Assimilation method for dual-polarization radar data

Abstract

The invention relates to the technical field of weather prediction, and discloses a dual-polarization radar data assimilation method, which comprises the following steps: S10, constructing a UNet hydrogel inversion model based on a deep learning algorithm of a UNet convolutional neural model and dual-polarization radar data, the UNet hydrogel inversion model is used for learning a nonlinear relationship between the dual-polarization radar data and a hydrogel mixing ratio; s20, a data set of dual-polarization radar data is obtained through a numerical forecasting simulation experiment, and corresponding dual-polarization parameters are calculated through a radar simulator and serve as input data of the UNet hydrogel inversion model; s30, processing the input data by using the UNet hydrogel inversion model to obtain the mixing ratio information of the four kinds of inversed hydrogel; and S40, based on an analysis increment updating method (IAU), using the hydrogel mixing ratio information output by the UNet hydrogel inversion model as an analysis increment, and introducing an integration process of a numerical mode within a preset time.

Description

Technical Field

[0001] The present invention relates to the technical field of weather prediction, and particularly to an assimilation method for dual-polarization radar data. Background Art

[0002] Numerical weather prediction plays a key role in weather process prediction. Its core is to describe the state of atmospheric motion and physical processes through mathematical equations and use computer simulation to predict future weather conditions. In numerical prediction, the accuracy of the model highly depends on the accuracy of the initial conditions. Providing a more accurate initial field can help the numerical model capture the actual state of the atmosphere more realistically. With the wide application of multi-source observation data such as ground observation stations, radars, and satellites, and the continuous development of data assimilation technology, the means and methods for improving the quality of the initial field have become increasingly rich.

[0003] As one of the important tools for atmospheric monitoring, dual-polarization radar plays an important role especially in hydrometeor monitoring and data assimilation. Compared with traditional Doppler weather radars, dual-polarization radars can provide polarization data including differential reflectivity (ZDR), co-polar correlation coefficient (ρhv), differential phase shift (ΦDP), and specific differential phase shift (KDP). These variables can reveal the microphysical characteristics of precipitation systems, such as the shape, type, and phase state of hydrometeors.

[0004] Currently, the application research of dual-polarization radar data mainly focuses on fields such as data quality control, raindrop size distribution inversion, hydrometeor phase classification, and quantitative precipitation estimation. In numerical models, researchers have developed an assimilation method for dual-polarization radar variables based on the Ensemble Kalman Filter (EnKF) algorithm and verified its effectiveness through simulation observation experiments and actual cases. However, the EnKF framework requires a large number of ensemble members to ensure the rationalization of the statistical results of background error covariance, which poses high requirements for computing resources. In actual operations, to improve computational efficiency, the three-dimensional variational data assimilation (3DVAR) scheme is usually adopted to replace the EnKF technology. Within the 3DVAR framework, researchers have developed an assimilation scheme for polarization quantities by constructing an observation operator. Related research shows that this scheme can significantly improve the quality of the model initial field and thus improve the prediction effect. However, the 3DVAR scheme also has certain limitations. On the one hand, the 3DVAR scheme needs to handle complex forward operators, which is difficult in theoretical modeling and code implementation; on the other hand, traditional dual-polarization radar inversion algorithms are usually based on simple statistical relationships and are difficult to fully describe the complex non-linear relationship between polarization quantities and hydrometeor variables. In addition, the observation information contained in these statistical relationships is relatively limited and it is difficult to fully explore and utilize the potential of dual-polarization radar data. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to solve the problem that the existing assimilation method of dual-polarization radar data is difficult to be used for accurately predicting weather conditions.

