A weak target detection method for sea polarimetric radar

Abstract

The present application belongs to the technical field of radar target detection, and particularly relates to a weak target detection method for a sea polarized radar, which comprises a radar echo signal training data set obtained; the training data set is preprocessed to obtain a training feature set, and the training feature set is substituted into a preset support vector machine to train a classifier capable of distinguishing weak target echoes and sea clutter; wherein after each round of training, according to the size relationship and difference value between the current training false alarm probability and the preset target false alarm probability, a false alarm cost factor in a corresponding penalty matrix of the support vector machine is adjusted, and the trained classifier is used for radar echo classification, so that the method can effectively realize complex structure target detection under the support of a preset false alarm probability.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection technology, specifically relating to the effective detection of complex-structured targets on the sea surface by processing and analyzing the recovered waves when a sea-polarized radar is interfered with by strong sea clutter. Background Technology

[0002] Maritime surveillance radar inevitably receives radar echoes reflected from the sea surface, known as sea clutter, during operation. The presence of sea clutter makes it difficult for radar to detect sea surface targets such as small boats, buoys, and drilling platforms. Especially when maritime surveillance radar faces strong sea clutter, traditional target detection methods perform poorly, making the detection of weak targets in strong sea clutter backgrounds a challenging problem.

[0003] Numerous attempts and studies have been conducted to address this challenge. Some existing technologies have improved upon the traditional adaptive constant false alarm (CFAR) detection method, resulting in CFAR detection methods combining clutter maps and CFAR detection methods combining fuzzy algorithms. These methods are often based on the statistical characteristics of sea clutter amplitude; however, classical statistical detection methods have performance limitations. Therefore, to overcome these limitations, existing technologies have proposed feature detection methods based on the differences between sea clutter and target features. For example, target detection methods based on fractal features achieve good detection results with long observation times; however, when the observation time is shortened, the detection performance deteriorates significantly, failing to meet the requirements of maritime search radar. Target detection methods based on Doppler features perform well in detecting moving targets; however, when sea clutter is strong, the target signal in the Doppler amplitude spectrum may be submerged in sea clutter, rendering the Doppler features ineffective.

[0004] In summary, while there are existing studies on target detection using multi-feature fusion, some features have high observation time requirements and are not suitable for low signal-to-clutter radar conditions. Furthermore, many studies only use field data with floating spheres as targets for validation, and the detection performance for complex targets still needs further verification. Therefore, a multi-feature fusion method for detecting weak sea surface targets that has been validated by field measurement data, is suitable for strong sea clutter environments, and is effective against complex targets has yet to emerge. Summary of the Invention

[0005] The purpose of this invention is to propose a method for detecting weak targets in sea polarization radar, which can effectively detect targets when the signals of complex target structures in the radar echo are almost completely submerged in strong sea clutter.

[0006] The technical solution of this invention is: a method for detecting weak targets using a sea-based polarization radar, specifically comprising the following steps:

[0007] Step 1: Obtain the radar echo signal training dataset;

[0008] Step 2: Preprocess the training dataset to obtain the training feature set, including:

[0009] The phase curve expansion method is used to preprocess the echo phase in the echo signal training data to obtain the expanded phase, and the mean value of the expanded phase difference is further calculated.

[0010] Calculate the phase mean of the same polarization ratio, the phase range of the same polarization ratio, and the amplitude-average ratio of the same polarization as polarization characteristics;

[0011] The mean phase difference and polarization characteristics of each echo signal data are used to form a feature vector, and then the training feature set is obtained.

[0012] Step 3: Substitute the training feature set into the preset support vector machine to train a classifier that can distinguish between weak target echoes and sea clutter. After each training round, adjust the false alarm cost factor in the corresponding penalty matrix of the support vector machine according to the difference and magnitude between the current training false alarm probability and the preset target false alarm probability. Then, substitute the training feature set again for the next round of training until the difference between the training false alarm probability and the set false alarm probability falls within the preset error range, and the training ends.

