SuperDARN radar net-based land and sea state identification method

Abstract

The invention discloses a land and sea state identification method based on a SuperDARN radar network. The method comprises the following steps: step 1, acquiring IQ signal data detected by a SuperDARN radar and preprocessing the IQ signal data; step 2, applying the preprocessed IQ signal data to obtain ACF data, extracting feature data from the ACF data, and constructing a data set; step 3, a clustering model is constructed and trained, the clustering model is formed by combining an auto-encoder and K-means clustering, and the auto-encoder comprises an encoder and a decoder; 4, evaluating the pre-trained clustering model by taking the test set as input; and step 5, carrying out land-sea state identification by using the evaluated clustering model. According to the method, the defects of a method for distinguishing ground echoes and sea surface echoes based on SuperDARN echo data in the prior art are overcome.

Description

Technical Field

[0001] The invention relates to the technical field of space weather forecasting, and in particular to a land and sea state recognition method based on a SuperDARN radar network. Background Art

[0002] The ionosphere is an important part of the near-Earth space environment. Many modern radio systems are inseparable from the state of the ionosphere, such as high-frequency long-distance radio communication systems, high-frequency radar detection systems, and satellite navigation systems. Changes in the state of the ionosphere may seriously affect the normal operation of these systems. The SuperDARN radar has achieved great success in studying the magnetosphere, ionosphere, thermosphere, and mesosphere. Its main goal is to study ionospheric inhomogeneities and global convection in the northern and southern hemispheres. Therefore, scholars using SuperDARN data are more concerned with the echoes reflected by the inhomogeneities in the E and F regions of the ionosphere, that is, the ionospheric echoes. However, its rich ground and sea surface scattered echoes have not been fully explored and applied, and these echoes can be used to divide the coastline and invert ionospheric parameters in real time. Therefore, in terms of the classification of SuperDARN data radar echoes, how to obtain different types of echoes is very important.

[0003] The study of SuperDARN radar echo data classification shows that standard processing techniques for radar data are effective in the case of a single source, but the situation becomes more complicated when the signal contains multiple sources, which are either the additional existence of ground scattering or multiple ionospheric lines associated with strongly non-uniform velocity fields. Scholars have previously proposed using a high-resolution spectrum analysis method, namely the multi-signal classification (MUSIC) method, to separate multiple echoes with different speeds and spectrum widths, but it is difficult to separate low-speed ionospheric echoes from ground echoes. Another scholar proposed a method for classifying backscatter to separate ionospheric scattering and ground scattering, but this is a point-by-point classification of data points without considering clusters, which is prone to bias. Another depth-first search model was proposed, which performs a depth-first search on the data in time / space, and classifies the clusters into ionospheric scattering (IS) and ground scattering (GS) according to the speed ratio. Compared with Blanchard, it reduces the bias and allows more low-speed ionospheric scattering to be correctly classified. However, this model uses boxcar filtering, which reduces the spatial and temporal resolution. Other scholars have proposed a model based on elevation angle and virtual height. This model requires relatively reliable elevation angle measurements, which are usually not available and the conditions taken into consideration are insufficient.

[0004] As an important branch of artificial intelligence, deep learning has made significant breakthroughs in recent years. Thanks to the development of the Internet and the Internet of Things, massive data has been generated. This big data environment makes deep learning based on statistical models more advantageous than traditional algorithms. In the field of space weather modeling, it has been used more and more frequently in recent years. For example, deep learning technology has been applied to the inversion of ionospheric parameters, the construction and prediction of ionospheric convective potential models, and the construction of ionospheric transpolar electric field models, and has achieved good results. This undoubtedly shows the great potential of deep learning technology in the field of space weather forecasting. Thanks to its powerful feature extraction and representation capabilities, deep learning has received widespread attention in combination with clustering tasks, and deep clustering has become one of the main research directions.