[0006] To solve the above technical problem, the present invention provides an assimilation method for dual-polarization radar data, and the method includes:

[0007] S10, constructing a UNet hydrometeor inversion model based on the deep learning algorithm of the UNet convolutional neural model and the dual-polarization radar data, where the UNet hydrometeor inversion model is used to learn the non-linear relationship between the dual-polarization radar data and the hydrometeor mixing ratio;

[0008] S20, obtaining a data set of dual-polarization radar data through a numerical weather prediction simulation experiment, where the data set includes the mixing ratio information and number concentration data of four hydrometeors including rain, snow, hail, and graupel, and calculating the corresponding dual-polarization parameters by using a radar simulator as the input data of the UNet hydrometeor inversion model;

[0009] S30, processing the input data by using the UNet hydrometeor inversion model to obtain the inverted mixing ratio information of the four hydrometeors;

[0010] S40, based on the analysis increment update method (IAU), taking the hydrometeor mixing ratio information output by the UNet hydrometeor inversion model as the analysis increment and introducing it into the integration process of the numerical model within a preset time.

[0011] Furthermore, the structure of the UNet hydrometeor inversion model includes two stages of downsampling and upsampling. The downsampling stage extracts features through continuous convolution and pooling operations, and the upsampling stage restores the spatial dimension and integrates features through deconvolution operations to generate a high-resolution output result.

[0012] Furthermore, the input data further includes air density and temperature, and the input data is normalized before being input into the model.

[0013] Furthermore, during the model training process, the PyTorch deep learning framework is adopted, GPU acceleration is enabled, the stochastic gradient descent method (SGD) is used as the optimizer, and the model is optimized in combination with the cosine annealing learning rate scheduler.

[0014] Furthermore, the analysis increment update method (IAU) decomposes the analysis increment into several equal small increments and gradually introduces them into the integration process of the numerical model within the relaxation time τ to implicitly utilize the dynamic-physical process of the model as a constraint.

[0015] Furthermore, the method further includes a random flipping data augmentation method.

[0016] Further, when constructing the UNet hydrometeor inversion model, the flipping probability of the data augmentation method using random flipping is 50%.

[0017] Further, the method further includes:

[0018] Preprocessing the dual-polarization radar data to remove noise and outliers in the dual-polarization radar data.

[0019] Compared with the prior art, the beneficial effect of the dual-polarization radar data assimilation method according to the embodiments of the present invention lies in:

[0020] By constructing a deep learning algorithm based on the UNet convolutional neural model, the present invention can more accurately learn the non-linear relationship between dual-polarization radar data and hydrometeor mixing ratio. The modeling method makes the inverted hydrometeor mixing ratio information more accurate, thereby improving the accuracy of weather prediction. By introducing a deep learning model based on UNet, the present invention realizes the ability to efficiently invert the mixing ratios of four hydrometeors from dual-polarization radar data. On this basis, using the analysis increment update (IAU) scheme, the hydrometeor analysis increment inverted by the UNet model is directly incorporated into the numerical model integration process, forming a complete set of dual-polarization radar data assimilation methods. The development of this method not only improves the utilization rate of radar data, but also significantly improves the calculation efficiency and prediction accuracy of the numerical model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of the dual-polarization radar data assimilation method provided by the embodiments of the present invention;

[0022] Figure 2 is an architecture diagram of the UNet model provided by the embodiments of the present invention;

[0023] Figure 3 is the hydrometeor inversion process of the UNet model provided by the embodiments of the present invention;

[0024] Figure 4 is a schematic diagram of the analysis increment update (IAU) scheme provided by the embodiments of the present invention, where τ is the relaxation time of the IAU scheme. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following describes exemplary embodiments of the present invention with reference to the accompanying drawings. Various details of the embodiments of the present invention are included to assist in understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present invention. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted below.

[0026] As Figures 1-4 shown, in an alternative embodiment of the present invention, the method for assimilating dual-polarization radar data includes:

[0027] S10. Construct a UNet hydrometeor inversion model based on a deep learning algorithm of the UNet convolutional neural model and dual-polarization radar data, where the UNet hydrometeor inversion model is used to learn the non-linear relationship between the dual-polarization radar data and the hydrometeor mixing ratio;

[0028] S20. Obtain a dataset of dual-polarization radar data through numerical prediction simulation experiments. The dataset includes the mixing ratio information and number concentration data of four hydrometeors, including rain, snow, hail, and graupel, and calculate the corresponding dual-polarization parameters using a radar simulator as the input data of the UNet hydrometeor inversion model;

[0029] S30. Process the input data using the UNet hydrometeor inversion model to obtain the inverted mixing ratio information of the four hydrometeors;

[0030] S40. Based on the analysis increment update method (IAU), use the hydrometeor mixing ratio information output by the UNet hydrometeor inversion model as the analysis increment and introduce it into the integration process of the numerical model within a preset time.

