CN114780911B - Ocean wide swath distance defuzzification method based on deep learning - Google Patents

Abstract

The invention discloses a deep learning-based ocean wide swath distance defuzzification method, which comprises the steps of firstly obtaining echo signals of a satellite-borne SAR system, constructing reference signal vectors according to radar parameters, and performing point multiplication on the echo signals by conjugation of the reference signal vectors to obtain a data matrix after distance pulse pressure; then carrying out inverse Fourier transform and Fourier transform; then constructing a Doppler compensation function, and performing dot multiplication with the matrix after Fourier transformation; and finally, reconstructing a deep neural network, taking the obtained vector as input and a label, and training to realize the range de-blurring of the ocean wide swath. The invention can effectively solve the problem of the distance defuzzification of the pattern swath of the spaceborne SAR system, obtain the ocean wide swath information without the distance fuzzification, and achieve the purpose of imaging the sea surface moving target without the distance fuzzification SAR.

Description

A deep learning-based distance defuzzification method for wide ocean mapping swaths

Technical field

The invention belongs to the technical field of radar signal processing, and specifically relates to a method for distance defuzzification of wide ocean surveying and mapping swaths.

Background technique

For spaceborne SAR systems, when high-resolution imaging of a certain scene is required, this can be achieved by increasing the revisit period of the satellite. However, the increase in satellite revisit cycles comes with significant costs. In order to reduce the revisit period of the satellite, the spaceborne SAR system can be used to transmit high pulse repetition frequency (PRF) chirp signals. Since the system emits high PRF signals, more echo information of scene targets can be obtained, thereby achieving high-resolution imaging of spaceborne SAR radar.

However, transmitting high PRF signals will also bring about the problem of distance ambiguity. As shown in Figure 1, when the difference in echo delays of different imaging scenes in the surveying zone is equal to an integer multiple of the pulse repetition period (PRT), the received wide Echoes from surveying and mapping belts can produce distance blur.

For solving range ambiguity problems in wide surveying swaths, the most commonly used methods are the look-up table method and the method of transmitting orthogonal coded signals. For the method of transmitting orthogonal encoded signals, commonly used quadrature phase number signals and orthogonal frequency encoded signals are used. The accuracy of defuzzification using quadrature phase encoding signals is difficult to meet the resolution requirements, while the quadrature frequency encoding signals can meet the resolution requirements, but the orthogonal encoding signals are more sensitive to Doppler. However, these methods all need to meet the condition of high echo signal-to-noise ratio.

In actual situations, the echo signal-to-noise ratio of ocean targets is relatively low, generally 0dB. At the same time, most spaceborne SAR systems obtain wide mapping swath information by transmitting high PRF signals, so they face the problem of range ambiguity in low signal-to-noise ratio echoes in wide mapping swaths.

Contents of the invention

In order to overcome the shortcomings of the existing technology, the present invention provides a deep learning-based ocean wide mapping swath distance defuzzification method. First, the echo signal of the spaceborne SAR system is obtained, a reference signal vector is constructed according to the radar parameters, and the echo signal is Dot multiply the conjugate of the reference signal vector to obtain the data matrix after distance pulse pressure; then perform the inverse Fourier transform and Fourier transform; then construct the Doppler compensation function and dot it with the matrix after Fourier transform Multiply; finally, a deep neural network is constructed, and the obtained vector is used as input and label. After training, the distance defuzzification of the wide ocean surveying swath is achieved. The invention can effectively solve the problem of distance deambiguation of the surveying and mapping swath of the spaceborne SAR system, and obtain the ocean wide surveying and mapping swath information without range ambiguity, so as to achieve the purpose of SAR imaging of moving targets on the sea surface without range ambiguity.

The technical solution adopted by the present invention to solve the technical problems includes the following steps:

Step 1: Obtain the echo signal of the spaceborne _SAR system. The echo signal is an n _r , n2); n _r represents the number of distance points, n _a represents the number of azimuth points;

Step 2: Construct a reference signal vector based on radar parameters s_r( _{n1 ) is an n r} /> Δf is the distance frequency domain interval,/> n1＝0,1,...,n _r -1;

Step 3: Dot-multiply each column of s(n1,n2) by the conjugate of the reference signal vector s_r(n1) to obtain the data matrix s(f _r ,x _n2 ) after distance pulse pressure; where x _n2 represents Azimuth time domain coordinates, L represents the synthetic aperture length, n2=0,1,...,n _a -1;

Step 4: Perform inverse Fourier transform on the matrix s(f _r ,x _n2 ) by column, and save the result in the matrix s(n _r , _na );