[0005] Based on the above status quo and the deficiencies in the prior art in the research of algorithms for distinguishing ground echoes from sea surface echoes based on SuperDARN echo data, the present invention proposes a method for identifying ground and sea states based on the SuperDARN radar network. Summary of the invention

[0006] In order to make up for the deficiency of the method for distinguishing ground echo and sea echo based on SuperDARN echo data in the prior art, the present invention proposes a method for identifying ground and sea states based on the SuperDARN radar network.

[0007] In order to solve the above technical problems, the technical solution of the present invention is:

[0008] A method for identifying land and sea conditions based on the SuperDARN radar network comprises the following steps:

[0009] Step 1: Obtain and pre-process the IQ signal data detected by the SuperDARN radar;

[0010] Step 2: Apply the preprocessed IQ signal data to obtain the raw autocorrelation function data (RAWACF), extract feature data from the RAWACF data, and construct a data set, dividing the data set into a training set and a test set;

[0011] Step 3: Build and train a clustering model. The clustering model is a combination of an autoencoder and k-means clustering. The autoencoder includes an encoder and a decoder. The autoencoder selects 32 as the dimension of the low-dimensional representation, builds an input layer, two fully connected layers, and uses ReLU as the activation function. A fully connected layer is created in the decoder, and Sigmoid is used as the activation function. The clustering model adds a K-means clustering loss function.

[0012] Step 4: Use the test set as input to evaluate the pre-trained clustering model;

[0013] Step 5: Apply the evaluated clustering model to identify land and sea states.

[0014] Preferably, in step 2, the IQ signal data is processed by an autocorrelation function to obtain RAWACF data.

[0015] Preferably, the IQ signal data is processed to obtain an autocorrelation function:

[0016] The complex IQ signal can usually be expressed as x[n]=I[n]+jQ[n], where j is an imaginary unit. For a long period of data, the local autocorrelation function can be obtained by overlapping sliding windows. Its autocorrelation function R x [k] can be expressed as: Since it is a complex signal, x[n]·x * [n+k] is the product of the real part and the imaginary part and the sum;

[0017] For discrete-time signals, it can be decomposed into:

[0018]

[0019] Where k represents the delay or lag time, N represents the total length of the signal, I[n+k] represents the real signal of the signal x[n] at index n+k, and Q[n+k] represents the imaginary signal of the signal x[n] at index n+k.

[0020] The time delay k usually starts from 0 and increases to half of the signal length (or according to actual needs). The autocorrelation value of each time delay k is calculated in turn, and finally the autocorrelation function R is obtained. x [k] Time delay k and autocorrelation function R x The relationship between [k] constitutes the RAWACF data.

[0021] Preferably, the autocorrelation function R x [k] Extract feature data, wherein the feature data includes time, Doppler velocity, Doppler spectrum width, echo intensity, frequency, velocity error, spectrum width error, beam, range gate, elevation angle, initial phase, and initial power, forming a data set with 12 features.

[0022] The extraction method of the characteristic data Doppler velocity and Doppler spectrum width is:

[0023] First, exclude the ACF lags that are contaminated by cross clustering and pulse overlap and find the lag zero noise R n [0] and R n (τ≠0), when R n (τ≠0) exceeds the set threshold, the noise is considered to have been subtracted from the RAWACF data;

[0024] Given an exponential model or a Gaussian model, a linear function or a quadratic function is fitted to the logarithm of the RAWACF power to estimate the Doppler spectrum width W;

[0025] Calculate the Doppler velocity using the following formula:

[0026] φ(τ)=ω D τ,

[0027] ω D =2kV LoS ,

[0028]

[0029] In the formula, is the RAWACF phase slope.

[0030] Preferably, the threshold is set as follows: expected fluctuation level σ R 60% of which:

[0031]

[0032] Where R(0) is the power at zero lag (i.e. the strongest signal part) in RAWACF, and N avg is the data average, that is, the average number of radar sampling points considered. Usually, this value is related to factors such as the average number of radar samples and observation time.

[0033] Preferably, the data set is divided into a data set and a test set in a 5:5 ratio.