[0013] Step 4: Use the trained classifier to classify the sample features obtained from radar echo signal processing in order to achieve effective detection of complex structural targets.

[0014] Further, in step 2, the unfolded phase is obtained as follows:

[0015] Step 2.1: Fix the phase of the first pulse echo sampling point within the same distance unit. The initial phase for unfolding the phase curve, i.e.

[0016] Step 2.2, from Start calculating the phase difference with the previous sampling point when hour, and Subsequent pulse echo sampling point phase values ​​are all subtracted by π; when hour, and π is added to the phase value of each subsequent pulse echo sampling point;

[0017] Step 2.3: Repeat the above steps until all sampling point phases have been traversed to obtain the expanded phase sequence φ.

[0018] Furthermore, in step 2, the average of the expanded phase results of multiple pulses is randomly selected and defined as the average expanded phase difference ξ1:

[0019]

[0020] Where: φ i The phase is the unfolded phase corresponding to the i-th pulse in the same distance unit, and N is the number of pulses selected.

[0021] Further, in step 2, the phase mean ξ2 of the same polarization ratio is calculated as follows:

[0022] For multiple randomly selected pulses, calculate the same polarization ratio and phase of the scattered echo from a single pulse. as follows:

[0023]

[0024] In the formula: S hh S vv , respectively, represent the scattering intensities of the HH and VV polarization channels, and Re(·) and Im(·) represent the real and imaginary parts of the complex number, respectively;

[0025] The average value of the phase mean of the same polarization ratio ξ2 obtained from the results is as follows:

[0026]

[0027] In the formula: The phase of the same polarization ratio corresponding to the i-th pulse is denoted as .

[0028] Furthermore, in step 2, the range of the obtained same polarization ratio phase result is taken, and normalized using the mean to obtain the same polarization ratio phase range ξ3 as follows:

[0029]

[0030] Further, in step 2, the average amplitude values ​​of the HH channel and VV channel are calculated for the extracted pulses, and the difference is used to obtain the same polarization amplitude ratio ξ4 as follows:

[0031]

[0032] In the formula: S hh,i S vv,i , respectively, represent the scattering intensities of the HH and VV polarization channels of the i-th pulse.

[0033] Furthermore, the specific operation of step 3 is as follows: First, initialize the false alarm cost factor C in the penalty matrix. 12 and the cost factor C for missed detection 21 ;

[0034] The classifier is trained based on the initial penalty matrix, and the training false alarm probability PF1 is obtained and compared with the target false alarm probability PF0. If the absolute value of the error between PF0 and PF1 is less than the product of the error ratio K and PF0, the classifier training is complete; otherwise, the false alarm cost factor C is modified. 12 And then, further retraining.

[0035] Furthermore, the false alarm cost factor C 12 and the cost factor C for missed detection 21 Initialize to 1.

[0036] Furthermore, the error ratio K is set to 0.1.

[0037] Furthermore, in step 3, the classifier is trained using the following steps:

[0038] Step 3.1: Set the target false alarm probability PF0, error ratio K, and penalty matrix. Weight cap C h and weight lower bound C l The penalty matrix The mathematical expression is:

[0039]

[0040] In the formula: C 12 The false alarm cost factor, which corresponds to the cost of the classifier classifying sea clutter as target echo and the missed detection cost factor, C 21 This is the cost factor for missed detection, which corresponds to the cost factor for the classifier to classify the target echo as sea clutter;

[0041] Step 3.2: Train a support vector machine classifier based on the training feature set and calculate the false alarm probability PF1 at this time;

[0042] Step 3.3, verify whether the training false alarm probability PF1 is within the error range:

[0043] |PF0-PF1|≤K·PF0

[0044] If the above conditions are met, proceed to step 3.6; otherwise, proceed to the next step.

[0045] Step 3.4: Determine the relationship between the target false alarm probability PF0 and the training false alarm probability PF1. If PF0 < PF1, let C l =C 12 C 12 =0.5(C h +C l Otherwise, let C h =C 12 C 12 =0.5(C h+C l );

[0046] Step 3.5, return to step 3.2, and based on the updated penalty matrix Retrain;

[0047] Step 3.6: Complete classifier training.

