Two-stage Storm Cell Extrapolation Method and System Based on Multi-scale and Attention

Abstract

This invention belongs to the field of storm cell extrapolation and discloses a two-stage storm cell extrapolation method and system based on multi-scale and attention. To obtain clear and accurate extrapolation results, the present invention proposes a two-stage framework that includes the application of multi-scale, attention, and frequency domain losses. In the first stage, multi-scale and attention modules are used to mine important features that influence prediction, and spatiotemporal long-short-term memory units are used for sequence prediction. In the second stage, bias correction is performed, and frequency domain losses are introduced to effectively learn high-frequency information. This method is suitable for storm cell extrapolation tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of storm cell extrapolation, and in particular relates to a two-stage storm cell extrapolation method and system based on multi-scale and attention. Background Art

[0002] Currently, radar echo extrapolation is an important tool for predicting storm cells. Extrapolation methods can be divided into two categories: traditional extrapolation algorithms and deep learning-based algorithms. Among traditional algorithms, the most commonly used is the cross-correlation method, which is inaccurate in severe convective weather events such as storms. Optical flow methods rely on the invariance assumption when calculating optical flow vectors, which radar echo images do not meet, affecting extrapolation accuracy. Among deep learning-based algorithms, the use of recurrent neural networks for storm cell extrapolation is currently the mainstream approach, but several challenges remain. First, storms are characterized by their small spatial scale. In radar echo images, the grayscale value of most pixels is zero. Furthermore, storms exhibit non-rigid motion characteristics such as rotation and scaling. This makes it difficult for prediction models to accurately locate key features relevant to the task, negatively impacting prediction performance. Second, due to spectral bias, the model struggles to fully learn high-frequency details in the echo images, resulting in blurred and incomplete extrapolation results.

[0003] Through the above analysis, the problems and defects of the existing technology are as follows:

[0004] (1) Traditional radar echo extrapolation methods such as the cross-correlation method are not very accurate in severe convective weather, such as storms.

[0005] (2) Although deep learning models such as recurrent neural networks have been used for storm cell extrapolation, due to the small spatial scale and non-rigid motion characteristics of storms, it is difficult for the models to accurately identify and locate key features relevant to the task, resulting in low accuracy. Deep learning models tend to learn low-frequency information, resulting in ambiguous extrapolation results. Summary of the Invention

[0006] In response to the problems existing in the prior art, the present invention provides a two-stage storm cell extrapolation method based on multi-scale and attention to improve the accuracy of storm cell extrapolation.

[0007] The present invention is implemented by constructing a two-stage prediction framework based on a multi-scale and attention-based two-stage Storm monomer extrapolation method by introducing key components such as a multi-scale module, an attention module, and a frequency domain loss. In the first stage, parallel dilated convolutions are used to extract multi-scale information, and improved channel and spatial attention mechanisms are used to mine important features that influence prediction. Then, spatiotemporal long short-term memory units are used for sequence prediction. In the second stage, the results of the first stage are input into a bias correction module to obtain the final prediction results, and a frequency domain loss is introduced to effectively learn high-frequency information, compensating for the model's shortcomings in learning high-frequency information.

[0008] The step 1 is specifically as follows:

[0009] (a) Extracting multi-scale feature information from input data: The multi-scale module uses four convolution kernels with different hole sizes to process the input radar echo feature map, forming multiple receptive field sizes to capture multi-granular representations of contextual information in different regions. The original radar echo data is input to these four branches. The feature maps obtained after the convolution operation are stacked in the channel dimension. 1×1 convolution is used to integrate the information of different scales in the stacked feature maps and compress them to the original number of channels. This effectively captures the multi-scale characteristics of the storm and integrates them into a unified feature representation, improving the accuracy of the prediction.

[0010] (b) Through improved channel and spatial attention mechanisms, important features that affect prediction are mined: two pooling operations are performed on the feature information obtained using the multi-scale module to generate channel features of size 1×1; global average pooling is used to extract the average characteristics of the overall activity level of each channel, and global maximum pooling focuses on the part with the highest activation level in the local area within each channel; these two pooling methods capture the complex structure and changing details of the storm. The two sets of channel description features obtained are respectively extracted through one-dimensional convolution to extract the complex relationship between features. The two channel features are then merged through element-wise addition operations to integrate the information between channels without reducing the dimensionality. Finally, the sigmoid function is used to normalize the addition results and output the weights of each channel in the range of 0 to 1. These weights are applied to the multi-scale features to obtain the channel attention output;

[0011] Different areas of the radar echo image have different levels of importance. Spatial attention is used to learn changes in storm structure and range. The spatial attention sub-block uses convolution and sigmoid functions to calculate spatial weights. Channel weights and spatial weights are applied to the input features, and then their results are added together to obtain the output of the attention module.

