Large-scale biological migration flight prediction method based on dual-polarization weather radar

Abstract

The invention relates to the technical field of weather radars, and discloses a large-scale biological migration flight prediction method based on a dual-polarization weather radar, and the method comprises the following steps: extracting polarization parameters of dual-polarization weather radar data; the radar data is converted into a unified 512 * 512 rectangular coordinate data format by adopting a method for converting polar coordinates into rectangular coordinates, and the spatial resolution is reconstructed; a biological echo classification model is constructed, and biological echo and meteorological echo classification is carried out on the polarization parameter data; based on the classified biological echo data, constructing a space-time convolutional neural network prediction model, and performing biological migration short-time proximity prediction; and outputting biological migration dynamic images of multiple times in the future, and providing a migration trend prediction result. Through the dual-polarization weather radar, multi-source data fusion and the generative adversarial network, high-precision monitoring and dynamic trend prediction of migratory organisms are realized, and data continuity and prediction reliability are ensured.

Description

Technical Field

[0001] The invention relates to the technical field of weather radars, and in particular to a large-scale biological migration prediction method based on dual-polarization weather radar. Background Art

[0002] Biological migration refers to the large-scale migration of organisms such as insects and birds driven by seasonal changes or environmental factors. Accurately predicting the dynamic distribution and trends of migratory organisms is of great significance for agricultural pest control, ecological protection and flight safety.

[0003] Current large-scale biological migration prediction methods mostly rely on meteorological radar and single-band insect radar, which extract the echo characteristics of the target (such as echo intensity, speed, etc.) and combine machine learning algorithms or physical modeling to predict biological migration trends. For example, the method based on recurrent neural network (RNN) can capture the temporal dynamic characteristics of migrating organisms and achieve short-term predictions in combination with historical data; while the method based on radar echo classification technology has high accuracy in distinguishing between migratory targets and meteorological targets. These methods can achieve effective monitoring of biological migration behavior in specific application scenarios, providing technical support for agricultural and forestry pest control and ecological monitoring.

[0004] However, existing methods have the following limitations: First, single-band radar or traditional meteorological radar is unable to distinguish between migratory organisms and complex meteorological targets (such as heavy rainfall), resulting in limited prediction accuracy; second, the lack of multi-source data fusion means makes it difficult to fully describe the dynamic environmental characteristics of the migration area, and the prediction effect is reduced in scenarios with heavy rainfall or missing data; third, the blank areas in the input data are not fully filled, resulting in insufficient spatiotemporal continuity, which in turn affects the prediction ability and reliability of the model. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a large-scale biological migration prediction method based on dual-polarization weather radar, which solves the problems existing in the existing technology of low accuracy in distinguishing migratory organisms from meteorological targets, incomplete dynamic description of migration areas due to lack of fusion of multi-source data, and insufficient spatiotemporal continuity due to failure to fill in blank areas of radar data.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a large-scale biological migration prediction method based on dual-polarization weather radar, comprising the following steps: Extract polarization parameters from dual-polarization weather radar data; The radar data is converted into a unified 512×512 rectangular coordinate data format using the polar coordinate to rectangular coordinate conversion method, and the spatial resolution is reconstructed; Construct a biological echo classification model to classify biological echoes and meteorological echoes based on polarization parameter data; Based on the classified biological echo data, a spatiotemporal convolutional neural network prediction model is constructed to perform short-term prediction of biological migration; Output dynamic images of biological migration at multiple time periods in the future and provide migration trend prediction results.

[0007] Preferably, the extracted polarization parameters include: Reflectivity Factor , used to describe the scattering intensity of the echo; Differential reflectivity , used to describe echo shape and directionality; Differential Phase , used to analyze the phase characteristics of the target scatterer; Correlation coefficient , used to evaluate the similarity of horizontal and vertical polarization echoes; Depolarization ratio , used to improve the distinction between biological echoes and meteorological echoes; Spectral width , which is used to characterize the distribution range of target motion speed and serves as an auxiliary parameter for the dynamic characteristics of biological targets.

[0008] Preferably, the depolarization ratio The calculation methods include: Perform initial calculations based on differential reflectance and correlation coefficients; Introducing the weight factor of differential phase To optimize, the formula includes: ; in, is the differential reflectivity, is the correlation coefficient, ranging from 0≤ ≤1.

