Meteorological radar echo structure labeling method based on semi-supervised and feedback reinforcement learning

Abstract

The application relates to a weather radar echo structure labeling method based on semi-supervised and feedback reinforcement learning, and belongs to the field of meteorological data processing. The method comprises the following steps: data acquisition and preprocessing; data labeling and division; semi-supervised learning model design and training; uncertainty sample screening and expert feedback; human feedback reinforcement learning optimization; and final labeled data set generation. The method can improve the accuracy of labeled data and the robustness of the model by using an improved 3D convolutional neural network to extract multi-scale features, introducing a consistency regularization method and an average teacher method in semi-supervised training, screening uncertainty samples through confidence, introducing expert feedback and designing a reward function, optimizing the labeling accuracy of key categories, using the REINFORCE algorithm to realize multi-round interaction optimization, making the model gradually approach the expert standard, and generating a high-quality labeled data set, thereby providing reliable data support for extreme weather warning and meteorological research.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological data processing, and in particular relates to a meteorological radar echo structure labeling method based on semi-supervised and feedback reinforcement learning. Background Art

[0002] With the advancement of meteorological radar observation technology, high-temporal and spatial resolution radar data plays a key role in weather forecasting and extreme weather warnings. However, accurate annotation of radar echo structures is crucial for improving the predictive capabilities of meteorological models. Traditional manual annotation methods, while highly accurate, are inefficient and difficult to adapt to the demands of large-scale radar observation data. While automated annotation methods based on fully supervised learning reduce manual intervention, their reliance on large amounts of labeled data and inadequate utilization of unlabeled data limit their practical applications. To further improve annotation efficiency and quality, semi-supervised learning has recently attracted attention due to its efficient utilization of unlabeled data even with limited labeled data. However, current methods are mostly limited to processing static data and fail to fully incorporate the dynamic characteristics of meteorological radar data. Furthermore, human-feedback reinforcement learning (RLHF), a method that optimizes models by incorporating expert feedback, has not yet been widely used in meteorological data annotation, and its potential remains to be explored. In summary, existing technologies have significant shortcomings in annotation efficiency, robustness, and utilization of unlabeled data. A new method that fully combines semi-supervised learning with human-feedback reinforcement learning is urgently needed to meet the practical needs of meteorological radar echo structure annotation. Summary of the Invention

[0003] In order to solve the above problems, the present invention provides a weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning.

[0004] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0005] The present invention provides a weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning, comprising the following steps:

[0006] S1. Data acquisition and preprocessing: Acquire high temporal and spatial resolution observation data from the meteorological radar system and standardize the data;

[0007] S2. Data Labeling and Partitioning: Select samples from the standardized dataset for manual labeling to construct labeled and unlabeled datasets.

[0008] S3. Semi-supervised learning model design and training: Use an improved 3D convolutional neural network to extract multi-scale features and combine supervised and unsupervised loss optimization models;

[0009] S4. Uncertainty sample screening and expert feedback: Screen uncertainty samples according to model confidence, construct a to-be-reviewed set, submit to expert review, obtain annotation correction and feedback;

[0010] S5. Human feedback reinforcement learning optimization: Design a reward function to distinguish the importance of different echo types, and use the REINFORCE algorithm to optimize the model parameters;

[0011] S6. Final labeled dataset generation: Generate a high-quality labeled dataset for meteorological prediction model training set and extreme weather warning.

[0012] Further, step S1 specifically includes:

[0013] Obtain high spatiotemporal resolution observation data from a meteorological radar system, including single-polarization and dual-polarization radar parameters, to obtain an original dataset , which is expressed by the following formula:

[0014] ,

[0015] wherein denotes the total number of samples, each sample is a four-dimensional tensor, expressed as , wherein and denote the number of grid points in the longitude and latitude directions, respectively, denotes the number of vertical layers, denotes the number of radar parameter channels, and the radar parameters include reflectivity factor , radial velocity , differential reflectivity , and correlation coefficient ;

[0016] Standardize the feature values of each radar parameter channel, and let the mean and standard deviation of the th channel be and , respectively. Standardize each sample , which is expressed by the following formula:

[0017] ,

[0018] wherein denotes the value of the th channel of sample at position , , , , , to obtain the standardized dataset .

