Ionospheric spread-f pattern prediction method and system based on time convolution and super-resolution

Abstract

The application belongs to the field of space target detection, and particularly relates to an ionospheric spread F pattern prediction method and system based on time sequence convolution and super resolution. The method comprises the following steps: inputting M continuous frequency-height patterns generated by an ionospheric altimeter before a certain moment into a pre-established and trained spread F radar pattern prediction generation model after preprocessing, so as to obtain N frequency-height patterns to be generated at a future moment; the network structure of the spread F radar pattern prediction generation model comprises a ConvGRU network and an EDSR network in sequence; analyzing the N frequency-height patterns, predicting whether an ionospheric spread F phenomenon occurs, and generating a pattern feature reflecting a spread F type; the spread F type comprises no spread F, a frequency type FSF, a regional type RSF, a mixed type MSF and a strong regional type SSF.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of space target detection, and relates to a radar pattern short-term prediction generation method, in particular to an ionospheric extended F pattern prediction method and system based on time series convolution and super-resolution. BACKGROUND

[0002] The ionospheric altimeter or vertical profiler is a remote sensing device for detecting the ionospheric space environment based on radar sweep frequency echo, and the generated image data is called frequency-altitude graph (frequency-altitude graph). A clear trace line reflects the change of electron density with height. In a certain height range of the ionospheric F layer (about 130 km-1000 km, most of the spacecraft flight area), it is not a stable layer, but there are some fine structures of plasma (density inhomogeneous body, or irregular body). They cause diffuse reflection of incident radio waves, and present not a clear line trace, but a diffuse piece. Therefore, the radar pattern of the altimeter contains the vertical profile information of the electron density of the ionospheric background and the image information related to the disturbance change of the ionospheric space environment. It is a key scientific instrument data for 7*24 hours continuous monitoring of the ionospheric space environment and space weather.

[0003] Extended F is a specific diffuse pattern produced in the frequency-altitude graph by the influence of radio wave propagation caused by natural phenomena such as ionospheric plasma disturbance and irregular body. Different forms correspond to different physical laws. The internationally accepted classification of extended F is the International Radio Science Union's 1978 revised "Ionogram Interpretation and Measurement Handbook", which 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) according to the pattern characteristics in the frequency-altitude graph. In practice, the researchers of the ionospheric research group of the National Space Science Center of the Chinese Academy of Sciences believe that the altimeter of the Hainan Fuk station (19.5°N, 109.1°E) in China's low latitude area almost does not appear branch type, and there is a pattern that extends from low frequency to high frequency, the height change is small (less than 100 km) with the increase of frequency, exceeds the ionospheric peak frequency, and the duration is more than half an hour. In the past, it has been divided into range type or mixed type extended F. After research, the researchers believe that it corresponds to a large-scale structure named ionospheric plasma bubble, and suggest to classify it as a separate type at low latitudes, named strong range spread F (Strong Range Spread F, abbreviated as SSF), and publish academic papers and get a lot of international researchers' recognition.

[0004] The ionospheric disturbances and plasma irregularities reflected by the spread-F phenomenon directly interfere with radio wave propagation. For example, the plasma bubble corresponding to the spread-F phenomenon can produce significant signal flickering. In addition, when spacecrafts pass through plasma irregularities, the effects of density changes must also be considered. Therefore, the occurrence, development, and characteristics of the spread-F phenomenon have important implications in both scientific research and applications, and the prediction of the spread-F phenomenon is an important research topic in space weather science.

[0005] Since the late 20th century, Abdu and other scholars have used the cubic spline interpolation method to predict the occurrence probability of the spread-F phenomenon in Brazil. This method has been successfully applied to the International Reference Ionosphere (IRI) model. By using parameters such as solar and geomagnetic activity indices, latitude, longitude, and time, this method can accurately predict the monthly average occurrence probability of the spread-F phenomenon. However, this model does not consider ionospheric vertical data in the Chinese region during its construction, so there is a large deviation in the prediction results in the Chinese region. Domestic scholars such as Xu Tong have used the Fourier series expansion algorithm to predict the occurrence probability of the spread-F phenomenon in the Chinese region and found that this model is superior to the IRI model in terms of prediction accuracy. Since then, some scholars have begun to use artificial neural network (ANN) algorithms to predict the spread-F phenomenon. This method can effectively handle the highly nonlinear and complex physical processes in the ionosphere, thus achieving higher prediction accuracy than traditional algorithms. The above prediction methods only provide data for the spread-F phenomenon. Since the information carried by the ionogram is much larger than that represented by a single data point, the prediction of ionospheric ionograms is relatively more difficult and has higher application value. In summary, domestic and foreign scholars have paid great attention to the prediction of the spread-F phenomenon in ionograms, but due to the complexity of ionogram graphs and the reliance on human eye recognition, there is no widely accepted prediction model with significant effects.