[0031] Among them, UNet is a deep learning architecture initially used for biomedical image processing, especially image segmentation tasks. Its characteristic is a symmetric upsampling and downsampling structure, which can effectively capture multi-scale features in images and is suitable for processing image data with complex spatial structures. A dual-polarization radar is a radar system that can transmit and receive electromagnetic waves in different polarization directions. By measuring the echo signals in different polarization states, more detailed information about precipitation particle types, sizes, shapes, and distributions can be obtained. The hydrometeor mixing ratio refers to the ratio of the mass or volume of hydrometeors (such as rain, snow, hail, graupel, etc.) in the atmosphere to the total mass or volume of air, and is an important parameter for describing the content of hydrometeors in the atmosphere. The analysis increment update method (IAU) is a data assimilation technique used to incorporate observational data into a numerical prediction model to improve the initial field of the model. By analyzing the differences between observational data and model predictions, the analysis increment is calculated and the initial state of the model is updated.

[0032] In S10, a UNet hydrometeor inversion model is constructed to learn the non - linear relationship between dual - polarization radar data and hydrometeor mixing ratios. The model can accurately invert the mixing ratio information of hydrometeors from dual - polarization radar data. A large number of sample data of dual - polarization radar data and hydrometeor mixing ratios are collected, and these data are used to train the UNet model, and the model parameters are adjusted to minimize the prediction error. In S20, a dataset of dual - polarization radar data is obtained, including the mixing ratio information and number concentration data of four types of hydrometeors, and the corresponding dual - polarization parameters are calculated to provide the input data required for training the UNet hydrometeor inversion model. A dataset containing the mixing ratio information and number concentration data of rain, snow, hail, and graupel is generated through a numerical prediction simulation experiment, and the radar simulator is used to calculate the dual - polarization parameters based on these data. In S30, the input data is processed using the UNet hydrometeor inversion model to obtain the inverted mixing ratio information of the four types of hydrometeors. The mixing ratio of hydrometeors is accurately inverted from the dual - polarization radar data. The dual - polarization parameters obtained in step S20 are input into the trained UNet model, and the model outputs the inverted mixing ratio information of the four types of hydrometeors. In S40, the mixing ratio information of hydrometeors output by the UNet hydrometeor inversion model is used as an analysis increment and introduced into the integration process of the numerical model to improve the initial field of the numerical prediction model and enhance the prediction accuracy. The IAU method is used to fuse the inverted mixing ratio information of hydrometeors with the initial field of the numerical prediction model, update the initial state of the model, and perform numerical integration within a preset time to obtain an improved prediction result.

[0033] Specifically, combined with Figures 2-4 A specific embodiment is further described as follows:

[0034] (1) UNet hydrometeor inversion model

[0035] In order to establish the relationship between radar polarization quantities and hydrometeor mixing ratios, the embodiment of the present invention proposes a deep - learning algorithm based on the UNet convolutional neural model. This algorithm can automatically learn the conversion relationship between dual - polarization radar data and the hydrometeor mixing ratios of the numerical model. UNet is a commonly used convolutional neural model, which is mainly divided into two stages in structure: downsampling and upsampling. In the downsampling stage, UNet gradually reduces the spatial dimension of the input image through continuous convolution and pooling operations, while extracting features at different abstraction levels; in the upsampling stage, UNet gradually restores the spatial dimension of the input image through deconvolution operations (also known as transposed convolutions) and integrates low - level and high - level features.

