A Radar Echo Prediction Method and System Based on Image Processing

Abstract

This invention discloses a radar echo forecasting method and system based on image processing, belonging to the field of meteorological forecasting technology. The system includes a data acquisition module, a data analysis module, an image processing module, and a visualization forecasting module. The data acquisition module collects model library data, radar echo data, and operating parameters. The data analysis module preprocesses the radar echo data, divides the data into regions, and calculates weight indices. The image processing module sets radar operating parameters based on the weight indices, establishes a forecasting model set, and generates forecast images. The visualization forecasting module displays the forecast images visually and performs automatic forecasting under set conditions.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, specifically to a radar echo forecasting method and system based on image processing. Background Technology

[0002] With the continuous improvement of radar data collection capabilities, research on automatic radar data interpretation has gradually attracted attention. As a sensor that uses microwaves for active sensing, radar has become an important means of Earth observation. Based on radar echo image data and combined with ERA5 numerical model data, using artificial intelligence technology to deeply mine weather pattern changes and achieve short-term near-term extrapolation of radar data has become an important research direction.

[0003] Currently, radar echo extrapolation methods typically employ traditional techniques such as optical flow and cross-correlation extrapolation, which have certain limitations. Firstly, under severe convective weather conditions, due to their strong locality and small scale, traditional radar echo extrapolation methods have low accuracy, making it difficult to meet the needs of future weather forecasting. Secondly, existing weather radars typically complete a volume scan in tens of minutes. This time interval often fails to capture the rapid development and changes in severe convective weather, and this scanning frequency limits the real-time nature of monitoring data. This not only affects timely responses to weather changes but may also lead to a failure to issue effective warnings during extreme weather events, thereby increasing potential safety risks. Therefore, it is necessary to further improve the accuracy and real-time performance of short-term extrapolation of radar echo images to provide a low-cost, high-efficiency monitoring method. Summary of the Invention

[0004] The purpose of this invention is to provide a radar echo prediction method and system based on image processing to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides a radar echo prediction method based on image processing, comprising:

[0006] S100, acquisition model library, and radar echo data and operating parameters.

[0007] S200: Preprocess the radar echo data, divide the data area, and calculate the weight index.

[0008] S300: Set radar operating parameters according to the weight index, establish a forecast model set and generate forecast images.

[0009] S400: Visualizes forecast images and allows for automatic forecasting based on set conditions.

[0010] In S100, the model library refers to a software library containing different models, which are used to predict the precipitation trend in the future based on radar echo data.

[0011] Operating parameters include beamwidth, pulse width, and scanning mode; scanning modes include fixed-direction scanning and rotational scanning.

[0012] S200 includes:

[0013] S201. When the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement.

[0014] The echo factor is calculated based on the processed echo signal and converted into a meteorological image Q. Using the radar position in the meteorological image Q as the starting point and the beamwidth as the angle, different data areas are divided within the meteorological image Q.

[0015] Meteorological images are generated by rotating radar to transmit and receiving echoes, so the displayable area is circular, with the radar located at the center. Based on the beamwidth as the angle, the displayable area in the meteorological image can be divided into several sector-shaped data areas. Data is collected at the same time within each data area, but at different times between different data areas.

[0016] S202. Analyze the precipitation distribution within meteorological image Q and set a threshold W. thr The precipitation is greater than W thr The area is designated as the region of interest. Regions of interest covering more than one data area are marked, and the data collection time in each data area is analyzed. Based on the collection time and historical data, the diffusion direction and diffusion speed of each marked region of interest are predicted.

[0017] In rotating scan mode, the acquisition times of adjacent data areas are usually sequential and at equal intervals. For a marker coverage area g within data area v1, the marker coverage area g in its adjacent data area v2 is not covered at the same time, but rather at a time after the acquisition time of data area v1.

[0018] Therefore, based on the time difference, the current meteorological image of data area v1, and the historical meteorological images of data areas v1 and v2, the existing prediction model is used to predict the diffusion direction and diffusion speed of each marked area of ​​interest.

[0019] S203. Analyze the time T between the current time and the next data acquisition based on the radar's operating parameters. next Based on the diffusion direction and diffusion rate, the duration T is analyzed. next The diffusion areas of each marked area of ​​interest are then analyzed. The area ratio of the areas of interest and diffusion areas within the data region, along with precipitation, are used to calculate the weight index QZ for each data region using the formula.

[0020]

[0021] In the formula, α is the weight coefficient of the region of interest, β is the weight coefficient of the diffusion region, r is the number of regions of interest, i is the number of diffusion regions, and GBF e and GW e These represent the area percentage and precipitation of the e-th area of ​​concern, respectively, KBF u and GW u These represent the area percentage and precipitation of the u-th diffusion zone, respectively.

[0022] The time for the next data acquisition is calculated by subtracting the operating time under the current beamwidth from the radar pulse width.

[0023] The settings for radar operating parameters in the S300 include:

[0024] S301. Obtain the weight index of all data regions and calculate the standard deviation, then set the threshold B. thr When the standard deviation is greater than B thr Then adjust the radar scanning mode to fixed-direction scanning. Select the data area S with the largest weight index. max As the area where radar first acquires data, the data area S in the meteorological image Q is updated based on the acquired data. max And recalculate the weight index of each data area.