[0019] Furthermore, step S2 specifically includes:

[0020] From the standardized dataset Some samples are selected for manual labeling. Meteorological experts determine the echo structure types for these samples based on their professional knowledge. The echo structure types include precipitation echoes, convection echoes, stratiform echoes, and potential extreme weather related echoes. The labeled data set is obtained. , the formula is as follows:

[0021] ,

[0022] in, Indicates labeled samples The echo type tag, Indicates the number of labeled samples; the standardized dataset The remaining data is constructed as an unlabeled data set , the formula is as follows:

[0023] ,

[0024] in, represents unlabeled samples.

[0025] Furthermore, step S3 specifically includes:

[0026] S31. The semi-supervised learning model uses an improved 3D convolutional neural network (3D-CNN) as its basic architecture and introduces a consistency regularization method and a mean teacher method (MeanTeacher) in semi-supervised training.

[0027] The improved 3D convolutional neural network includes a multi-channel branch convolution layer, an intermediate fusion layer, a multi-scale 3D convolution and pooling layer, and an output layer; the sample in step S1 is The standardized four-dimensional tensor is input into the improved 3D convolutional neural network for category prediction to obtain the predicted probability of various echo structures. ,in, represents the student model parameters, Represents the model prediction sample Belong to category probability;

[0028] S32. For labeled data sets , using supervised loss For model training, the formula is as follows:

[0029] ,

[0030] in, Indicates the number of labeled samples, represents the echo category set, represents the student model parameters, Represents the model's response to the sample Predicted as class probability;

[0031] S33. For unlabeled data sets , introduced a semi-supervised learning strategy, and used the consistency regularization method and the average teacher method to complete the definition and optimization of the unlabeled loss, specifically:

[0032] S331. For unlabeled samples Enhanced samples obtained through data augmentation ,Data augmentation includes adding noise, random cropping and slight spatiotemporal perturbations;

[0033] S332. Using the teacher model in the Mean Teacher method right Make predictions and evaluate the student model exist The prediction requirements on is the student model parameter The exponential sliding average during training, the unsupervised loss is a measure of the distance between the teacher model and the student model prediction results, and the formula is as follows:

[0034] ,

[0035] in, Indicates the number of unlabeled samples;

[0036] S34. The loss function of the final semi-supervised learning model is expressed as follows:

[0037] ,

[0038] in, represents the balance coefficient, represents the student model parameters;

[0039] S35. Use the stochastic gradient descent SGD optimization algorithm to optimize and update the student model parameters in step S34.

[0040] Furthermore, step S4 specifically includes:

[0041] S41. Screening uncertain samples based on confidence threshold technique;

[0042] S411. For each sample , let student model output the predicted probability distribution of each class ;

[0043] S412. Set a fixed threshold , if the predicted probability of the most likely class of the student model is less than , collect all samples with classification confidence less than as the set of samples to be reviewed , and request feedback from experts for the samples in

[0044] S42. Expert annotation and correction: for samples in the set of samples to be reviewed , experts give annotated labels based on meteorological radar echo characteristics and professional knowledge .

[0045] Further, step S5 specifically includes:

[0046] S51. The reward function is expressed as follows:

[0047] ,

[0048] wherein represents the weight of the true class, represents the weight of the predicted class, represents the class predicted by the model, represents the annotated label corrected by expert feedback;

[0049] S52. Update the model parameters using the RLHF parameter update strategy for reinforcement learning

[0050] S521. Strategy representation and target: the RLHF parameter update strategy aims to maximize the expected reward, and the expected optimization target is expressed as follows:

[0051] ,

[0052] wherein represents the expected reward value of the sample ;

[0053] S522. Update by unbiased estimation of parameter gradient using the REINFORCE algorithm, and the update rule is:

[0054] ​ ,

[0055] wherein, represents a learning rate, represents a sample reward value, represents a policy gradient;

[0056] S53. Multiple rounds of interaction and gradual improvement

[0057] S531. Model annotates new samples: using parameters updated model to predict the remaining samples in the unlabeled data set , and the samples with a classification confidence lower than are added to the review set , and the review set is updated;

[0058] S532. Expert feedback iteration: re-annotating the newly added samples in the review set request expert annotation and correction again, and obtain the annotation label of the new sample;

[0059] S533. Based on the annotation label of the new sample, the parameters are updated again through steps S51 and S52; through multiple rounds of interaction and updating, the model converges to the expert standard in each round;

[0060] Further, step S6 specifically comprises:

[0061] When the model reaches a stable state after multiple rounds of RLHF correction, if the average reward improvement in consecutive rounds of iteration is lower than the threshold , the iteration is stopped, and the final annotated data set is output, which is used to train the weather prediction model or provide reliable data support for extreme weather warning and decision-making.