[0034] Preferably, the clustering function is trained by selecting an Adam optimizer, selecting a combination of the reconstruction loss of the autoencoder and the K-means clustering loss as a comprehensive loss function, combining the autoencoder with the K-means clustering model, alternately optimizing the parameters and clustering centers of the autoencoder network model during training, first performing K-means clustering on the low-dimensional representation of the features passed through the encoder to obtain the clustering centers and clustering losses, optimizing the network parameters of the autoencoder through the clustering losses and the reconstruction losses of the autoencoder, and then obtaining a new low-dimensional representation of the features, and then performing K-means clustering on the new low-dimensional feature representation to obtain the optimized clustering centers and clustering losses. This process is repeated until convergence is reached.

[0035] Preferably, in step 4, three evaluation indicators, namely, the within-cluster sum of squares WSS, the overall silhouette coefficient and the adjusted Rand index ARI, are used to evaluate the model.

[0036] The present invention has the following characteristics and beneficial effects:

[0037] Using the above technical solution, the Fourier transform FFT algorithm, Burg power spectral density estimation algorithm, and Welch power spectral density estimation algorithm are used to extract and process the features of FITACF data, and preliminarily distinguish between ground scattered echoes and sea surface scattered echoes. On this basis, deep learning technology is pioneeringly used to extract effective information on feature classification after SuperDARN radar detection data fitting. Deep learning technology and SuperDARN radar observation data are pioneeringly applied to land and sea state identification with a high distinction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 It is a schematic diagram and block diagram of the flow chart of an embodiment of the present invention.

[0040] Figure 2 Schematic diagram of the principle of the clustering model in an embodiment of the present invention.

[0041] Figure 3 It is a schematic diagram of the clustering effect of an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram of the identification effect in the open source Radar Software Toolkit (RST) radar data processing tool.

[0043] Figure 5 Schematic diagram of the traditional method. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0045] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0046] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0047] The present invention provides a method for identifying land and sea conditions based on the SuperDARN radar network, such as Figure 1 and Figure 2 As shown, the following steps are included:

[0048] Step 1: Obtain and pre-process the IQ signal data detected by the SuperDARN radar.

[0049] In this embodiment, the observation data and IQ signal data of the SuperDARN radar FIR station in January 2021 are obtained.

[0050] Step 2: Apply the preprocessed IQ signal data to obtain RAWACF data, extract feature data from the RAWACF data, and construct a data set, which is divided into a training set and a test set at a ratio of 5:5.

[0051] Specifically, the method for calculating RAWACF using the original IQ signal data of the SuperDARN radar is:

[0052] The complex IQ signal can usually be expressed as x[n]=I[n]+jQ[n], where j is an imaginary unit. For a long period of data, the local autocorrelation function can be obtained by overlapping sliding windows. Its autocorrelation function Rx [k] can be expressed as: Since it is a complex signal, x[n]x * [n+k] is the product of the real part and the imaginary part and the sum;

[0053] For discrete-time signals, it can be decomposed into:

[0054]

[0055] Where k represents the delay or lag time, N represents the total length of the signal, I[n+k] represents the real signal of the signal x[n] at index n+k, and Q[n+k] represents the imaginary signal of the signal x[n] at index n+k.

[0056] The time delay k usually starts from 0 and increases to half of the signal length (or according to actual needs). The autocorrelation value of each time delay k is calculated in turn, and finally the autocorrelation function R is obtained. x [k] Time delay k and autocorrelation function R x The relationship between [k] constitutes the RAWACF data.

[0057] Then, RAWACF data is used to extract feature data, including time, Doppler velocity, Doppler spectrum width, echo intensity, frequency, velocity error, spectrum width error, beam, gate, elevation angle, initial phase, and initial power, forming a data set with 12 feature data.

[0058] Specifically, time information is usually stored in the metadata of the IQ file, usually recorded in UTC (Coordinated Universal Time) format. Each record usually contains a timestamp indicating the observation time corresponding to the record, and the time information can be obtained by parsing its header metadata.

