A method and system for screening ionospheric frequency-height map extension F phenomena radar patterns

Abstract

The present application relates to a kind of ionospheric frequency height chart extension F phenomenon radar pattern screening method and system, the method comprises: the frequency height chart image of height finder output is preprocessed and input into the extension F phenomenon radar pattern recognition model that is established and trained in advance, obtain whether there is the identification result of extension F phenomenon, and obtain corresponding extension F type;The extension F type includes: no extension F, frequency type FSF, regional type RSF, mixed type MSF and strong regional type SSF;The extension F phenomenon radar pattern recognition model is resnet34Net network, improved resnet34Net network or residual_attention_Net network, and is obtained by using the method of supervised learning training.This application can automatically judge the occurrence and type of extension F phenomenon in Hainan height finder frequency height chart, give the identification result, and the accuracy of judgment result is extremely high.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of space target detection, and relates to a radar pattern automatic identification method, in particular to a screening method and system for radar patterns of ionospheric spread F phenomenon. BACKGROUND

[0002] An ionospheric altimeter is a remote sensing device for detecting the ionospheric space environment based on radar echoes, and the generated image data is called a frequency-altitude map (frequency-altitude map). A clear trace line reflects the change of electron density with height. When the ionospheric F layer (about 130 km-1000 km, most of the spacecraft flight area) is not stable in a certain height range, there are some fine structures of plasma (density inhomogeneous body, or irregular body). They cause diffuse reflection of incident radio waves, and the presented is not a clear line trace, but a diffuse piece.

[0003] Spread F is a specific diffuse pattern produced in the frequency-altitude map by the influence of the natural phenomenon of ionospheric plasma irregularity on radio wave propagation. Different forms correspond to different physical laws. The currently internationally accepted classification of spread F is the International Radio Science Union's 1978 revised "Ionogram Interpretation and Measurement Handbook". According to the pattern characteristics in the frequency-altitude map, it is divided into frequency type (Frequency Spread F, abbreviated as FSF), range type (Range Spread F, abbreviated as RSF), mixed type (Mixed Spread F, abbreviated as MSF), and branch type (Branch Spread F, abbreviated as BSF). In view of the great differences in the characteristics of spread F phenomenon in various stations around the world, it is suggested that each station can use its own targeted classification. The four types are: (1) Frequency type, clear F layer trace line in low frequency range, and spread in high frequency range, corresponding to disturbance structure near the peak height of F layer; (2) Range type, clear F layer trace line in high frequency range, and spread in low frequency range, corresponding to inhomogeneous plasma density structure near the bottom of F layer; (3) Mixed type, having the characteristics of frequency type and range type, and the mechanism is also complex; (4) Branch type, having spread near the peak frequency of F layer, and having different F layer trace lines, corresponding to horizontally distributed and different density plasma structures, which may be associated with ion sedimentation in high latitude area.

[0004] In practice, the scientific researchers of the ionospheric research group of the National Space Science Center of the Chinese Academy of Sciences (hereinafter referred to as the Space Center of the Chinese Academy of Sciences) believe that the spread F detected by the altimeter of the Hainan Fuk station (19.5°N, 109.1°E) located in the low latitude area of China almost does not appear to be a type, and there is a pattern that the spread extends from low frequency to high frequency, and the height changes less (less than 100 km) as the frequency increases, exceeds the ionospheric peak frequency, and lasts more than half an hour. In the past, it has been divided into regional type or mixed type spread F. Such patterns often occur at 20:00-22:00 local time in spring and autumn, and the occurrence probability is usually more than 20% per year. The researchers believe that it corresponds to a large-scale structure named ionospheric plasma bubble, and suggest to be classified as a separate type at low latitudes, named Strong Range Spread F (SSF) for short, and publish academic papers and get a lot of international researchers' recognition.

[0005] Due to the scientificity and complexity of the frequency-height graph, in the past, only the human eye could make an empirical judgment. A major defect of this method in scientific research is that it mixes human subjective judgment, and different scientific researchers have different standards for judging the type of spread F. Even the same scientific researcher will change the judgment standard in the process of working for many years, with the change of years, seasons, local time, etc.

[0006] With the development of China's space technology, especially the further construction of the Meridian Project Phase II led by the Space Center of the Chinese Academy of Sciences, more than ten digital altimeters will be added in China by 2023, which work 24 hours a day, with a resolution of about 5-15 minutes, and even some engineers have developed a high-precision detection network with a resolution of 1 minute. In this case, the method of relying on manual interpretation of frequency-height graph will be detrimental to real-time monitoring of the space environment, and it is impossible to manually identify all stations with uniform standards 24 hours a day, so from the application point of view, it is necessary to develop an artificial intelligence identification method for the spread F phenomenon of ionospheric frequency-height graph.