[0009] Preferably, the polar coordinate data conversion includes: Use the following formula to convert polar coordinate data to rectangular coordinate data: ; ; in, is the distance library, is the azimuth, is the elevation angle; The converted data are reconstructed into spatial resolution by bilinear interpolation to ensure data spatial consistency and standardization.

[0010] Preferably, the constructed bio-echo classification model includes: After extracting polarization parameters, the XGBoost algorithm is used to classify biological echoes and meteorological echoes; Use convolutional neural networks to extract spatial texture features; The output results of XGBoost and convolutional neural network are fused based on the weighted average method to improve the classification accuracy.

[0011] Preferably, the specific calculation formula of the weighted average method is: ; in, is the weight coefficient, and the optimal value is determined by cross-validation; : represents the final classification result after weighted fusion, which is used to determine the probability of the input data being a biological echo or a meteorological echo; : represents the output result of the XGBoost classification model, which is the probability value of the biological echo category; : represents the output result of the enhanced convolutional neural network, which is the probability value of the biological echo category; : Represents the weight coefficient of CNN classification results, and Complementarity is used to balance the contributions of the two classification methods.

[0012] Preferably, the constructed spatiotemporal convolutional neural network model includes: Input radar echo images of multiple times in the past as time-space series data; Use a 5-layer spatiotemporal convolutional neural network module to perform spatiotemporal dynamic modeling and predict radar echo images at multiple times in the future; The state update of a spatiotemporal convolutional neural network unit consists of the following steps: ; in, , , are the states of the forget gate, input gate, and output gate respectively. is the unit state, In hidden state.

[0013] Preferably, the spatiotemporal convolutional neural network model further introduces an attention mechanism, which includes: ; ; in, and is the learnable weight matrix, is the bias, Score for attention, is the normalized attention weight.

[0014] Preferably, the migration trend prediction result is optimized by fusion of multi-source data, and the multi-source data includes: Wind field data: Extract wind speed and direction information and combine it with dual-polarization weather radar data to correct the spatial path of migration trends; Satellite remote sensing data: Use high-resolution satellite images to fill in blank areas in radar echoes caused by precipitation elimination, and enhance the description of environmental characteristics of the migration area.

[0015] Preferably, the method further comprises: For the blank areas in the input data, a generative adversarial network is used to generate highly credible simulated data to complete the missing data.

[0016] The present invention provides a large-scale biological migration prediction method based on dual-polarization weather radar. It has the following beneficial effects: 1. The present invention realizes accurate monitoring and dynamic trend prediction of migratory organisms by combining XGBoost classification model, spatiotemporal convolutional neural network and attention mechanism. The system can capture the characteristic changes of migratory organisms in time and space, meet high-precision requirements, and is suitable for real-time ecological monitoring of large-scale areas.

[0017] 2. The present invention corrects the migration path through wind field data and uses satellite remote sensing data to fill in the radar blank area caused by heavy rainfall, further optimizing the migration trend prediction. The synergistic effect of multi-source data makes up for the limitations of a single data source, ensuring a complete description of the migration area and the environmental adaptability of the path prediction, and showing higher reliability and accuracy under complex weather conditions.

[0018] 3. The present invention introduces a generative adversarial network to generate high-confidence simulated data for blank areas in radar data and restore the spatial continuity of the input data. The generated data has physical properties consistent with real radar echoes, which enhances the robustness and applicability of trend prediction and can provide high-quality monitoring results even in heavy rainfall and data missing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Please see attached Figure 1 The embodiment of the present invention provides a large-scale biological migration prediction method based on dual-polarization weather radar, comprising the following steps: Extracting polarization parameters from dual-polarization weather radar data can fully capture the characteristics of migratory organisms, effectively distinguish biological echoes from meteorological echoes, and provide high-quality input data for subsequent classification and prediction; The method of polar coordinate to rectangular coordinate conversion is adopted to convert radar data into a unified 512×512 rectangular coordinate data format, and the spatial resolution is reconstructed. By unifying the converted rectangular coordinate data format and spatial resolution, the present invention solves the resolution inconsistency problem caused by the original format of radar data, ensures the uniformity and effectiveness of subsequent model input, and thus improves the accuracy and stability of prediction; A biological echo classification model is constructed to classify biological echoes and meteorological echoes based on polarization parameter data. Through multi-model fusion classification, the present invention can effectively improve the accuracy of biological echo and meteorological echo classification, while reducing noise and misclassification risks, and provide high-precision classification results for subsequent biological migration dynamics prediction; Based on the classified biological echo data, a spatiotemporal convolutional neural network (ConvLSTM) prediction model is constructed to perform short-term prediction of biological migration. The spatiotemporal convolutional neural network can capture the dynamic characteristics of migrating organisms in time and space, and achieve accurate prediction of migration trends in the next 1 to 2 hours. The root mean square error is less than 0.035, which meets the actual application needs. The invention outputs dynamic images of biological migration at multiple times in the future and provides migration trend prediction results. By outputting dynamic migration results, the invention provides real-time monitoring and early warning capabilities of biological migration, and can provide technical support for agricultural pest control, ecological protection and biological migration research.