[0059] For beams and range gates, the SuperDARN radar works with a specific beam scanning strategy. The beam direction is adjusted in turn for each observation, and the beam number usually increases from 0. The range gate corresponds to the range segment on the radar beam. Both are stored in the metadata part of the file. The beam and range gate information can be directly obtained by reading the metadata part of each data record.

[0060] For noise processing, the RAWACF lag term caused by cross-interference and pulse overlap is eliminated to extract the zero-lag noise R n (0) and non-zero lag noise R n (τ≠0), where τ represents the delay or lag time. nIf the amplitude of (τ≠0) exceeds the set threshold, it can be considered that the noise in the autocorrelation function data has been effectively suppressed. Otherwise, this part of the noise is considered as the noise ACF and will be deducted from all RAWACF.

[0061] In this embodiment, the threshold is set as follows: expected fluctuation level σ R 60% of which:

[0062]

[0063] Where R(0) is the power at zero lag (i.e. the strongest signal part) in RAWACF, and N avg is the data mean, i.e. the average number of radar sampling points considered.

[0064] It should be noted that this value is usually related to factors such as the average sampling number and observation time of the radar.

[0065] If the noise ACF is successfully subtracted, the zero-lag noise R is recalculated. n (0) and non-zero lag noise R n (τ≠0), and re-perform these calculations based on the cleaned RAWACF data;

[0066] Autocorrelation function R x There is a Fourier transform relationship between [k] and the spectrum of the signal. By performing Fourier transform on the autocorrelation function, the power spectrum of the signal (i.e., Doppler spectrum) can be obtained, which is expressed as follows:

[0067]

[0068] In the formula, S x (f) is the power spectrum of the signal, f is the frequency, R x (τ) is the autocorrelation function.

[0069] The spectrum calculated from RAWACF is fitted into an exponential model or a Gaussian model through a linear function and a quadratic function, respectively, and the Doppler spectrum width W is estimated by calculating the 3dB bandwidth. The calculation formula is as follows;

[0070]

[0071] B 3dB =2f0 ln2≈1.386f0

[0072]

[0073] Where S e (f) is the exponential model of the spectrum, S g(f) is the Gaussian model of the spectrum, A is a constant (related to the amplitude), f0 is the center frequency, representing the center position of the signal's spectrum. σ is the standard deviation, which determines the degree of spectrum expansion and is directly related to the spectrum width. 3dB It is the frequency range corresponding to when the spectrum amplitude drops to half of its maximum value, which can be used to represent the Doppler spectrum width.

[0074] Calculate the spectral width error and fit the autocorrelation function using the least squares method to obtain the spectral width estimate

[0075]

[0076] According to the error propagation law, the spectrum width error Δσ is:

[0077]

[0078] Calculate the Doppler velocity ω D For the SuperDARN radar, when the signal frequency moves down, the phase slope of the autocorrelation function is determined only by the Doppler shift. At the same time, the Doppler shift is related to the line-of-sight component V of the plasma drift velocity. LoS It is proportional to the signal wave number k, and the calculation formula is as follows:

[0079] φ(τ)=ω D τ,

[0080] ω D =2kV LoS ,

[0081]

[0082] Where τ is the lag time, ω D is the Doppler velocity, k is the signal wave number, V LoS is the line-of-sight component of the plasma drift velocity, is the RAWACF phase slope.

[0083] Calculate the echo strength, which is the energy or power of the signal returned by the target, which is usually related to the target's reflection characteristics (such as radar cross-section), distance attenuation, and transmission power. The echo strength is determined by the hysteresis zero point R n (0), and then the echo intensity is normalized to remove the influence of system parameters. The formula for calculating echo intensity is:

[0084]

[0085] Where NES (Normalized Echo Strength) is the normalized echo strength, |R n (0)| is the amplitude of the lag zero point, PTx is the radar transmit power.

[0086] Calculate the Doppler frequency, which is the frequency shift caused by the relative motion between the target and the radar. x (f) Signal power spectrum. The frequency corresponding to the main peak position is the Doppler frequency.