[0025] S302. On the meteorological image Q, plan the center line for each data area. Taking data area S... max Using the centerline as the basic line, analyze the angle AJ between the centerline and the basic line of each other data area. p , with weighted index QZ p Substitute them together into the formula to calculate the pointing coefficient ZX of each data area:

[0026]

[0027] In the formula, z is a constant greater than 1.

[0028] The center line and the basic line of each data area share a common vertex, and the included angle is less than or equal to 180 degrees.

[0029] S303. Select the data area with the highest pointing coefficient as the radar's next acquisition area. Update the corresponding data area in the meteorological image Q again based on the acquired data, and recalculate the weighting index and the pointing coefficients of other data areas. Continue this process to continuously plan the next acquisition area for the radar.

[0030] In fixed-direction scanning mode, the weight index of all data areas is dynamically updated synchronously with each update of the meteorological image. Except for the initial selection of the data area with the highest weight index, subsequent selections of the next data acquisition area are based on the pointing coefficient. This avoids the radar continuously acquiring data over a large area across angles, thus preventing resource waste.

[0031] The process of establishing a forecast model set and generating forecast images in S300 includes:

[0032] First, in the initial sampling area S max Once determined, extract data region S from meteorological image Q. max The feature information is used to generate predicted images for each model in the model library. After the radar completes data acquisition, the data area S is updated. max The similarity between the updated image and the predicted image of each model is calculated separately.

[0033] Before generating predicted images, the model needs to be trained. A deep learning technique is used to construct an echo evolution prediction scheme. The spatiotemporal characteristics of radar data are learned using the Diffusion architecture, and a generative adversarial GAN ​​Loss is added for supervised learning to enhance the model's ability to learn complex echo patterns.

[0034] The generator network receives radar time-series data as input and outputs predicted time-series data. The discriminator network receives the predicted time-series data and the corresponding real data as input, determines whether the predicted data is similar to the real data, and calculates a loss function. Then, the generator network performs backpropagation based on the feedback from the discriminator network, updating its parameters. Simultaneously, the discriminator network also performs backpropagation based on the calculated loss function, updating its parameters.

[0035] Evaluation of Prediction Performance: Evaluation metrics such as CSI or SSIM are used to assess the prediction performance, thereby determining the generator network's capabilities. Since the generator and discriminator are competitive and mutually reinforcing, multiple iterations of training are required to achieve better prediction results. During training, the generator gradually learns how to predict radar images, while the discriminator progressively improves its ability to distinguish between predicted and real images. After training, the predicted images generated by the generator will be as close as possible to the real images and will retain the main features of convection morphology; that is, visually, the predicted images will resemble real radar images.

[0036] Secondly, a threshold h is set to filter out models with a similarity greater than h to establish a prediction model set. After the next acquisition area is determined, the prediction images and similarity are generated again using the above method, and models with a similarity less than h are removed from the prediction model set.

[0037] Finally, the accuracy is calculated by averaging the historical similarities of each model in the forecast model ensemble, and weights are assigned based on the accuracy of different models. Using an ensemble forecasting method, the prediction results of all models in the ensemble are integrated according to their weights to obtain the final forecast image. The weights are adjusted in real time based on changes in the accuracy of each model, continuously improving the performance of the forecast model ensemble.

[0038] Error correction is achieved through bias learning of the deviation between a single prediction result and the target, primarily involving forward propagation and backdiffusion. Forward propagation involves adding Gaussian noise to the predicted image to corrupt it. A Markov chain is used to progressively add noise to the data at each time step to obtain the posterior probability. Ultimately, the initial ground truth image is gradually transformed into a pure Gaussian noise image after the forward propagation process. Backdiffusion, on the other hand, involves traversing the Markov chain backward, regenerating the input image from the pure noise image.

[0039] The training of the diffusion model mainly focuses on the inverse diffusion process. The training objective is to learn the inverse diffusion process, that is, to train the parameters related to the posterior probability distribution. Ultimately, these parameters can be used to connect the feature relationships between radar predicted images and real images, correcting the error accumulation caused by long-term prediction.

[0040] In S400, meteorological and forecast images are displayed in real time on a visualization screen in the data center, with data areas, diffusion areas, and labeled areas marked, as well as the weight index and pointing coefficient for each data area. Precipitation forecast conditions are set, and forecasts are automatically generated if the forecast image meets the conditions.

[0041] A radar echo forecasting system based on image processing includes a data acquisition module, a data analysis module, an image processing module, and a visualization forecasting module.

[0042] The data acquisition module is used to collect echo data and operating parameters from the model library and radar.

[0043] The data analysis module is used to preprocess radar echo data, divide data areas, and calculate weight indices.

[0044] The image processing module is used to set radar operating parameters according to the weight index, establish a set of forecast models, and generate forecast images.

[0045] The visualization forecast module is used to display forecast images visually and to set conditions for automatic forecasting.

[0046] The data acquisition module includes a radar data acquisition unit and a model library acquisition unit.

[0047] The radar data acquisition unit is used to acquire radar echo data and operating parameters. Operating parameters include beamwidth, pulse width, and scanning mode. Scanning modes include fixed-direction scanning and rotating scanning.

[0048] The model library acquisition unit is used to acquire model libraries, specifically software libraries containing different models. These models are used to predict precipitation trends over a future period based on radar echo data.