Step 5: Perform Fourier transform on the matrix s(n _r , _na ) row by row, and store the result in the s(n _r , _fa ) matrix; where _fa represents the azimuth frequency domain coordinate, PRF is the azimuth adopted frequency, Δf _a is the azimuth frequency domain interval,

Step 6: Construct Doppler compensation function based on radar parameters s_h(n1) is a vector _of n _r The operating wavelength, g represents the acceleration of gravity;

Step 7: Perform dot multiplication of the conjugate of s(n _r , f _a ) and the Doppler compensation function s_h(n1) vector, and perform inverse FFT processing on the multiplication result row by row to obtain the intrapulse Doppler compensation The data matrix ss(n _r ,n _a );

Step 8: According to the spatial filtering theory, construct a matrix Ω composed of the target scattering coefficients, where Ω is an n _r × n _a- dimensional matrix;

Step 9: Divide the matrix ss(n _r , _na ) into two channel data ss_real(n _r , _na ) and ss_imag(n _r , _na ) according to the real part and the imaginary part; divide the matrix Ω according to the real part and The imaginary part is divided into two channel data Ω_real(n _r , _na ) and Ω_i mag(n _r , _na );

Step 10: Construct a deep learning network based on the two-channel input and two-channel output of the convolutional neural network; use the matrix ss_real(n _r ,n _a ) and the matrix ss_i mag(n _r ,n _a ) as the input data of the deep learning network, and The matrix Ω_real(n _r , _na ) and the matrix Ω_i mag(n _r , _na ) are used as training label data to train the deep learning network;

Step 11: Input the data that needs to be defuzzified into the deep learning network trained in step 10, and obtain the distance defuzzified data matrix y_real(n _r , _na ) and matrix y_i mag(n _r , _na );

Step 12: Combine matrix y_real(n _r , _na ) and matrix y_i mag(n _r , _na ) according to the relationship between real and imaginary parts to synthesize matrix y(n _r , _na );

Step 13: Perform Doppler parameter estimation on the matrix y(n _r , _na ) to obtain the estimated Doppler parameters

Step 14: Perform range migration correction on the matrix y(n _r , _na ) according to the Doppler parameter f _dc , and store the result in the matrix y′(n _r , _na ).

Step 15: Construct the reference signal vector according to the radar parameters s_a(n2) is a 1×n _a vector; among them, f _c represents the carrier frequency of the radar transmission signal, c is the propagation speed of electromagnetic waves, V is the satellite speed, θ is the oblique angle of the satellite, and t _n2 is the azimuth slow time;

Step 16: Perform FFT processing on the matrix y′(n _r , _na ) row by row to obtain y(n _r , _fa ); take out each row of y(n _r , _fa ) and dot-multiply the reference signal vector s_a( The conjugate of n2) is used to obtain the data matrix y(n _r ,f _a ) after distance pulse pressure;

Step 17: Perform inverse FFT processing on the matrix y(n _r , _fa ) row by row to obtain the SAR focusing result matrix z(n _r , _na ).

Further, the specific steps of step 8 are as follows:

The spaceborne SAR system transmits multiple code patterns of coded signals, and the echo signal received by a multi-antenna digital receiver is expressed as A=[s(1),…,s(m),…,s(M)], m=1,2,…,M, M is the number of mapping zones, s(m) represents the echo signal of the mth mapping zone received by the multi-antenna digital receiver; the spatial filtering vector is used to perform spatial filtering on the signal, which can The scattering coefficient of the target is recovered from the blurred echo signal, that is, the corresponding scattering coefficient is 1 when there is a target, and the corresponding scattering coefficient is 0 when there is no target; the weight vector is Ω=[Ω ₁ ,...,Ω _m ,...,Ω _M ]; for the m-th mapping zone A ^T Ω _M =Y _m , where, Y _m =[0,...,1,...0]; the matrix composed of the target scattering coefficient is Ω _m =inv (A) Y _m , solving Ω completes the entire defuzzification process.

The beneficial effects of the present invention are as follows:

The invention can effectively solve the problem of distance deambiguation of the surveying and mapping swath of the spaceborne SAR system, and obtain the ocean wide surveying and mapping swath information without range ambiguity, so as to achieve the purpose of SAR imaging of moving targets on the sea surface without range ambiguity.

Description of drawings

Figure 1 is a schematic diagram of distance blurring in a wide surveying swath according to the present invention.