[0007] From the scientific research point of view, the complex patterns in the ionospheric frequency-height graph contain deep information related to the physical mechanism, and through the judgment of the type, the key information is screened, which is of great significance for scientific researchers to study the physical principles behind the spread F phenomenon. SUMMARY

[0008] The purpose of the present application is to overcome the problem that in the past, only the human eye could make an empirical judgment due to the scientificity and complexity of the frequency-height graph. The present application develops an automatic intelligent identification software that can automatically judge the occurrence and type of spread F phenomenon in Hainan altimeter frequency-height graph by machine, and gives the identification result, and achieves the following purposes:

[0009] (1) Real-time and automation of scientific detection instruments of ground ionosphere monitoring network are realized;

[0010] (2) Unified machine standard eliminates interference of personnel subjective factors;

[0011] (3) It is helpful for researchers to study deep ionosphere characteristics reflected in ionosphere frequency-altitude map.

[0012] In order to achieve the above purposes, the present application realizes through the following technical solutions.

[0013] The present application provides a screening method for ionosphere frequency-altitude map extended F phenomenon radar pattern, which comprises:

[0014] The frequency-altitude map image output by the altimeter is input into the pre-established and trained extended F phenomenon radar pattern recognition model after pretreatment, to obtain the recognition result of whether there is an extended F phenomenon, and to obtain the corresponding extended F type; the extended F type includes: no extended F, frequency type FSF, regional type RSF, mixed type MSF and strong regional type SSF;

[0015] The extended F phenomenon radar pattern recognition model is resnet34Net network, improved resnet34Net network or residual_attention_Net network, which is trained by a supervised learning method.

[0016] As one of the improvements of the above technical solutions, the pretreatment comprises:

[0017] According to the setting requirements, the coordinates, pixels and formats of the frequency-altitude map image output by the altimeter are unified;

[0018] The picture is cropped to remove the surrounding information part and coordinate axis information, and the main part of the frequency-altitude map in the picture is retained;

[0019] The picture is stretched to standardize the pixels.

[0020] As one of the improvements of the above technical solutions, the resnet34Net network model adopts a classic ResNet34 model, and in the last layer of convolution and operation, the arithmetic average of every 4 pixels is replaced by adaptive arithmetic average; in terms of learning rate, 100 epochs, and the learning rate is reduced to 1 / 5 every 20 epochs;

[0021] In terms of learning rate, the improved resnet34Net network sets 200 epochs, and the learning rate is reduced to 1 / 2 every 20 epochs; the convolution and maximum downsampling with a step of 2 are integrated and replaced by a convolution with a step of 4; the ratio of block is adjusted from 3:4:6:3 to 3:3:9:3;

[0022] The residual_attention_Net network is a residual attention model, and the learning rate is reduced to 1 / 5 every 25 epochs.

[0023] As one of the improvements of the above technical solutions, the method further comprises an extended F phenomenon radar pattern recognition model training step; specifically comprising:

[0024] Obtain the ionospheric altimeter output frequency height map, and perform normalization preprocessing on the image;

[0025] Label and record each preprocessed image, use the label assigned to the image to classify and sample the image according to different extended F type characteristics, and perform classification calibration; the classification method used here is based on the internationally accepted non-extended F, frequency type FSF, regional type RSF, mixed type MSF and bifurcated type BSF, which are all based on the position of the diffuse extended F layer in the frequency height map. The description of the pattern characteristics is classified, and other sub-classifications such as the strong regional type SSF used in this method are all within the same principle of this classification method;

[0026] Downsample or upsample the images of each category to reach the set number, and allocate training set and test set;

[0027] Based on the resnet34Net network, the improved resnet34Net network or the residual_attention_Net network model, an identification model is established to extract and identify the characteristics of the extended F type;

[0028] Use the training set to train the identification model to obtain the trained extended F phenomenon radar pattern recognition model.

[0029] As one of the improvements of the above technical solutions, the obtained ionospheric altimeter output frequency height map contains high, medium and low solar activity years, and contains all seasons and local time; at the same time, the frequency height map includes image data obtained continuously every set time.

[0030] As one of the improvements of the above technical solutions, when classifying different extended F type characteristics, the artificial discrimination method is used, and the characteristics of the front and rear ionization map are also considered, and the continuity of the ionosphere phenomenon is also considered, and then the judgment and classification are performed;

[0031] If the pattern in the image is blurred or similar to the diffuse or individual blank pattern due to ionospheric absorption, it is necessary to determine whether the extended F phenomenon really appears, specifically:

[0032] If the picture pattern is related to the characteristics of the place, the picture is classified and sampled;

[0033] If the picture appears completely blank pattern for more than ten hours or even longer, the picture is not classified and sampled.