[0022] Please see attached Figure 1 In a preferred embodiment of the present invention, the extracted polarization parameters include: Reflectivity Factor , used to describe the scattering intensity of the echo. The specific extraction process includes: Echo acquisition: Extract horizontal polarization echo intensity from radar base data , and calculate its distribution based on the distance library and azimuth; Data cleaning: remove Because the signal in this range is usually noise or invalid data, the reflectivity factor It provides basic features for identifying biological targets. Its value range can distinguish low-intensity biological echoes from high-intensity precipitation echoes, providing an important basis for subsequent classification; differential reflectivity , used to describe the echo shape and directivity, through the differential reflectivity , can accurately describe the shape characteristics of migratory biological targets, and can be used in combination with the reflectivity factor to effectively distinguish migratory biological targets from precipitation targets; Differential Phase , used to analyze the phase characteristics of the target scatterer, differential phase It provides unique information for identifying the internal scattering characteristics of the target, and can further improve the accuracy of target differentiation when combined with other polarization parameters; Correlation coefficient , used to evaluate the similarity between horizontal and vertical polarization echoes, using the correlation coefficient , can clearly distinguish between regularly shaped precipitation targets and irregularly shaped migratory organisms, improving the reliability of target recognition; Depolarization ratio , used to improve the distinction between biological echoes and meteorological echoes, depolarization ratio By combining information from multiple polarization parameters, the ability to distinguish between biological targets and meteorological targets is further enhanced, especially in complex weather conditions (such as areas where precipitation and migratory organisms intersect). Spectral width , used to characterize the distribution range of target motion speed, as an auxiliary parameter of the dynamic characteristics of biological targets, spectral width It provides a quantitative description of the dynamic characteristics of target motion, which helps to further distinguish migratory organisms from static targets (such as ground clutter or precipitation) and enhance the performance of classification models.

[0023] Please see attached Figure 1 In a preferred embodiment of the present invention, the depolarization ratio The calculation methods include: The initial calculation is based on the differential reflectivity and correlation coefficient, and the initial calculation formula is: ; Through the preliminary combination of differential reflectance and correlation coefficient, the depolarization ratio can effectively capture the irregular shape and consistency characteristics of migratory biological targets, providing basic calculations for further optimization; Introducing the weight factor of differential phase To optimize, the formula includes: ; in, is the differential reflectivity, is the correlation coefficient, ranging from 0≤ ≤1; Optimization process: Calculate the differential phase change : ; in, Represents a time point, and quantifies the dynamic change characteristics of the target by taking the absolute value of the differential phase of adjacent time points. By introducing the differential phase and weight factor The calculation formula of the depolarization ratio is more flexible and can adapt to the distinction between biological echoes and meteorological echoes in complex scenes, especially in areas with high precipitation intensity, which can effectively improve the classification performance.

[0024] Please see attached Figure 1 In a preferred embodiment of the present invention, polar coordinate data conversion includes: Use the following formula to convert polar coordinate data to rectangular coordinate data: ; ; in, is the distance library, is the azimuth, The elevation angle can be accurately mapped to the rectangular coordinate system through formula conversion, providing a consistent spatial basis for subsequent data reconstruction and analysis, and avoiding the problem of polar coordinate resolution decreasing with increasing distance; The converted data are reconstructed into spatial resolution through bilinear interpolation to ensure data spatial consistency and standardization; The converted data are reconstructed into spatial resolution through bilinear interpolation to ensure data spatial consistency and standardization; Interpolation formula: For any point , whose value is obtained by passing through the values ​​of the four adjacent grid points The calculation results are: ; ; By reconstructing and standardizing the data resolution through bilinear interpolation, the spatial consistency of the output data can be ensured, providing a unified input format for subsequent deep learning model training, thereby improving the accuracy of classification and prediction.