[0087] Calculate the elevation angle. The SuperDARN radar uses an array of multiple antennas to calculate the elevation angle by comparing the phase difference of the received signals. The SuperDARN radar calculates the autocorrelation function R for each pair of antennas. ij (τ), where i and j are antenna indices and τ is the delay. First, extract the phase of the complex signal. For each pair of antennas (i, j), calculate the phase of the autocorrelation function φ ij , and get the phase information:

[0088] φ ij =arg(R ij (τ)

[0089] Next, we need to find the phase difference between the two channels. We select two receiving antennas, such as A and B, with a distance d between them. We calculate the phase difference between the two antennas:

[0090] Δφ AB =φ B -φ A

[0091] For multiple lag times τ, the Δφ at different lag times can be averaged to reduce noise interference. According to the geometric arrangement of the antenna and the measured phase difference, the elevation angle is calculated using the formula:

[0092]

[0093] Where θ is the elevation angle, λ is the wavelength of the antenna, and d is the distance between the two antennas.

[0094] Calculate the initial phase φ0, which is the phase of the radar signal at the reference time point (usually zero lag time). The calculation formula is as follows:

[0095]

[0096] Where φ0 is the initial phase, R im (0) is the imaginary part of the autocorrelation function at zero lag point, R re (0) Real part of the autocorrelation function at zero lag.

[0097] Calculate the initial frequency f0, extract the phase of each lag time point in RAWACF, fit the relationship between the phase φ(τ) and the lag time τ, and obtain the slope k. Use the slope k to calculate the initial frequency. The calculation formula is as follows:

[0098]

[0099] φ(τ)=k·τ+φ0

[0100]

[0101] In the formula, R im (τ) is the imaginary part of the autocorrelation function, R re (τ) is the real part of the autocorrelation function, φ0 is the initial phase, and f0 is the initial frequency.

[0102] Calculate the speed error, calculate the frequency error through the influence of phase noise on frequency estimation, and calculate the speed error through the frequency error. The calculation formula is as follows:

[0103]

[0104] Where Δv is the velocity error, λ is the radar signal wavelength, and σ φ is the phase noise, Δτ is the interval of the lag time points, the larger the interval, the smaller the speed error. For multiple lag time points, the weighted average method is used to synthesize the error.

[0105] Step 3: Build and train a clustering model, such as Figure 2 As shown, the clustering model is a combination of autoencoder and K-means clustering. The autoencoder includes an encoder and a decoder, wherein the encoder selects 32 as the dimension of the low-dimensional representation, constructs an input layer, two fully connected layers, and uses ReLU as the activation function. A fully connected layer is created in the decoder, and the activation function uses Sigmoid. K-means clustering is performed on the low-dimensional representation of the features passed through the encoder to obtain the cluster center and cluster loss. The network parameters of the autoencoder are optimized through the cluster loss and the reconstruction loss of the autoencoder, and then a new low-dimensional representation of the features is obtained. The new low-dimensional feature representation is then subjected to K-means clustering to obtain the optimized cluster center and cluster loss, and this is repeated until convergence.

[0106] It is understandable that since the dataset contains a lot of feature data, using autoencoders and K-means clustering for classification will have a good effect. The main idea of ​​combining autoencoders with K-means clustering is: first use the autoencoder to learn the low-dimensional representation of the data, and then use K-means clustering to cluster in the low-dimensional space. In order to make the clustering effect more ideal, a joint optimization strategy is used to simultaneously optimize the reconstruction error of the autoencoder and the error of the K-means clustering.

[0107] The encoder of the autoencoder is constructed as follows; 32 is selected as the dimension of the low-dimensional representation, and an input layer is constructed with the original dimension of the data. The first fully connected layer is used as an intermediate layer to compress the input data into 512 dimensions, and the activation function uses ReLU, which allows the network to better learn nonlinear features. The second fully connected layer further compresses the data to 32 dimensions, and the activation function still uses ReLU.