[0049] The data analysis module includes an echo processing unit and a weight analysis unit.

[0050] The echo processing unit is used to preprocess radar echo data and divide it into data areas and areas of interest.

[0051] First, when the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement.

[0052] After acquiring the radar base data, the data is cleaned to remove some missing frames or to create mosaic data, and then traditional quality control algorithms are used to remove clutter.

[0053] By using a general denoising model trained on radar data and traditional denoising models such as fuzzy algorithms, ground clutter suppression is performed on radar reflectivity data to ensure the stability of echo data and prevent abrupt and abnormal high-intensity wave interference caused by ground clutter from interfering with network training.

[0054] Secondly, the echo factor is calculated based on the processed echo signal and converted into a meteorological image Q. Using the radar position in the meteorological image Q as the starting point and the beamwidth as the angle, different data areas are divided within the meteorological image Q.

[0055] Then, the precipitation distribution within meteorological image Q is analyzed, and a threshold W is set. thr The precipitation is greater than W thr The area is designated as the area of ​​interest.

[0056] Finally, the areas of interest that cover more than one data region are marked, the data collection time in each data region is analyzed, and the diffusion direction and diffusion speed of each marked area of ​​interest are predicted based on the collection time and historical data.

[0057] The weighting analysis unit is used to calculate the weight index for each data area.

[0058] First, analyze the radar's operating parameters to determine the time T between the current time and the next data acquisition. next Based on the diffusion direction and diffusion rate, the duration T is analyzed. next The diffusion areas of each marked area of ​​interest are then analyzed, along with the area ratio of the areas of interest and diffusion areas within the data area, as well as the precipitation.

[0059] Then according to the formula Calculate the weight index for each data region. Where α is the weight coefficient for the region of interest, β is the weight coefficient for the diffusion region, r is the number of regions of interest, i is the number of diffusion regions, and GBF... e and GW e These represent the area percentage and precipitation of the e-th area of ​​concern, respectively, KBF u and GW u These represent the area percentage and precipitation of the u-th diffusion zone, respectively.

[0060] The image processing module includes a parameter setting unit and an image generation unit.

[0061] The parameter setting unit is used to calculate the pointing coefficient of each data area, thereby setting the order of radar data acquisition.

[0062] First, obtain the weight index of all data regions and calculate the standard deviation, then set the threshold B. thr When the standard deviation is greater than B thr The radar scanning mode is then adjusted to fixed-direction scanning.

[0063] Select the data region S with the largest weight index. max As the area where radar first acquires data, the data area S in the meteorological image Q is updated based on the acquired data. max And recalculate the weight index of each data area.

[0064] Secondly, center lines are planned for each data region on the meteorological image Q. Taking data region S... max Using the center line as the basic line, the weight index QZ of other data areas is analyzed. p And the angle AJ between the center line and the basic line. p .

[0065] Then, according to the formula Calculate the pointing coefficient of each data area, and select the data area with the largest pointing coefficient as the radar's next acquisition area. In the formula, z is a constant greater than 1.

[0066] Finally, based on the collected data, the corresponding data area in the meteorological image Q is updated again, and the weighting index and the pointing coefficients of other data areas are recalculated. This process is repeated to continuously plan the next data acquisition area for the radar.

[0067] The image generation unit is used to build a set of forecast models and generate forecast images.

[0068] First, in the initial sampling area S max Once determined, extract data region S from meteorological image Q. max Based on the feature information, each model in the model library generates a predicted image according to the feature information.

[0069] The model needs to be trained before generating predicted images. During training, supervised learning is performed using historical echo data and observed precipitation data to optimize model parameters and ensure its predictive ability on new data. Simultaneously, cross-validation is used to verify the model's performance on different datasets to prevent overfitting.

[0070] The main network adopts a Vid2Vid network architecture. Data encoding and quality control are performed in the preprocessing stage, which completes the encoding of radar data and the completion of missing data. The main modules of the algorithm will provide additional information, and these modules are also performed in the preprocessing stage.

[0071] Information is input into the main network as a new channel, and the Diffusion model is used to learn the deviation between the single forecast result and the target to correct the error accumulation caused by long-term forecasts.

[0072] Secondly, after the radar finishes collecting data, it updates the data area S. max The similarity between the updated image and the predicted image of each model is calculated. A threshold h is set to filter out models with a similarity greater than h to build a prediction model set.

[0073] After the next acquisition area is determined, continue to generate predicted images and calculate similarity using the above method, and remove models in the prediction model set with a similarity of no more than h.

[0074] Then, the accuracy is calculated by averaging all historical similarities of each model in the prediction model set, and weights are set for weighted averages based on the accuracy of different models.

[0075] During the forecast generation and optimization phase, the adjusted model is used to perform near-term extrapolation forecasts on newly acquired radar echo data. Through dynamically updated models, the evolution of echoes and precipitation distribution over a future period are predicted in real time.

[0076] To improve forecast accuracy, ensemble forecasting is used, which involves weighted averaging of forecasts from multiple models to reduce the error of a single model and enhance forecast reliability.

[0077] Finally, an ensemble forecasting method is used to integrate the prediction results of all models in the forecast model set according to their weights to obtain the final forecast image.

