A SAR moving target detection method based on intelligent learning

Abstract

The present invention relates to a SAR moving target detection method based on intelligent learning, comprising: obtaining a first training sample set or a second training sample set after the original echo data of the pulse system SAR is subjected to range pulse compression processing, or range pulse compression processing and azimuth Fourier transformation; obtaining a third training sample set or a fourth training sample set after the original echo data of the frequency modulated continuous wave SAR is subjected to range Fourier transformation, or range Fourier transformation and azimuth Fourier transformation; using each training sample set to train the deep learning artificial neural network respectively to obtain a corresponding neural network model; collecting the echo data to be detected; selecting the neural network model, performing corresponding processing or transformation on the echo data to be detected, and inputting it into the neural network model for moving target detection, thereby detecting the moving target. The present invention realizes the detection of low-speed motion in the clutter band under the SAR working system and weak moving target under low signal-to-clutter ratio.

Description

Technical Field

[0001] The invention belongs to the field of radar technology, and in particular relates to a SAR moving target detection method based on intelligent learning. Background Art

[0002] Ground moving target detection is to install airborne radar on an aerial platform to detect ground moving targets. It is an important function of airborne radar. Using synthetic aperture radar (SAR) to detect ground moving targets and estimate motion parameters is an important function of modern multi-function SAR.

[0003] SAR ground moving target detection method is closely related to the SAR system: the system is relatively simple, and the traditional frequency domain filtering method cannot effectively suppress the wide-band ground clutter formed by azimuth motion, resulting in poor ground moving target detection performance. Other detection methods in the image domain or range Doppler domain, such as WVD, RDM, single-view image sequence method, etc., have relatively complex algorithms and high requirements for signal-to-clutter ratio. In addition, the non-ideal motion of the SAR platform will have a greater impact on the algorithm, affecting its use in practical application systems; multi-antenna systems combined with DPCA, STAP and other detection methods can eliminate ground clutter and improve ground motion detection performance, but multi-antenna SAR systems are complex and the flight platforms that can be installed are greatly limited.

[0004] Therefore, there is a need in the art for a new method that can overcome the above-mentioned defects or shortcomings of the traditional SAR ground moving target detection method. Summary of the invention

[0005] In view of the above analysis, the present invention aims to disclose a SAR moving target detection method based on intelligent learning to overcome the defects of traditional SAR ground moving target detection.

[0006] The present invention discloses a SAR moving target detection method based on intelligent learning, comprising:

[0007] Model training step: constructing a training data set for SAR moving target detection, the training data set includes a training sample set obtained by processing or transforming one or more SAR original echo data, including a first training sample set composed of two-dimensional data obtained by performing distance-direction pulse compression processing on the original echo data of the pulse system SAR; a second training sample set composed of two-dimensional data obtained by performing distance-direction pulse compression processing and azimuth Fourier transform on the original echo data of the pulse system SAR; a third training sample set composed of two-dimensional data obtained by performing distance-direction Fourier transform on the original echo data of the frequency modulated continuous wave SAR; and a fourth training sample set composed of two-dimensional data obtained by performing distance-direction Fourier transform and azimuth Fourier transform on the original echo data of the frequency modulated continuous wave SAR; using each training sample set to train a deep learning artificial neural network respectively to obtain a corresponding neural network model;

[0008] Ground moving target detection steps: collect the echo data to be detected obtained during the flight of SAR in imaging mode; select a neural network model, perform corresponding processing or transformation on the echo data to be detected according to the selected neural network model, and input it into the neural network model for moving target detection to detect the moving target.

[0009] Furthermore, the process of constructing training samples in the first training sample set established by the pulse SAR includes:

[0010] The pulse SAR raw echo data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction to form two-dimensional data;

[0011] After the two-dimensional data is pulse compressed in the range direction, that is, matched filtering in the range direction is performed, the obtained two-dimensional data array of range-azimuth is used as a training sample.