[0108] The decoder construction process of the autoencoder is as follows: the first fully connected layer expands the 32-dimensional representation to a 512-dimensional one, using the ReLU activation function to introduce nonlinearity. The second fully connected layer restores the data to the original input dimension, and the activation function uses Sigmoid. The Sigmoid function can limit the output to the range of [0,1], which is suitable for binary cross entropy loss.

[0109] The loss function uses joint optimization to combine the reconstruction loss of the autoencoder and the loss based on K-means clustering. The reconstruction loss of the autoencoder uses binary cross entropy loss. The goal of the autoencoder is to minimize the reconstruction error, while the binary cross entropy is suitable for cases where the input data is binary or normalized data. The clustering loss of K-means is the sum of the squares of the distances from the data points in the cluster to the cluster center.

[0110]

[0111] In the formula, is the reconstruction loss of the autoencoder, is the loss based on K-means clustering, α is the weight of clustering loss, which controls the balance between autoencoder reconstruction loss and K-means clustering loss. X is the original input data, X′ is the output of the autoencoder, x i is the i-th data point, is the data point x i The cluster center of the cluster to which it belongs, and N is the total number of data points.

[0112] The optimizer is the Adam optimizer because it combines momentum and adaptive learning rate adjustment, and can effectively handle sparse gradients and large-scale data. Adam's update rule adjusts the learning rate of each parameter based on the mean and variance of the gradient.

[0113] During the training process, first only the autoencoder is trained, with the goal of minimizing the reconstruction loss so that the network can effectively compress the data into low-dimensional representations and reconstruct the original data from these representations. After the autoencoder is trained, its encoder part is used to obtain the low-dimensional representation of each sample. These low-dimensional representations will be used as input for K-means clustering. The low-dimensional representations are clustered using the K-means algorithm. K-means will look for cluster centers in these low-dimensional representation spaces and assign samples to the closest clusters. During the training process, the model parameters and cluster centers are optimized alternately. The cluster centers are first updated by the embedding calculated by the network, and then the cluster centers are fixed and the network parameters are updated by minimizing the clustering loss. This process is repeated until convergence is reached. Finally, a clustering model for land-sea state recognition is obtained.

[0114] Step 4: Use the test set as input to evaluate the pre-trained clustering model.

[0115] Specifically, in this embodiment, three evaluation indicators are used to evaluate the model, namely, the intra-cluster sum of squares WSS, the overall silhouette coefficient, and the adjusted Rand index ARI. The calculation formulas of WSS, CH, and ARI are as follows:

[0116]

[0117]

[0118] In the WSS formula, K represents the number of clusters, C k represents the sample set in the kth cluster, μ k is the mean of the kth cluster, x i is the sample within the cluster, the overall silhouette coefficient, n is the total number of samples, s(i) is the silhouette coefficient of sample i, and the value of s(i) is between -1 and 1, indicating the balance between the closeness and separation of the data point relative to the cluster. ARI, RI is the Rand Index, which calculates the classification consistency of all sample pairs, and E[RI] is the expected value of the classification consistency of random selections.

[0119] Step 5: Apply the evaluated clustering model to identify land and sea states.

[0120] The test data set is clustered using the land-sea state recognition clustering model to obtain classification labels at different longitudes and latitudes. The labels of adjacent areas are then compared, the intersection lines of different labels of adjacent areas are marked, and the coastline is drawn.

[0121] For example, Figure 3The figure shows the schematic diagram of the clustering results. The figure shows the fan-shaped distribution of FIR stations, the green part represents the actual land area, and the white part represents the actual ocean area. In the clustering results, the blue part corresponds to the area classified as ocean scattered echo, and the brown part corresponds to the area classified as land scattered echo. Overall, the clustering results are consistent with the actual situation, and the recognition accuracy of ground and sea echo data within the detection range of a single radar reaches 80%.