[0034] As one of the improvements of the above technical solutions, the picture sample is expanded by adding one or more noises in the picture during the up-sampling process; the noise is "Poisson noise", "Gaussian noise", "salt and pepper noise", "row salt and pepper noise" or "column salt and pepper noise".

[0035] The application also provides a screening system for ionospheric frequency-height map expansion F phenomenon radar patterns, which is based on the screening method for ionospheric frequency-height map expansion F phenomenon radar patterns described above, and can identify and screen the expansion F type of the frequency-height map.

[0036] The data preparation module is used for pre-processing the imported frequency-height map and inputting the pre-processed frequency-height map into the expansion identification module.

[0037] The expansion identification module is established based on the trained expansion F phenomenon radar pattern identification model, and is used for identifying and screening the expansion F phenomenon type of the imported frequency-height map.

[0038] As one of the improvements of the above technical solutions, the system further comprises:

[0039] The data output module is used for displaying and exporting the screening result.

[0040] The display form of the screening result includes "by file" and "by time":

[0041] Each row of "by file" represents a picture and the corresponding identification result of the picture.

[0042] "By time" is used for displaying the result according to the time period of the expansion F phenomenon.

[0043] Correspondingly,

[0044] When the screening result is exported by file, each row contains the path of the picture and the identification result.

[0045] When the screening result is exported by time, each row contains a time period of the expansion F.

[0046] As one of the improvements of the above technical solutions, the system further comprises:

[0047] The identification result checking module is used for checking and correcting the displayed screening result, including checking the picture name, the identification result, the picture without zooming and the serial number recorded in the "file list", and correcting the identification result by reselecting a result if the identification result is problematic.

[0048] Technical effects of the present application:

[0049] From the test results and the software functions of the present application, based on the ionospheric frequency-altitude map expansion F phenomenon of Hainan height finder data, the artificial intelligence identification software developed by the present application can automatically judge the occurrence and type of the expansion F phenomenon in the Hainan height finder frequency-altitude map, and give the identification result. The accuracy of the judgment result is very high, close to the artificial judgment of professional researchers. The software is convenient and fast to apply, and can realize real-time and automation of the scientific detection instrument of the ground ionospheric monitoring network. The software not only greatly saves the manpower of scientific research and operation and control personnel, but also significantly improves the working efficiency of scientific instruments. The unified machine standard eliminates the interference of personnel subjective factors, and is also helpful for researchers to study the deep ionospheric characteristics reflected in the ionospheric frequency-altitude map, and has scientific and application prospects.

[0050] Compared with the prior art, the present application has the following advantages:

[0051] (1) Based on the artificial identification of professional researchers for many years, supervised learning is carried out, and the classification method is original and belongs to the international scientific research frontier;

[0052] (2) The latest image depth convolution model is applied to the use of ionospheric digital height finder data for the first time;

[0053] (3) The automatic discrimination analysis of the ionospheric expansion F physical phenomenon is realized for the first time. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the ionospheric frequency-altitude map expansion F phenomenon radar pattern screening method of the present application;

[0055] Figure 2 The module function diagram of the system structure of the present application;

[0056] Figure 3 The "file" identification result graph of the system of the present application;

[0057] Figure 4 The "time" identification result graph of the system of the present application. DETAILED DESCRIPTION

[0058] The technical solutions provided by the present application are further illustrated below in combination with embodiments.

[0059] Example 1

[0060] As Figure 1 shown, it is a method flow chart of a screening method of ionospheric frequency-height map extended F phenomenon radar pattern according to an embodiment of the present application.

[0061] Embodiment 2

[0062] The screening method of ionospheric frequency-height map extended F phenomenon radar pattern based on the method according to the embodiment 1 of the present application is used to identify and screen the extended F type of the frequency-height map, and the system comprises a data import module, a data preparation module, a data output module, an identification result checking module and an extended identification module.

[0063] The data import module is used to import the frequency-height map output by the altimeter and needing to identify the extended F phenomenon type.

[0064] The data preparation module is used to pre-process the imported frequency-height map and input the pre-processed frequency-height map to the extended identification module.

[0065] The extended identification module is established based on the trained extended F phenomenon radar pattern identification model and is used to identify and screen the extended F phenomenon type of the imported frequency-height map.

[0066] The data output module is used to display and export the result after the screening is completed.

[0067] The display form of the screening result comprises "by file" and "by time":

[0068] Each row of "by file" represents a picture and the identification result corresponding to the picture.

[0069] "By time" is used to display the result according to the time period of the extended F phenomenon.