[0012] Furthermore, in the ground moving target detection step, for the neural network model trained by the first training sample set, the detection step includes:

[0013] Collect the echo data to be detected during the flight of pulse SAR in imaging mode;

[0014] Arranging the echo data in two dimensions, i.e., the distance receiving time and the azimuth receiving time, to form two-dimensional data;

[0015] Performing pulse compression processing on the two-dimensional data in the range direction to obtain a two-dimensional data array of range-azimuth;

[0016] The distance-azimuth two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

[0017] Furthermore, the process of constructing training samples in the second training sample set established by the pulse SAR includes:

[0018] The raw echo data of pulse SAR is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction to form two-dimensional data;

[0019] The two-dimensional data is pulse compressed in the range direction to obtain a range-azimuth two-dimensional data array, and then the range-azimuth two-dimensional data array is Fourier transformed along the azimuth direction to obtain a range-Doppler two-dimensional data array as a training sample.

[0020] Furthermore, in the ground moving target detection step, for the neural network model trained by the second training sample set, the detection step includes:

[0021] Collect the echo data to be detected during the flight of pulse SAR in imaging mode;

[0022] Arrange the echo data in two dimensions, i.e., distance receiving time and azimuth receiving time, to form two-dimensional data;

[0023] The two-dimensional data is pulse compressed in the range direction to obtain a two-dimensional data array of range-azimuth, and then the two-dimensional data array of range-azimuth is Fourier transformed in the azimuth direction to obtain a two-dimensional data array of range-Doppler.

[0024] The range-Doppler two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

[0025] Furthermore, the process of constructing training samples in the third training sample set established by the FMCW SAR includes:

[0026] The original echo data of FMCW SAR is arranged in two dimensions according to the receiving time in range and the receiving time in azimuth to form two-dimensional data of range sweep-azimuth.

[0027] The range-frequency sweep-azimuth two-dimensional data is Fourier transformed along the range direction, and the obtained range-azimuth two-dimensional data array is used as a training sample.

[0028] Furthermore, in the ground moving target detection step, for the neural network model trained by the third training sample set, the detection step includes:

[0029] Collect the echo data to be detected during the flight in the imaging mode of FMCW SAR;

[0030] Arrange the echo data in two dimensions, i.e., the receiving time in range and the receiving time in azimuth, to form two-dimensional data of range frequency sweep and azimuth;

[0031] Performing Fourier transform on the range sweep-azimuth two-dimensional data in the range direction to obtain a range-azimuth two-dimensional data array,

[0032] The distance-azimuth two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

[0033] Furthermore, the process of constructing training samples in the fourth training sample set established by the FMCW SAR includes:

[0034] The original echo data of FMCW SAR is arranged in two dimensions according to the receiving time in range and the receiving time in azimuth to form two-dimensional data of range sweep-azimuth.

[0035] The range sweep-azimuth two-dimensional data is Fourier transformed along the range direction to form a range-azimuth two-dimensional data array, and then the range-azimuth two-dimensional data array is Fourier transformed along the azimuth direction, and the obtained range-Doppler two-dimensional data array is used as a training sample.

[0036] Furthermore, in the ground moving target detection step, for the neural network model trained by the fourth training sample set, the detection step includes:

[0037] Collect the echo data to be detected during the flight in the imaging mode of FMCW SAR;

[0038] Arrange the echo data in two dimensions, i.e., the receiving time in range and the receiving time in azimuth, to form two-dimensional data of range sweep-azimuth;

[0039] Performing Fourier transformation on the two-dimensional data of range frequency sweep-azimuth along the range direction to obtain a two-dimensional data array of range-azimuth, and then performing Fourier transformation on the two-dimensional data array of range-azimuth along the azimuth direction to obtain a two-dimensional data array of range-Doppler;

[0040] The range-Doppler two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

[0041] Furthermore, before using each training sample set to train the deep learning artificial neural network, on the training samples of the training sample set, the point traces of the ground moving target data are marked as positive samples, and the point traces of the ground stationary strong target data are marked as negative samples;

[0042] The neural network model trained with each training sample set can correctly classify the ground moving target points and ground stationary strong target points in the sample data.

