X-band weather radar insect and bird identification method based on multi-scale information fusion

Through the insect and bird identification method of multi-scale information fusion, combined with pixel-level and image classification models, S-band radar data is used for joint monitoring, which solves the problem of insect and bird echo recognition confusion in X-band weather radar, improves the recognition accuracy and realizes the More accurate monitoring of insect and bird migration.

CN118918457APending Publication Date: 2024-11-08BEIJING INST OF TECH
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202410887733.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The X-band weather radar has a similar wavelength to the insect and bird scale, which leads to confusion in echo identification, and traditional methods are difficult to accurately monitor insect and bird migration.

Method used

Using insect and bird identification method based on multi-scale information fusion, the pixel-level classifier and image classification model are constructed, combined with point scale and global scale features, and S-band radar data is used for joint monitoring and sample data set construction, using Gaussian Bayes Classifiers and ResNet50 algorithms are used to identify and class probability correction is carried out to improve identification accuracy.

Benefits of technology

The accuracy of insect and bird recognition in X-band weather radar has been significantly improved, the misjudgment rate has been reduced, and more accurate insect and bird migration monitoring has been achieved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918457A_ABST
    Figure CN118918457A_ABST
Patent Text Reader

Abstract

The invention discloses an X-band weather radar insect and bird identification method based on multi-scale information fusion. According to the method, identification is carried out from point and local scales based on a pixel-level classifier, identification is carried out from global scales based on an image classification model, and then point and local scale information is corrected based on global scale information; through combination of point, local and global multi-scale information, insect and bird scattering features and spatial distribution features are fully utilized, and the accuracy of X-band weather radar insect and bird identification is greatly improved. Compared with a model before improvement, the method has the advantages that the recognition accuracy of the X-band weather radar is improved by 9.33%, and an effective means is provided for accurately realizing insect and bird migration monitoring of the X-band weather radar in China.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of weather radar and biometric identification technology, and in particular to an insect and bird identification method of an X-band weather radar based on multi-scale information fusion. Background Art

[0002] Migratory animals such as insects and birds have always migrated on a large scale and over long distances in my country to obtain sufficient food and water, thereby maintaining the health and reproduction of the population. my country has four migratory bird routes and two East Asian insect routes, making it the world's most important animal migration site. Animal migration is often accompanied by bird strikes, virus transmission, agricultural pests and diseases, and is closely related to human daily life. Therefore, it is urgent to develop an effective means to monitor the migration of insects and birds. However, traditional observation methods are difficult to achieve large-scale and effective monitoring of migratory animals. Weather radar has significant advantages such as large scale, high precision, and business operation. A large number of studies have confirmed that weather radar is an important tool for monitoring the migration of insects and birds, and the contribution of S-band weather radar in this field is particularly obvious. However, due to the influence of the earth's curvature and beam broadening effects, weather radar cannot achieve long-distance low-altitude monitoring. Although the detection range of X-band weather radar is much smaller than that of S-band, it is low in cost, and multiple X-band radars can be deployed to make up for the shortcomings of S-band low-altitude monitoring. Unlike S-band weather radar, the wavelength of X-band radar (about 3.2 cm) is close to the scale of insects and birds, and no longer meets the Rayleigh scattering requirements. At this time, the backscatter cross section (RCS) depends largely on the angle of illumination between the radar beam and the scatterer, and has little to do with the size of the scatterer itself. This means that even insects in groups can easily reach the RCS of birds, and the RCS of birds will become the same size as the RCS of insects due to inappropriate angles, which will lead to confusion in the identification of insect and bird echoes. Therefore, a separate insect and bird identification algorithm needs to be developed for X-band weather radar for aerial ecological monitoring. Summary of the invention

[0003] In view of this, the present invention provides an X-band weather radar insect and bird identification method based on multi-scale information fusion, which can effectively improve the accuracy of X-band weather radar insect and bird identification.

[0004] The X-band weather radar insect and bird identification method based on multi-scale information fusion of the present invention comprises:

[0005] Step 1, constructing a pixel-level classifier and an image classification model; wherein the input of the pixel-level classifier and the image classification model is the point-scale features of the X-band weather radar monitoring data, including reflectivity factor, radial velocity, spectral width, differential reflectivity, correlation coefficient and differential phase, and the standard deviation texture of the point-scale features;

[0006] The output of the pixel-level classifier is the identification category of each pixel in the X-band weather radar image; the output of the image classification model is the main identification category in the X-band weather radar image; the main identification category is the category with the highest probability of identification;

[0007] Step 2: Based on the sample data set, the pixel-level classifier and image classification model are trained;

[0008] Step 3, identifying the monitoring data of the X-band weather radar based on the pixel-level classifier and image classification model trained in step 2;

[0009] Step 4: Based on the output category of the image classification model in step 3, the probability of each identification category output by the pixel-level classifier in step 3 is corrected, specifically: the probability of the category that is the same as the output category of the image classification model is enhanced, and the probability of other categories is suppressed;

[0010] The output result of the corrected pixel-level classifier is the final X-band weather radar insect and bird identification result.