[0078] During the model evaluation and feedback phase, an evaluation system is established to comprehensively analyze and verify the forecast results. Multiple evaluation indicators, such as root mean square error, accuracy, or false alarm rate, are used to comprehensively evaluate the model's forecast performance.

[0079] By comparing the model's performance with real-time observation data, we can analyze its performance under different meteorological conditions and identify potential problems and areas for improvement. Simultaneously, a feedback mechanism is established to incorporate real-time monitoring data and user feedback into model updates, ensuring the model can continuously adapt to changing meteorological environments.

[0080] The weights are adjusted in real time based on changes in the accuracy of each model, and the performance of the forecast model set is continuously improved.

[0081] The implementation of this integrated technical approach can significantly improve the accuracy of radar echo nowcasting, providing strong support for short-term weather forecasting and meteorological services.

[0082] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0083] Data Preprocessing and Dynamic Adjustment: This scheme includes denoising and correction steps in the data preprocessing stage, and dynamically adjusts the data acquisition area by setting weighting indices and pointing coefficients. This feature can more effectively handle interfering data and improve forecast accuracy.

[0084] Parameter optimization: By analyzing the weighting index and standard deviation, this method can automatically adjust the radar scanning mode when significant differences are detected between data areas. This makes the forecasting process more flexible and improves the ability to respond to extreme weather changes.

[0085] Application of Deep Learning and GAN Models: This technical solution utilizes deep learning and Generative Adversarial Network (GAN) techniques to enhance the model's ability to learn complex echo patterns. Through supervised learning, the model can more accurately capture the evolution of precipitation patterns, thereby improving prediction accuracy.

[0086] Adaptive forecast model ensemble: This approach establishes a forecast ensemble comprising multiple models and integrates them using weighted averages based on their historical performance. This not only improves forecast accuracy but also updates model weights in real time, avoiding the limitations of relying on a single model.

[0087] Real-time visualization and automatic forecasting: This solution displays forecast images in real time through a data center visualization system and allows for the setting of automatic forecasting conditions. This not only improves the user experience but also enables timely issuance of weather warnings, helping to cope with weather changes.

[0088] In summary, this technical solution effectively improves the accuracy and response speed of radar echo prediction by integrating multiple advanced technologies such as deep learning, flexible parameter adjustment, and real-time visualization, and has significant advantages over existing technologies. Attached Figure Description

[0089] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0090] Figure 1 This is a flowchart illustrating a radar echo prediction method based on image processing according to the present invention.

[0091] Figure 2 This is a schematic diagram of the structure of a radar echo prediction system based on image processing according to the present invention. Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] Please see Figure 1 This invention provides a radar echo prediction method based on image processing, comprising:

[0094] S100, acquisition model library, and radar echo data and operating parameters.

[0095] S200: Preprocess the radar echo data, divide the data area, and calculate the weight index.

[0096] S300: Set radar operating parameters according to the weight index, establish a forecast model set and generate forecast images.

[0097] S400: Visualizes forecast images and allows for automatic forecasting based on set conditions.

[0098] In S100, the model library refers to a software library containing different models, which are used to predict the precipitation trend in the future based on radar echo data.

[0099] Operating parameters include beamwidth, pulse width, and scanning mode; scanning modes include fixed-direction scanning and rotational scanning.

[0100] S200 includes:

[0101] S201. When the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement.

[0102] The echo factor is calculated based on the processed echo signal and converted into a meteorological image Q. Using the radar position in the meteorological image Q as the starting point and the beamwidth as the angle, different data areas are divided within the meteorological image Q.

[0103] Meteorological images are generated by rotating radar to transmit and receiving echoes, so the displayable area is circular, with the radar located at the center. Based on the beamwidth as the angle, the displayable area in the meteorological image can be divided into several sector-shaped data areas. Data is collected at the same time within each data area, but at different times between different data areas.

[0104] S202. Analyze the precipitation distribution within meteorological image Q and set a threshold W. thr The precipitation is greater than W thr The area is designated as the region of interest. Regions of interest covering more than one data area are marked, and the data collection time in each data area is analyzed. Based on the collection time and historical data, the diffusion direction and diffusion speed of each marked region of interest are predicted.

[0105] In rotating scan mode, the acquisition times of adjacent data areas are usually sequential and at equal intervals. For a marker coverage area g within data area v1, the marker coverage area g in its adjacent data area v2 is not covered at the same time, but rather at a time after the acquisition time of data area v1.

[0106] Therefore, based on the time difference, the current meteorological image of data area v1, and the historical meteorological images of data areas v1 and v2, the existing prediction model is used to predict the diffusion direction and diffusion speed of each marked area of ​​interest.

[0107] S203. Analyze the time T between the current time and the next data acquisition based on the radar's operating parameters. next Based on the diffusion direction and diffusion rate, the duration T is analyzed. next The diffusion areas of each marked area of ​​interest are then analyzed. The area ratio of the areas of interest and diffusion areas within the data region, along with precipitation, are used to calculate the weight index QZ for each data region using the formula.

[0108]

[0109] In the formula, α is the weight coefficient of the region of interest, β is the weight coefficient of the diffusion region, r is the number of regions of interest, i is the number of diffusion regions, and GBF e and GW e These represent the area percentage and precipitation of the e-th area of ​​concern, respectively, KBF u and GW u These represent the area percentage and precipitation of the u-th diffusion zone, respectively.