[0043] The present invention can achieve at least one of the following beneficial effects:

[0044] (1) The SAR system is used to detect moving targets. The difference between the echo signals of ground moving targets and stationary targets in the range-Doppler domain or the range-azimuth domain is used to train a deep learning artificial neural network. The echo signals of ground moving targets are detected through intelligent recognition, and low-speed moving target detection in the clutter band under the SAR working system and weak moving target detection under low signal-to-clutter ratio are achieved.

[0045] (2) Using deep learning-based artificial neural networks to perform intelligent detection of ground moving targets can reduce the impact of non-ideal motion of the SAR platform on the detection algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0047] Figure 1 is a flow chart of a SAR moving target detection method in an embodiment of the present invention;

[0048] Figure 2 It is a flow chart of a SAR moving target detection method based on intelligent learning in the range-azimuth domain of a pulse SAR in an embodiment of the present invention;

[0049] Figure 3 It is a flow chart of a SAR moving target detection method based on intelligent learning in the pulse SAR range-Doppler domain in an embodiment of the present invention;

[0050] Figure 4 It is a flow chart of a method for detecting SAR moving targets in a frequency modulated continuous wave SAR range-azimuth domain based on intelligent learning in an embodiment of the present invention;

[0051] Figure 5 The present invention is a flowchart of a method for detecting SAR moving targets in a frequency modulated continuous wave SAR range-Doppler domain based on intelligent learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0053] An embodiment of the present invention discloses a SAR moving target detection method based on intelligent learning, such as Figure 1 As shown, including:

[0054] Model training step: constructing a training data set for SAR moving target detection, the training data set includes a training sample set obtained by processing or transforming one or more SAR original echo data, including a first training sample set composed of two-dimensional data obtained by performing distance-direction pulse compression processing on the original echo data of the pulse system SAR; a second training sample set composed of two-dimensional data obtained by performing distance-direction pulse compression processing and azimuth Fourier transform on the original echo data of the pulse system SAR; a third training sample set composed of two-dimensional data obtained by performing distance-direction Fourier transform on the original echo data of the frequency modulated continuous wave SAR; and a fourth training sample set composed of two-dimensional data obtained by performing distance-direction Fourier transform and azimuth Fourier transform on the original echo data of the frequency modulated continuous wave SAR; using each training sample set to train a deep learning artificial neural network respectively to obtain a corresponding neural network model;

[0055] Ground moving target detection steps: collect the echo data to be detected obtained during the flight of SAR in imaging mode; select a neural network model, perform corresponding processing or transformation on the echo data to be detected according to the selected neural network model, and input it into the neural network model for moving target detection to detect the moving target.

[0056] Specifically, in a specific solution of this embodiment, a SAR moving target detection method after intelligent learning based on a first training sample set of raw echo data of a pulse SAR is disclosed, such as Figure 2 As shown,

[0057] Specifically, during the model training process,

[0058] 1) Constructing a first training sample set for SAR moving target detection after performing range pulse compression processing on the original echo data; during the construction process, including:

[0059] (1) SAR emits detection pulse signals during flight;

[0060] (2) Receive the original echo data containing the moving target, perform range pulse compression processing, and use it as training samples in the training set;

[0061] The pulse SAR raw data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and then processed by the range pulse compression process (i.e., the range matched filter process) to form a two-dimensional data array of range and azimuth.

[0062] (3) Use multiple training samples to form a training sample set.