[0006] With the continuous advancement of China's space technology, especially the second phase of the Meridian Project led by the National Space Science Center of the Chinese Academy of Sciences, 10 ionospheric altimeters will be newly deployed nationwide by 2023. There are more than 20 dedicated altimeter scientific instruments and a large number of similar high-frequency radio wave radars in China. These altimeters work at a resolution of 5-15 minutes, 7*24 hours, or even 1 minute resolution developed by engineers. These instruments will generate a large amount of ionograms for ionospheric research in different latitudes in China, and there is an urgent need for real-time monitoring of scientific research and applications for short-term prediction. For this aspect, the ionosphere research group of the Key Laboratory of Solar and Space Weather of the Space Center has manually completed the classification and labeling of the ionograms at the Hainan Fuk station (19.5°N, 109.1°E) from 2002 to 2015, which provides a research basis for the feature extraction and prediction of various types of spread-F.

[0007] The radar pattern of the ionospheric altimeter continuous observation can be regarded as a time sequence image sequence, the ionospheric density and height reflected in the pattern are gradually changed, and various spread F phenomena also have a process of occurrence-development-extinction, and the 15-minute time resolution of the frequency-height pattern ensures that these phenomena are not changed suddenly. Therefore, based on the time sequence convolution model, it is expected to extract the change characteristics according to the change trend of the pattern, generate a frequency-height pattern that can reflect the key feature information, and predict whether the spread F phenomenon will appear in the pattern and the characteristics of different types of spread F in the pattern. However, due to the limitation of the model, the generated image has low definition and does not conform to the characteristics of the real frequency-height pattern, and the blurred pattern is easy to be confused with the spread F pattern, which interferes with the prediction result, so the high definition is realized by using the super-resolution method, so as to better and accurately predict the spread F phenomenon. SUMMARY

[0008] The purpose of the present application is to overcome the problem that due to the scientificity and complexity of the frequency-height pattern, only partial information can be predicted in the past, such as only predicting the ionospheric parameters such as foF2 or the monthly occurrence rate of spread F, and especially the appearance and characteristics of the spread F phenomenon in the short-term frequency-height pattern cannot be effectively predicted. The system developed by combining the generation model and the super-resolution model can perform short-term prediction on the Hainan altimeter frequency-height pattern, especially the appearance and type of the spread F phenomenon, and achieve the following purposes:

[0009] (1) Realize the short-term prediction and early warning of the ionospheric disturbance and plasma irregularity represented by the spread F detected by the ground-based altimeter;

[0010] (2) Help to automatically identify and extract the precursor information about the spread F phenomenon in the ionospheric frequency-height pattern;

[0011] (3) Help researchers to study the physical process and mechanism before the appearance of the ionospheric disturbance and plasma irregularity represented by the spread F.

[0012] In order to achieve the above purposes, the present application realizes the technical scheme as follows.

[0013] The present application provides an ionospheric spread F pattern prediction method based on time sequence convolution and super-resolution, comprising:

[0014] M continuous frequency-height patterns generated before a certain time of the ionospheric altimeter are preprocessed and input into a pre-established and trained spread F radar pattern prediction generation model to obtain N frequency-height patterns to be generated at a future time; the network structure of the spread F radar pattern prediction generation model comprises a ConvGRU network and an EDSR network in sequence;

[0015] The N frequency-height maps are analyzed to predict whether ionospheric spread F phenomenon occurs, and a graph feature reflecting the spread F type is generated; the spread F type includes: no spread F, frequency type FSF, region type RSF, mixed type MSF, and strong region type SSF.

[0016] As one of the improvements of the above technical solutions, the preprocessing includes:

[0017] According to the setting requirements, the coordinates, pixels and formats of the frequency-height map images are unified;

[0018] The picture is cropped to remove the auxiliary information part and coordinate axis information around the main graph, and the main graph part of the frequency-height map is retained; the auxiliary information part includes text description and color bar;

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

[0020] As one of the improvements of the above technical solutions, the ConvGRU network includes an encoding network, a decoding network and a convolutional layer;

[0021] The encoding network includes 6 blocks, each block including a convolutional Conv layer and a ConvGRU layer; the convolutional Conv layer is used to compress the input frequency-height map sequence into a hidden state; and the ConvGRU layer is used to learn the space-time features;

[0022] The decoding network includes 6 blocks, each block including a deconvolutional layer and a ConvGRU layer; the deconvolutional layer and the ConvGRU layer are used to expand the hidden state information;

[0023] The convolutional layer is used to output the predicted frequency-height map.

[0024] As one of the improvements of the above technical solutions, the method further includes training the ConvGRU network; the training process includes:

[0025] The frequency-height map output by the ionospheric altimeter is obtained, and the image is normalized and preprocessed;

[0026] Each preprocessed picture is labeled and recorded, and the label assigned to the picture is used; M+N images are taken as a short sequence, M as input, N as output, and the spread F type in the M+N image as the spread F type of the output image;

[0027] Taking all spread F sequences as the center, Y sequences before and after the center are taken to join the sample set, and repeated spread F sequences are removed; from the remaining short sequence set, "background" sequences are randomly selected to supplement the sample set to obtain a first data set;

[0028] The first data set is divided into a training set and a test set, and the ConvGRU network is trained to obtain the trained ConvGRU network.