[0110] The time for the next data acquisition is calculated by subtracting the operating time under the current beamwidth from the radar pulse width.

[0111] The settings for radar operating parameters in the S300 include:

[0112] S301. Obtain the weight index of all data regions and calculate the standard deviation, then set the threshold B. thr When the standard deviation is greater than B thr Then adjust the radar scanning mode to fixed-direction scanning. Select the data area S with the largest weight index. max As the area where radar first acquires data, the data area S in the meteorological image Q is updated based on the acquired data. max And recalculate the weight index of each data area.

[0113] S302. On the meteorological image Q, plan the center line for each data area. Taking data area S... max Using the centerline as the basic line, analyze the angle AJ between the centerline and the basic line of each other data area. p , with weighted index QZ p Substitute them together into the formula to calculate the pointing coefficient ZX of each data area:

[0114]

[0115] In the formula, z is a constant greater than 1.

[0116] The center line and the basic line of each data area share a common vertex, and the included angle is less than or equal to 180 degrees.

[0117] S303. Select the data area with the highest pointing coefficient as the radar's next acquisition area. Update the corresponding data area in the meteorological image Q again based on the acquired data, and recalculate the weighting index and the pointing coefficients of other data areas. Continue this process to continuously plan the next acquisition area for the radar.

[0118] In fixed-direction scanning mode, the weight index of all data areas is dynamically updated synchronously with each update of the meteorological image. Except for the initial selection of the data area with the highest weight index, subsequent selections of the next data acquisition area are based on the pointing coefficient. This avoids the radar continuously acquiring data over a large area across angles, thus preventing resource waste.

[0119] The process of establishing a forecast model set and generating forecast images in S300 includes:

[0120] First, in the initial sampling area S max Once determined, extract data region S from meteorological image Q. max The feature information is used to generate predicted images for each model in the model library. After the radar completes data acquisition, the data area S is updated. maxThe similarity between the updated image and the predicted image of each model is calculated separately.

[0121] Before generating predicted images, the model needs to be trained. A deep learning technique is used to construct an echo evolution prediction scheme. The spatiotemporal characteristics of radar data are learned using the Diffusion architecture, and a generative adversarial GAN ​​Loss is added for supervised learning to enhance the model's ability to learn complex echo patterns.

[0122] The generator network receives radar time-series data as input and outputs predicted time-series data. The discriminator network receives the predicted time-series data and the corresponding real data as input, determines whether the predicted data is similar to the real data, and calculates a loss function. Then, the generator network performs backpropagation based on the feedback from the discriminator network, updating its parameters. Simultaneously, the discriminator network also performs backpropagation based on the calculated loss function, updating its parameters.

[0123] Evaluation of Prediction Performance: Evaluation metrics such as CSI or SSIM are used to assess the prediction performance, thereby determining the generator network's capabilities. Since the generator and discriminator are competitive and mutually reinforcing, multiple iterations of training are required to achieve better prediction results. During training, the generator gradually learns how to predict radar images, while the discriminator progressively improves its ability to distinguish between predicted and real images. After training, the predicted images generated by the generator will be as close as possible to the real images and will retain the main features of convection morphology; that is, visually, the predicted images will resemble real radar images.

[0124] Secondly, a threshold h is set to filter out models with a similarity greater than h to establish a prediction model set. After the next acquisition area is determined, the prediction images and similarity are generated again using the above method, and models with a similarity less than h are removed from the prediction model set.

[0125] Finally, the accuracy is calculated by averaging the historical similarities of each model in the forecast model ensemble, and weights are assigned based on the accuracy of different models. Using an ensemble forecasting method, the prediction results of all models in the ensemble are integrated according to their weights to obtain the final forecast image. The weights are adjusted in real time based on changes in the accuracy of each model, continuously improving the performance of the forecast model ensemble.

[0126] Error correction is achieved through bias learning of the deviation between a single prediction result and the target, primarily involving forward propagation and backdiffusion. Forward propagation involves adding Gaussian noise to the predicted image to corrupt it. A Markov chain is used to progressively add noise to the data at each time step to obtain the posterior probability. Ultimately, the initial ground truth image is gradually transformed into a pure Gaussian noise image after the forward propagation process. Backdiffusion, on the other hand, involves traversing the Markov chain backward, regenerating the input image from the pure noise image.

[0127] The training of the diffusion model mainly focuses on the inverse diffusion process. The training objective is to learn the inverse diffusion process, that is, to train the parameters related to the posterior probability distribution. Ultimately, these parameters can be used to connect the feature relationships between radar predicted images and real images, correcting the error accumulation caused by long-term prediction.

[0128] In S400, meteorological and forecast images are displayed in real time on a visualization screen in the data center, with data areas, diffusion areas, and labeled areas marked, as well as the weight index and pointing coefficient for each data area. Precipitation forecast conditions are set, and forecasts are automatically generated if the forecast image meets the conditions.

[0129] Please see Figure 2 The present invention provides a radar echo prediction system based on image processing, including a data acquisition module, a data analysis module, an image processing module, and a visualization prediction module.

[0130] The data acquisition module is used to collect echo data and operating parameters from the model library and radar.

[0131] The data analysis module is used to preprocess radar echo data, divide data areas, and calculate weight indices.

[0132] The image processing module is used to set radar operating parameters according to the weight index, establish a set of forecast models, and generate forecast images.