[0070] Correspondingly,

[0071] When the screening result is exported by "by file", each row contains the path of the picture and the identification result.

[0072] When the screening result is exported by "by time", each row contains a time period of the extended F.

[0073] The identification result checking module is used to check and correct the displayed screening result, which comprises checking the name of the picture, the identification result, the picture without scaling and the serial number recorded in the "by file" list, and correcting the identification result by reselecting a result if the identification result is problematic.

[0074] Embodiment 3

[0075] As Figure 2As shown, it is the module composition structure diagram of the automatic intelligent identification software developed according to the method or system of the application, which is embodiment 3 of the application.

[0076] In order to overcome the shortcomings of the existing analysis technology of the space (ionosphere) environment monitoring data of the ground-based ionosonde, the application provides a machine learning-based ionosonde frequency-height map extension F phenomenon artificial intelligence identification software, which uses a deep convolutional network to supervise the learning of the artificial calibration of the Hainan Fuk station (19.5°N, 109.1°E) ionosonde graph, extracts the features of the natural phenomenon in the data graph, develops an automatic intelligent identification software, and can automatically judge the occurrence and type of the extension F phenomenon in the Hainan ionosonde frequency-height map, and give the identification result.

[0077] The technical solution adopted by the application to solve its technical problems comprises the following steps:

[0078] (1) Image normalization preprocessing: the application is based on the graph features in the ionosphere frequency-height map, so the frequency-height map as a sample needs to be standardized first to ensure that the picture size and the horizontal and vertical coordinates are consistent.

[0079] (2) Generating a sample library with artificial labeling results: scientific researchers have manually labeled and recorded each picture of the Hainan Fuk station (19.5°N, 109.1°E) ionosonde from 2002 to 2015 based on their own experience, and used these records to assign labels to the pictures of (1), which are divided into no extension F (label 0), frequency type FSF (label 1), regional type RSF (label 2), mixed type MSF (label 3), and strong regional type SSF (label 4), and are subjected to secondary proofreading to meet the standard extension F type feature definition. After labeling, a total of 46,5493 samples are obtained.

[0080] Extracting training set and test set: the obtained sample set has the problem of uneven sample distribution, with the number of label 0 reaching 426,555, and the number of labels 1-4 being several thousand to more than 10,000. Therefore, 20,000 is selected as the final number of images of each category, the label 0 is down-sampled, and the labels 1-4 are up-sampled, to obtain a total of 5 categories, 100,000 picture samples, which are randomly divided into training set and test set according to the ratio of 8:2, thus completing the preprocessing and arrangement of the samples.

[0081] In the up-sampling process, in view of the characteristics of the ionosonde scientific pictures, the original images are retained, and different categories of sample images are increased to 20,000 by adding "Poisson noise", "Gaussian noise", "salt and pepper noise", and "row salt and pepper noise" and "column salt and pepper noise" in the images.

[0082] (3) Standardized feature extraction and identification of extended F phenomenon: After obtaining the effective sample data set, a supervised learning method is used for training.

[0083] Machine learning uses Python language, and uses VGG16, VGG19, ResNet34, ResNet50, EfficientNetV2, ConvNeXt and ResNet34_Attention, ResNet34_Modified and other various types of classification networks for processing.

[0084] After the machine learning training parameters are obtained in the training set (80,000 samples), the test set (20,000 samples) is tested, and the best three models with the highest accuracy in the obtained results are selected, which are ResNet34_20_5_100 (resnet34Net) (93.20%), resNet34-modified-20-2-200 (93.50%), and residual_attention_Net_old_25_5 (last_model_92_sgd_25_5.pkl) (93.53%). The trained weight parameters are encapsulated into software.

[0085] The beneficial effects of the present application are: based on the ionospheric extended F phenomenon type manually identified by researchers based on their own experience in 2002-2015, an automatic intelligent recognition software is developed using artificial intelligence technology, which can automatically judge the occurrence and type of the extended F phenomenon in the Hainan ionospheric frequency-altitude chart, and give the recognition result, with an accuracy of more than 93%, close to the effect of manual judgment by experienced researchers. This method can realize real-time and scale automation of scientific detection instruments of the ground ionospheric monitoring network, and the judgment standard is unified, which eliminates the interference of personnel subjective factors. The machine judgment standard refined in the research is helpful for researchers to study the deep ionospheric characteristics reflected in the ionospheric frequency-altitude chart.

[0086] The present application realizes the real-time and scale automation of the ionospheric monitoring instrument by supervised learning of different types of extended F phenomenon in the frequency-altitude chart data of the scientific instrument.