[0062] The advantages of the present application are:

[0063] This paper combines semi-supervised learning with human-feedback reinforcement learning (RLHF) to propose a weather radar echo structure annotation method based on semi-supervised and feedback-based reinforcement learning. Leveraging the multi-scale feature extraction capabilities of an improved 3D convolutional neural network, this method significantly improves the accuracy of labeled data and the robustness of the model through consistency regularization and a mean teacher framework. By filtering uncertain samples based on confidence, introducing expert feedback, and designing a reward function, the annotation accuracy of key categories is optimized. Furthermore, the REINFORCE algorithm is used to implement multiple rounds of interactive optimization, gradually approaching expert standards and generating high-quality annotated datasets. This method provides reliable data support for extreme weather warnings and meteorological research, possessing significant practical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0065] Figure 1 is a flow chart of the steps of the method of the present invention;

[0066] Figure 2 Graphs showing experimental results of the method of the present invention and other methods. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] Example 1

[0069] In this embodiment, Figure 1 As shown, the present invention provides a weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning, and the specific steps include:

[0070] S1. Data acquisition and preprocessing: Acquire high temporal and spatial resolution observation data from the meteorological radar system and standardize the data;

[0071] Specifically, high temporal and spatial resolution observation data are obtained from the weather radar system, including single-polarization and dual-polarization radar parameters, to obtain the original data set. , the formula is as follows:

[0072] ,

[0073] wherein, denotes the total number of samples, each sample is a four-dimensional tensor, denoted as , wherein and denote the number of grid points in the longitude and latitude directions, respectively, denotes the number of vertical layers, denotes the number of radar parameter channels, the radar parameters including reflectivity factor , radial velocity , differential reflectivity , and correlation coefficient ;

[0074] The eigenvalues of each radar parameter channel are standardized, and the mean and standard deviation of the i-th channel are denoted as and , respectively. Each sample is standardized, and the formula is as follows:

[0075] ,

[0076] wherein, is the i-th channel value of sample at position , , , , , , and a standardized data set is obtained.

[0077] S2. Data labeling and division: select samples from the standardized data set for manual labeling to construct a labeled data set and an unlabeled data set;

[0078] Specifically, some samples are selected from the standardized data set for manual labeling. Meteorological experts determine the echo structure type for these samples according to professional knowledge, and the echo structure type includes precipitation echo, convective echo, stratiform echo, and potential extreme weather related echo, to obtain a labeled data set after labeling, and the formula is as follows:

[0079] ,

[0080] wherein, denotes the echo type label of the labeled sample , and denotes the number of labeled samples; the remaining data in the standardized data set is constructed as an unlabeled data set​ , the formula is as follows:

[0081] ,

[0082] in, represents unlabeled samples.

[0083] S3. Semi-supervised learning model design and training: Use an improved 3D convolutional neural network to extract multi-scale features and combine supervised and unsupervised loss optimization models;

[0084] Specifically, S31. The semi-supervised learning model uses an improved 3D convolutional neural network (3D-CNN) as a basic architecture, and introduces a consistency regularization method and a mean teacher method in semi-supervised training;

[0085] The improved 3D convolutional neural network includes a multi-channel branch convolution layer, an intermediate fusion layer, a multi-scale 3D convolution and pooling layer, and an output layer. First, a multi-channel branch convolution layer is used to perform preliminary feature extraction on different parameter channels, with each branch responsible for a different parameter combination, thereby retaining the heterogeneous characteristics of multiple radar parameters. Then, the branch feature maps are fused and convolved in the intermediate fusion layer, enabling the model to obtain multi-parameter joint representation capabilities. Furthermore, in the deep feature extraction stage, multi-scale 3D convolution and pooling operations are used to extract local fine-grained features and large-scale structural features, respectively, enabling the model to capture local features such as strong convective cores and analyze the distribution of large-scale features such as layered cloud systems.