[0012] (c) Using spatiotemporal long short-term memory units for sequence prediction: In LSTM, gate structures and cell states are introduced to control the flow of information; in ConvLSTM, the matrix calculations of LSTM are converted into convolution calculations; in ST-LSTM, the information flow from the last layer to the first layer of the next moment is increased;

[0013] The forget gate determines what information to discard from the cell at the current moment. For the input information x at the current moment t and the output h at the previous moment t-1 The weight matrix W fh and W fx , after nonlinear mapping of s, we get the matrix f t ,This matrix determines which information in the cell should be retained and which information should be forgotten;

[0014] Furthermore, the specific formula of step (c) is as follows:

[0015] f t =s(W f [h t-1 ,x t ]+b f ) (1)

[0016] The input gate allows cells to selectively add new memories, which is implemented as shown in formula (2):

[0017] i t =s(W i [h t-1 ,x t ]+b i ) (2)

[0018] Use formula (3) to get the candidate cell state g t , keep x t and h t-1 information;

[0019] g t =tanh(W c [h t-1 ,x t ]+b c ) (3)

[0020] With the input gate and forget gate, the new cell state can be obtained as follows

[0021] C t =f t *C t-1 +i t *g t (4)

[0022] The output gate determines what the final output is and acts on C t The final output is h t :

[0023] O t =s(W o [h t-1 ,x t ]+b o ) (5)

[0024] h t =O t *tanh(C t ) (6)

[0025] ConvLSTM changes the matrix multiplication from (4) to (6) into a convolution operation.

[0026] The ST-LSTM unit uses the spatiotemporal memory unit M to transfer memory between layers, while retaining the memory unit C already present in ConvLSTM. In this way, the spatiotemporal memory unit M is used to realize the flow of information from the last layer to the first layer of the next period, allowing memory to flow vertically in the hierarchical structure and horizontally on the timeline. The output of the attention module is input into the ST-LSTM unit, and a 4-layer ST-LSTM unit is used for prediction to obtain a one-stage prediction result.

[0027] Further, the step 2 is specifically as follows:

[0028] The results of the first stage are input into the bias correction module to obtain the final prediction results: the bias correction module of the second stage uses the prediction results of the first stage as input, and fine-tunes the previous predictions by learning the deviation pattern of the first stage prediction; at the same time, the second stage uses the true values in the original sequence to further improve the model.

[0029] Furthermore, the deviation correction module introduces Fourier transform to further improve the prediction effect. The image is converted to the frequency domain through Fourier transform, and the frequency domain difference between the predicted image and the real image is calculated. The frequency domain difference of each pixel is normalized and used as a weight matrix to act on each pixel. The frequency domain loss obtained in this way can accurately control the training time of high-frequency and low-frequency components. For the two-dimensional discrete Fourier transform, the formula is as follows

[0030]

[0031] Among them, f(x,y) represents the grayscale value of the image at the coordinate (x,y). After the above transformation, the frequency domain vector F(u,v) at the frequency domain (u,v) is obtained.

[0032] The expression of the weight matrix is obtained as formula (8):

[0033]

[0034] Among them, F pd (i, j) represents the frequency domain vector of the predicted image, F gt (i, j) is the frequency domain vector of the real image, and the weight matrix w(i, j) is to normalize the frequency domain difference;

[0035] After obtaining the frequency domain weight matrix from formula (8), it is multiplied by the frequency domain difference to obtain the frequency domain loss. The frequency domain loss is shown in formula (9):

[0036]

[0037] Another object of the present invention is to provide a two-stage storm monomer extrapolation system based on multi-scale and attention, comprising:

[0038] Multi-scale feature extraction module: This module is configured with convolution kernels of different hole sizes to process the input radar echo feature map to capture multi-granularity contextual information in different areas. This module can form multiple receptive field sizes through convolution operations of different scales and integrate the obtained feature maps using 1×1 convolution;

[0039] Channel and spatial attention module: It includes channel attention submodule and spatial attention submodule. The channel attention submodule extracts and fuses channel features through global average pooling and global maximum pooling operations, and uses the Sigmoid function to generate channel weights. The spatial attention submodule uses convolution and Sigmoid functions to calculate spatial weights and perform weighted processing on the input features.

[0040] Spatiotemporal long short-term memory (ST-LSTM) unit: This unit contains a series of gate structures and cell states, as well as a spatiotemporal memory unit M for memory transfer between layers. This unit is responsible for processing sequence data and learning the spatiotemporal dependencies in the sequence.