[0087] The present application adopts the following technical solutions:

[0088] (1) Image normalization preprocessing

[0089] The original file of the ionospheric monitoring instrument is exported as a frequency-altitude chart by SAO-X software, the coordinates are unified as 0-17MHz and 90-800km, and the standard PNG picture is 700 pixels wide and 600 pixels high.

[0090] Crop the picture, remove the surrounding information part and coordinate axis information, and keep the main part of the image, that is, take the pixels of each picture from the top left corner (150, 61) to the bottom right corner (645, 520);

[0091] After stretching, the specification is 448*448 pixels, which is beneficial to the subsequent classification network processing (the existing network mainly processes 224*224 images).

[0092] (2) Generate a sample library with artificial calibration results, extract test set and training set

[0093] (2-1) Generate a sample library with artificial calibration results:

[0094] Based on their own experience, researchers have manually labeled and recorded each picture of the altimeter at Hainan Fuk station (19.5°N, 109.1°E) from 2002 to 2015. Using these records to assign labels to pictures in (1), they are divided into no extension F (label 0), frequency type FSF (label 1), regional type RSF (label 2), mixed type MSF (label 3), and strong regional type SSF (label 4).

[0095] However, due to the nature of scientific research, researchers are used to considering "complete event of the same type" when recording, that is, a spatial weather event that lasts for several hours. If the features of the dozens of pictures during this period are ambiguous, they are identified as the same type. But this work requires samples with distinct features, so the artificial secondary classification calibration is carried out according to the standard extended F type feature definition. The sample quantities of each type after the second calibration are shown in the following table:

[0096] Table 1 Sample quantities of each type of extended F in the labeled sample library

[0097]

[0098] During sample classification, based on scientific principles and experience, there are the following special cases and processing methods:

[0099] - The influence of season, year, solar activity week, etc. on ionospheric parameter characteristics:

[0100] According to previous statistics, the occurrence rate of different types of ionospheric extended F changes with local time, season, etc., and due to the change of foF2 with time, the characteristics of the image will also change.

[0101] The data used for training and verification in this work is more than a solar cycle (11 years), so it contains years of high, medium and low solar activity, and contains all seasons and local times, and contains data every 15 minutes / 5 minutes (24 hours continuously). Therefore, the process of unified machine learning covers all changing factors and results.

[0102] - Occasional E layer (Es layer) effects:

[0103] Es layer is a natural phenomenon near the height of 100 kilometers of ionosphere, with thin thickness (about several hundred meters) and large electron density, showing as a straight line near the height of 100 kilometers in the frequency-height graph of the ionosonde, sometimes accompanied by some similar extended patterns.

[0104] The Es layer trace may obscure the extended F pattern of the F layer at a height of about 200-300 kilometers, or the Es layer and the multiple ground echoes appear at a height of several hundred kilometers, overlapping with the ionospheric F layer trace, or cause the dispersion of ionosonde echo signals at lower frequencies (about 1-3 Mhz), which will interfere with the identification of the extended F phenomenon.

[0105] For such phenomena, when manually discriminated, the characteristics of the ionospheric charts before and after are referred to as much as possible, the continuity of ionospheric phenomena is considered, and judgment and classification are made, so as to improve the reliability of the training samples.

[0106] - Effects caused by instrument detection capability:

[0107] When foF2 is very small (about 1-1.5 MHz, mainly in the morning of local time), due to factors such as echo signal-to-noise ratio, the pattern may be blurred or appear similar to the dispersion of extended F, or even individual blank patterns due to ionospheric absorption, and whether it is really an extended F is determined according to experience. Such patterns are related to the characteristics of the ionosphere at different local times, and should still be included in the machine learning samples.

[0108] Some continuous 10-odd hours or even longer completely blank patterns are related to instrument maintenance, etc., and are not included in the machine learning samples.

[0109] - Effects of multiple echoes, noise and others:

[0110] The radio waves reflected by the ionosphere are reflected by the ground and then reflected by the ionosphere and received by the ionosonde antenna, which will appear multiple echo patterns at integer multiples of the height of the ionosphere in the frequency-height graph. Due to the interference of reflection on radio waves, there are often no radio wave dispersion patterns (no extended F) for single echo, but there are obvious dispersion patterns for multiple echoes, which are easily misidentified as extended F.

[0111] In addition, the altimeter based on the principle of radio wave radar often has noise patterns such as horizontal lines, vertical lines and color blocks in the echo pattern.

[0112] The samples that exhibit these phenomena are manually judged by experience during marking, and the model is trained to master the characteristics through machine learning, so as to finally eliminate these disturbances as much as possible.

[0113] (2-2) Extract the test set and the training set:

[0114] After obtaining the sample library as shown in Table 1, it can be seen that the number of image data of each type is different, in order to avoid uneven distribution of samples (machine learning is not friendly to small samples), 20,000 is selected as the final number of images of each type.