[0036] Figure 2The UNet model structure adopted by the present invention is shown. Before being input into the UNet model, the training and test data are uniformly converted into images with a resolution of 512×512. Our method is based on the UNet model with 1×1 convolutions, where both the downsampling and upsampling parts are composed of 1×1 convolutional layers and normalization layers. The one-dimensional convolutional kernel can learn the point-to-point relationship, that is, the correspondence between the value at a certain position in the model input and the corresponding position in the output. The UNet model adopts an encoder-decoder structure: the encoder is responsible for feature extraction, gradually increasing the number of channels and refining features as the model deepens; the decoder aims to restore features to generate high-resolution outputs. Since convolution and pooling operations will cause partial information loss, the decoder integrates the corresponding information from the encoder through skip connections, thereby restoring the spatial information lost during the downsampling process. Finally, the output size in the upsampling stage matches the input image to generate the final result. The UNet structure is simple and efficient, and can well solve the problem of feature extraction.

[0037] Before model training, sufficient preparation of the dataset is crucial. To obtain the required data (such as the number concentration and mixing ratio of hydrometeors), we conducted multiple batches of simulation experiments through numerical prediction, and extracted and saved the mixing ratio and number concentration data of four types of hydrometeors including rain, snow, hail, and graupel. Based on these data, we further calculated the corresponding dual-polarization parameters using a radar simulator. The embodiment of the present invention mainly focuses on the southeastern coastal area, so the dual-polarization radar simulator proposed by Wang Hong in 2016 is selected. This simulator takes the mixing ratio and number concentration of hydrometeors as inputs and calculates dual-polarization parameters such as horizontal reflectivity, differential reflectivity, specific differential phase, and co-polar correlation coefficient. The reason for choosing this simulator is that it is compatible with the S-band dual-polarization radars commonly used in South China and has been tested and verified through typical weather cases, and can accurately reproduce the main characteristics of the weather systems in the South China region.

[0038] The overall workflow based on the UNet model is as Figure 2 shown. First, data such as the mixing ratio and number concentration of four types of hydrometeors are obtained through numerical experiments. Then, four dual-polarization radar variables (horizontal reflectivity, differential reflectivity, specific differential phase, and co-polar correlation coefficient) are calculated using a polarization radar simulator. The input of the UNet model includes ten parameters: four dual-polarization radar variables, four hydrometeor number concentrations, air density, and temperature. All input parameters are normalized before prediction, and the mixing ratios of the four types of hydrometeors obtained from numerical simulations are used as labels.

[0039] To optimize the model, we used the PyTorch deep learning framework and enabled GPU acceleration. The model was trained for 100 epochs with both the input and output images having a resolution of 512×512, and then restored to the original size after training. To increase data diversity, we adopted a data augmentation method of random flipping with a flipping probability of 50%. The optimizer selected was Stochastic Gradient Descent (SGD), combined with the Cosine Annealing learning rate scheduler. This scheduler accelerates the model convergence in the early stage of training and ensures stable fluctuations near the optimal solution in the later stage. This strategy is commonly used in the training of complex deep models, which helps to prevent overfitting and improve the model performance and generalization ability.

[0040] (2) Analyze the incremental update method

[0041] Next, we combined the hydrometeor information obtained from the UNet model with the background field in the numerical model. To improve the forecasting performance of the numerical model, especially the short-term forecasting quality for the first 0 to 6 hours, we introduced the hydrometeor information output by the UNet model into the integration process of the numerical model by analyzing the Incremental Analysis Update (IAU) scheme. The IAU technique decomposes the analysis increment into several equal small increments and gradually introduces them into the model integration within the relaxation time τ.

[0042] The IAU method can implicitly utilize the dynamic-physical processes of the model as constraints to achieve adjustments between different variable fields. This method can better balance the incremental information and the background field, significantly reduce and suppress gravity wave noise, shorten the model adjustment time, and thus improve the quality of short-term numerical weather forecasting. Figure 3 The schematic diagram of the IAU technique adopted in the present invention is shown. Within the time window [-τ / 2, τ / 2], the IAU method introduces the information of four hydrometeor mixing ratios into the model integration process to improve the forecasting accuracy.