[0063] 2) On the training samples, the points of ground moving targets in the two-dimensional data of distance and azimuth are marked as positive samples, and the points of ground stationary strong targets in the two-dimensional data of distance and azimuth are marked as negative samples;

[0064] 3) Using the training sample set to train the deep learning artificial neural network, the trained deep learning artificial neural network can correctly classify the ground moving target point traces and the ground stationary strong target point traces in the distance-azimuth two-dimensional data;

[0065] The deep learning artificial neural network can adopt a deep learning artificial neural network for image detection, including adopting an R-CNN network, or a Fast R-CNN network, or a Faster R-CNN network, or a YOLO network, or an SSD network, or an RCNN network.

[0066] Specifically, the ground moving target detection process includes:

[0067] 1) SAR emits detection pulse signals during flight;

[0068] 2) Receive ground echo signals containing moving targets;

[0069] 3) The acquired echo data are arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and then processed by pulse compression in the range direction (matched filter processing) to form a two-dimensional data array of range and azimuth;

[0070] 4) Inputting the distance-azimuth two-dimensional data array into the trained deep learning artificial neural network model;

[0071] The deep learning artificial neural network model outputs ground moving target point traces to complete ground moving target detection.

[0072] Specifically, in another specific solution of the present embodiment, a second training sample set consisting of two-dimensional data obtained by performing range-direction pulse compression processing and azimuth Fourier transform on the original echo data of a pulse SAR is disclosed, and a SAR moving target detection method after intelligent learning is performed, such as Figure 3 As shown,

[0073] Specifically, during the model training process,

[0074] 1) Constructing a second training sample set for SAR moving target detection, which is composed of two-dimensional data obtained by performing pulse compression processing in the range direction and Fourier transform in the azimuth direction on the original echo data; in the construction process, including:

[0075] (1) SAR emits detection pulse signals during flight;

[0076] (2) Receive the original echo data containing the moving target; perform pulse compression processing in the range direction, and then perform Fourier transform in the azimuth direction to obtain training samples;

[0077] The original echo data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and then processed by pulse compression in the range direction (matched filter processing) to form a two-dimensional data array of range and azimuth.

[0078] Performing Fourier transform on the range-azimuth two-dimensional data array along the azimuth direction to obtain a range-Doppler two-dimensional data array;

[0079] After pulse compression processing in the range direction, targets at different distances correspond to different frequencies, which are converted into different times for targets at different distances. Then, by performing Fourier transform along the azimuth direction, the difference in Doppler processes between moving targets and stationary backgrounds in the azimuth frequency domain (combined with the range direction) can be used for training and detection.

[0080] (3) Use multiple training samples to form a training sample set.

[0081] 2) In the training samples, the points of ground moving targets in the range-Doppler two-dimensional data are marked as positive samples, and the points of ground stationary strong targets in the range-Doppler two-dimensional data are marked as negative samples.

[0082] 3) Use the training sample set to train the deep learning artificial neural network so that the trained deep learning artificial neural network can correctly classify ground moving target points and ground stationary strong target points in the range-Doppler two-dimensional data.

[0083] The deep learning artificial neural network can adopt a deep learning artificial neural network for image detection, including adopting an R-CNN network, or a Fast R-CNN network, or a Faster R-CNN network, or a YOLO network, or an SSD network, or an RCNN network.

[0084] Specifically, the ground moving target detection process includes:

[0085] 1) SAR emits detection pulse signals during flight;

[0086] 2) Receive ground echo signals containing moving targets;

[0087] The collected SAR raw echo data are arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction.

[0088] 3) After the range pulse compression processing (matched filter processing), a range-azimuth two-dimensional data array is formed; then the range-azimuth data array is Fourier transformed along the azimuth direction to form a range-Doppler two-dimensional data array;

[0089] 4) Inputting the range-Doppler two-dimensional data array into the trained deep learning artificial neural network model;

[0090] The deep learning artificial neural network model outputs ground moving target point traces to complete ground moving target detection.