Figure 2 is the wide mapping swath distance deblurring result of the spaceborne SAR system according to the embodiment of the present invention, wherein (a) the deblurring result of 100 iterations; (b) the deblurring result of 200 iterations; (c) the deblurring result of 300 iterations; (d) The defuzzification result after 500 iterations; (e) The defuzzification result after 900 iterations; (f) The defuzzification result after 1500 iterations.

Figure 3 is the imaging result after solving the distance blur according to the embodiment of the present invention, wherein: (a) the imaging result of three moving targets; (b) the distance profile of the first point; (c) the azimuth profile of the first point; (d) ) The distance profile of the second point; (e) The azimuth profile of the second point; (f) The distance profile of the third point; (g) The azimuth profile of the third point.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and examples.

In view of the distance ambiguity problem of the ocean wide surveying swath of the spaceborne SAR system, based on the existing distance defuzzification algorithm, the purpose of this invention is to combine the discrete frequency coding signal and deep learning technology to propose a deep learning-based ocean wide mapping swath. Distance defuzzification method for surveying and mapping swaths, which can achieve range defuzzification for low signal-to-noise ratio echoes.

A deep learning-based distance defuzzification method for wide ocean mapping swaths, including the following steps:

Step 1: Obtain the echo signal of the spaceborne _SAR system. The echo signal is an n _r , n2); n _r represents the number of distance points, n _a represents the number of azimuth points;

Step 2: Construct a reference signal vector based on radar parameters s_r( _{n1 ) is an n r} /> Δf is the distance frequency domain interval,/> n1＝0,1,...,n _r -1;

Step 3: Dot-multiply each column of s(n1,n2) by the conjugate of the reference signal vector s_r(n1) to obtain the data matrix s(f _r ,x _n2 ) after distance pulse pressure; where x _n2 represents Azimuth time domain coordinates, L represents the synthetic aperture length, n2=0,1,...,n _a -1;

Step 4: Perform inverse Fourier transform on the matrix s(f _r ,x _n2 ) by column, and save the result in the matrix s(n _r , _na );

Step 5: Perform Fourier transform on the matrix s(n _r , _na ) row by row, and store the result in the s(n _r , _fa ) matrix; where _fa represents the azimuth frequency domain coordinate, PRF is the azimuth adopted frequency, Δf _a is the azimuth frequency domain interval,

Step 6: Construct the Doppler compensation function s_h(n1) according to the radar parameters, s_h(n1) is a vector of n _r ×1;

Step 7: Perform dot multiplication of the conjugate of s(n _r , f _a ) and the Doppler compensation function s_h(n1) vector, and perform inverse FFT processing on the multiplication result row by row to obtain the intrapulse Doppler compensation The data matrix ss(n _r ,n _a );

Step 8: According to the spatial filtering theory, construct a matrix Ω composed of the target scattering coefficients, where Ω is an n _r × n _a- dimensional matrix;

Step 9: Divide the matrix ss(n _r , _na ) into two channel data ss_real(n _r , _na ) and ss_imag(n _r , _na ) according to the real part and the imaginary part; divide the matrix Ω according to the real part and The imaginary part is divided into two channel data Ω_real(n _r , _na ) and Ω_i mag(n _r , _na );

Step 10: Construct a deep learning network based on the two-channel input and two-channel output of the convolutional neural network; use the matrix ss_real(n _r ,n _a ) and the matrix ss_i mag(n _r ,n _a ) as the input data of the deep learning network, and The matrix Ω_real(n _r , _na ) and the matrix Ω_i mag(n _r , _na ) are used as training label data to train the deep learning network;

Step 11: Input the data that needs to be defuzzified into the deep learning network trained in step 10, and obtain the distance defuzzified data matrix y_real(n _r , _na ) and matrix y_i mag(n _r , _na );

Step 12: Combine matrix y_real(n _r , _na ) and matrix y_i mag(n _r , _na ) according to the relationship between real and imaginary parts to synthesize matrix y(n _r , _na );

Step 13: Perform Doppler parameter estimation on the matrix y(n _r , _na ) to obtain the estimated Doppler parameters

Step 14: Perform range migration correction on the matrix y(n _r , _na ) according to the Doppler parameter f _dc , and store the result in the matrix y′(n _r , _na ).