[0086] The step S1 sample The standardized four-dimensional tensor is input into the improved 3D convolutional neural network for category prediction to obtain the predicted probability of various echo structures. ,in, represents the student model parameters, Represents the model prediction sample Belong to category probability;

[0087] S32. For labeled data sets , using supervised loss For model training, the formula is as follows:

[0088] ,

[0089] in, Indicates the number of labeled samples, represents the echo category set, represents the student model parameters, Represents the model's response to the sample Predicted as class probability;

[0090] S33. For unlabeled data sets , introduced a semi-supervised learning strategy, and used the consistency regularization method and the average teacher method to complete the definition and optimization of the unlabeled loss, specifically:

[0091] S331. For unlabeled samples Enhanced samples obtained through data augmentation ,Data augmentation includes adding noise, random cropping and slight spatiotemporal perturbations;

[0092] S332. Using the teacher model in the Mean Teacher method right Make predictions and evaluate the student model exist The prediction requirements on the above are consistent, and it is hoped that the prediction results of the student model are as similar as possible to the prediction results of the teacher model. The teacher model parameters is the student model parameter The exponential sliding average during training ensures that the teacher prediction is more stable and robust. The unsupervised loss is a measure of the distance between the teacher model and the student model prediction results. The formula is as follows:

[0093] ,

[0094] in, Indicates the number of unlabeled samples;

[0095] S34. By minimizing the unsupervised loss, the student model’s predictions on the augmented data are kept consistent with the teacher model’s predictions on the original data, improving the model’s ability to discriminate against unlabeled data. The loss function of the final semi-supervised learning model is expressed as follows:

[0096] ,

[0097] in, represents the student model parameters, Represents the balance coefficient, which is used to control the relative influence of labeled and unlabeled losses on the training process. and optimization strategies, so that the model can take into account the characteristics of unlabeled data under the guidance of labeled data, so that the echo structure characteristics can still be effectively learned when the amount of labeled data is limited;

[0098] S35. Use the stochastic gradient descent SGD optimization algorithm to optimize and update the student model parameters in step S34. The optimization process is implemented as follows: In each iteration, the labeled data With unlabeled data Extract small batches of samples and calculate supervised loss on labeled data , using the teacher-student model structure and Consistency Regularization to calculate the unsupervised loss for unlabeled data , then according to

[0099] ,

[0100] right Perform gradient descent update:

[0101] ,

[0102] in, is the learning rate; when updating the student model parameters Then, the EMA strategy is used to adjust the teacher model parameters Update accordingly:

[0103] ,

[0104] in, It is a smoothing factor, which is usually set to 0.999 to ensure that the changes in the teacher model parameters are slow and stable.

[0105] S4. Uncertainty sample screening and expert feedback: Screen uncertainty samples based on model confidence, build a review set, submit to experts for review, and obtain annotation corrections and feedback;

[0106] Specifically, S41. screening uncertain samples based on confidence threshold technology;

[0107] S411. For each sample , let the students model Output the predicted probability distribution of various echo structures ;

[0108] S412. Set fixed threshold =0.6, if the student model predicts the probability of the most likely category , all classification confidences lower than The sample set is recorded as the review set ,right The middle sample requested feedback from the experts;

[0109] S42. Expert annotation and correction: For the set to be reviewed samples in the training set , the expert gives the labeled label based on the meteorological radar echo characteristics and professional knowledge .

[0110] S5. Human feedback reinforcement learning optimization: design a reward function to distinguish the importance of different echo types, and optimize the model parameters using the REINFORCE algorithm;

[0111] Specifically, S51. The reward function is expressed as follows:

[0112] ,

[0113] wherein, represents the true category weight, represents the predicted category weight, represents the model predicted category, represents the expert feedback corrected labeled label;

[0114] S52. Update the model parameters using the RLHF parameter update strategy

[0115] S521. Strategy representation and target: the RLHF parameter update strategy aims to maximize the expected reward, and the expected optimization target is expressed as follows:

[0116] ,

[0117] wherein, represents the expected reward value of the sample ;

[0118] S522. Update by unbiased estimation of parameter gradient using the REINFORCE algorithm, and the update rule is:

[0119] ,

[0120] wherein, represents the learning rate, represents the reward value of the sample , represents the policy gradient; the present application adopts the REINFORCE algorithm as the reinforcement learning optimization method. The REINFORCE algorithm is a classic policy gradient method that directly updates the parameters of the policy, and has the characteristics of simplicity and directness, which is suitable for the needs of the present application.