[0043] In an alternative embodiment of the present invention, the structure of the UNet hydrometeor retrieval model includes two stages: downsampling and upsampling. The downsampling stage extracts features through consecutive convolution and pooling operations, and the upsampling stage restores the spatial dimension and integrates features through deconvolution operations to generate a high-resolution output result.

[0044] Among them, in the downsampling stage, the spatial dimension of the input image is gradually reduced through successive convolution and pooling operations. This stage aims to extract features at different levels of abstraction, laying a foundation for subsequent feature integration and classification / regression tasks. Contrary to the downsampling stage, in the upsampling stage, the spatial dimension of the input image is gradually restored through deconvolution operations (also known as transposed convolutions). In this stage, the model integrates low-level and high-level features to generate high-resolution output results. The convolution operation is a commonly used operation in image processing. By sliding a convolution kernel over the input image and calculating the dot product, features of the image are extracted. The convolution operation can capture local patterns and textures in the image. The pooling operation is a downsampling technique that reduces the spatial dimension of the image by selecting the maximum value (max pooling) or the average value (average pooling) within a region. The pooling operation can reduce the redundancy of data and improve the robustness of the model. The deconvolution operation: also known as transposed convolution, is an operation in the upsampling process. It gradually restores the spatial dimension of the image by transposing the convolution kernel and convolving the input feature map.

[0045] In the embodiment of the present invention, by gradually reducing the spatial dimension of the image, the redundancy of data is reduced, and the efficiency of feature extraction is improved; high-resolution output results are generated, the detailed information in the image is retained, and the prediction accuracy of the model is improved.

[0046] In an optional embodiment of the present invention, the input data further includes air density and temperature, and the input data is normalized before being input into the model.

[0047] Specifically, the input data not only includes dual-polarization radar data, but also extends to meteorological parameters such as air density and temperature. These data together constitute the input of the model for training and optimizing the UNet hydrometeor inversion model. The normalization process scales the values of the input data to a specific range, making the magnitudes of different features consistent, which helps the model better learn and understand the data, accelerates the training process of the model, improves the convergence speed and prediction accuracy of the model. At the same time, the normalization process also helps prevent the model from overfitting during the training process.

[0048] In an optional embodiment of the present invention, the PyTorch deep learning framework is adopted during the model training process, GPU acceleration is enabled, the stochastic gradient descent method (SGD) is used as the optimizer, and the model is optimized in combination with a cosine annealing learning rate scheduler.

[0049] In the embodiment of the present invention, by adopting measures such as the PyTorch deep learning framework, enabling GPU acceleration, using the SGD optimizer, and the cosine annealing learning rate scheduler, the training efficiency, convergence speed, and generalization ability of the UNet hydrometeor inversion model are further improved.

[0050] In an alternative embodiment of the present invention, the analysis incremental update method (IAU) decomposes the analysis increment into several equal small increments and gradually introduces them into the integration process of the numerical model within the relaxation time τ to implicitly utilize the dynamic-physical processes of the model as constraints.

[0051] Among them, the analysis incremental update method (IAU, Incremental Analysis Update) is a data assimilation technique aimed at gradually introducing the difference between the observed data and the analysis result (i.e., the analysis increment) into the numerical weather prediction model to improve the initial conditions. The analysis increment refers to the difference between the observed data and the model analysis result, which is used to update the initial state of the numerical model. The integration process of the numerical model is the process by which the numerical weather prediction model predicts the future weather state through time integration. During the integration process, the model state evolves over time and is affected by the initial conditions, boundary conditions, and internal physical processes of the model. The relaxation time τ is a control parameter that defines the time length during which the analysis increment is gradually introduced into the integration process of the numerical model. Within the relaxation time, the analysis increment gradually affects the model state in a certain way (such as linearly increasing).

[0052] In the embodiment of the present invention, by decomposing the analysis increment and gradually introducing it into the numerical model, the instability phenomenon of the model state caused by sudden changes in the initial conditions is reduced; implicitly utilizing the dynamic-physical processes of the numerical model as constraints improves the data assimilation ability of the numerical model; by optimizing the initial conditions, the accuracy of numerical weather prediction is improved, providing more powerful support for meteorological forecasting and services.