[0115] Downsample the samples of others (no extension): 426,555 samples, generate a random number of 0,1, select 20,000 png images corresponding to 1 as the final samples of the category;

[0116] Upsample the samples of the four types of extension F: In each category of images, keep the original image, and according to the characteristics of the ionization map, add "Poisson noise", "Gaussian noise", "salt and pepper noise", and "row salt and pepper noise", "column salt and pepper noise", etc. to increase the number of sample images of different categories to 20,000.

[0117] Assign the training set and the test set: for the 5 categories, a total of 100,000 photos, according to the ratio of 8:2, randomly divided into the training set and the test set, thus completing the preprocessing and sorting of the samples.

[0118] (3) Standardize the feature extraction and recognition of the extension F phenomenon, and establish a recognition model

[0119] The recognition model is the technical core of the business indicators of this project, and the recognition requirements are realized through the artificial marking data provided by the scientific research personnel.

[0120] The focus of machine supervised learning in this work is to extract the key areas reflecting the characteristics of the extension F phenomenon from the artificially marked pictures, judge the layout or trend of the data points, and summarize the rules. The weight parameters obtained by training are used for the automatic recognition process of the test set and future new data.

[0121] · Considering that the analysis is based on images, machine learning mainly uses deep convolutional neural models, and the ViT model is considered according to the experimental results;

[0122] · Since the characteristics and positions of the 5 types of extension F (4 types appear at Hainan Fuk station) and the basic area of the dispersion pattern distribution are determined, a suitable attention model is sought. · Considering that the analysis is based on images, machine learning mainly uses deep convolutional neural models, and the ViT model is considered according to the experimental results;

[0122] · Since the characteristics and positions of the 5 types of extension F (4 types appear at Hainan Fuk station) and the basic area of the dispersion pattern distribution are determined, a suitable attention model is sought.

[0123] • For the sample distribution and its imbalance, the ratio of positive and negative samples is more than 1:10, and even higher. The classification model is targeted, and the loss function is selected and improved.

[0124] (3-1) After training parameters are obtained by machine learning on the training set (80,000 samples), the test set (20,000 samples) is tested, and the effects of some models are shown in Table 2:

[0125] Table 2 Basic method of selected model and judgment accuracy

[0126]

[0127]

[0128]

[0129] As shown in the table, the three best models with the highest accuracy are ResNet34_20_5_100(resnet34Net) (93.20%), resNet34-modified-20-2-200 (93.50%), and residual_attention_Net_old_25_5 (last_model_92_sgd_25_5.pkl) (93.53%). The automatic judgment of the expansion F results in an accuracy of more than 93% compared to the results of the human eye judgment based on the experience of researchers.

[0130] The specific implementation of these three models is as follows:

[0131] 1) ResNet34_20_5_100(resnet34Net)

[0132] In terms of learning rate, 100 epochs, and every 20 epochs, the learning rate is reduced to 1 / 5;

[0133] The overall model is a classic ResNet34 model, and in the last layer of convolution and operation, the arithmetic mean of every 4 pixels is replaced by adaptive arithmetic mean.

[0134] 2) resNet34-modified-20-2-200

[0135] In terms of learning rate, 200 epochs, and every 20 epochs, the learning rate is reduced to 1 / 2;

[0136] Integrate the convolution and maximum downsampling with a step of 2, and replace it with a convolution with a step of 4 and a size of 4*4;

[0137] Adjust the ratio of the block from 3:4:6:3 to 3:3:9:3.

[0138] 3) residual_attention_Net_old_25_5

[0139] Learning rate, every 25 epochs, the learning rate is reduced to 1 / 5;

[0140] Residual attention model belongs to, machine learning focuses on the changing part.

[0141] Due to the difference in the complexity of the graphics, take the training results of the attention residual ResNet34_Attention model as an example, its overall accuracy on the test set is 93.525% (recognition ability is basically the same as human judgment):

[0142] - The accuracy of F judgment without expansion is 96%, that is, it can accurately identify whether there is a disturbance phenomenon (expansion F) in the ionosphere;

[0143] - The accuracy of frequency type FSF judgment is 93%, and the accuracy of regional type RSF judgment is 99%, which are the most basic and most studied types of global station expansion F, and the model judges accurately;

[0144] - The accuracy of strong regional type SSF judgment is 98%, which is a natural phenomenon in China's low-latitude region led by Hainan, which has a strong correlation with the equatorial plasma bubble and can strongly interfere with the stability of electromagnetic wave signals. It is the most concerned in science and application, and the model judges very accurately;

[0145] - The accuracy of mixed type MSF judgment is 82%, which contains a large number of ambiguous graphics that do not have obvious expansion F characteristics, so it is reasonable that the accuracy is lower, and 82% accuracy is within the acceptable range.