[0121] S53. Multi-round interaction and gradual improvement

[0122] S531. Model labeling new samples: use the parameters The updated model predicts the remaining samples in the updated set of unannotated data The samples with a classification confidence lower than are added to the set of samples to be reviewed The set of samples to be reviewed is updated ;

[0123] S532. Expert feedback iteration: the newly added samples in the set of samples to be reviewed are again requested for expert annotation and correction, and the annotation labels of the new samples are obtained;

[0124] S533. Based on the annotation labels of the new samples, the parameters are updated again through steps S51 and S52; through multiple rounds of interaction and updating, the model converges to the expert standard in each round, reducing the occurrence of incorrect annotation.

[0125] S6. Final annotated data set generation: generate a high-quality annotated data set for meteorological prediction model training set and extreme weather warning.

[0126] Specifically, when the model reaches a stable state after multiple rounds of RLHF correction, if the average reward improvement in consecutive rounds of iteration is lower than the threshold , the iteration is stopped, and the final annotated data set is output, which is used to train the meteorological prediction model or provide reliable data support for extreme weather warning and decision-making.

[0127] Embodiment 2

[0128] In this embodiment, as shown in Figure 2 , in order to verify the effectiveness of the present application, a small amount of annotated data is combined with a large amount of unannotated data under different annotation ratios, and the following three methods are compared: (1) only using annotated data (traditional supervised learning); (2) using traditional semi-supervised learning to annotate unannotated data; (3) using the present application to annotate unannotated data, and training an SVM model to classify meteorological radar observation data echo structure. The results show that under the same annotation ratio, traditional semi-supervised learning can improve the accuracy to a certain extent compared with only supervised learning, while the present application fuses semi-supervised learning and expert feedback reinforcement, and the accuracy is significantly improved, fully proving the effectiveness and advancement of the present application in utilizing a small amount of annotated data, deeply mining the value of unannotated data, and integrating expert feedback to improve the annotation quality of meteorological radar echo structure.

[0129] ​Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that the technical solutions described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalent ones. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: Acquire high temporal and spatial resolution observation data from the meteorological radar system and standardize the data; S2. Data Labeling and Partitioning: Select samples from the standardized dataset for manual labeling to construct labeled and unlabeled datasets. S3. Semi-supervised learning model design and training: Use an improved 3D convolutional neural network to extract multi-scale features and combine supervised and unsupervised loss optimization models. The specific steps are as follows: S31. The semi-supervised learning model uses an improved 3D convolutional neural network (3D-CNN) as its basic architecture and introduces a consistency regularization method and a mean teacher method (MeanTeacher) in semi-supervised training. The improved 3D convolutional neural network includes a multi-channel branch convolution layer, an intermediate fusion layer, a multi-scale 3D convolution and pooling layer, and an output layer; the sample in step S1 is The standardized four-dimensional tensor is input into the improved 3D convolutional neural network for category prediction to obtain the predicted probability of various echo structures. ,in, represents the student model parameters, Represents the model prediction sample Belong to category probability; S32. For labeled data sets , using supervised loss For model training, the formula is as follows: ， in, Represents a labeled data set, Indicates the number of labeled samples, Represents a sample The echo type tag, represents the echo category set, represents the student model parameters, Represents the model for the sample Predicted as class probability; S33. For unlabeled data sets , introduced a semi-supervised learning strategy, and used the consistency regularization method and the average teacher method to complete the definition and optimization of the unlabeled loss, specifically: S331. For unlabeled samples Enhanced samples obtained through data augmentation ,Data augmentation includes adding noise, random cropping and slight spatiotemporal perturbations; S332. Using the teacher model in the Mean Teacher method right Make predictions and evaluate the student model exist The prediction requirements on is the student model parameter Exponential moving average during training, unsupervised loss It is the distance metric between the prediction results of the teacher model and the student model. The formula is as follows: ， in, represents an unlabeled data set, Indicates the number of unlabeled samples; S34. Loss function of the final semi-supervised learning model , the formula is as follows: ， in, represents the balance coefficient; S35. Optimize and update the student model parameters in step S34 using the stochastic gradient descent SGD optimization algorithm; S4. Uncertainty sample screening and expert feedback: Screen uncertainty samples based on model confidence, build a review set, submit to experts for review, and obtain annotation corrections and feedback; S5. Human Feedback Reinforcement Learning Optimization: Design a reward function to distinguish the importance of different echo types and use the REINFORCE algorithm to optimize model parameters; S6. Final annotated dataset generation: Generate high-quality annotated datasets for use in meteorological forecast model training and extreme weather warnings.