[0025] Please see attached Figure 1 In a preferred embodiment of the present invention, the constructed bio-echo classification model includes: After extracting the polarization parameters, the XGBoost algorithm is used to classify biological echoes and meteorological echoes. The specific implementation process is as follows: Input features: Select the extracted polarization parameters, including reflectivity factor , differential reflectivity , differential phase , correlation coefficient , depolarization ratio and spectral width , forming the feature vector: ; The historical data collected by the radar are annotated to generate training sample sets of biological echoes and meteorological echoes. The feature distribution is as follows: Bioecho: The eigenvalues ​​usually have smaller , lower and larger ; Weather echo: The characteristic value usually has a higher , higher and smaller ; The gradient boosting decision tree is constructed by the XGBoost algorithm. The objective function is to minimize the classification error. The loss function is defined as: ; in, and are the true value and the predicted value respectively, It is a regularization term used to prevent overfitting. XGBoost's ability to process multi-dimensional features enables the classification model to efficiently capture the characteristic differences between biological echoes and meteorological echoes. The classification results have high accuracy and high computational efficiency, which is suitable for large-scale radar data processing scenarios. Use convolutional neural networks (CNNs) to extract spatial texture features. CNNs automatically extract spatial texture features of radar echoes through deep learning, which can capture subtle differences in the spatial distribution of biological echoes and improve the overall performance of the classification model. Based on the weighted average method, the output results of XGBoost and convolutional neural network are fused to improve the classification accuracy. The classification probabilities of XGBoost and CNN are weighted averaged. The final classification probability calculation formula after fusion is: ; in: and These are the classification probabilities of biological echoes by XGBoost and CNN respectively; is the fusion weight, with an initial value of 0.5, which is optimized and adjusted through cross-validation; by weighted fusion of the output results of XGBoost and CNN, the present invention greatly improves the accuracy of biological echo classification and enhances the robustness of the model in complex scenarios.

[0026] Please see attached Figure 1 In a preferred embodiment of the present invention, the specific calculation formula of the weighted average method is: ; in, is the weight coefficient, and the optimal value is determined by cross-validation; : represents the final classification result after weighted fusion, which is used to determine the probability of the input data being a biological echo or a meteorological echo; : represents the output result of the XGBoost classification model, which is the probability value of the biological echo category; : represents the output result of the enhanced convolutional neural network, which is the probability value of the biological echo category; : Represents the weight coefficient of CNN classification results, and Complementarity, used to balance the contributions of the two classification methods; Through cross-validation, the weight distribution of XGBoost and CNN can be dynamically adjusted so that the fusion model can achieve the best classification effect in different feature scenarios. ,The fusion model has the ability to adapt to scenes, which greatly improves the flexibility and robustness of classification.

[0027] Please see attached Figure 1 In a preferred embodiment of the present invention, the constructed spatiotemporal convolutional neural network (ConvLSTM) model includes: Input radar echo images from multiple times in the past as spatiotemporal series data. By inputting radar echo images from multiple times in the past, the model can capture the temporal dynamic characteristics and spatial variation laws of biological migration, providing high-quality input data for subsequent spatiotemporal modeling. A five-layer ConvLSTM module is used for spatiotemporal dynamic modeling to predict radar echo images at multiple times in the future. By stacking five layers of ConvLSTM modules, the model can fully capture the dynamic characteristics of migratory organisms in time and space, thereby improving the prediction accuracy of future echo images. The state update of the ConvLSTM unit includes the following steps: ; in, , , are the states of the forget gate, input gate, and output gate respectively. is the unit state, It is a hidden state. Through the state update of the forget gate, input gate and output gate, the ConvLSTM unit can flexibly capture the changes in the temporal and spatial characteristics of migratory organisms, ensuring the efficiency and accuracy of the model in processing spatiotemporal series data. By predicting the radar echo images for multiple time periods in the future, the model realizes the short-term and near-term prediction of migratory organisms, providing real-time early warning capabilities for agricultural prevention and control and ecological protection.

[0028] Please see attached Figure 1 In a preferred embodiment of the present invention, the ConvLSTM model further introduces an attention mechanism, which includes: ; ; in, and is the learnable weight matrix, is the bias, Score for attention, The attention weights are normalized. By introducing the attention mechanism, the model can automatically allocate computing resources, give priority to the spatiotemporal regions that contribute more to the prediction of biological migration, and significantly improve the model's feature extraction capabilities and prediction accuracy. The attention mechanism improves the robustness and adaptability of the model, ensuring the reliability and accuracy of the migration trend prediction results. Through the expansion of the attention mechanism, the model can maintain high performance in more diverse and complex scenarios, providing more extensive technical support for the monitoring and prediction of migratory organisms.