[0091] Specifically, in another specific solution of the present embodiment, a SAR moving target detection method after intelligent learning is disclosed based on a third training sample set consisting of two-dimensional data obtained by performing range Fourier transform on the original echo data of the FMCW SAR, such as Figure 4 As shown,

[0092] 1) Constructing a third training sample set for SAR moving target detection, which is composed of two-dimensional data obtained by Fourier transforming the original echo data in the range direction; in the construction process, including:

[0093] (1) SAR transmits a frequency modulated continuous wave signal during flight;

[0094] (2) Receive the original echo data containing the moving target; perform range Fourier transform to obtain training samples;

[0095] In the FMCW SAR receiving process, de-skewing is performed in the analog signal domain, and then the SAR original echo data is acquired. The data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction to obtain a two-dimensional data array of range sweep frequency and azimuth.

[0096] The range sweep-azimuth two-dimensional data array of the original echo data is Fourier transformed along the range direction to form a range-azimuth two-dimensional data array; by performing Fourier transform along the range direction, the different frequencies corresponding to targets at different distances after the simulated de-skewing processing can be converted into different times corresponding to targets at different distances.

[0097] (3) Use multiple training samples to form a training sample set.

[0098] 2) In the training samples, the points of ground moving targets in the range-Doppler two-dimensional data are marked as positive samples, and the points of ground stationary strong targets in the range-Doppler two-dimensional data are marked as negative samples.

[0099] 3) Using the training sample set to train the deep learning artificial neural network, the trained deep learning artificial neural network can correctly classify the ground moving target points and the ground stationary strong target points in the range-Doppler two-dimensional data;

[0100] The deep learning artificial neural network can adopt a deep learning artificial neural network for image detection, including adopting an R-CNN network, or a Fast R-CNN network, or a Faster R-CNN network, or a YOLO network, or an SSD network, or an RCNN network.

[0101] Specifically, the ground moving target detection process includes:

[0102] 1) SAR transmits detection continuous wave signals during flight;

[0103] 2) Receive ground echo signals containing moving targets;

[0104] The collected SAR raw data are arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and a two-dimensional data array of range sweep-azimuth is obtained;

[0105] 3) Performing Fourier transform on the range-frequency sweep-azimuth two-dimensional data array along the range direction to form a range-azimuth two-dimensional data array;

[0106] 4) Inputting the distance-azimuth two-dimensional data array into the trained deep learning artificial neural network;

[0107] The trained deep learning artificial neural network outputs ground moving target traces to complete ground moving target detection.

[0108] Specifically, in another specific solution of the present embodiment, a fourth training sample set consisting of two-dimensional data obtained by performing range Fourier transform and azimuth Fourier transform on the original echo data of continuous wave SAR is disclosed, and a SAR moving target detection method after intelligent learning is performed, such as Figure 5 As shown,

[0109] 1) Constructing a fourth training sample set for SAR moving target detection, which is composed of two-dimensional data obtained by performing distance Fourier transform and azimuth Fourier transform on the original echo data; during the construction process, including:

[0110] (1) SAR transmits continuous wave detection signals during flight;

[0111] (2) receiving raw echo data containing moving targets; performing Fourier transform in range and Fourier transform in azimuth to obtain training samples;

[0112] The collected SAR raw echo data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and a two-dimensional data array of range sweep frequency and azimuth is obtained;

[0113] Performing Fourier transformation on the range sweep-azimuth two-dimensional data array of the original echo data along the range direction to form a range-azimuth two-dimensional data array; then performing Fourier transformation on the range-azimuth two-dimensional data array along the azimuth direction to form a range-Doppler two-dimensional data array;

[0114] By performing Fourier transform along the range direction, the different frequencies corresponding to targets at different distances after simulated de-skewing can be converted into different times corresponding to targets at different distances. By performing Fourier transform along the azimuth direction, the difference in Doppler processes between moving targets and stationary backgrounds in the azimuth frequency domain (combined with the range direction) can be used for training and detection;

[0115] (3) Use multiple training samples to form a training sample set.

[0116] 2) In the training samples, the points of ground moving targets in the range-Doppler two-dimensional data are marked as positive samples, and the points of ground stationary strong targets in the range-Doppler two-dimensional data are marked as negative samples.