[0029] As one of the improvements of the above technical solutions, the EDSR network sequentially comprises a convolution layer, an activation function layer, a residual scaling layer and an up-sampling layer; wherein,

[0030] The convolution layer is used to extract image features of the ConvGRU network predicted frequency height map;

[0031] The activation function layer and the residual scaling layer are used to learn the image features of the frequency height map;

[0032] The up-sampling layer is used for reconstruction of the frequency height map.

[0033] As one of the improvements of the above technical solutions, the method further comprises: training the ConvGRU network by using a supervised learning method; and the training process comprises:

[0034] The ionospheric altimeter historical output frequency height map is obtained, and the image is normalized and preprocessed;

[0035] X images are randomly extracted, a degradation method of combining different noise types and mean or median filtering is adopted, a degradation algorithm is selected to perform degradation operation on each frequency height map, so that the frequency height map is degraded into a blurred pattern, and the original frequency height map is used as a pair, to obtain a sample library with X pairs of low resolution and corresponding high resolution frequency height maps, as a second data set;

[0036] The second data set is divided into a training set and a test set, and the EDSR network is trained to obtain the trained ConvGRU network.

[0037] As one of the improvements of the above technical solutions, the degradation operation sequentially comprises:

[0038] ① adding salt and pepper noise with a noise density of 0.005 and using mean filtering with a convolution kernel of 5*5,

[0039] ② adding Gaussian noise with a variance of 0.005 and using mean filtering with a convolution kernel of 5*5,

[0040] ③ adding salt and pepper noise with a noise density of 0.01 and using mean filtering with a convolution kernel of 5*5,

[0041] ④ adding salt and pepper noise with a noise density of 0.01 and using mean filtering with a convolution kernel of 7*7,

[0042] ⑤ adding Poisson noise and using mean filtering with a convolution kernel of 3*3,

[0043] ⑥ adding Poisson noise and using mean filtering with a convolution kernel of 5*5.

[0044] As one of the improvements of the above technical solutions, the ionospheric altimeter historical output frequency height map is continuously acquired every set time; the historical period contains a whole solar activity week, high, medium and low years of solar activity, and all seasons and local time; the sequence of the frequency height map covers the sequence of all types of extended F type occurrence, a period of time before and after all types of extended F type occurrence, and a sequence randomly extracted irrelevant to the extended F phenomenon.

[0045] The application also provides an ionospheric extended F radar pattern prediction system based on time series convolution and super-resolution, comprising:

[0046] A data processing module is configured to input M continuous frequency height maps generated before a certain moment of the ionospheric altimeter into a pre-established and trained extended F radar pattern prediction generation model after preprocessing, to obtain N frequency height maps to be generated at a future moment; the network structure of the extended F radar pattern prediction generation model comprises a ConvGRU network and an EDSR network in sequence; and

[0047] A type prediction module is configured to analyze the N frequency height maps, predict whether an ionospheric extended F phenomenon occurs, and predict and generate pattern features reflecting the 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.

[0048] Technical effects of the application:

[0049] (1) The application predicts the background density trace and the extended F pattern in the frequency height map according to the trend of the image features changing with time, thereby predicting the occurrence of ionospheric disturbance and plasma irregularity, exploring the ability of the image to reflect the signs of disturbance phenomenon, and the idea is innovative;

[0050] (2) The application introduces the accuracy rate of predicting the occurrence of the ionospheric extended F phenomenon as a judgment basis to judge the accuracy rate of the image generated by the frequency height map prediction model;

[0051] (3) The application provides a method for generating a frequency height map at a future moment based on a convolution time series model ConvGRU, compares the ability of the method to reflect the extended F pattern features in the frequency height map, and explores the best model parameters; six image degradation algorithms are designed and used to train the EDSR super-resolution model; the EDSR super-resolution model is connected to the back of the frequency height map prediction generation model, and a complete frequency height map prediction model is formed. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A flowchart of an ionospheric extended F radar pattern prediction generation method and system based on a time series convolution model and a super-resolution method of the application;

[0053] Figure 2 Module function diagram of system structure of the present application;

[0054] Figure 3 Interface of software prediction expansion F of the system of the present application;

[0055] Figure 4 Function of the system of the present application, viewing large image of input sequence image;

[0056] Figure 5 Function of the system of the present application, viewing large image of comparison of prediction generation image and actual frequency height image;

[0057] Fig. 6(a) is an original frequency height image, and Fig. 6(b) is a frequency height image after cutting and normalization;

[0058] Figure 7 Sequence diagram of frequency height image;

[0059] Figure 8 Quantity statistical diagram of different types of image sequences;

[0060] Figure 9 Result diagram of simulation of frequency height image degradation;

[0061] Figure 10 ConvGRU network structure diagram;

[0062] Figure 11 EDSR network structure diagram;

[0063] Figure 12 Confusion matrix diagram;

[0064] Figure 13 Super-resolution effect comparison diagram. DETAILED DESCRIPTION

[0065] The technical solutions provided by the present application are further described below in combination with embodiments.