[0053] In an alternative embodiment of the present invention, the method further includes a data augmentation method of random flipping.

[0054] The data augmentation method of random flipping is a commonly used technique in the fields of machine learning and computer vision, mainly used to increase the diversity of training data and improve the generalization ability of the model. Specifically, random flipping includes two methods: horizontal flipping and vertical flipping.

[0055] In the embodiment of the present invention, through data augmentation, the generalization ability of the model can be significantly improved. During the training process, the model will learn more diverse features, enabling it to perform better on unseen data. In addition, data augmentation can also reduce the model's dependence on the training data and reduce the risk of overfitting.

[0056] In an alternative embodiment of the present invention, when constructing the UNet hydrometeor inversion model, the flipping probability of the data augmentation method of random flipping is 50%.

[0057] In an alternative embodiment of the present invention, the method further includes:

[0058] Preprocess the dual-polarization radar data to remove noise and outliers from the dual-polarization radar data.

[0059] In the embodiments of the present invention, preprocessing can significantly reduce or eliminate the noise components in the radar data. These noises may originate from equipment errors, environmental factors, or interference during data transmission. By removing outliers, abnormal data points caused by equipment failures, data recording errors, or extreme weather conditions can be corrected, thereby ensuring the accuracy of the data.

[0060] The above specific embodiments do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A dual polarization radar data assimilation method, characterized in that: The method comprises: S10, constructing a UNet hydrometeor inversion model based on a deep learning algorithm of a UNet convolutional neural model and dual-polarization radar data, wherein the UNet hydrometeor inversion model is used to learn a nonlinear relationship between the dual-polarization radar data and a hydrometeor mixing ratio; S20, obtaining a data set of dual-polarization radar data through a numerical forecast simulation experiment, wherein the data set includes mixing ratio information and number concentration data of four types of hydrometeors, including rain, snow, hail, and graupel, and calculating corresponding dual-polarization parameters using a radar simulator as input data of a UNet hydrometeor inversion model; S30, using the UNet hydrometeor inversion model to process the input data to obtain inverted mixing ratio information of four hydrometeors; S40, based on an analysis increment update method (IAU), the hydrometeor mixing ratio information output by the UNet hydrometeor inversion model is used as an analysis increment, and an integration process of a numerical model is introduced within a preset time.

2. The dual polarization radar data assimilation method according to claim 1, characterized in that: The structure of the UNet hydrometeor inversion model includes two stages: downsampling and upsampling. The downsampling stage extracts features through continuous convolution and pooling operations, and the upsampling stage restores spatial dimensions and integrates features through deconvolution operations to generate high-resolution output results.

3. The dual polarization radar data assimilation method according to claim 2, characterized in that: The input data also includes air density and temperature, and the input data is normalized before being input into the model.

4. The dual polarization radar data assimilation method according to claim 2, characterized in that: The PyTorch deep learning framework is used in the model training process, and GPU acceleration is enabled. Stochastic gradient descent (SGD) is used as the optimizer, and the cosine annealing learning rate scheduler is combined for model optimization.

5. The dual polarization radar data assimilation method according to claim 1, characterized in that: The analytical incremental update method (IAU) decomposes the analytical increment into a number of equal small increments, gradually introducing the integration process of the numerical model within the relaxation time τ to implicitly utilize the dynamic-physical process of the model as a constraint.

6. The dual polarization radar data assimilation method according to claim 1, characterized in that: The method also includes a random flipping data augmentation method.

7. The dual polarization radar data assimilation method according to claim 6, characterized in that: When constructing the UNet hydrometeor inversion model, the flip probability of the data enhancement method using random flipping is 50%.

8. The dual polarization radar data assimilation method according to any one of claims 1 to 7, characterized in that: The method further comprises: The dual-polarization radar data is preprocessed to remove noise and outliers in the dual-polarization radar data.

Images