[0041] Bias correction module: uses the prediction results of the first stage as input, performs fine-tuning and correction through a small number of convolutional structures and residual connections, and uses the true values in the original sequence to optimize the prediction;

[0042] Frequency domain loss calculation module: Fourier transform is introduced to convert the image into the frequency domain, and the frequency domain difference between the predicted image and the real image is calculated. This module can accurately control the training of high-frequency and low-frequency components, thereby improving the quality of the prediction results.

[0043] Another object of the present invention is to provide a synthetic aperture radar, characterized in that the synthetic aperture radar applies a two-stage storm monomer extrapolation method based on multi-scale and attention.

[0044] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a two-stage storm monomer extrapolation method based on multi-scale and attention.

[0045] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of a two-stage storm monomer extrapolation method based on multi-scale and attention.

[0046] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0047] First, the present invention aims to solve the problems of existing storm cell extrapolation methods in terms of accuracy and fuzzy extrapolation results. The core of the present invention is the use of a two-stage framework of multi-scale and attention to improve the accuracy of storm cell extrapolation. Existing methods have difficulties in capturing key features that affect predictions, and the present invention, by introducing multi-scale modules and attention modules, can comprehensively mine features from the perspectives of different scales and attention levels. The multi-scale module captures contextual information in different areas of the storm echo image by using convolution kernels with different hole sizes, and integrates feature representations of multiple receptive field sizes. This can more accurately locate and predict key features related to the task.

[0048] At the same time, the present invention introduces improved channel and spatial attention mechanisms to explore the contribution of important features to prediction. Channel attention extracts channel features through pooling operations, capturing the implicit information about channel relationships in the radar echo feature map. This allows for a better understanding of the importance and weighting between different channels, improving prediction accuracy. Spatial attention is used to capture spatial correlations in the feature map, further understanding the relationships between different pixels. The introduction of these attention mechanisms enables the prediction model to more accurately capture the key features that influence storm extrapolation.

[0049] Furthermore, the present invention uses a bias correction module and frequency domain loss to further optimize prediction results. The bias correction module takes the prediction results from the first stage as input and corrects and optimizes them to obtain more accurate final prediction results. The frequency domain loss is used to learn the high-frequency information of the image, compensating for the model's shortcomings in learning high-frequency details. This improves the ability to express detailed information in the prediction results, making them clearer and more accurate.

[0050] Second, based on solving the existing technical problems, the present invention has achieved significant technological progress in the field of storm cell extrapolation, especially in improving prediction accuracy, improving image quality and enhancing model adaptability.

[0051] Improved prediction accuracy: The combination of multi-scale feature extraction and attention mechanism significantly improves the accuracy of storm extrapolation, especially in capturing the detailed changes of storms.

[0052] Improved image quality: The introduction of the bias correction module and frequency domain loss improves the clarity and detail of the predicted image, making the predicted results closer to the real image.

[0053] Enhanced model adaptability: The introduction of the ST-LSTM unit enables the model to better capture the long-term dependencies of storms and improves the model's adaptability in complex dynamic environments.

[0054] Third, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0055] Competitive Advantages: The multi-scale and attention-based two-stage storm cell extrapolation method provided by this invention offers unique technical advantages, improving the accuracy and predictive effectiveness of storm cell extrapolation. Compared to traditional methods and other deep learning-based approaches, this method can better capture the temporal and spatial correlations of storms, identify key features, and learn high-frequency information, thereby achieving more accurate storm cell extrapolation predictions. This gives this invention a competitive advantage in the field of storm cell extrapolation, attracting the attention and selection of more users and partners.

[0056] Application expansion potential: Storm cell extrapolation has broad application prospects in the meteorological field, involving multiple fields such as weather forecasting, disaster prevention, agriculture, aerospace, etc. The technical solution of the present invention has good application expansion potential. It can be applied to meteorological forecasting agencies to provide more accurate storm cell extrapolation predictions and improve the accuracy and reliability of weather forecasts. At the same time, this technology can also provide important early warning information for disaster prevention and help relevant departments take timely and effective measures. In addition, the agricultural field can use this technology to predict precipitation and provide a scientific basis for the planting and management of crops. The aerospace field can use this technology to predict and evaluate flight safety and improve the safety and efficiency of flights.

[0057] Resource optimization and cost savings: Accurate storm cell extrapolation forecasts can help relevant departments and businesses optimize resources and save costs. For example, in disaster preparedness, accurate predictions of a storm's path and intensity can help deploy rescue teams, supplies, and equipment, allowing for early prevention and preparation, thereby reducing rescue costs and losses.