[0066] The present application divides the continuously collected frequency height image sequence into short sequences one by one, and uses the short sequence sample set of the frequency height image to train a time series convolution model by using a supervised learning method to generate a future time frequency height image. Then, a super-resolution model is connected behind the time series convolution model to improve the clarity of the generated image. Whether the clear frequency height image output by the image feature recognition appears expansion F and the type of the appearing expansion F are determined, so as to determine the possibility of various types of expansion F phenomena of the ionosphere in the future and the physical mechanism behind it. That is, the present application realizes intelligent prediction and type interpretation of the frequency height image of the future time of the ionosonde.

[0067] Embodiment 1

[0068] As Figure 1 shown, it is a flow chart of the ionospheric spread F radar pattern prediction generation method and system based on the time sequence convolution model and super-resolution method of the application; Figure 2 It is a module function diagram of the system structure of the application; Figure 3 It is the interface of the software prediction of the system of the application; Figure 4 It is the function of the system of the application, viewing the large image of the input sequence image; Figure 5 It is the function of the system of the application, viewing the large image of the comparison between the predicted generation image and the actual height image. The specific embodiment of the application is as follows:

[0069] I. Image normalization preprocessing

[0070] The raw file of the height finder is exported to the SAO-X software to generate the height image, the coordinates are unified to 0-17MHz and 90-800km, and the standard PNG picture is 700 pixels wide and 600 pixels high;

[0071] Crop the picture, remove the surrounding information part and coordinate axis information, and keep the main part of the height image in the picture, that is, take the pixels of each picture from the top left corner (150, 61) to the bottom right corner (645, 520); Then stretch, all the cropped images are unified to 448*448 pixels, which is conducive to the subsequent processing of the time sequence convolution network and the super-resolution network. As shown in Fig. 6(a), it is the original height image; As shown in Fig. 6(b), it is the cropped and normalized height image.

[0072] II. Sample library generation

[0073] 1. Height image sequence sample library

[0074] All data from 2002 to 2015 are regarded as a super-long time sequence data, then according to the time sequence of the height image generation, 10 images are taken as a short sequence, of which 8 are input and 2 are output. Each time, move one picture backward along the time axis, if the interval between the collection time of the new image and the collection time of the last image of the previous short sequence is more than 15 minutes, then take this image as the first image of a new short sequence, and start moving backward again. The purpose of this is to ensure that the time interval between any two adjacent images in all the last intercepted height image short sequences is less than or equal to 15 minutes. Figure 1 It is a display of the height image sequence generated by the data in 2014. Each row has 10 images, of which the first 8 are input to the model, and the last 2 are output, which are used for supervised learning training of the model. As Figure 7 shown, it is a height image sequence diagram;

[0075] After processing all the frequency-height map data from 2002 to 2015 according to the above method, a short sequence set of frequency-height maps is obtained. Then, the last image in the short sequence of frequency-height maps is taken as the basis for judgment. If the extended F phenomenon appears in this image, then this short sequence of frequency-height maps is marked as an extended F sequence of this type; if the extended F phenomenon does not appear, then this short sequence of frequency-height maps is marked as a "background" sequence. As shown in Figure Three , in this way, the number of all extended F sequences can be obtained: frequency type (Freq): 10585; range type (Range): 4857; mixed type (Mix): 14197; enhanced range type (Srange): 6796; subtotal: 36435. The number of background (Others) sequences is: 52051.

[0076] Then, the short sequence set for training of the present application is established. Taking all the extended F sequences as the center, the previous and next 2 sequences adjacent to them are added to the sample set, and the repeated extended F sequences are removed. This is done to ensure that the model can learn the changes of the frequency-height map sequence before and after the occurrence of the extended F.

[0077] The last image of the short sequence is defined as a "background" sequence if it is not an extended F phenomenon. After the above operation and de-duplication, 23657 "background" sequences are added to the short sequence sample library. As shown in Figure 8 , it is a bar chart of the number of different types of image sequences. From Figure 8 , it can be seen that the number of short sequences of the "background" type is too large. In order to solve the problem of imbalance between the number of "background" and "extended F" sequence samples and enable the model to fully learn the spatiotemporal features of the "background" sequence, the number of "background" sequences needs to be appropriately increased. The specific number is determined as twice the maximum number of the four types of sequences (mixed type: 14197), i.e., 28394 "background" sequences are randomly selected from the remaining short sequence set to supplement to the sample set. Therefore, in the sample set, the number of "background" sequences is finally determined as 23657 + 28394 = 52051.

[0078] In general, the finally established sequence sample library contains 88486 short sequences, including:

[0079] Frequency type: 10585

[0080] Range type: 4857

[0081] Mixed type: 14197

[0082] Enhanced range type: 6796

[0083] Background: 52051

[0084] The sequence sample library is divided into training set: validation set in the ratio of 8:2. After the data in 2016 is cut into short sequences in the above manner, it is used as the test set of the model.