[0029] Please see attached Figure 1 In a preferred embodiment of the present invention, the migration trend prediction result is optimized by fusion of multi-source data, and the multi-source data includes: Wind field data: Extract wind speed and direction information and combine it with dual-polarization weather radar data to correct the spatial path of migration trends; Satellite remote sensing data: Use high-resolution satellite images to fill in the blank areas in radar echoes caused by precipitation elimination, and enhance the description of environmental characteristics of the migration area; The fusion formula includes: ; in, , and These are the contributions of radar data, wind field data and satellite remote sensing data to migration trends; , and is the fusion weight, satisfying + + =1, determined through experimental tuning, through multi-source data fusion, the present invention can effectively make up for the limitations of a single data source and improve the accuracy and stability of migration trend prediction results; The introduction of wind field data effectively captures the impact of external meteorological factors on biological migration behavior, significantly improving the accuracy and credibility of migration route predictions; the introduction of satellite remote sensing data fills the gaps in radar data in areas of heavy precipitation, and enhances the description of environmental characteristics of migration areas, providing key data support for more accurate trend predictions.

[0030] Please see attached Figure 1In a preferred embodiment of the present invention, the method further comprises: For the blank areas in the input data, high-confidence simulated data is generated through generative adversarial networks to fill in the missing data. The sources of the blank areas include: in heavy rainfall areas, since it is difficult to distinguish the polarization characteristics of the echo from the biological echo, the signal in this area is usually removed, resulting in a large area of ​​blank data; in the close range of radar detection, the data is unavailable due to the interference of strong echo clutter; due to radar hardware or network transmission failures, some data is lost at certain times. By filling in the data in the blank areas, the spatial continuity of the input data can be restored, providing complete and consistent input for subsequent models, and improving the accuracy of migration trend prediction; Through adversarial training of the generator and the discriminator, the GAN model makes the generated simulated data have characteristics that are highly consistent with the real radar echo data, ensuring the credibility and physical rationality of the completed data. Through adversarial training of the generator and the discriminator, the generated simulated data can highly restore the radar echo characteristics of the blank area, providing a high-credibility solution for input data completion.

[0031] In order to better understand the present invention, the above contents are described in detail below in conjunction with specific embodiments.

[0032] Example 1: Prediction of migratory organisms based on dual-polarization radar data and traditional classification methods In this embodiment, only dual-polarization weather radar data is used, and the biological echoes and meteorological echoes are classified by traditional machine learning classification methods (such as random forests or support vector machines); The specific steps include: Extract the basic polarization parameters (reflectivity factor, differential reflectivity, etc.) of dual-polarization radar and classify them; Classify radar data using random forest algorithm; No multi-source data or supplementation methods were introduced, and the biological migration trend was predicted based on single radar data. Effect: This method can classify biological echoes in simple scenarios, but the classification effect is poor under complex weather conditions. It cannot fill in blank areas in heavy rainfall areas, the prediction accuracy of migration trends is insufficient, and its applicability is relatively limited.

[0033] Example 2: Prediction of migratory organisms by introducing multi-source data fusion In this embodiment, in addition to dual-polarization weather radar data, wind field data and satellite remote sensing data are also introduced to perform multi-source data fusion; The specific steps include: Extract polarization parameters of dual-polarization radar and correct migration paths in combination with wind field data; Use satellite remote sensing images to fill in radar blank data in heavy rainfall areas; Use XGBoost algorithm combined with CNN to classify radar data, and perform weighted fusion of multi-source data to optimize migration trend prediction; Effect: This method enhances the dynamic environmental description of the migration area through multi-source data fusion and improves the classification and prediction accuracy under complex weather conditions, but it still does not solve the problem of insufficient input spatiotemporal continuity caused by missing data.