[0011] Preferably, in step 2, S-band weather radar and X-band weather radar are used for joint monitoring to establish a sample data set; wherein the monitoring range of the S-band radar completely covers the monitoring range of the X-band radar;

[0012] Insect and bird identification is performed on the monitoring data of the S-band weather radar; the S-band radar identification results are used as the classification labels for the X-band weather radar sample data.

[0013] Preferably, in the pixel-level classifier, the identification categories include: weather, insects and birds;

[0014] In the image classification model, the identification categories include: weather, insects, birds, weather + birds, weather + insects, and birds + insects.

[0015] Preferably, the pixel-level classifier is a Gaussian Bayes classifier, XGBoost, random forest or fuzzy logic; and the image classification model is ResNet50, VGG16, GoogLeNet or DenseNet.

[0016] Preferably, in step 4, for each identification category output by the pixel-level classifier in step 3, the probability of the same category as the category output by the image classification model is increased to the (1 / 4 to 1 / 2)th power of the original prediction probability, and the probability of other categories of the pixel-level classifier is reduced to the (2 to 4)th power of the original prediction probability.

[0017] Beneficial effects:

[0018] (1) The present invention uses a pixel-level classifier to identify at point and local scales, and uses an image classification model to identify at a global scale, and then corrects the point and local scale information based on the global scale information; by combining point, local, and global multi-scale information, the scattering characteristics and spatial distribution characteristics of insects and birds are fully utilized, greatly improving the accuracy of X-band weather radar insect and bird identification. Compared with the previous model, the X-band weather radar identification accuracy of the present invention is improved by 9.33%, providing an effective means for accurately realizing the migration monitoring of insects and birds by X-band weather radar in my country.

[0019] (2) By conducting joint observation experiments with S-band weather radar and X-band weather radar, we can make full use of the data from the widely deployed S-band weather radar and construct a sample set with authentic and reliable data and accurate classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the identification method of the present invention.

[0021] Figure 2 Schematic diagram of radar site locations for joint observation verification.

[0022] Figure 3 It is the characteristic distribution map of insect and bird echoes; among them, (a) is the point scale characteristic distribution map of polarization data; (b) is the local scale characteristic distribution map of polarization data.

[0023] Figure 4 Visualize the results of the ResNet50 training process.

[0024] Figure 5 This is a schematic diagram of the effect of the category enhancement module. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] The present invention provides an X-band weather radar insect and bird identification method based on multi-scale information fusion. The point and local scale information of X-band weather radar insect and bird echoes are classified based on a pixel-level classifier to obtain preliminary insect and bird echo identification results. At the same time, the global scale information of X-band weather radar data is judged based on an image classification model to obtain the main categories of echoes. Then, the point and local scale information are enhanced based on the global scale information to obtain the final result of insect and bird identification. The overall algorithm flow chart is as follows: Figure 1 shown.

[0027] Step 1: Build a sample data set.

[0028] In this embodiment, a joint monitoring test of S-band weather radar and X-band weather radar is carried out, requiring that the monitoring range of S-band radar can completely cover the monitoring range of X-band radar; the insect and bird echo results identified by S-band weather radar are used as classification labels, the point-scale information data of X-band radar is marked, and an X-band weather radar sample data set is constructed.

[0029] Considering that the radar beam scattering effect will cause a significant decrease in the reliability of long-distance monitoring data, this embodiment sets the S-band monitoring range to 100km and the X-band monitoring range to 35km. Combined with the existing weather radar station network layout, the Beijing area was selected for the experiment, where the S-band radar selected Beijing Daxing Station Z9010, and the X-band radar selected Fangshan Station ZA001, Changping Station ZA002 and Shunyi Station ZA003. Figure 2 Among them, X-band radars are all dual-polarization radars. In addition to the three traditional single-polarization variables of reflectivity factor, radial velocity, and spectral width, they can also provide dual-polarization variables such as differential reflectivity, correlation coefficient, and differential phase. These dual-polarization variables can reflect the horizontal and vertical differences of target scatterers in the sampling space, providing necessary information for distinguishing rainfall from organisms.

[0030] Step 2: Construct point-scale features of X-band weather radar monitoring data.