[0085] 2. Super-resolution model sample library

[0086] To train the super-resolution model, there must be corresponding clear images and blurred images. The clear frequency-height images can be easily obtained from the height measuring instrument. For blurred images, that is, the blurred images generated by the time series convolution model, a clear-to-blurred degradation model must be established to simulate the predicted images generated by the time series convolution model.

[0087] From all the frequency-height images from 2002 to 2015, 100,000 images were randomly selected, and different noise types and mean / median filter combinations were used for degradation. After testing 32 degradation methods, the final 6 were selected to simulate image degradation based on visual effects. As shown in Figure 9 , the results of simulating frequency-height image degradation are shown; they are:

[0088] ① Salt and pepper noise with a noise density of 0.005 + mean filter with a convolution kernel of 5*5;

[0089] ② Gaussian noise with a variance of 0.005 + mean filter with a convolution kernel of 5*5;

[0090] ③ Salt and pepper noise with a noise density of 0.01 + mean filter with a convolution kernel of 5*5;

[0091] ④ Salt and pepper noise with a noise density of 0.01 + mean filter with a convolution kernel of 7*7;

[0092] ⑤ Poisson noise + mean filter with a convolution kernel of 3*3;

[0093] ⑥ Poisson noise + mean filter with a convolution kernel of 5*5.

[0094] For the selected 100,000 images, a random one is selected from the 6 degradation algorithms and added to each frequency-height image for degradation operation, thus establishing a sample library with 100,000 pairs of low-resolution (LR) and corresponding high-resolution (SR) frequency-height images. The sample library is divided into training set: validation set in the ratio of 8:2.

[0095] III. Time series convolution model ConvGRU

[0096] The time series convolution network is mainly used to process the time series problem of spatial data. The generation of the frequency map has obvious time series, and considering the composition of the RGB three colors and the picture form, the time series convolution model is a very intuitive choice for processing the prediction generation problem. The invention adopts the classic ConvGRU model, and the network structure is as shown in Figure 5 The network is composed of an encoding network and a decoding network. The encoding network has six blocks, each of which is composed of a convolution layer (Conv) and a ConvGRU layer, which are respectively used to compress the input frequency map sequence into a hidden state and learn the space-time features. The decoder network also has six blocks, which are deconvolution (Deconv) layers and ConvGRU layers, respectively, for expanding the hidden state information. The final decoding network output is sent to a 1x1 convolution layer to obtain the predicted frequency map.

[0097] As Figure 10 The ConvGRU network structure diagram is shown in the figure.

[0098] Four, super-resolution model EDSR

[0099] The predicted frequency map generated by the time series convolution network alone can reflect the extended F image features we are interested in, but the image clarity is low, which is far from the real frequency map, and the blurred part of the image is similar to the extended F diffuse image, which interferes with the prediction and reflection of the real extended F; in order to effectively improve the clarity of the generated image, the invention introduces a super-resolution (Super-Resolution, SR) network to solve this problem. The essence of the super-resolution network is to learn the degradation process of a clear (high-resolution) image to a blurred (low-resolution) image or its inverse process. The super-resolution network used in the invention is EDSR, which stands for Enhanced Deep Super-Resolution Network (Enhanced Deep Super-Resolution Network). The network structure is shown in Figure 6. EDSR can well learn the super-resolution relationship between the blurred frequency map and the corresponding clear frequency map, thereby restoring the high-frequency information of the predicted generated image, greatly improving the image clarity. In the method of the invention, EDSR will first extract image features after image input, then learn the high-frequency features of the picture through the activation function layer (RELU layer) and the residual scaling layer, and then perform the final super-clearness reconstruction of the image through the upsample.

[0100] The EDSR network has two advantages of using the ResNet residual idea and removing the BN layer, so that the model is more lightweight, saving 40% of the storage resources, increasing the expressiveness of the model, and effectively processing the relatively blurred picture predicted by the time series model, improving the image clarity.

[0101] As shown in Figure 11 Figure 1 is a schematic diagram of an EDSR network structure.

[0102] Five, training the convolutional model

[0103] 1, loss function

[0104] From the previous work, it is known that only using L2 loss function, the generated image will be more blurred. In order to solve this problem, the present application attempts to use multiple loss functions and their combinations to consider improving the clarity of the generated image. Specifically, the present application uses three different loss schemes to train and compare the ConvGRU model:

[0105] Scheme one: only using L1 Loss

[0106] Scheme two: only using L2 Loss

[0107] Scheme three: 0.15*L1 Loss+0.8*L2 Loss+0.05*PerceptualLoss

[0108] Among them, L1 Loss is also called mean absolute error (MAE), which is the sum of the absolute value of the difference between the target image pixel value and the predicted image pixel value; L2 Loss is also called mean square error (MSE), which is the square sum of the difference between the target image pixel value and the predicted image pixel value; Perceptual Loss is also called perceptual loss, which is the difference between the features of two images extracted by a pre-trained neural network (VGG16, etc.). Unlike traditional pixel-level loss function (MAE, MSE), perceptual loss is not calculated in the pixel domain, but in the image feature domain. Compared with the traditional pixel-level loss function (MAE, MSE), perceptual loss pays more attention to the perceptual quality of the image, and is more in line with the feeling of human eyes to the quality of the image. In order to compare the performance of the three loss functions, other parameters remain the same when training the model, as shown in Table 1 below.