[0058] The technical solution of the present invention solves a technical problem that people have long desired to solve but have never been able to solve successfully: in the field of storm cell extrapolation, both traditional methods and deep learning-based methods have certain limitations and difficulties. The cross-correlation method, the most commonly used traditional method, has shortcomings in accuracy, especially in severe convective weather, where the accuracy of prediction results is low. Although deep learning-based methods can use recurrent neural networks to extrapolate storm cells, problems such as low accuracy and fuzzy extrapolation results still exist.

[0059] The technical solution of the present invention effectively overcomes the aforementioned technical difficulties by introducing a two-stage storm cell extrapolation method that combines multi-scale and attention. In the first stage, a multi-scale module and an attention module are used to process the input time series data. The multi-scale module extracts feature information at different scales, while the attention module identifies important features that influence prediction. This combination enables the model to better capture the key characteristics of storms, improving prediction accuracy.

[0060] The present invention introduces a second stage, introducing a bias correction module and frequency domain loss, to further optimize the prediction results. The bias correction module uses the rough prediction results from the first stage as input and improves prediction accuracy through correction and optimization. The introduction of frequency domain loss helps the model learn high-frequency information in the image, thereby avoiding spectral bias and improving the ability of the prediction results to express detailed information.

[0061] Through this two-stage storm monomer extrapolation method based on multi-scale and attention, the present invention provides an innovative solution, which effectively solves the technical difficulties existing in the storm monomer extrapolation of traditional methods and deep learning-based methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a diagram of the overall implementation framework of the algorithm provided by the embodiment of the present invention.

[0063] Figure 2 Schematic diagram of a multi-scale module provided by an embodiment of the present invention.

[0064] Figure 3 Schematic diagram of the attention module provided by an embodiment of the present invention.

[0065] Figure 4 This is a diagram of an ST-LSTM unit provided by an embodiment of the present invention.

[0066] Figure 5 Schematic diagram of a 4-layer ST-LSTM provided by an embodiment of the present invention.

[0067] Figure 6 Schematic diagram of a deviation correction module provided in an embodiment of the present invention.

[0068] Figure 7 This is the experimental comparison result provided by the embodiment of the present invention Figure 1 .

[0069] Figure 8 This is the experimental comparison result provided by the embodiment of the present invention Figure 2 . DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0071] Two specific application examples of the present invention are:

[0072] Example 1: Weather Radar Storm Tracking

[0073] This embodiment is used to monitor storm development in real time, provide early warnings to potentially affected areas, and assist emergency management departments in preparing for and responding to potential disasters caused by storms.

[0074] Data processing: The echo data generated by the weather radar is used as input and processed using a multi-scale module to capture storm characteristics at different scales.

[0075] Feature extraction: Through channel and spatial attention mechanisms, the system deeply analyzes radar echo data and emphasizes important features such as the storm center and edge areas.

[0076] Prediction model: Use ST-LSTM units to process time series data and combine them with the output of the attention mechanism to accurately predict the storm's movement trajectory.

[0077] Bias Correction: Based on the predictions from the first stage, the bias correction module fine-tunes the preliminary results to enhance the prediction accuracy.

[0078] Output: The final output provides the expected path and intensity of the storm in the future, assisting meteorological departments in making timely warnings and decisions.

[0079] Example 2: Storm Evacuation Route Planning in the Aviation Field

[0080] Data collection: Acquire real-time echo data of the flight area from weather radar.

[0081] Multi-scale analysis: The system analyzes radar data through multi-scale modules to identify storm characteristics in different areas.

[0082] Attention weighting: Channel and spatial attention mechanisms enhance the identification and analysis of storm activities near flight paths.

[0083] Storm prediction: The ST-LSTM unit combines the output of the attention mechanism to predict the future movement trajectory and intensity of the storm.

[0084] Path planning: Based on the forecast results, the system provides the aircraft with the optimal route to avoid the storm.

[0085] In view of the problems existing in the prior art, the technical solution adopted by the present invention is:

[0086] Multi-scale feature extraction: This method uses a multi-scale module to extract features from storm radar echo data, effectively addressing the small spatial scale and rapid changes of storms. By processing with dilated convolution kernels of varying sizes, it captures multi-granular contextual information, addressing the shortcomings of traditional methods in processing storm details.

[0087] Attention Mechanism: By using channel-wise and spatial-wise attention mechanisms, we more effectively exploit features in radar echo signatures that are crucial for prediction. This approach improves the model's ability to discern the complex structure and changing details of storms, addressing challenges with traditional prediction models in identifying key features.