[0117] 3) Use the training sample set to train the deep learning artificial neural network so that the trained deep learning artificial neural network can correctly classify ground moving target points and ground stationary strong target points in the range-Doppler two-dimensional data.

[0118] The deep learning artificial neural network can adopt a deep learning artificial neural network for image detection, including adopting an R-CNN network, or a Fast R-CNN network, or a Faster R-CNN network, or a YOLO network, or an SSD network, or an RCNN network.

[0119] Specifically, the ground moving target detection process includes:

[0120] 1) SAR transmits detection continuous wave signals during flight;

[0121] 2) Receive ground echo signals containing moving targets;

[0122] The collected SAR raw echo data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction, and a two-dimensional data array of range sweep frequency and azimuth is obtained;

[0123] 3) Performing Fourier transform on the range sweep-azimuth two-dimensional data array along the range direction to form a range-azimuth two-dimensional data array; then performing Fourier transform on the range-azimuth two-dimensional data array along the azimuth direction to form a range-Doppler two-dimensional data array;

[0124] 4) Inputting the range-Doppler two-dimensional data array into the trained deep learning artificial neural network;

[0125] The trained deep learning artificial neural network outputs ground moving target traces to complete ground moving target detection.

[0126] In summary, the SAR moving target detection method based on intelligent learning in this embodiment adopts a SAR system to detect moving targets, and utilizes the difference between the echo signal of a ground moving target and the echo of a stationary target in the range-Doppler domain or the range-azimuth domain to train a deep learning artificial neural network. The echo signal of the ground moving target is detected by intelligent recognition, which can realize the detection of low-speed moving targets in the clutter band under the SAR working system, and the detection of weak moving targets under low signal-to-clutter ratio.

[0127] The artificial neural network based on deep learning is used to perform intelligent detection of ground moving targets. Combined with the motion compensation processing of the collected SAR raw echo data in the range direction, the range motion error caused by non-ideal motion can be compensated, which can further reduce the impact of non-ideal motion of the SAR platform on the detection algorithm.

[0128] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A SAR moving target detection method based on intelligent learning, characterized in that: include: Model training step: constructing a training data set for SAR moving target detection, wherein the training data set includes a training sample set obtained by processing or transforming one or more SAR original echo data, including a first training sample set consisting of two-dimensional data obtained by performing range-direction pulse compression processing on the pulse SAR original echo data; A second training sample set is composed of two-dimensional data obtained by performing distance-wise pulse compression processing and azimuth-wise Fourier transform on the original echo data of the pulse system SAR; a third training sample set is composed of two-dimensional data obtained by performing distance-wise Fourier transform on the original echo data of the frequency modulated continuous wave SAR; and a fourth training sample set is composed of two-dimensional data obtained by performing distance-wise Fourier transform and azimuth-wise Fourier transform on the original echo data of the frequency modulated continuous wave SAR; each training sample set is used to train the deep learning artificial neural network respectively to obtain the corresponding neural network model; Ground moving target detection step: collecting echo data to be detected obtained during the flight of the SAR in the imaging mode; selecting a neural network model, performing corresponding processing or transformation on the echo data to be detected according to the selected neural network model, and inputting the data into the neural network model for moving target detection to detect the moving target; The process of constructing training samples in the first training sample set established by the pulse SAR includes: The pulse SAR raw echo data is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction to form two-dimensional data; After pulse compression processing is performed on the two-dimensional data in the range direction, that is, matched filtering processing is performed in the range direction, the obtained two-dimensional data array of range-azimuth is used as a training sample; In the ground moving target detection step, for the neural network model trained with the first training sample set, the detection step includes: Collect the echo data to be detected during the flight of pulse SAR in imaging mode; Arranging the echo data in two dimensions, i.e., the distance receiving time and the azimuth receiving time, to form two-dimensional data; Performing pulse compression processing on the two-dimensional data in the range direction to obtain a two-dimensional data array of range-azimuth; Inputting the distance-azimuth two-dimensional data array into the neural network model to perform moving target detection, thereby detecting the moving target; The process of constructing training samples in the second training sample set established by the pulse SAR includes: The raw echo data of pulse SAR is arranged in two dimensions according to the receiving time in the range direction and the receiving time in the azimuth direction to form two-dimensional data; Performing pulse compression processing on the two-dimensional data in the range direction to obtain a range-azimuth two-dimensional data array, and then performing Fourier transform on the range-azimuth two-dimensional data array along the azimuth direction to obtain a range-Doppler two-dimensional data array as a training sample; The process of constructing training samples in the third training sample set established by the FMCW SAR includes: The original echo data of FMCW SAR is arranged in two dimensions according to the receiving time in range and the receiving time in azimuth to form two-dimensional data of range sweep-azimuth. Performing Fourier transform on the range-frequency sweep-azimuth two-dimensional data along the range direction, and obtaining the range-azimuth two-dimensional data array as a training sample; In the ground moving target detection step, for the neural network model trained with the third training sample set, the detection step includes: Collect the echo data to be detected during the flight in the imaging mode of FMCW SAR; Arrange the echo data in two dimensions, i.e., the receiving time in range and the receiving time in azimuth, to form two-dimensional data of range frequency sweep and azimuth; Performing Fourier transform on the range sweep-azimuth two-dimensional data in the range direction to obtain a range-azimuth two-dimensional data array, The distance-azimuth two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