[0109] The shape of each frequency-height map in the input sequence is "448x448x3", which means the height and width are both 448 pixels, and "3" means the input image is a three-channel RGB image. Due to the limitation of GPU memory, the number of sequences input into the model at each training is limited to 8. The number of blocks in the encoder / decoder network is represented by "block_num", and each "block" consists of a ConvGRU layer and a Conv / Deconv layer. During the training process selected by the loss function, block_num = 5. The shape of the convolution kernel is set to 5*5. A maximum of 500 epochs is trained. Adaptive moment estimation (Adam) is used as the optimizer for gradient updates, with a learning rate of 1e-4. At the same time, the ReduceLROnPlateau decay strategy is adopted, where factor is 0.5, indicating that the new learning rate = the original learning rate * factor; patience is 4, i.e. if the index does not improve, the learning rate will be reduced after 4 epochs. To prevent gradient explosion or disappearance, EarlyStopping is also used during training, with patience set to 20 epochs. If the performance of the validation set does not improve for patience epochs, training will be stopped.

[0110]

[0111] After obtaining the frequency-height maps generated by the ConvGRU model with different loss schemes, we can calculate the classification accuracy of the predicted image by identifying and classifying whether a certain type of extended F appears in the predicted frequency-height map and comparing it with the true frequency-height map at the same time. The classification accuracy of different loss schemes is shown in Table 2. Due to the sparsity of the frequency-height map, the model using only L1 Loss is still significantly better than the other two models, with an average accuracy of more than 93.20% for each year. Therefore, in this invention, L1 Loss is finally selected as the loss function for all subsequent training and the final model.

[0112]

[0113] 2. Comparative experiment

[0114] In the ConvGRU network, there are two very important hyperparameters: the number of blocks in the ConvGRU model and the size of the convolution kernel in the ConvGRU layer. We conducted a large number of comparative experiments to determine these two parameters. To better evaluate the performance of the ConvGRU network under different parameter combinations, we use the confusion matrix to analyze the performance of the predicted generated image under the classification model.

[0115] Confusion Matrix, also known as Possibility Matrix or Error Matrix, is a visualization tool for supervised learning, which can be used to quantitatively analyze the performance of the prediction picture in the classification model. The numerical value of the classification result is displayed in a confusion matrix, that is, each column of the confusion matrix represents the predicted class, and the total number of each column represents the number of data predicted as the class. Each row represents the true belonging class of the data, and the total number of data in each row represents the number of data instances of the class. The value in each column represents the number of true data predicted as the class.

[0116] By calculating the confusion matrix of the prediction picture classification result, the secondary indicators such as precision, recall, accuracy and F1 Score of the extended F phenomenon can be further calculated, and the definitions of each indicator required in the evaluation process are as follows.

[0117] True Positive: The true class of the sample is positive, and the result of the model recognition is also positive. (For example: the GT picture is Freq type, and the classification result of the prediction picture is also Freq type).

[0118] False Negative: The true class of the sample is positive, but the model identifies it as negative.

[0119] False Positive: The true class of the sample is negative, but the model identifies it as positive.

[0120] True Negative: The true class of the sample is negative, and the model identifies it as negative.

[0121] Precision: The number of samples with true class positive / the number of samples identified as positive by the model.

[0122] Recall: The number of samples correctly identified by the model as positive / the total number of positive samples.

[0123] Accuracy: The number of correct identifications by the model / the total number of samples (generally, the higher the accuracy of the model,

[0124] the better the effect of the model).

[0125] F1 Score: 2 * Precision * Recall / (Precision + Recall).

[0126] Prediction Accuracy: Prediction accuracy of the extended F phenomenon.

[0127] Overall Accuracy: Prediction accuracy of the extended F classification.

[0128] As Figure 12 shown, it is a schematic diagram of a confusion matrix.

[0129] The following Table 3 lists the results of different super parameter combinations. As can be seen from the experimental results, the network with larger convolution kernel size and deeper layer can achieve better results. Therefore, in the present application, the final time sequence convolution model determines to use 6 blocks, and the size of the convolution kernel is determined to be 7*7.

[0130]

[0131] Six, training super-resolution model

[0132] In order to solve the problem of ionospheric map generated by ConvGRU prediction, we introduce the super-resolution network EDSR, which is expected to solve this problem to some extent. The training of EDSR is based on the super-resolution sample library established before. The super parameters used in the training process are shown in the following Table 4.