[0088] Spatiotemporal Long Short-Term Memory (ST-LSTM) units allow for the consideration of temporal information and spatial dependencies within sequences in the prediction model. This addresses the shortcomings of traditional LSTMs in processing three-dimensional information and improves the accuracy of predictions for the non-rigid motion characteristics of storms.

[0089] Bias Correction Module: The introduced bias correction module provides more accurate and detailed prediction results by analyzing the deviation patterns of the first-stage prediction results. This addresses the limitations of the single-stage model in terms of image quality and prediction accuracy.

[0090] Frequency domain loss calculation: Through Fourier transform and frequency domain loss calculation, the present invention overcomes the shortcomings of traditional time series image prediction models in capturing high-frequency details and improves the quality of prediction results.

[0091] like Figure 1 Shown is a flow chart of the method of the present invention, and its specific implementation steps are as follows.

[0092] Step 1: Extract multi-scale feature information of input data. In view of the characteristics of small spatial scale and fast change of storm, a multi-scale module structure is adopted. Figure 2As shown in the figure, it uses four convolution kernels with different hole sizes to process the input radar echo feature map, forming multiple receptive field sizes to capture multi-granular representations of contextual information in different regions. The original radar echo data is input to these four branches. The feature maps generated after the convolution operation are stacked along the channel dimension. A 1×1 convolution is used to integrate the information at different scales in the stacked feature maps and compress them to the original number of channels. This effectively captures the multi-scale characteristics of the storm and integrates them into a unified feature representation, thereby improving prediction accuracy.

[0093] Step 2: Through the improved channel and spatial attention mechanism, the important features that affect the prediction are mined. Figure 3 The upper branch of the function represents channel attention. To exploit the implicit information about channel relationships in the radar echo feature map, two pooling operations are performed on the feature information obtained using the multi-scale module to generate 1×1 channel features. Global average pooling extracts the average overall activity of each channel, while global maximum pooling focuses on the most activated local region within each channel. These two pooling methods effectively capture the complex structure and changing details of the storm. The two sets of channel descriptors are each subjected to one-dimensional convolution to extract the complex relationships between features. The two channel features are then combined using element-wise addition, which better integrates the information between channels while avoiding dimensionality reduction. Finally, the sigmoid function is used to normalize the summed results, outputting a weight for each channel in the range of 0 to 1. This weight is then applied to the multi-scale features to generate the channel attention output.

[0094] like Figure 3 The lower branch of the model shows spatial attention. Different regions of the radar echo image have different levels of importance, and using spatial attention can learn changes in storm structure and range. The spatial attention sub-block uses convolution and sigmoid functions to calculate spatial weights. Channel weights and spatial weights are applied to the input features, and their results are summed to produce the output of the attention module.

[0095] Step 3: Use spatiotemporal long short-term memory units for sequence prediction. In LSTM, researchers introduced a series of gate structures and cell states to control the flow of information. In ConvLSTM, the matrix calculations inherent in LSTM were transformed into convolutional calculations. In ST-LSTM, the information flow from the last layer to the first layer input at the next moment was increased.

[0096] The forget gate determines what information to discard from the cell at the current moment. For the input information x at the current moment t and the output h at the previous moment t-1 The weight matrix W fh and W fx, after nonlinear mapping of s, we get the matrix f t , which determines which information in the cell should be retained and which information should be forgotten. The specific formula is as follows:

[0097] f t =s(W f [h t-1 ,x t ]+b f ) (1)

[0098] The input gate allows cells to selectively add new memories, and its implementation is shown in formula (2).

[0099] i t =s(W i [h t-1 ,x t ]+b i ) (2)

[0100] Use formula (3) to get the candidate cell state g t , keep x t and h t-1 information.

[0101] g t =tanh(W c [h t-1 ,x t ]+b c ) (3)

[0102] With the input gate and forget gate, the new cell state can be obtained as follows

[0103] C t =f t *C t-1 +i t *g t (4)

[0104] The output gate determines what the final output is and acts on C t The final output is h t .

[0105] O t =s(W o [h t-1 ,x t ]+b o ) (5)

[0106] h t =O t *tanh(C t ) (6)

[0107] In order to overcome the shortcomings of LSTM in processing three-dimensional information, ConvLSTM changes the matrix multiplication from (4) to (6) into a convolution operation, which not only takes into account the temporal information but also learns the spatial dependency in the sequence. Figure 4 As shown in Figure 2, the ST-LSTM unit uses the spatiotemporal memory unit M to transfer memory between layers, while retaining the memory unit C that ConvLSTM already has. In this way, the spatiotemporal memory unit M is used to transfer information from the last layer to the first layer of the next period, allowing memory to flow vertically in the hierarchical structure and horizontally on the time axis. The output of the attention module is input to the ST-LSTM unit and used Figure 5 The 4-layer ST-LSTM unit shown in the figure is used for prediction to obtain the first-stage prediction result.

