprocessed multi-source data, the trained neural network outputs future radar echo extrapolation data to complete the extrapolation processing of short-term radar echoes. Specifically:

[0139] The system is initialized by loading the trained neural network and acquiring multi-source input data. Among the multi-source data, radar data is required, while satellite data and meteorological data are optional.

[0140] The multi-source data is preprocessed to obtain aligned multi-source data, and then the aligned multi-source data is input into the neural network for inference to obtain future radar echo extrapolation data.

[0141] The future radar echo extrapolation data will be saved in a gridded data storage format;

[0142] Based on the saved radar echo extrapolation data, a base map of latitude and longitude range is obtained;

[0143] The latest predicted data is retrieved based on the timestamp, and the plotting library is called in conjunction with the retrieved base map to draw graphics, thus completing the extrapolation processing of short-term radar echoes.

[0144] In this embodiment, for the overall operation flow, the system first initializes by loading the trained neural network into memory or the GPU. Then, through a timed function or asynchronous framework, it automatically acquires newly input radar data or other meteorological data, and checks file integrity and parameter availability. Radar data is required, while other data is optional. The system reads the data and performs data preprocessing steps to obtain spatiotemporally aligned radar and other meteorological data. This data is then fed into the neural network for inference, producing extrapolated future radar echo data, which is saved as gridded data storage formats such as nc, npy, or tif. Finally, it calls plotting libraries such as cartopy or matplotlib to obtain a base map of the corresponding latitude and longitude range, and draws pseudo-color maps, contour maps, etc., on the base map as required.

[0145] The data preprocessing process involves first reading and parsing radar data to determine latitude and longitude ranges, resolution, and other information. Then, other meteorological data is acquired and spatiotemporally aligned. Based on requirements, the radar data at different altitude levels is combined to obtain complete input time-series data.

[0146] The automated deployment process is as follows: First, the pre-trained neural network is initialized and loaded into memory or the GPU. Then, using a timed function or asynchronous framework, it automatically checks for new data, starts data preprocessing, feeds the input data into the model for computation, writes to the log, and displays a notification in the command line or pop-up window indicating that the task is complete.

[0147] The post-output processing procedure is as follows: check whether a new work has been successfully completed, then read the latitude and longitude range, time and other information left by the data, find the latest output prediction data according to the timestamp, and then call plotting libraries such as cartopy or matplotlib, and draw pseudo-color maps, outline maps and other plots according to the requirements of the corresponding range map base map.

[0148] Example 2

[0149] like Figure 4 As shown, the present invention provides a short-range radar echo extrapolation system, which is used to execute the short-range radar echo extrapolation method described in Embodiment 1, including:

[0150] The first processing module is used to acquire multi-source data and preprocess the multi-source data;

[0151] The second processing module is used to construct a neural network that combines multi-source data input and multi-branch feature extraction structure, and to train the neural network using a training method based on interval optimal strategy.

[0152] The third processing module is used to output future radar echo extrapolation data based on the preprocessed multi-source data and the trained neural network, thereby completing the extrapolation processing of short-term radar echoes.

[0153] like Figure 4 The short-range radar echo extrapolation system provided in the embodiment shown can execute the technical solution shown in the short-range radar echo extrapolation method of the above-described method embodiment. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0154] In this embodiment, the functional units can be divided according to the short-range radar echo extrapolation method. For example, each function can be divided into its own functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this invention is illustrative and represents only a logical division; in actual implementation, other division methods may be used.