2. The SAR moving target detection method according to claim 1, characterized in that: In the ground moving target detection step, for the neural network model trained with the second training sample set, the detection step includes: Collect the echo data to be detected during the flight of pulse SAR in imaging mode; Arrange the echo data in two dimensions, i.e., distance receiving time and azimuth receiving time, to form two-dimensional data; The two-dimensional data is pulse compressed in the range direction to obtain a two-dimensional data array of range-azimuth, and then the two-dimensional data array of range-azimuth is Fourier transformed in the azimuth direction to obtain a two-dimensional data array of range-Doppler. The range-Doppler two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

3. The SAR moving target detection method according to claim 1, characterized in that: The process of constructing training samples in the fourth training sample set established by the frequency modulated continuous wave SAR system includes: The original echo data of FMCW SAR is arranged in two dimensions according to the receiving time in range and the receiving time in azimuth to form two-dimensional data of range sweep-azimuth. The range sweep-azimuth two-dimensional data is Fourier transformed along the range direction to form a range-azimuth two-dimensional data array, and then the range-azimuth two-dimensional data array is Fourier transformed along the azimuth direction, and the obtained range-Doppler two-dimensional data array is used as a training sample.

4. The SAR moving target detection method according to claim 3, characterized in that: In the ground moving target detection step, for the neural network model trained with the fourth training sample set, the detection step includes: Collect the echo data to be detected during the flight in the imaging mode of FMCW SAR; Arrange the echo data in two dimensions, i.e., the receiving time in range and the receiving time in azimuth, to form two-dimensional data of range sweep-azimuth; Performing Fourier transformation on the two-dimensional data of range frequency sweep-azimuth along the range direction to obtain a two-dimensional data array of range-azimuth, and then performing Fourier transformation on the two-dimensional data array of range-azimuth along the azimuth direction to obtain a two-dimensional data array of range-Doppler; The range-Doppler two-dimensional data array is input into the neural network model to perform moving target detection, thereby detecting the moving target.

5. The SAR moving target detection method according to any one of claims 1 to 4, characterized in that: Before using each training sample set to train the deep learning artificial neural network, on the training samples of the training sample set, the point traces of ground moving target data are marked as positive samples, and the point traces of ground stationary strong target data are marked as negative samples; The neural network model trained with each training sample set can correctly classify the ground moving target points and ground stationary strong target points in the sample data.

Images