[0108] Step 4: Input the results of the first stage into the bias correction module to obtain the final prediction results. Due to the complexity of the storm cell extrapolation task itself, using only a one-stage model cannot obtain sufficiently accurate and high-quality prediction results. Figure 6 As shown, the second-stage bias correction uses the first-stage prediction results as input. By learning the deviation patterns of the first-stage predictions, it fine-tunes and corrects the previous predictions. Furthermore, the second stage can further refine the model and improve prediction quality by utilizing the true values in the original sequence. The module's implementation utilizes only a small number of convolutional structures and residual connections. This simple yet effective design fully leverages the bias correction module's role in time series forecasting tasks.

[0109] Traditional time-series image prediction models tend to capture low-frequency spatial information, such as overall structure and color, during the learning process, making it difficult to fully learn high-frequency details, resulting in frequency domain differences between the predicted and true images. This paper introduces the Fourier transform to further improve prediction performance. This transform converts the image to the frequency domain, calculates the frequency domain difference between the predicted and true images, and normalizes the frequency domain difference for each pixel as a weight matrix, which is then applied to each pixel. The resulting frequency domain loss can precisely control the training time for high- and low-frequency components, improving the quality of prediction results.

[0110] For the two-dimensional discrete Fourier transform, the formula is as follows

[0111]

[0112] Among them, f(x,y) represents the grayscale value of the image at the coordinate (x,y). After the above transformation, the frequency domain vector F(u,v) at the frequency domain (u,v) is obtained.

[0113] The expression of the weight matrix is obtained as formula (8):

[0114]

[0115] Among them, F pd (i, j) represents the frequency domain vector of the predicted image, F gt (i, j) is the frequency domain vector of the real image, and the weight matrix w(i, j) is used to normalize the frequency domain difference.

[0116] After obtaining the frequency domain weight matrix from formula (8), it is multiplied by the frequency domain difference to obtain the frequency domain loss. The frequency domain loss is shown in formula (9):

[0117]

[0118] The specific application examples of the present invention are:

[0119] Storm Cell Extrapolation 3D Visualization Software: Based on the technical solution of this invention, radar echoes are extrapolated at different elevation angles and visualized using a 3D meteorological data reconstruction algorithm. This smooth and intuitive 3D visualization interface overcomes the human comprehension limitations of single data or text, presenting complex, multidimensional forecast results to decision makers. Users can view possible trends in different regions in real time, providing a reference for response planning. In actual use, users can select parameters such as the input meteorological data file and the forecast duration to obtain a 3D visualization of the forecast results.

[0120] Meteorological early warning system: The technical solution of the present invention can be applied to the meteorological early warning system. By extrapolating the prediction of storm cells, more accurate early warning information can be provided to help people take protective measures in time and reduce casualties and property losses.

[0121] Weather radar data analysis software: Based on the technical solutions of this invention, software specifically designed for analyzing radar echo data can be developed. This software can accurately predict the future location, intensity, and shape of storms, providing powerful support for meteorological researchers and forecasters.

[0122] Weather perception system for autonomous vehicles: The technical solution of this invention can be applied to the weather perception system of autonomous vehicles. By extrapolating and predicting storms, the system can sense the location and intensity of storms in advance, thereby adopting appropriate driving strategies to ensure driving safety.

[0123] Meteorological Big Data Analysis Platform: Based on the technical solutions of this invention, a meteorological big data analysis platform can be built. This platform uses deep learning algorithms to analyze and predict historical meteorological data, providing more accurate storm extrapolation results and providing strong support for meteorological research, meteorological services, and meteorological disaster management. To verify the performance of this invention in storm cell extrapolation prediction, an experimental analysis was conducted using a radar echo dataset from the Xiamen Radar Station.

[0124] The Xiamen radar station records data every four minutes, covering a radius of 460 kilometers, an altitude of 150 kilometers, and a radial resolution of 250 meters. The raw radar base data determines the target position using slant range, azimuth, and elevation. After interpolation of the base data, the radar intensity is mapped to grayscale values and represented as a 920×920 grayscale image. Each sequence contains 10 frames, with the first five frames serving as input data and the remaining five frames as predicted targets. The training set consists of 990 consecutive frames, while the test set consists of the remaining 90 frames. Due to the large image size, each frame was split into 16 230×230 images in the experiment and concatenated along the channel dimension. After the prediction results were obtained, the images were restored to their original size.