[0133] The input shape of each frequency height map is the same as the output of the ConvGRU model, and the output size of the EDSR model is enlarged to "448x448x3". The EDSR network is composed of 32 residual blocks, and is trained using the ADAM optimizer, and the momentum is set to 0.9. This means that when updating the parameters, we will consider the gradient direction of the previous step, and make its influence greater than the current gradient. Usually, this helps to speed up the training of the model and helps the model to better jump out of the local optimal solution. The learning rate is initialized to 1e-4, and the MultiStepLR decay strategy is adopted, and the Decay is 200. That is, after every 200 batch updates, the learning rate will be halved.

[0134]

[0135] Since the graphic content of the ionospheric map is relatively simple, we only use 10 epochs to stop training and achieve good visual improvement effect. The visual effect comparison of the original frequency height map image (GT) and the image before and after super-resolution is as follows Figure 13 shown, is a super-resolution effect comparison diagram.

[0136]

[0137] The training results are further verified on the 2016 test set. The confusion matrix index comparison before and after super-resolution is shown in the following Table 5. After super-resolution, the prediction accuracy of the extended F phenomenon is improved by an average of 0.5%, and the accuracy of the extended F classification is also improved by an average of 0.2%.

[0138] Based on the comprehensive experimental results, the technical route proposed in the application can accurately predict the ionospheric frequency-altitude map of the Hainan height finder and identify the extended F phenomenon, can provide accurate short-term forecast for businesses greatly affected by ionospheric changes, can provide some assistance for studying the occurrence mechanism of the extended F phenomenon of the ionosphere, and has clear practical application value and theoretical application prospect.

[0139] Finally, the application encapsulates the trained model weight parameters as an intelligent prediction software system of the Hainan station ionospheric frequency-altitude map, for the convenience of researchers. The function modules and use interface of the software are as shown in the above Figure 2 、 3 The main functions of the intelligent prediction software of the Hainan station ionospheric frequency-altitude map are as follows:

[0140] 1. The extended F prediction function includes:

[0141] 1) Using the ConvGRU model, the input ionospheric frequency-altitude map sequence of the previous 8 continuous time points is used to predict and generate the ionospheric frequency-altitude map of the next 2 time points.

[0142] 2) Using the EDSR model, the ionospheric map generated by prediction is super-resolution processed to improve the definition and visual effect of the image.

[0143] 2. The extended F identification function includes:

[0144] An accepted extended F identification model is used, the model is initialized, and the default pre-training weight file is loaded to analyze the super-resolution frequency-altitude map and identify whether the extended F phenomenon exists, and give specific classification.

[0145] 3. The result display function includes:

[0146] a) The input frequency-altitude map sequence, the prediction generation result and the actual observation image (if existing) are displayed in the main window

[0147] b) Double-clicking the image of the input sequence part can view the large image, image path and classification model identification result of the selected image. The large image, image path and classification model identification result of the input sequence (8 images) and the prediction image (2 images) can be viewed.

[0148] c) Double-clicking the image of the prediction result part can view the comparison chart, image time and classification model identification result of the selected image and the actual observation image (if existing), and can switch the previous group / next group for viewing.

[0149] 5. The result export function includes:

[0150] a) The frequency-altitude image generated by prediction after super-resolution is output.

[0151] Technical effects of the present application: From the test results and the software functions of the present application, the artificial intelligence prediction software developed by the present application can use the frequency-height graph of the previous 8 consecutive time points (usually the past 2 hours) to predict the frequency-height graph of the next 2 time points (usually the next half hour), and according to the predicted frequency-height graph, judge whether the ionosphere will expand F phenomenon in the future (accuracy rate 94.82%) and the accuracy rate of the expansion F phenomenon (accuracy rate 92.21%). This can not only play a role in predicting short-term ionospheric phenomena for human activities that are greatly affected by ionospheric changes, such as spacecraft launch, satellite navigation, etc., but also provide some auxiliary information for researchers to study the occurrence mechanism of the expansion F phenomenon, which helps researchers better understand the evolution process and physical laws of the ionosphere.

[0152] Embodiment 2

[0153] The ionospheric expansion F radar pattern prediction generation system of the present application comprises:

[0154] The data processing module is used for inputting M continuous frequency-height graphs generated by the ionospheric altimeter before a certain time into the pre-established and trained expansion F radar pattern prediction generation model after preprocessing, to obtain N frequency-height graphs to be generated in the future time.

[0155] The type prediction module is used for analyzing the N frequency-height graphs, predicting whether the ionospheric expansion F phenomenon occurs, and predicting the pattern characteristics reflecting the expansion F type; the expansion F type includes: no expansion F, frequency type FSF, regional type RSF, mixed type MSF and strong regional type SSF.