[0031] The point-scale information data of the X-band radar includes six features: reflectivity factor, radial velocity, spectral width, differential reflectivity, correlation coefficient, and differential phase. In addition, the present invention finds that the polarization characteristics of the rainfall echo, such as the reflectivity factor, correlation coefficient, and differential phase, are often more uniform in spatial distribution. However, due to factors such as different radar irradiation angles, different flight postures, and different volume sizes in space, migratory animals have large differences in the monitoring results of adjacent sampling units. This provides theoretical support for identifying insect and bird echoes from local scale spatial distribution. This means that texture features not only reflect the uniformity of the spatial distribution around the sampling space, but can also be used to enhance the feature dimension, helping the X-band weather radar to obtain more accurate insect and bird identification results. Therefore, the present invention also calculates the standard deviation texture of these six point-scale features to reflect the local scale information. The sample distribution of the final data set is shown in Table 1, and the distribution of different categories in point-scale features and local-scale features is shown in Table 1. Figure 3 shown.

[0032] Table 1. Sample distribution of point and local scale feature datasets

[0033]

[0034] Step 3: X-band weather radar insect and bird identification based on pixel-level classifier.

[0035] Gaussian Bayesian classifier, XGBoost, random forest, fuzzy logic and other pixel-level classifiers can be used; the input of the pixel-level classifier is the polarization characteristics and standard deviation texture of X-band weather radar data, a total of 12 features; the output is the identification category of each pixel point. Based on the point / local scale insect and bird echo data set obtained in step 2, the training of the pixel-level classifier is completed.

[0036] Since the 12 feature distributions are very consistent with the Gaussian distribution in terms of morphology, this embodiment uses the Gaussian Bayes classifier as the pixel-level classifier. Specifically: First, the prior probabilities of the three categories of rainfall, birds, and insects are calculated, as shown in formula (1). Then the conditional probability of each feature attribute is calculated to obtain the likelihood probability of each sample, as shown in formula (2). Finally, the probability that each set of data belongs to each category can be obtained, as shown in formula (3).

[0037]

[0038] Where N c represents the number of samples belonging to category c, and N represents the total number of samples.

[0039] p(x|c)=p(x1|c)p(x2|c)…p(x n |c) (2)

[0040] Among them, p(x|c) represents the probability of sample x appearing under the condition of specified category c; p(x i |c) means that under the condition of the specified category c, the i-th feature x i Probability of occurrence.

[0041]

[0042] Step 4: X-band weather radar insect and bird identification based on image classification model.

[0043] The image classification model is a global scale classification, and ResNet50, VGG16, GoogLeNet, DenseNet, etc. can be used. The original 6 polarization data of the X-band weather radar and the corresponding standard deviation texture, a total of 12 features, are drawn into an image as the input of the image classification model. This step is to mark the main categories of the radar echo, that is, to determine whether the overall echo is dominated by rainfall or caused by the migration of animals such as birds or insects. Therefore, the output of the image classification model is the main identification category in the X-band weather radar image, that is, the category with the highest probability of identification.

[0044] Referring to step 2, the results of S-band insect and bird identification are used as labels to annotate the X-band radar data. A total of 4367 radar data are manually annotated as data sets for the global scale image classification model, and the specific distribution is shown in Table 2. Based on the global scale insect and bird echo data set in Table 2, the training of the image classification model is completed.

[0045] Considering that the residual network is one of the most advanced classification models and can solve the classification tasks of most data, this embodiment uses ResNet50 to fit the obtained sample data set, of which 80% is used for training and 20% is used for testing. Finally, the recognition accuracy rate of the test set is close to 100%, as shown in the figure. Figure 4 shown.

[0046] In this step, considering the wide coverage of weather radar monitoring, sometimes some areas may experience rainfall, while other areas may experience insect and bird migration. In addition to meteorology, insects, and birds, three more categories have been added: meteorology + birds, meteorology + insects, and birds + insects.

[0047] Table 2. Sample distribution of global scale feature dataset

[0048]

[0049] Step 5: Information fusion of multi-scale recognition results based on category enhancement module.