Step 15: Construct the reference signal vector according to the radar parameters s_a(n2) is a 1×n _a vector; among them, f _c represents the carrier frequency of the radar transmission signal, c is the propagation speed of electromagnetic waves, V is the satellite speed, θ is the oblique angle of the satellite, and t _n2 is the azimuth slow time;

Step 16: Perform FFT processing on the matrix y′(n _r , _na ) row by row to obtain y(n _r , _fa ); take out each row of y(n _r , _fa ) and dot-multiply the reference signal vector s_a( The conjugate of n2) is used to obtain the data matrix y(n _r ,f _a ) after distance pulse pressure;

Step 17: Perform inverse FFT processing on the matrix y(n _r , _fa ) row by row to obtain the SAR focusing result matrix z(n _r , _na ).

Further, the specific steps of step 8 are as follows:

The spaceborne SAR system transmits multiple code patterns of coded signals, and the echo signal received by a multi-antenna digital receiver is expressed as A=[s(1),…,s(m),…,s(M)], m=1,2,…,M, M is the number of mapping zones; using the spatial filtering vector to perform spatial filtering on the signal, the scattering coefficient of the target can be recovered from the blurred echo signal, that is, the scattering corresponding to the presence of the target The coefficient is 1, and the corresponding scattering coefficient when there is no target is 0; the weight vector is Ω=[Ω ₁ ,…,Ω _m ,…,Ω _M ]; for the mth mapping zone A ^T Ω _m =Y _m , Among them, Y _m =[0,...,1,...0]; the matrix composed of the target scattering coefficient is Ω _m =inv(A)Y _m . Solving Ω completes the entire defuzzification process.

Specific examples:

The effectiveness of the present invention is further verified through simulation experimental data below.

(1) Simulation experiment

1. Measured parameters

In order to verify the effectiveness of the method of the present invention, the measured data parameters in Table 1 are given here.

Table 1 Simulation data parameters

Carrier frequency 5.3GHz Sampling frequency 200MHz Speed of spaceborne SAR system 7100m/s Signal bandwidth 160MHz Scene center line distance 850km Double blur distance 75km Pulse Repetition Frequency (PRF) 2000Hz Number of mapping swaths 3 Number of sports goals 3 Moving target speed -8m/s,4m/s,10m/s

2.Experimental content

Figure 2 illustrates the simulation data processing results obtained by using a deep learning-based ocean wide mapping swath distance defuzzification algorithm proposed by the present invention. Among them, (a) the defuzzification result of 100 iterations; (b) the defuzzification result of 200 iterations; (c) the defuzzification result of 300 iterations; (d) the defuzzification result of 500 iterations; (e) the defuzzification result of 900 iterations Results; (f) Deblurring results after 1500 iterations. It can be seen from the figure that the distance defuzzification effect of the method of the present invention is used for training using a deep learning network. When the number of iterations reaches 300, better distance defuzzification results can be obtained.

Figure 3 illustrates the results of SAR imaging using the range deblurred data of the present invention, in which (a) the imaging results of three moving targets; (b) the range profile of the first point; (c) the first point The azimuth profile of the second point; (d) the distance profile of the second point; (e) the azimuth profile of the second point; (f) the distance profile of the third point; (g) the azimuth profile of the third point. It can be seen from the figure that the data after the range blur is solved by the method of the present invention can realize two-dimensional and high-resolution SAR imaging of ocean moving targets at different speeds.

Therefore, the method of the present invention can effectively solve the wide mapping swath distance deambiguation problem of the spaceborne SAR system.

Claims (2)

1. The ocean wide swath distance defuzzification method based on deep learning is characterized by comprising the following steps of:

step 1: acquiring an echo signal of a satellite-borne SAR system, wherein the echo signal is n _r ×n _a The dimension matrix is used for carrying out Fourier transform processing on the echo signal matrix according to columns, and storing the result in a matrix s (n 1, n 2); n is n _r Represents the distance to the point number, n _a Indicating azimuth points;

step 2: constructing reference signal vectors based on radar parameterss_r (n 1) is n _r X 1 vector; wherein k is _r Represents the tuning frequency, k _r ＝B/T _p B represents the bandwidth of the transmitted signal, T _p Representing the width of the transmitted pulse, f _r Expressed as distance to frequency domain coordinates>Δf is distance frequency domain interval, +.>

Step 3: multiplying the conjugate of the reference signal vector s_r (n 1) by the average point of each column of s (n 1, n 2) to obtain a data matrix s (f) after distance pulse compression _r ，x _n2 ) The method comprises the steps of carrying out a first treatment on the surface of the Wherein x is _n2 Representing the azimuth time domain coordinates,l is expressed as the synthetic aperture length, n2=0, 1,.. _a -1；