[0133] The visualization forecast module is used to display forecast images visually and to set conditions for automatic forecasting.

[0134] The data acquisition module includes a radar data acquisition unit and a model library acquisition unit.

[0135] The radar data acquisition unit is used to acquire radar echo data and operating parameters. Operating parameters include beamwidth, pulse width, and scanning mode. Scanning modes include fixed-direction scanning and rotating scanning.

[0136] The model library acquisition unit is used to acquire model libraries, specifically software libraries containing different models. These models are used to predict precipitation trends over a future period based on radar echo data.

[0137] The data analysis module includes an echo processing unit and a weight analysis unit.

[0138] The echo processing unit is used to preprocess radar echo data and divide it into data areas and areas of interest.

[0139] First, when the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement.

[0140] After acquiring the radar base data, the data is cleaned to remove some missing frames or to create mosaic data, and then traditional quality control algorithms are used to remove clutter.

[0141] By using a general denoising model trained on radar data and traditional denoising models such as fuzzy algorithms, ground clutter suppression is performed on radar reflectivity data to ensure the stability of echo data and prevent abrupt and abnormal high-intensity wave interference caused by ground clutter from interfering with network training.

[0142] Secondly, the echo factor is calculated based on the processed echo signal and converted into a meteorological image Q. Using the radar position in the meteorological image Q as the starting point and the beamwidth as the angle, different data areas are divided within the meteorological image Q.

[0143] Then, the precipitation distribution within meteorological image Q is analyzed, and a threshold W is set. thr The precipitation is greater than W thr The area is designated as the area of ​​interest.

[0144] Finally, the areas of interest that cover more than one data region are marked, the data collection time in each data region is analyzed, and the diffusion direction and diffusion speed of each marked area of ​​interest are predicted based on the collection time and historical data.

[0145] The weighting analysis unit is used to calculate the weight index for each data area.

[0146] First, analyze the radar's operating parameters to determine the time T between the current time and the next data acquisition. next Based on the diffusion direction and diffusion rate, the duration T is analyzed. next The diffusion areas of each marked area of ​​interest are then analyzed, along with the area ratio of the areas of interest and diffusion areas within the data area, as well as the precipitation.

[0147] Then according to the formula Calculate the weight index for each data region. Where α is the weight coefficient for the region of interest, β is the weight coefficient for the diffusion region, r is the number of regions of interest, i is the number of diffusion regions, and GBF... e and GW e These represent the area percentage and precipitation of the e-th area of ​​concern, respectively, KBF u and GW u These represent the area percentage and precipitation of the u-th diffusion zone, respectively.

[0148] The image processing module includes a parameter setting unit and an image generation unit.

[0149] The parameter setting unit is used to calculate the pointing coefficient of each data area, thereby setting the order of radar data acquisition.

[0150] First, obtain the weight index of all data regions and calculate the standard deviation, then set the threshold B. thr When the standard deviation is greater than B thr The radar scanning mode is then adjusted to fixed-direction scanning.

[0151] Select the data region S with the largest weight index. max As the area where radar first acquires data, the data area S in the meteorological image Q is updated based on the acquired data. max And recalculate the weight index of each data area.

[0152] Secondly, center lines are planned for each data region on the meteorological image Q. Taking data region S... max Using the center line as the basic line, the weight index QZ of other data areas is analyzed. p And the angle AJ between the center line and the basic line. p .

[0153] Then, according to the formula Calculate the pointing coefficient of each data area, and select the data area with the largest pointing coefficient as the radar's next acquisition area. In the formula, z is a constant greater than 1.

[0154] Finally, based on the collected data, the corresponding data area in the meteorological image Q is updated again, and the weighting index and the pointing coefficients of other data areas are recalculated. This process is repeated to continuously plan the next data acquisition area for the radar.

[0155] The image generation unit is used to build a set of forecast models and generate forecast images.

[0156] First, in the initial sampling area S max Once determined, extract data region S from meteorological image Q. max Based on the feature information, each model in the model library generates a predicted image according to the feature information.

[0157] The model needs to be trained before generating predicted images. During training, supervised learning is performed using historical echo data and observed precipitation data to optimize model parameters and ensure its predictive ability on new data. Simultaneously, cross-validation is used to verify the model's performance on different datasets to prevent overfitting.

[0158] The main network adopts a Vid2Vid network architecture. Data encoding and quality control are performed in the preprocessing stage, which completes the encoding of radar data and the completion of missing data. The main modules of the algorithm will provide additional information, and these modules are also performed in the preprocessing stage.

[0159] Information is input into the main network as a new channel, and the Diffusion model is used to learn the deviation between the single forecast result and the target to correct the error accumulation caused by long-term forecasts.

[0160] Secondly, after the radar finishes collecting data, it updates the data area S. max The similarity between the updated image and the predicted image of each model is calculated. A threshold h is set to filter out models with a similarity greater than h to build a prediction model set.

[0161] After the next acquisition area is determined, continue to generate predicted images and calculate similarity using the above method, and remove models in the prediction model set with a similarity of no more than h.

[0162] Then, the accuracy is calculated by averaging all historical similarities of each model in the prediction model set, and weights are set for weighted averages based on the accuracy of different models.