[0146] (3-2) In order to verify the accuracy and effectiveness of the three models in automation, further use the frequency height pattern samples of Hainan Fukang ionospheric sounding station from 2013 to 2016 for verification.

[0147] Among them, more than 46,000 samples per year from 2013 to 2015, only a few are subjected to previous up-sampling and down-sampling into the training set and sample set of machine learning. Test all samples in a year (mainly check if the expansion F phenomenon is misidentified), compare the inference results of the three models with the GT-SF results, and find that the comparison difference is about 2.5% in 2013, about 2.3% in 2014, and about 5.2 in 2015. The consistency is very good.

[0148] The data in 2016 is not artificially judged and identified, and does not participate in machine learning. The parameters trained by the three models are used to automatically judge the samples in 2016, and the difference between the inference results of the three models is within 5%.

[0149] In summary, the test shows that the machine learning three models in the application have excellent accuracy in automatically judging the ionospheric frequency height map expansion F phenomenon of the Hainan height meter data, and are very excellent in automation, can greatly improve the efficiency of the height meter in monitoring the ionospheric space environment, and have practical application prospects.

[0150] Therefore, the application uses the weight parameters of the above three trained models to encapsulate and manufacture intelligent identification software, which can be directly used in real work scenes. Figure 2

[0151] In use, the original PNG picture output by the height meter can be directly imported, which is divided into single file import and folder-data time range selection, and the picture file to be judged is directly read through the date and time in the file name;

[0152] Confirm the selected weight file (corresponding to three models), click the "identify" button, and the system will pop up a progress bar and start running. After testing, the average processing time of 100 pictures is: in the GPU mode, the model loading time is 3226ms, and the average model identification time is 80.6ms; in the CPU mode, the model loading time is 795ms, and the average model identification time is 439.4ms;

[0153] After the identification is completed, the statistical results will be displayed, that is, how many pictures each type of expansion F occupies;

[0154] The results will be displayed in the "by file" and "by time" tabs in the lower part of the software:

[0155] Each row of "by file" represents a file and the identification result, as shown in Figure 3 ;

[0156] "By time" displays the results according to the time period when the expansion F phenomenon occurs, as shown in Figure 4 .

[0157] The software supports checking and correcting the identification results. Click the result directly to open the dialog box. The dialog box displays the picture name, identification result, picture without zoom, and current record number in the "by file" list. If the user feels that the identification result is wrong, the identification result can be corrected. Select a new result from the identification result list and click "correct". The software will pop up a confirmation box. Click yes, and the result will be updated;

[0158] ​The user can export the recognition result to a text file, click the "export" button on the "by file" or "by time" tab, and after a dialog box is popped up to select the saving location and file name, the system starts to execute and prompts a dialog box for completion;

[0159] The export file of the "by file" tab contains the path of the picture and the recognition result in each line;

[0160] The export file of the "by time" tab contains a time period of the occurrence of the extended F in each line.

[0161] As can be seen from the above detailed description of the present application, the artificial intelligence recognition software developed by the present application can automatically judge the occurrence and type of the extended F phenomenon in the Hainan altimeter height-frequency chart, and give the recognition result, which has very high accuracy.

[0162] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the examples, those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present application do not deviate from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for screening a radar pattern of an ionospheric frequency-height map expansion F phenomenon, characterized by, The method comprises: The method comprises: The method comprises: The method further comprises a training step of the extended F phenomenon radar pattern recognition model, and specifically comprises: Obtaining the frequency-altitude map output by the ionospheric altimeter and performing standardized preprocessing on the image; Labeling and recording each preprocessed image, using the label assigned to the image to classify and sample the image according to different extended F type features, and performing classification calibration; Downsampling or upsampling the images of each category to reach a set number and allocate training sets and test sets; Based on the resnet34Net network, the improved resnet34Net network or the residual_attention_Net network model, an identification model is established to extract and identify the features of the extended F type; Using the training set to train the identification model to obtain the trained extended F phenomenon radar pattern recognition model; When classifying different extended F type features, an artificial judgment method is used, the features of the front and rear ionospheric maps are referred to, and the continuity of the ionospheric phenomenon is considered for judgment and classification; If the pattern in the image is blurred or similar to the extended F phenomenon, or individual blank patterns appear due to ionospheric absorption, it is necessary to determine whether the extended F phenomenon really appears, and specifically: If the pattern in the image is related to the characteristics of the local time, the image is classified and sampled; If the image appears completely blank for more than ten hours or even longer, the image is not classified and sampled.