[0034] Example 3: Prediction of migratory organisms based on generative adversarial networks (GANs) In this embodiment, a generative adversarial network (GAN) is further introduced on the basis of Embodiment 2 to complete the blank areas of radar data with high confidence; The specific steps include: Use the GAN generator to generate simulated data consistent with real radar echoes to fill in blank areas caused by heavy precipitation or missing data; Integrate the completed radar data with multi-source data (wind field and satellite remote sensing) to further optimize the classification and trend prediction models; The ConvLSTM model with attention mechanism is used to dynamically capture key spatiotemporal features and accurately predict migration trends. Effect: This embodiment uses GAN to complete the blank areas of radar data, improves the continuity of data input and the robustness of the model, ensures the high-precision performance of migration trend prediction in heavy rainfall or data missing scenarios, and provides comprehensive technical support for large-scale migratory biological monitoring.

[0035] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-scale biological migration prediction method based on dual-polarization weather radar, characterized in that: The following steps are involved: Extract polarization parameters from dual-polarization weather radar data; The radar data is converted into a unified 512×512 rectangular coordinate data format using the polar coordinate to rectangular coordinate conversion method, and the spatial resolution is reconstructed; Construct a biological echo classification model to classify biological echoes and meteorological echoes based on polarization parameter data; Based on the classified biological echo data, a spatiotemporal convolutional neural network prediction model is constructed to perform short-term prediction of biological migration; Output dynamic images of biological migration at multiple time periods in the future and provide migration trend prediction results.

2. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The extracted polarization parameters include: Reflectivity Factor , used to describe the scattering intensity of the echo; Differential reflectivity , used to describe echo shape and directionality; Differential Phase , used to analyze the phase characteristics of the target scatterer; Correlation coefficient , used to evaluate the similarity of horizontal and vertical polarization echoes; Depolarization ratio , used to improve the distinction between biological echoes and meteorological echoes; Spectral width , which is used to characterize the distribution range of target motion speed and serves as an auxiliary parameter for the dynamic characteristics of biological targets.

3. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 2 is characterized in that: The depolarization ratio The calculation methods include: Perform initial calculations based on differential reflectance and correlation coefficients; Introducing the weight factor of differential phase To optimize, the formula includes: ； in, is the differential reflectivity, is the correlation coefficient, ranging from 0≤ ≤1.

4. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The polar coordinate data conversion includes: Use the following formula to convert polar coordinate data to rectangular coordinate data: ； ； in, is the distance library, is the azimuth, is the elevation angle; The converted data are reconstructed into spatial resolution by bilinear interpolation to ensure data spatial consistency and standardization.

5. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The constructed biological echo classification model includes: After extracting polarization parameters, the XGBoost algorithm is used to classify biological echoes and meteorological echoes; Use convolutional neural networks to extract spatial texture features; The output results of XGBoost algorithm and convolutional neural network are fused based on weighted average method to improve the classification accuracy.

6. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 5 is characterized in that: The specific calculation formula of the weighted average method is: ； in, is the weight coefficient, and the optimal value is determined by cross-validation; : represents the final classification result after weighted fusion, which is used to determine the probability of the input data being a biological echo or a meteorological echo; : represents the output result of the XGBoost classification model, which is the probability value of the biological echo category; : represents the output result of the enhanced convolutional neural network, which is the probability value of the biological echo category; : Represents the weight coefficient of CNN classification results, and Complementarity is used to balance the contributions of the two classification methods.

7. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The constructed spatiotemporal convolutional neural network model includes: Input radar echo images of multiple times in the past as time-space series data; Use a 5-layer spatiotemporal convolutional neural network module to perform spatiotemporal dynamic modeling and predict radar echo images at multiple times in the future; The state update of a spatiotemporal convolutional neural network unit consists of the following steps: ； in, , , are the states of the forget gate, input gate, and output gate, respectively. is the unit state, In hidden state.

8. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 7 is characterized in that: The spatiotemporal convolutional neural network model further introduces an attention mechanism, which includes: ； ； in, and is the learnable weight matrix, is the bias, Score for attention, is the normalized attention weight.

9. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The migration trend prediction result is optimized by fusion of multi-source data, and the multi-source data includes: Wind field data: Extract wind speed and direction information and combine it with dual-polarization weather radar data to correct the spatial path of migration trends; Satellite remote sensing data: Use high-resolution satellite images to fill in blank areas in radar echoes caused by precipitation elimination, and enhance the description of environmental characteristics of the migration area.

10. The large-scale biological migration prediction method based on dual-polarization weather radar according to claim 1 is characterized in that: The method further comprises: For the blank areas in the input data, a generative adversarial network is used to generate highly credible simulated data to complete the missing data.

Images