[0156] Finally, it should be noted that the above embodiments 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 embodiments, 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 predicting ionospheric extension F-graphs based on temporal convolution and super-resolution, comprising: After preprocessing the M consecutive frequency-elevation images generated by the ionospheric altimeter before a certain moment, the images are input into the pre-established and trained extended F radar image prediction generation model to obtain the N frequency-elevation images to be generated at future moments. The network structure of the extended F radar graphic prediction generation model includes, in sequence, a ConvGRU network and an EDSR network; Analyze N frequency-height maps to predict whether the ionospheric extension F phenomenon will occur, and generate graphic features reflecting the type of extension F; the types of extension F include: no extension F, frequency-type FSF, regional type RSF, mixed MSF, and strong regional type SSF; The ConvGRU network includes: an encoder network, a decoder network, and a convolutional layer; The encoding network comprises six blocks, each block including: a convolutional Conv layer and a ConvGRU layer; the convolutional Conv layer is used to compress the input frequency-highness map sequence into a hidden state; the ConvGRU layer is used to learn spatiotemporal features; The decoding network comprises 6 blocks, each block including: a deconvolutional layer and a ConvGRU layer; the deconvolutional layer and the ConvGRU layer are used to expand the hidden state information; The convolutional layer is used to output the predicted frequency-height map; The EDSR network comprises, in sequence: a convolutional layer, an activation function layer, a residual scaling layer, and an upsampling layer; wherein... The convolutional layer is used to extract image features of the frequency height map predicted by the ConvGRU network; The activation function layer and residual scaling layer are used to learn the image features of the frequency height map; The upsampling layer is used to reconstruct the frequency-height map.

2. The ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution according to claim 1, characterized in that, The preprocessing includes: In accordance with the settings requirements, unify the coordinates, pixels, and format of the frequency-height map image; The image is cropped to remove auxiliary information and coordinate axis information around the main graphic, retaining only the main graphic portion of the frequency-height map; the auxiliary information includes text descriptions and color bars; Stretch the image and standardize the pixels.

3. The ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution according to claim 1, characterized in that, The method further includes: training the ConvGRU network; the training process includes: Obtain the frequency-altitude map of the historical output of the ionospheric altimeter and perform normalization preprocessing on the image; Each preprocessed image is labeled and recorded, and the labels assigned to the images are used. M+N images are used as a short sequence, where M images are used as input and N images are used as output. The extended F type of the M+Nth image is used as the extended F type of the output image. Centered on all extended F sequences, the Y sequences preceding and following them are added to the sample set, and duplicate extended F sequences are removed; then, "background" sequences are randomly selected from the remaining short sequence set to supplement the sample set, thus obtaining the first dataset. The first dataset is divided into a training set and a test set, and the ConvGRU network is trained to obtain the trained ConvGRU network.

4. The ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution according to claim 3, characterized in that, The method further includes: training the ConvGRU network using supervised learning; the training process includes: Obtain the frequency-altitude map of the historical output of the ionospheric altimeter and perform normalization preprocessing on the image; X images are randomly selected, and different noise types and combinations of mean or median filtering are used to degrade them. A degradation algorithm is selected to perform degradation operations on each frequency height image, so that the frequency height image is degraded into a blurry image. This image is then paired with the original frequency height image to obtain a sample library containing X pairs of low-resolution and corresponding high-resolution frequency height images, which serves as the second dataset. The second dataset is divided into a training set and a test set. The EDSR network is then trained to obtain the trained ConvGRU network.

5. The ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution according to claim 4, characterized in that, The degradation operation includes, in sequence: ① Add salt and pepper noise with a noise density of 0.005 and apply mean filtering with a 5*5 convolution kernel. ② Gaussian noise with a variance of 0.005 is added, and mean filtering with a 5*5 convolution kernel is applied. ③ Add salt-and-pepper noise with a noise density of 0.01 and apply mean filtering with a 5*5 convolution kernel. ④ Add salt-and-pepper noise with a noise density of 0.01 and apply mean filtering with a 7*7 convolution kernel. ⑤ Add Poisson noise and apply mean filtering with a 3x3 convolution kernel. ⑥ Add Poisson noise and use a mean filter with a 5*5 convolution kernel.

6. The ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution according to claim 3 or 4, characterized in that, The frequency-height map output by the ionospheric altimeter is continuously acquired at set intervals; the historical period includes the years with high, medium, and low solar activity in a whole solar activity cycle, and includes all seasons and local times; the frequency-height map sequence covers the sequence when all types of extended F-types occur, the sequence before and after the occurrence of all types of extended F-types, and the randomly selected sequence unrelated to the extended F phenomenon.

7. A system based on the ionospheric extended F-pattern prediction method based on temporal convolution and super-resolution as described in claim 1, characterized in that, include: The data processing module is used to preprocess the M consecutive frequency-elevation maps generated by the ionospheric altimeter before a certain moment and input them into the pre-established and trained extended F radar image prediction generation model to obtain the N frequency-elevation maps to be generated at future moments. The network structure of the extended F radar image prediction generation model includes, in sequence: a ConvGRU network and an EDSR network; and The type prediction module is used to analyze N frequency height maps, predict whether the ionospheric extension F phenomenon will occur, and generate graphic features that reflect the type of extension F. The types of extension F include: no extension F, frequency-type FSF, regional type RSF, hybrid type MSF, and strong regional type SSF.

Images