2. The method of claim 1, wherein the ionospheric frequency-height map extension F-phenomenon radar pattern is selected from the group consisting of: The preprocessing comprises: According to the setting requirements, the coordinates, pixels and formats of the frequency-altitude map image output by the altimeter are unified; The image is cropped to remove the surrounding information and coordinate axis information, and the main part of the frequency-altitude map in the image is retained; The image is stretched to standardize the pixels.

3. The method of claim 1, wherein the ionospheric frequency-height map extension F-phenomenon radar pattern is selected from the group consisting of: The resnet34Net network model uses the classic ResNet34 model, and in the last layer of convolution and operation, the arithmetic average of every 4 pixels is replaced by adaptive arithmetic average. In terms of learning rate, 100 epochs are used, and the learning rate is reduced to 1 / 5 every 20 epochs; ​ In the improved resnet34Net network, in terms of learning rate, 200 epochs are set, and the learning rate is reduced to 1 / 2 every 20 epochs; the convolution with a step of 2 and the maximum downsampling are integrated and replaced by a convolution with a step of 4; the ratio of the block is adjusted from 3:4:6:3 to 3:3:9:

3. The residual_attention_Net network is a residual attention model, and the learning rate is reduced to 1 / 5 every 25 epochs.

4. The method for filtering radar patterns of the extended F-phenomenon in the ionospheric frequency-height map according to claim 1, characterized in that, The obtained ionospheric altimeter output frequency height map contains high, medium and low solar activity years, and contains all seasons and local time; meanwhile, the frequency height map includes image data obtained continuously at a set time interval.

5. The method for filtering radar patterns of the extended F-phenomenon in the ionospheric frequency-height map according to claim 1, characterized in that, In the upsampling process, the picture samples are expanded by adding one or more noises to the picture; the noise is "Poisson noise", "Gaussian noise", "salt and pepper noise", "row salt and pepper noise" or "column salt and pepper noise".

6. A system for screening ionospheric spread-F radar patterns, based on the method for screening ionospheric spread-F radar patterns according to any one of claims 1 to 5, wherein the spread-F type of the ionospheric spread-F radar patterns is identified and screened. The system comprises a data import module, a data preparation module and an expansion identification module; the data import module is used to import the frequency height map output by the altimeter, which needs to identify the expansion F phenomenon type; The data preparation module is used to pre-process the imported frequency height map and input the pre-processed frequency height map into the expansion identification module; The expansion identification module is established based on the trained expansion F phenomenon radar pattern recognition model, and is used to identify and screen the expansion F phenomenon type of the imported frequency height map, and is also used to obtain the frequency height map output by the ionospheric altimeter and to pre-process the image; each picture after pre-processing is labeled and recorded, the picture is classified and sampled according to different expansion F type characteristics by using the label assigned to the picture, and the picture is classified and calibrated; the pictures of each category are down-sampled or up-sampled so as to reach a set number and to be distributed into a training set and a test set; the recognition model is respectively established based on a resnet34Net network, an improved resnet34Net network or a residual_attention_Net network model to extract and identify the characteristics of the expansion F type; the training set is used to train the recognition model to obtain the trained expansion F phenomenon radar pattern recognition model; when classifying different expansion F type characteristics, an artificial discrimination method is adopted, the characteristics of the ionospheric maps before and after are referred to, and the continuity of the ionospheric phenomenon is considered, and then the judgment and classification are performed; if the pattern in the picture is blurred or similar to the dispersion of the expansion F phenomenon or individual blank patterns appear due to ionospheric absorption, it is necessary to determine whether the expansion F phenomenon really occurs, and specifically, if the pattern of the picture relates to the characteristics of the local time, the picture is classified and sampled; if the picture appears a completely blank pattern for more than ten hours or even longer, the picture is not classified and sampled.

7. The ionospheric frequency-height map extension F-phenomenon radar pattern screening system according to claim 6, characterized by, The system further comprises: A data output module is used to display and export the screened results; The display form of the screening results comprises "by file" and "by time"; Each row of "by file" represents a picture and the corresponding identification result of the picture; "By time" is to display the results according to the time period of the expansion F phenomenon; Correspondingly, When the screening results are exported "by file", each row contains the path of the picture and the identification result; When the screening results are exported "by time", each row contains an occurrence time period of the expansion F.

8. The ionospheric frequency-height map extension F-phenomenon radar pattern screening system according to claim 7, characterized by, The system further comprises: A recognition result checking module is used to check and correct the displayed screening result, including checking the picture name, recognition result, non-scaled picture and the serial number recorded in the "by file" list. If the recognition result is problematic, the recognition result is corrected by reselecting a result.

Images