[0163] During the forecast generation and optimization phase, the adjusted model is used to perform near-term extrapolation forecasts on newly acquired radar echo data. Through dynamically updated models, the evolution of echoes and precipitation distribution over a future period are predicted in real time.

[0164] To improve forecast accuracy, ensemble forecasting is used, which involves weighted averaging of forecasts from multiple models to reduce the error of a single model and enhance forecast reliability.

[0165] Finally, an ensemble forecasting method is used to integrate the prediction results of all models in the forecast model set according to their weights to obtain the final forecast image.

[0166] During the model evaluation and feedback phase, an evaluation system is established to comprehensively analyze and verify the forecast results. Multiple evaluation indicators, such as root mean square error, accuracy, or false alarm rate, are used to comprehensively evaluate the model's forecast performance.

[0167] By comparing the model's performance with real-time observation data, we can analyze its performance under different meteorological conditions and identify potential problems and areas for improvement. Simultaneously, a feedback mechanism is established to incorporate real-time monitoring data and user feedback into model updates, ensuring the model can continuously adapt to changing meteorological environments.

[0168] The weights are adjusted in real time based on changes in the accuracy of each model, and the performance of the forecast model set is continuously improved.

[0169] The implementation of this integrated technical approach can significantly improve the accuracy of radar echo nowcasting, providing strong support for short-term weather forecasting and meteorological services.

[0170] Example 1:

[0171] Assuming there are four data areas in the meteorological image: A1, A2, A3, and A4, with the center line of data area A1 as the baseline, the angles between the center lines of data areas A2, A3, and A4 and the baseline are 90°, 180°, and 90°, respectively.

[0172] When the weight exponents of A2, A3, and A4 are 1.2, 1.4, and 1.6 respectively, and the constant z is 1.2, substitute these values ​​into the formula to calculate the pointing coefficients of the data areas A2, A3, and A4 respectively:

[0173] A2 pointing coefficient:

[0174] A3 pointing coefficient:

[0175] A4 pointing coefficient:

[0176] Then select data area A4 as the radar's next acquisition area.

[0177] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0178] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A radar echo prediction method based on image processing, characterized in that: The method includes: S100, acquisition model library, and radar echo data and operating parameters; S200. Preprocessing the radar echo data, dividing the data area, and calculating the weight index; specifically including: S201. When the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement; the echo factor is calculated based on the processed echo signal and converted into a meteorological image Q; different data areas are divided in the meteorological image Q with the radar position in the meteorological image Q as the starting point and the beamwidth as the angle; S202. Analyze the precipitation distribution within meteorological image Q and set a threshold. The precipitation is greater than The area is designated as the area of ​​interest; the area of ​​interest covering more than one data area is marked, the data collection time in each data area is analyzed, and the diffusion direction and diffusion speed of each marked area of ​​interest are predicted based on the collection time and historical data. S203. Analyze the time elapsed between the current time and the next data acquisition based on the radar's operating parameters. Analysis of duration based on diffusion direction and diffusion rate The diffusion areas of each marked area of ​​interest are then analyzed; the area ratio of the areas of interest and diffusion areas within the data region, as well as the precipitation, are substituted into the formula to calculate the weight index of each data region. : ； In the formula, For the weighting coefficients of the region of interest, For the diffusion region weighting coefficient, For the number of areas of interest, For the number of diffusion zones, and The first The area percentage and precipitation of each area of ​​concern and The first The area proportion and precipitation of each diffusion zone; S300: Set radar operating parameters according to the weight index, establish a set of forecast models and generate forecast images; The settings for radar operating parameters include: S301. Obtain the weight index of all data regions and calculate the standard deviation, then set the threshold. When the standard deviation is greater than The radar scanning mode was then adjusted to fixed-direction scanning; the data area with the highest weighting index was selected. As the area where radar first acquires data, the data region in meteorological image Q is updated based on the acquired data. And recalculate the weight index for each data region; S302. On the meteorological image Q, plan the center line for each data area; based on the data area... Using the centerline as the basic line, analyze the angle between the centerline and the basic line of other data areas. , with weight index Substitute them together into the formula to calculate the pointing coefficient of each data area. : ； In the formula, It is a constant greater than 1; S303. Select the data area with the largest pointing coefficient as the radar’s next acquisition area. Update the corresponding data area in the meteorological image Q again based on the acquired data, and recalculate the weight index and the pointing coefficients of other data areas. Continue to plan the next acquisition area for the radar in this way. S400: Visualizes forecast images and allows for automatic forecasting based on set conditions.

2. The radar echo prediction method based on image processing according to claim 1, characterized in that: In S100, the model library refers to a software library containing different models. These models are used to predict the future precipitation trend based on radar echo data. The operating parameters include beamwidth, pulse width, and scanning mode. The scanning modes include fixed-direction scanning and rotating scanning.

3. The radar echo prediction method based on image processing according to claim 1, characterized in that: The process of establishing a forecast model set and generating forecast images in S300 includes: First, in the initial sampling area Once determined, extract the data area from meteorological image Q. The feature information is used to generate predicted images for each model in the model library; after the radar collects data, the data area is updated. Calculate the similarity between the updated image and the predicted image of each model; Secondly, a threshold h is set to filter out models with similarity greater than h to establish a prediction model set; after the next collection area is determined, the prediction image and similarity are generated and calculated again according to the above method, and models with similarity less than h are removed from the prediction model set. Finally, the accuracy is calculated by averaging the historical similarities of each model in the forecast model set, and weights are set by weighting the accuracy of different models. The ensemble forecasting method is used to integrate the prediction results of all models in the forecast model set according to their weights to obtain the final forecast image. The weights are adjusted in real time according to the changes in the accuracy of each model to continuously improve the performance of the forecast model set.