[0125] The proposed method is compared with PredRNN-V2, PredRNN-V2 + Attention, PredRNN-V2 + Attention + Two-Stage, PredRNN-V2 + Attention + Multi-Scale, SimVP, and TAU to verify its effectiveness in the Storm single-unit extrapolation task. The experiment uses the Adam optimizer for training at a learning rate of 0.0001, a batch_size of 1, a 4-layer ST-LSTM with 64 hidden states per layer, and 10,000 training rounds for 21 hours.

[0126] Figure 7 The first row shows the predicted images from each model, and the second row shows the difference between the predicted and true images. Subjectively, using both the attention module and the multi-scale module improves PredRNN-V2's predictions. Our model, building on this foundation, uses a two-stage model, achieving the best prediction results with minimal error. Figure 8 The prediction results of SimVP and TAU models are shown in Figure 2. They can both predict the changing trend of the echo image, but the prediction results of TAU model are clearer and have smaller errors than SimVP.

[0127]

[0128] Table 1

[0129] Table 1 provides an objective evaluation of the prediction results of each model. It can be seen that PredRNN-V2 performs better than SimVP on the radar echo dataset, but worse than the TAU model. Using the attention module and the multi-scale module, PredRNN-V2 achieves a significant 7% reduction in mean squared error. Using the two-stage model for correction further improves prediction quality, reducing mean squared error by 11.4%, and outperforming the TAU model in all metrics. This means that the technical solution of the present invention can produce higher-quality storm cell extrapolation results.

[0130] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0131] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A two-stage Storm monomer extrapolation method based on multi-scale and attention, characterized by: include: Step 1: Perform a rough prediction on the input time series data. The input radar echo data first passes through the multi-scale module to extract multi-scale features, then enters the three attention modules to mine key features, uses residual connections for feature fusion, and performs prediction through the ST-LSTM unit to obtain the prediction result; Step 2: Take the prediction result as input, use the deviation correction module for correction optimization, and use the frequency domain loss to learn the high-frequency information of the image to obtain the final output.

2. The two-stage storm monomer extrapolation method based on multi-scale and attention as claimed in claim 1 is characterized in that: The step 1 is specifically as follows: (a) Extracting multi-scale feature information from input data: The multi-scale module uses four convolution kernels with different hole sizes to process the input radar echo feature map, forming multiple receptive field sizes to capture multi-granular representations of contextual information in different regions. The original radar echo data is input to these four branches. The feature maps obtained after the convolution operation are stacked in the channel dimension. 1×1 convolution is used to integrate the information of different scales in the stacked feature maps and compress them to the original number of channels. This effectively captures the multi-scale characteristics of the storm and integrates them into a unified feature representation, improving the accuracy of the prediction. (b) Through the improved channel and spatial attention mechanism, important features that affect prediction are mined: two pooling operations are performed on the feature information obtained using the multi-scale module to generate channel features of size 1×1; Global average pooling is used to extract the average overall activity characteristics of each channel, while global maximum pooling focuses on the most activated part of the local area within each channel. These two pooling methods capture the complex structure and changing details of the storm. The two sets of channel description features obtained are respectively extracted through one-dimensional convolution to extract the complex relationship between features. The two channel features are then combined through element-wise addition to integrate the information between channels without reducing dimensionality. Finally, the sigmoid function is used to normalize the addition result and output the weight of each channel in the range of 0 to 1. This weight is applied to the multi-scale features to obtain the channel attention output. Different areas of the radar echo image have different levels of importance. Spatial attention is used to learn changes in storm structure and range. The spatial attention sub-block uses convolution and sigmoid functions to calculate spatial weights. Channel weights and spatial weights are applied to the input features, and then their results are added together to obtain the output of the attention module. (c) Using spatiotemporal long short-term memory units for sequence prediction: In LSTM, gate structures and cell states are introduced to control the flow of information; in ConvLSTM, the matrix calculations of LSTM are converted into convolution calculations; in ST-LSTM, the information flow from the last layer to the first layer of the next moment is increased; The forget gate determines what information to discard from the cell at the current moment. For the input information x at the current moment t and the output h at the previous moment t-1 The weight matrix W fh and W fx , after nonlinear mapping of σ, we get the matrix f t , which determines which information in the cell should be retained and which information should be forgotten.