[0122] In the open source Radar Software Toolkit (RST) radar data processing tool, data is mainly divided into ionospheric echo data and ground echo data, where ground echo data includes ground scattered echo and sea surface scattered echo. Figure 4 As shown, the data used is the FIR station data from 22:00 to 24:00 on January 9, 2021. In the figure, the gray area represents the ground echo data, and the colored area represents the ionospheric echo data according to the Doppler velocity. Compared with the existing classification algorithm, the present invention focuses on further subdividing the ground echo data (including ground scattered echo and sea surface scattered echo), dividing it into ground scattered echo and sea surface scattered echo, and realizing the accurate identification of the ground and sea state.

[0123] like Figure 5 As shown, a data set is constructed based on the IQ signal data and FITACF data observed by the SuperDARN radar; feature selection is performed according to the traditional method analysis, and the ground scattered echo and the sea scattered echo are distinguished by spectral analysis algorithms such as Fourier transform, Burg power spectrum density estimation algorithm, and Welch power spectrum density estimation algorithm; the training data set is trained and adjusted according to the deep learning network model to obtain a clustering model; the test data set is used as the input data of the clustering model to obtain a clustering result, and the clustering result is output. The present invention utilizes the spectral difference between the sea scattered echo and the ground scattered echo and the nonlinear mapping ability of the deep learning network and the information extraction ability of the historical data to realize the recognition of the ground and sea state based on the SuperDARN radar observation data.

[0124] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A method for identifying land and sea conditions based on the SuperDARN radar network, characterized in that: The steps include: Step 1: Obtain and pre-process the IQ signal data detected by the SuperDARN radar; Step 2: Apply the preprocessed IQ signal data to obtain RAWACF data, extract feature data from the RAWACF data, and construct a data set, dividing the data set into a training set and a test set; Step 3, construct and train a clustering model, the clustering model is a combination of an autoencoder and K-means clustering, the autoencoder includes an encoder and a decoder, the autoencoder selects 32 as the dimension of the low-dimensional representation, constructs an input layer, two fully connected layers, and uses ReLU as the activation function. A fully connected layer is created in the decoder, and the activation function uses Sigmoid. The loss function of the clustering model includes the reconstruction loss of the autoencoder and the K-means clustering loss; Step 4: Use the test set as input to evaluate the pre-trained clustering model; Step 5: Apply the evaluated clustering model to identify land and sea states.

2. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 1, characterized in that: In the step 2, the IQ signal data is processed by an autocorrelation function to obtain RAWACF data.

3. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 2, characterized in that: In step 2, the IQ signal data is processed to obtain an autocorrelation function: The complex IQ signal data can usually be expressed as x[n]=I[n]+jQ[n], where j is an imaginary unit. For a segment of IQ signal data, the local autocorrelation function is obtained by overlapping sliding windows. The autocorrelation function R x [k] can be expressed as: Since it is a complex signal, x[n]·x * [n+k] is the product of the real part and the imaginary part and the sum; For discrete-time signals, it can be decomposed into: Where k represents the delay or lag time, N represents the total length of the signal, I[n+k] represents the real part of the signal x[n] at index n+k, and Q[n+k] represents the imaginary part of the signal x[n] at index n+k; The time delay k starts from 0 and increases to half of the signal length. The autocorrelation value of each time delay k is calculated in turn, and finally the autocorrelation function R is obtained. x [k], time delay k and autocorrelation function R x The relationship between [k] constitutes the RAWACF data.

4. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 3, characterized in that: Feature data is extracted from the RAWACF data, and the feature data includes time, Doppler velocity, Doppler spectrum width, echo intensity, frequency, velocity error, spectrum width error, beam, gate, elevation angle, initial phase, and initial power.

5. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 4, characterized in that: The method for extracting the feature data is: For the extraction of Doppler velocity and Doppler spectrum width: First, exclude the ACF lags that are contaminated by cross clustering and pulse overlap and find the lag zero noise R n [0] and R n (τ≠0), when R n (τ≠0) exceeds the set threshold, the noise is considered to have been subtracted from the RAWACF data; Given an exponential model or a Gaussian model, a linear function or a quadratic function is fitted to the logarithm of the RAWACF power to estimate the Doppler spectrum width W; Calculate the Doppler velocity using the following formula: φ(τ)=ω D t, oh D =2kV LoS , In the formula, is the RAWACF phase slope; For the extraction of beam and range gate: SuperDARN radar works with a specific beam scanning strategy. The beam direction is adjusted in sequence for each observation. The beam number usually starts from 0 and increases. The range gate corresponds to the range segment on the radar beam. Both are stored in the metadata part of the file. The beam and range gate information can be directly obtained by reading the metadata part of each data record. For the extraction of Doppler spectrum: there is a Fourier transform relationship between the autocorrelation function ACF and the spectrum of the signal. By performing Fourier transform on the autocorrelation function, the Doppler spectrum of the signal can be obtained. For the calculation of Doppler spectrum width: the spectrum calculated from RAWACF is fitted into an exponential model or a Gaussian model through a linear function and a quadratic function respectively, and the Doppler spectrum width is estimated by calculating the 3dB bandwidth; For the spectrum width error, the autocorrelation function is fitted by the least square method to obtain the spectrum width estimate, and then according to the error propagation law, the spectrum width error; For echo intensity: the echo intensity is determined by the hysteresis zero point R n (0) and then the echo intensity is normalized to remove the influence of system parameters; Calculate the Doppler frequency. The Doppler frequency is the frequency offset caused by the relative motion between the target and the radar. The frequency corresponding to the main peak position is determined by the above Doppler spectrum, which is the Doppler frequency. Calculate the elevation angle. SuperDARN radar uses an array of multiple antennas to calculate the elevation angle by comparing the phase difference of the received signal. SuperDARN radar calculates the autocorrelation function R for each pair of antennas. ij (τ), where i and j are antenna indices and τ is the delay. First, the phase of the complex signal is extracted. For each pair of antennas (i, j), the phase of the autocorrelation function is calculated by ij , and get the phase information: f ij =arg(R ij (t)) Next, we need to find the phase difference between the two channels. We select two receiving antennas with a spacing of d and calculate the phase difference between the two antennas: Df AB =φ B -f A For multiple lag times τ, the Δφ at different lag times can be averaged to reduce noise interference. According to the geometric arrangement of the antenna and the measured phase difference, the elevation angle is calculated using the formula: Where θ is the elevation angle, λ is the wavelength of the antenna, and d is the distance between the two antennas; Calculate the initial phase, that is, the phase of the radar signal at the reference time point; Calculate the initial frequency, extract the phase of each lag time point in RAWACF, fit the relationship between phase and lag time, obtain the slope, and use the slope k to calculate the initial frequency; Calculate the speed error, calculate the frequency error through the effect of phase noise on frequency estimation, and calculate the speed error through the frequency error.

6. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 5, characterized in that: The threshold is set as follows: expected volatility level σ R 60% of which: Where R(0) is the power at zero lag (i.e. the strongest signal part) in RAWACF, and N avg is the data mean, i.e. the average number of radar sampling points considered.

7. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 1, characterized in that: The dataset is divided into a dataset and a test set at a ratio of 5:

5.

8. A method for identifying land and sea conditions based on SuperDARN radar network according to any one of claims 1 to 7, characterized in that: The training method of the clustering function is as follows: selecting the Adam optimizer, selecting the reconstruction loss of the autoencoder and the K-means clustering loss as the loss function, combining the autoencoder and the K-means clustering model, alternately optimizing the parameters and clustering centers of the autoencoder network model during training, first performing K-means clustering on the low-dimensional representation of the features passed through the encoder to obtain the clustering centers and clustering losses, optimizing the network parameters of the autoencoder through the clustering losses and the reconstruction losses of the autoencoder, and then obtaining a new low-dimensional representation of the features, and then performing K-means clustering on the new low-dimensional feature representation to obtain the optimized clustering centers and clustering losses, and repeating this process until convergence is reached.

9. The method for identifying land and sea conditions based on the SuperDARN radar network according to claim 8, characterized in that: In step 4, the model is evaluated using three evaluation indicators: intra-cluster sum of squares WSS, overall silhouette coefficient, and adjusted Rand index ARI.

Images