2. The weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning according to claim 1 is characterized in that: Step S1 specifically includes: Acquire high temporal and spatial resolution observation data from weather radar systems, including single-polarization and dual-polarization radar parameters, to obtain raw data sets , the formula is as follows: ， in, Indicates the total number of samples, each sample is a four-dimensional tensor, represented as ,in and Represents the number of grid points in longitude and latitude respectively, Indicates the number of vertical layers, Indicates the number of radar parameter channels, including reflectivity factor , radial velocity , differential reflectivity and the correlation coefficient ; The characteristic value of each radar parameter channel is standardized. The mean and standard deviation of each channel are and , for each sample For standardization, the formula is as follows: ， in, For samples In position The first Channel value, , , , , get the standardized data set .

3. The weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning according to claim 2 is characterized in that: Step S2 specifically includes: From the standardized dataset Some samples are selected for manual labeling. Meteorological experts determine the echo structure types for these samples based on their professional knowledge. The echo structure types include precipitation echoes, convection echoes, stratiform echoes, and potential extreme weather related echoes. The labeled data set is obtained. , the formula is as follows: ， The standardized dataset The remaining data is constructed as an unlabeled data set , the formula is as follows: ， in, represents unlabeled samples.

4. The weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning according to claim 3 is characterized in that: Step S4 specifically includes: S41. Screening of uncertain samples based on confidence threshold technology; S411. For each sample , let the students model Output the predicted probability distribution of various echo structures ; S412. Set fixed threshold , if the predicted probability of the most likely category of the student model is , all classification confidences lower than The sample set is recorded as the review set ,right The middle sample requested feedback from the experts; S42. Expert annotation and correction: For the set to be reviewed Samples in Experts give annotation labels based on weather radar echo characteristics and professional knowledge .

5. The weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning according to claim 4 is characterized in that: Step S5 specifically includes: S51. The reward function The formula is as follows: ， in, represents the true category weight, represents the predicted category weight, represents the category predicted by the model, Annotation labels indicating revisions to expert feedback; S52. Using RLHF parameter update strategy to update model parameters Perform reinforcement learning updates S521. Strategy representation and goal: The RLHF parameter update strategy goal is to maximize the expected reward, the expected optimization goal The formula is as follows: ， in, Represents a sample Expected reward value; S522. Use the REINFORCE algorithm to update the unbiased estimate of the parameter gradient. The update rule is: ， in, represents the learning rate, Represents a sample The reward value, represents the policy gradient; S53. Multiple rounds of interaction and gradual improvement S531. Model annotation of new samples: using parameters The updated model performs well on the unlabeled dataset. The remaining samples in the prediction are classified as The samples are added to the set to be reviewed Update the set to be reviewed ; S532. Expert Feedback Iteration: Treating the Review Set The newly added samples are again requested to be annotated and corrected by experts to obtain the annotation labels of the new samples; S533. Based on the label of the new sample, the parameters are adjusted again through steps S51 and S52. Update; through multiple rounds of interaction and update, the model converges to the expert standard in each round.

6. The weather radar echo structure annotation method based on semi-supervised and feedback reinforcement learning according to claim 5 is characterized in that: Step S6 specifically includes: When the model reaches a stable state after multiple rounds of RLHF correction, if In each round of iteration, the average reward increase is lower than the threshold , the iteration stops and the final labeled dataset is output. This dataset is used to train meteorological prediction models or provide reliable data support for extreme weather warnings and decision-making.

Images