3. The two-stage storm monomer extrapolation method based on multi-scale and attention as claimed in claim 2 is characterized in that: The specific formula of step (c) is as follows: f t =s(W f [h t-1 ,x t ]+b f ) (1) The input gate allows cells to selectively add new memories, which is implemented as shown in formula (2): i t =s(W i [h t-1 ,x t ]+b i ) (2) Use formula (3) to get the candidate cell state g t , keep x t and h t-1 information; g t =tanh(W c [h t-1 ,x t ]+b c ) (3) With the input gate and forget gate, the new cell state can be obtained as follows C t =f t *C t-1 +i t *g t (4) The output gate determines what the final output is and acts on C t The final output is h t : O t =s(W o [h t-1 ,x t ]+b o ) (5) h t =O t *tanh(C t ) (6) ConvLSTM changes the matrix multiplication from (4) to (6) into a convolution operation; The ST-LSTM unit uses the spatiotemporal memory unit M to transfer memory between layers, while retaining the memory unit C already present in ConvLSTM. In this way, the spatiotemporal memory unit M is used to realize the flow of information from the last layer to the first layer of the next period, allowing memory to flow vertically in the hierarchical structure and horizontally on the timeline. The output of the attention module is input into the ST-LSTM unit, and a 4-layer ST-LSTM unit is used for prediction to obtain a one-stage prediction result.

4. The two-stage storm monomer extrapolation method based on multi-scale and attention as claimed in claim 1 is characterized in that: The step 2 is specifically as follows: The results of the first stage are input into the bias correction module to obtain the final prediction results: the bias correction module of the second stage uses the prediction results of the first stage as input, and fine-tunes the previous predictions by learning the deviation pattern of the first stage prediction; at the same time, the second stage uses the true values in the original sequence to further improve the model.

5. The two-stage storm monomer extrapolation method based on multi-scale and attention as claimed in claim 4 is characterized in that: The deviation correction module introduces Fourier transform to further improve the prediction effect. The image is converted to the frequency domain through Fourier transform, and the frequency domain difference between the predicted image and the real image is calculated. The frequency domain difference of each pixel is normalized and used as a weight matrix to act on each pixel. The frequency domain loss obtained in this way can accurately control the training time of high-frequency and low-frequency components. For the two-dimensional discrete Fourier transform, the formula is as follows Among them, f(x,y) represents the grayscale value of the image at the coordinate (x,y). After the above transformation, the frequency domain vector F(u,v) at the frequency domain (u,v) is obtained. The expression of the weight matrix is obtained as formula (8): Among them, F pd (i, j) represents the frequency domain vector of the predicted image, F gt (i, j) is the frequency domain vector of the real image, and the weight matrix ω(i, j) is to normalize the frequency domain difference; After obtaining the frequency domain weight matrix from formula (8), it is multiplied by the frequency domain difference to obtain the frequency domain loss. The frequency domain loss is shown in formula (9):

6. A two-stage storm monomer extrapolation system based on multi-scale and attention for implementing the method of claim 1, characterized in that: include: Multi-scale feature extraction module: This module is configured with convolution kernels of different hole sizes to process the input radar echo feature map to capture multi-granularity contextual information in different areas. This module can form multiple receptive field sizes through convolution operations of different scales and integrate the obtained feature maps using 1×1 convolution; Channel and spatial attention module: It includes channel attention submodule and spatial attention submodule. The channel attention submodule extracts and fuses channel features through global average pooling and global maximum pooling operations, and uses the Sigmoid function to generate channel weights. The spatial attention submodule uses convolution and Sigmoid functions to calculate spatial weights and perform weighted processing on the input features. Spatiotemporal long short-term memory (ST-LSTM) unit: This unit contains a series of gate structures and cell states, as well as a spatiotemporal memory unit M for memory transfer between layers. This unit is responsible for processing sequence data and learning the spatiotemporal dependencies in the sequence. Bias correction module: uses the prediction results of the first stage as input, performs fine-tuning and correction through a small number of convolutional structures and residual connections, and uses the true values in the original sequence to optimize the prediction; Frequency domain loss calculation module: Fourier transform is introduced to convert the image into the frequency domain, and the frequency domain difference between the predicted image and the real image is calculated. This module can accurately control the training of high-frequency and low-frequency components, thereby improving the quality of the prediction results.

7. A synthetic aperture radar, characterized in that: The synthetic aperture radar applies the two-stage storm monomer extrapolation method based on multi-scale and attention as described in any one of claims 1 to 5.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the two-stage storm monomer extrapolation method based on multi-scale and attention as described in claim 1.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the two-stage storm monomer extrapolation method based on multi-scale and attention as described in claim 1.

Images