[0155] In this embodiment, the short-range radar echo extrapolation system, in order to realize the principle and beneficial effects of the short-range radar echo extrapolation method, includes hardware structures and / or software modules corresponding to the execution of various functions. Those skilled in the art should readily recognize that, in conjunction with the illustrative units and algorithm steps described in the embodiments disclosed herein, this invention can be implemented in hardware and / or a combination of hardware and computer software. Whether a function is executed by hardware or computer software depends on the specific application and design constraints of the technical solution. Different methods can be used to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for extrapolating short-range radar echoes, characterized in that, Includes the following steps: Acquire multi-source data and preprocess the multi-source data; A neural network with both multi-source data input and multi-branch feature extraction structure is constructed, and the neural network is trained using a training method based on interval optimal strategy. The neural network includes: The backbone encoder is used to extract primary features of radar spatiotemporal sequences. A dual-branch intermediate network is used to learn the spatiotemporal evolution of primary features of radar spatiotemporal sequences and extract spatiotemporal features. The multi-data processing branch is used to extract initial features from satellite or meteorological grid data using the branch encoder to obtain primary features of satellite / meteorological data, and to obtain deep spatiotemporal latent features based on the primary features of satellite / meteorological data. The output size of the deep spatiotemporal latent features is adjusted by sampling, and the adjusted deep spatiotemporal latent features are merged with the spatiotemporal features extracted by the dual-branch intermediate network. The multi-data processing branch includes a branch encoder and a branch spatiotemporal feature extraction network. The backbone decoder is used to reconstruct and generate prediction sequences based on the spatiotemporal latent features of merged radar, satellite, and meteorological grid data. The dual-branch intermediate network includes a backbone spatiotemporal feature branch and an attention enhancement branch; The main spatiotemporal feature branch is used to learn features based on the primary features of the radar spatiotemporal sequence output by the main encoder, using the spatiotemporal latent feature space to learn the spatiotemporal change trend information of the primary features of the radar spatiotemporal sequence in the latent feature space, and to extract spatiotemporal features. The attention enhancement branch is used to extract and learn detailed information of radar spatiotemporal sequence primary feature data in spatiotemporal changes, wherein a meta-transformer framework is used for feature learning. Based on the preprocessed multi-source data, the trained neural network outputs future radar echo extrapolation data to complete the extrapolation processing of short-term radar echoes.

2. The short-range radar echo extrapolation method according to claim 1, characterized in that, The acquisition and preprocessing of multi-source data specifically involves: The radar data format is read, which includes binary files and image formats. When reading radar data, each element must be in dBz. If it is echo signal intensity, it needs to be converted to dBz. If terrain data is available, the grid data of the corresponding area in the radar data is read. D ,make d The number of layers in the radar data is used to combine the radar data to obtain fused radar grid data. Read satellite data formats, including satellite-specific storage files and image formats; Read meteorological data, where each meteorological data point includes longitude, latitude, and time information; Time-aligned processing is performed on satellite data, meteorological data, and radar data; Spatial alignment processing is performed on satellite data, meteorological data, and radar data; A training set is constructed by offline processing of historical offline data that has been aligned spatially and temporally, where satellite data and meteorological data are optional.

3. The short-range radar echo extrapolation method according to claim 2, characterized in that, The expression for the fused radar grid data is as follows: ； in, This represents the fused radar grid data. This represents the normalization constant after weighted summation of grid points within the neighborhood. i and j These represent the discretized latitude and longitude coordinates of the grid points within the neighborhood, respectively. , , and Both represent the neighborhood range. x and y These represent the planar spatial coordinates of the current grid point after discretization. Indicates the height layer of the radar data. This indicates the total number of height layers in the radar data. Representation and height layer d The relevant weighting coefficients, Represents pixel value, This represents the weight set based on the current pixel value range, used to filter out outliers. This indicates the weighting of the geographical area.

4. The short-range radar echo extrapolation method according to claim 2, characterized in that, The time alignment process specifically involves: Let the time of the current radar data be... t The signal values ​​of satellite data and meteorological data are I ; and will be with t The time series of the two most recent meteorological data is set to... t 1 and t 2; like Using the following formula, in time t Interpolation is performed at the point: in, Indicates time t The discrete coordinates of latitude and longitude are i, j The height layer discretized coordinates are d signal strength, Indicates time t Signal strength at point 1 Indicates time t Signal strength at point 2; For those unable to obtain t 2, then use directly. t 1 or according to the t 2. Make predictions and complete time alignment processing.

5. The short-range radar echo extrapolation method according to claim 2, characterized in that, The spatial alignment process specifically involves: Based on the resolution and size of the radar grid data, the actual coverage area of ​​the radar grid data is obtained; Based on the resolution of satellite data and meteorological data, the offset and scale of satellite grid data and meteorological data within the actual coverage area of ​​radar grid data are determined respectively. Based on the actual coverage area, offset, and scale of the radar grid data, cropping is performed; If the resolution of satellite grid data and the resolution of radar grid data are not multiples of 2 or fractions of each other, then the satellite grid data is sampled so that the resolution of the satellite grid data is a multiple of 2 or fractions of the resolution of the radar grid data.