[0050] In order to further improve the accuracy of insect and bird identification, the present invention corrects the point / local scale identification results of the pixel-level classifier based on the global scale identification results of the image classification model. Specifically, the point scale and local scale data can obtain the probability of belonging to each category through the Gaussian Bayes classification model. This link needs to be input pixel by pixel to obtain the category probability of each pixel; and the global scale data is input into the ResNet50 classification model to obtain the main category and non-main category of the radar data; this step performs an enhancement function on the predicted probability of the main category and a suppression function on the predicted probability of the non-main category, as shown in formula (4), and inputs the results of the two function outputs into the category enhancement module. The specific function of this module is to combine the global scale classification results and react to the point / local scale prediction probability, thereby realizing the correction of the original prediction results. Figure 5 For example, Figure 5 (a) is a typical bird migration data. Figure 5 (d) is the manually annotated label; the recognition results of radar data pixel by pixel using Gaussian Bayes classifier are as follows: Figure 5As shown in (b), it can be seen that there are a lot of bird echoes that are misclassified as insect echoes in the identification results. The classification result of ResNet50 on the global scale data believes that the echo is caused by bird migration. Therefore, it is necessary to perform an enhancement function on the bird probability of the Gaussian Bayes classifier prediction result, and perform a suppression function on the insect probability and rainfall probability. After the category enhancement module, the echo category corresponding to the maximum pixel-by-pixel probability is selected as the identification category. The final identification result is shown in Figure 5 As shown in (d), it can be seen that after the category enhancement module, the recognition accuracy of insect and bird echoes is significantly improved by combining the recognition results of multi-scale feature data, which can effectively solve the problem of misjudgment of insect and bird echoes in X-band radar.

[0051]

[0052] In the formula, represents the corrected prediction probability; y represents the probability predicted by the Gaussian Bayes classifier for each category; when y is the main category predicted by ResNet50, (1 / 3 is taken in this embodiment); when y is a non-main category predicted by ResNet50, n=2-4 (3 is taken in this embodiment). Compared with the original Gaussian naive Bayes classification model, the insect-bird identification algorithm based on multi-scale information fusion proposed in the present invention not only makes full use of the scattering characteristics of migrating insects and birds, but also incorporates the spatial distribution characteristics of migrating insects and birds, so that the accuracy of insect-bird identification is increased from the original 78.35% to 87.68%, an increase of 9.33%.

[0053] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying insects and birds using X-band weather radar based on multi-scale information fusion, characterized in that: include: Step 1, constructing a pixel-level classifier and an image classification model; wherein the input of the pixel-level classifier and the image classification model is the point-scale features of the X-band weather radar monitoring data, including reflectivity factor, radial velocity, spectral width, differential reflectivity, correlation coefficient and differential phase, and the standard deviation texture of the point-scale features; The output of the pixel-level classifier is the identification category of each pixel in the X-band weather radar image; the output of the image classification model is the main identification category in the X-band weather radar image; the main identification category is the category with the highest probability of identification; Step 2: Based on the sample data set, the pixel-level classifier and image classification model are trained; Step 3, identifying the monitoring data of the X-band weather radar based on the pixel-level classifier and image classification model trained in step 2; Step 4: Based on the output category of the image classification model in step 3, the probability of each identification category output by the pixel-level classifier in step 3 is corrected, specifically: the probability of the category that is the same as the output category of the image classification model is enhanced, and the probability of other categories is suppressed; The output result of the corrected pixel-level classifier is the final X-band weather radar insect and bird identification result.

2. The method according to claim 1, characterized in that In the step 2, S-band weather radar and X-band weather radar are used for joint monitoring to establish a sample data set; wherein the monitoring range of the S-band radar completely covers the monitoring range of the X-band radar; Insect and bird identification is performed on the monitoring data of the S-band weather radar; the S-band radar identification results are used as the classification labels for the X-band weather radar sample data.

3. The method according to claim 1, characterized in that In the pixel-level classifier, the identification categories include: weather, insects and birds; In the image classification model, the identification categories include: weather, insects, birds, weather + birds, weather + insects, and birds + insects.

4. The method according to claim 1, 2 or 3, characterized in that: The pixel-level classifier is Gaussian Bayesian classifier, XGBoost, random forest or fuzzy logic; the image classification model is ResNet50, VGG16, GoogLeNet or DenseNet.

5. The method according to claim 1, characterized in that In step 4, for each identification category output by the pixel-level classifier in step 3, the probability of the same category as the category output by the image classification model is increased to the (1 / 4 to 1 / 2)th power of the original prediction probability, and the probability of other categories of the pixel-level classifier is reduced to the (2 to 4)th power of the original prediction probability.

Citation Information

Cited By

  • Bird detection radar target recognition system and method based on multi-sensor data fusion

    CN119148126A

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

    CN119740134A

  • A large-scale biological migration forecasting method based on dual-polarized weather radar

    CN119740134B

  • X-band and S-band radar joint cross validation-based migrant bird tracking method

    CN121856926A

  • Migratory bird tracking method based on X-band and S-band radar joint cross-validation

    CN121856926B