4. The radar echo prediction method based on image processing according to claim 3, characterized in that: In S400, meteorological and forecast images are displayed in real time through a visualization screen in the data center, with data areas, diffusion areas, and labeled areas marked, as well as the weight index and pointing coefficient of each data area; precipitation forecast conditions are set, and forecasts are automatically made if the forecast image meets the conditions.

5. A radar echo prediction system based on image processing, applied to the radar echo prediction method based on image processing as described in claim 1, characterized in that: The system includes a data acquisition module, a data analysis module, an image processing module, and a visualization forecasting module; The data acquisition module is used to collect echo data and operating parameters from the model library and radar. The data analysis module is used to preprocess radar echo data, divide data areas, and calculate weight indices; The image processing module is used to set radar operating parameters according to the weight index, establish a set of forecast models, and generate forecast images; The visualization forecast module is used to display forecast images visually and to set conditions for automatic forecasting.

6. The radar echo prediction system based on image processing according to claim 5, characterized in that: The data acquisition module includes a radar data acquisition unit and a model library acquisition unit; The radar data acquisition unit is used to acquire radar echo data and operating parameters; the operating parameters include beamwidth, pulse width and scanning mode, and the scanning modes include fixed-direction scanning and rotating scanning; The model library acquisition unit is used to acquire model libraries, specifically software libraries containing different models. These models are used to predict precipitation trends over a future period based on radar echo data.

7. A radar echo prediction system based on image processing according to claim 6, characterized in that: The data analysis module includes an echo processing unit and a weight analysis unit; The echo processing unit is used to preprocess radar echo data and divide it into data areas and areas of interest; First, when the radar scanning mode is rotating scan, the radar echo data is preprocessed after acquisition, including noise reduction, correction and enhancement. Secondly, the echo factor is calculated based on the processed echo signal and converted into a meteorological image Q; Starting from the radar position in meteorological image Q and using the beamwidth as the angle, different data areas are divided in meteorological image Q. Then, the precipitation distribution within meteorological image Q is analyzed, and a threshold is set. The precipitation is greater than The area is designated as the area of ​​interest; Finally, the areas of interest that cover more than one data region are marked, the data collection time in each data region is analyzed, and the diffusion direction and diffusion speed of each marked area of ​​interest are predicted based on the collection time and historical data. The weighting analysis unit is used to calculate the weight index for each data region; First, analyze the time elapsed between the current time and the next data acquisition based on the radar's operating parameters. Analysis of duration based on diffusion direction and diffusion rate The diffusion areas of each marked area of ​​interest are then analyzed, along with the area ratio of the areas of interest and diffusion areas and the precipitation within the data area. Then according to the formula Calculate the weight index for each data region; where, For the weighting coefficients of the region of interest, For the diffusion region weighting coefficient, For the number of areas of interest, For the number of diffusion zones, and The first The area percentage and precipitation of each area of ​​concern and The first The area proportion and precipitation of each diffusion zone.

8. A radar echo prediction system based on image processing according to claim 7, characterized in that: The image processing module includes a parameter setting unit and an image generation unit; The parameter setting unit is used to calculate the pointing coefficient of each data area, thereby setting the order of radar data acquisition; First, obtain the weight index for all data regions and calculate the standard deviation, then set the threshold. When the standard deviation is greater than The radar scanning mode is then adjusted to fixed-direction scanning; Select the data area with the largest weight index. As the area where radar first acquires data, the data region in meteorological image Q is updated based on the acquired data. And recalculate the weight index for each data region; Secondly, center lines are planned for each data region in the meteorological image Q; based on the data region Using the center line as the basic line, we analyze the weighting indices of other data regions. and the angle between the center line and the basic line. ; Then, according to the formula Calculate the pointing coefficient of each data area, and select the data area with the largest pointing coefficient as the radar's next acquisition area; where, It is a constant greater than 1; Finally, based on the collected data, the corresponding data area in the meteorological image Q is updated again, and the weight index and the pointing coefficients of other data areas are recalculated; and so on, the next collection area is continuously planned for the radar. The image generation unit is used to build a set of forecast models and generate forecast images; First, in the initial sampling area Once determined, extract the data area from meteorological image Q. Based on the feature information, each model in the model library generates a predicted image according to the feature information. Secondly, the radar updates the data area after collecting the data. Calculate the similarity between the updated image and the predicted image of each model; Set a threshold h and filter out models with similarity greater than h to build a set of prediction models; After the next acquisition area is determined, continue to generate predicted images and calculate similarity using the above method, and remove models with similarity no greater than h from the prediction model set; Then, the accuracy is calculated as the average of all historical similarities for each model in the prediction model set, and weights are set according to the accuracy of different models. Finally, the ensemble forecasting method is used to integrate the prediction results of all models in the forecast model set according to their weights to obtain the final forecast image. The weights are adjusted in real time based on changes in the accuracy of each model, and the performance of the forecast model set is continuously improved.

Images