Step 4: for matrix s (f _r ，x _n2 ) The inverse fourier transform processing is performed for each column, and the result is stored in a matrix s (n _r ，n _a ) In (a) and (b);

step 5: for matrix s (n _r ，n _a ) Fourier transforming the result in terms of rows, storing the result in s (n _r ，f _a ) In the matrix; wherein f _a Represents the azimuth frequency domain coordinates,PRF is azimuth frequency, Δf _a For azimuth frequency domain interval, +.>

Step 6: constructing Doppler compensation functions from radar parameters s_h (n 1) is n _r A vector of x 1; wherein n is _r The number of the distance direction points is represented, V is the radial speed of a moving target, V is the moving speed of a radar platform, θ is the azimuth angle of the sight of the inclined plane radar, λ is the working wavelength of the radar, and g represents the gravitational acceleration;

step 7: will s (n) _r ，f _a ) Dot multiplication is carried out on the conjugate of the vector of the Doppler compensation function s_h (n 1), and the multiplication result is processed by inverse FFT according to the line to obtain a data matrix ss (n) _r ，n _a )；

Step 8: constructing a matrix omega consisting of target scattering coefficients according to a spatial filtering theory, wherein omega is n _r ×n _a A matrix of dimensions;

step 9: the matrix ss (n _r ，n _a ) Two channel data ss_real (n _r ，n _a ) And ss_imag (n) _r ，n _a ) The method comprises the steps of carrying out a first treatment on the surface of the Will beThe matrix Ω is divided into two channel data Ω_real (n _r ，n _a ) And Ω_imag (n) _r ，n _a )；

Step 10: constructing a deep learning network based on two-channel input and two-channel output of a convolutional neural network; the matrix ss_real (n _r ，n _a ) Sum matrix ss_imag (n) _r ，n _a ) As deep learning network input data, matrix Ω_real (n _r ，n _a ) And matrix Ω_imag (n) _r ，n _a ) Training the deep learning network as training tag data;

step 11: inputting the data to be deblurred into the deep learning network trained in the step 10 to obtain a distance deblurred data matrix y_real (n) _r ，n _a ) And matrix y_imag (n) _r ，n _a )；

Step 12: the matrix y_real (n _r ，n _a ) And matrix y_imag (n) _r ，n _a ) Synthesizing matrix y (n) according to real-part and imaginary-part relation _r ，n _a )；

Step 13: for matrix y (n _r ，n _a ) Estimating Doppler parameter to obtain estimated Doppler parameter

Step 14: according to Doppler parameter f _dc For matrix y (n _r ，n _a ) Performing range migration correction and storing the result in a matrix y' (n) _r ，n _a ) In (a) and (b);

step 15: constructing reference signal vectors based on radar parameters s_a (n 2) is 1×n _a Vector; wherein f _c Representing the carrier frequency of a radar transmitted signal, c is the propagation velocity of electromagnetic waves, V is the satellite velocity, θ is the angle of view of the satellite, t _n2 For slow time of azimuth；

Step 16: matrix y' (n) _r ，n _a ) FFT processing is performed on a row-by-row basis to obtain y (n) _r ，f _a ) The method comprises the steps of carrying out a first treatment on the surface of the Take out y (n) _r ，f _a ) Is multiplied by the conjugate of the reference signal vector s_a (n 2) to obtain a data matrix y (n) after pulse pressure _r ，f _a )；

Step 17: the matrix y (n _r ，f _a ) Performing inverse FFT processing according to the rows to obtain SAR focusing result matrix z (n _r ，n _a )。

2. The deep learning-based ocean wide swath distance deblurring method of claim 1, wherein the step 8 comprises the specific steps of:

the space-borne SAR system transmits a plurality of groups of code-type coded signals, and echo signals received by a multi-antenna digital receiver are expressed as A= [ s (1), …, s (M), …, s (M)]M=1, 2, …, M is the number of swaths, s (M) represents the echo signal of the mth swath received with the multi-antenna digital receiver; the signal is subjected to spatial filtering by using the spatial filtering vector, and the scattering coefficient of the target can be recovered from the blurred echo signal, namely, the scattering coefficient corresponding to the presence of the target is 1, and the scattering coefficient corresponding to the absence of the target is 0; the weight vector is Ω= [ Ω ] ₁ ，…，Ω _m ，…，Ω _M ]The method comprises the steps of carrying out a first treatment on the surface of the For the mth mapping band A ^T Ω _m ＝Y _m Wherein Y is _m ＝[0，…，1，…0]The method comprises the steps of carrying out a first treatment on the surface of the The matrix consisting of the scattering coefficients of the target is Ω _m ＝inv(A)Y _m And solving omega to complete the whole deblurring process.