6. The short-range radar echo extrapolation method according to claim 1, characterized in that, The expression for the loss function of the neural network is as follows: ； ； ； ； ； ； ； ； ； Among them, threshold τ have K Each value corresponds to P Threshold of the function τ k , i , j and d This represents the discretized X, Y, and Z axis coordinates on the latitude and longitude plane of the prediction range. t Indicates time, The loss function of a neural network. This represents the mean squared error loss function. k Indicating signal strength k Each level corresponds to a threshold. τ k , K This indicates the total number of signal strength levels. Indicates different thresholds τ Weighted loss function This represents the intensity threshold used to assess prediction accuracy. This represents a multi-threshold loss function. T Indicates the length of the predicted time series. D The height layer discretized coordinates represent the prediction range. This represents the discretized Y-axis length on the prediction range plane. This represents the length of the discretized X-axis on the prediction range plane. Indicates the weighting coefficient. Indicates the predicted value. Indicates the gold standard value. This indicates that the predicted value is greater than the threshold. τ k The probability, β Indicates hyperparameters, This indicates that the gold standard value is greater than the threshold. τ k The probability is 1. This indicates that when the threshold is set to τ k The critical accuracy of the prediction area. This indicates that both the prediction region and the gold standard region are greater than the threshold. τ k The overlap probability, Indicates control variables, β and eps Used to control the range of the region with the lowest loss function value. This indicates that the predicted region is smaller than the threshold. τ k However, the corresponding gold standard area is greater than the threshold. τ k The overlap probability, This indicates that the predicted region is larger than the threshold. τ k However, the corresponding gold standard area is less than the threshold. τ k The probability of overlap.

7. The short-range radar echo extrapolation method according to claim 1, characterized in that, The process of extrapolating future radar echoes based on preprocessed multi-source data and using a trained neural network to output such data completes the extrapolation of short-range radar echoes. Specifically, this involves: The system is initialized by loading the trained neural network and acquiring multi-source input data. Among the multi-source data, radar data is required, while satellite data and meteorological data are optional. The multi-source data is preprocessed to obtain aligned multi-source data, and then the aligned multi-source data is input into the neural network for inference to obtain future radar echo extrapolation data. The future radar echo extrapolation data will be saved in a gridded data storage format; Based on the saved radar echo extrapolation data, a base map of latitude and longitude range is obtained; The latest predicted data is retrieved based on the timestamp, and the plotting library is called in conjunction with the retrieved base map to draw graphics, thus completing the extrapolation processing of short-term radar echoes.

8. A short-range radar echo extrapolation system, wherein the short-range radar echo extrapolation system is used to execute the short-range radar echo extrapolation method according to any one of claims 1-7, characterized in that, include: The first processing module is used to acquire multi-source data and preprocess the multi-source data; The second processing module is used to construct a neural network that combines multi-source data input and multi-branch feature extraction structure, and to train the neural network using a training method based on interval optimal strategy. The neural network includes: The backbone encoder is used to extract primary features of radar spatiotemporal sequences. A dual-branch intermediate network is used to learn the spatiotemporal evolution of primary features of radar spatiotemporal sequences and extract spatiotemporal features. The multi-data processing branch is used to extract initial features from satellite or meteorological grid data using the branch encoder to obtain primary features of satellite / meteorological data, and to obtain deep spatiotemporal latent features based on the primary features of satellite / meteorological data. The output size of the deep spatiotemporal latent features is adjusted by sampling, and the adjusted deep spatiotemporal latent features are merged with the spatiotemporal features extracted by the dual-branch intermediate network. The multi-data processing branch includes a branch encoder and a branch spatiotemporal feature extraction network. The backbone decoder is used to reconstruct and generate prediction sequences based on the spatiotemporal latent features of merged radar, satellite, and meteorological grid data. The dual-branch intermediate network includes a backbone spatiotemporal feature branch and an attention enhancement branch; The main spatiotemporal feature branch is used to learn features based on the primary features of the radar spatiotemporal sequence output by the main encoder, using the spatiotemporal latent feature space to learn the spatiotemporal change trend information of the primary features of the radar spatiotemporal sequence in the latent feature space, and to extract spatiotemporal features. The attention enhancement branch is used to extract and learn detailed information of radar spatiotemporal sequence primary feature data in spatiotemporal changes, wherein a meta-transformer framework is used for feature learning. The third processing module is used to output future radar echo extrapolation data based on the preprocessed multi-source data and the trained neural network, thereby completing the extrapolation processing of short-term radar echoes.

Images