[0048] Furthermore, the upper limit of the weight C h and lower limit C l 15 and 0.1 respectively.

[0049] The beneficial effects of this invention are that its technical solution addresses the target detection problem under low signal-to-clutter conditions at sea, which is of great significance. Specifically, the classifier training incorporates a mechanism for finding an appropriate false alarm cost factor, enabling the trained classifier to adapt to a preset false alarm probability, i.e., the classifier is trained specifically according to the required false alarm probability. Furthermore, the computational complexity of the implementation steps is low, making this invention easy to implement in engineering. Verification using field measurement data proves the feasibility of this invention. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the specific implementation and performance testing of a weak target detection method for sea polarization radar in this invention embodiment;

[0051] Figure 2 This is a comparison diagram of the original phase and the expanded phase obtained after processing in an embodiment of the present invention.

[0052] Figure 3 This is a comparison diagram of the mean probability distribution of the phase difference between complex structural targets and sea clutter obtained from the processing of field measurement data, provided in an embodiment of the present invention.

[0053] Figure 4 This is a comparison diagram of the mean phase distribution of the same polarization ratio between complex structural targets and sea clutter, obtained from processing field measurement data, provided in an embodiment of the present invention.

[0054] Figure 5 This is a comparison diagram of the probability distribution of the phase difference between the same polarization ratio of complex structural targets and sea clutter, obtained from the processing of field measurement data, provided in an embodiment of the present invention.

[0055] Figure 6 This is a comparison diagram of the probability distribution of the amplitude-average ratio of co-polarization of complex structural targets and sea clutter, obtained from the processing of field measurement data provided in this embodiment of the invention.

[0056] Figure 7 This is a comparison chart of the detection performance of the weak target detection method of the present invention for sea polarization radar and some traditional target detection methods on field measurement data. Detailed Implementation

[0057] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0058] Figure 1 This is a flowchart illustrating the specific implementation and performance testing of the weak target detection method used in this example for sea polarization radar, including:

[0059] Step 1: Obtain the radar echo signal training dataset. In this example, the training dataset is based on field measurement data.

[0060] Step 2: Preprocess the training dataset to obtain the training feature set, including:

[0061] The echo phase is preprocessed using the phase curve expansion method to obtain the expanded phase, and the mean value of the expanded phase difference is further calculated. The specific steps include the following:

[0062] Step 2.1: Fix the phase of the first pulse echo sampling point within the same distance unit. The initial phase for unfolding the phase curve, i.e.

[0063] Step 2.2, from Start calculating the phase difference with the previous sampling point when hour, and Subsequent pulse echo sampling point phase values ​​are all subtracted by π; when hour, and π is added to the phase value of each subsequent pulse echo sampling point;

[0064] Step 2.3: Repeat the above steps until all sampling point phases have been traversed to obtain the expanded phase sequence φ;

[0065] Step 2.4: Randomly select multiple pulses, process the results, and take the average value, which is defined as the mean of the expanded phase difference ξ1:

[0066]

[0067] Where: φ i The phase is the unfolded phase corresponding to the i-th pulse in the same distance unit; N is the number of pulses taken, and in this example, the total number of pulses in one frame is 32.

[0068] It also includes calculating polarization characteristics; in this example, polarization characteristics include the mean phase of the same polarization ratio, the phase range of the same polarization ratio, and the amplitude-average ratio of the same polarization.

[0069] For the aforementioned randomly selected multiple pulse-scattered echoes, calculate the same polarization ratio and phase of each individual pulse-scattered echo. as follows:

[0070]

[0071] In the formula: S hh S vv , respectively, represent the scattering intensities of the HH and VV polarization channels; Re(·) and Im(·) represent the real and imaginary parts of the complex number, respectively;

[0072] The average value of the phase of the same polarization ratio ξ2 is obtained by taking the average value of the results:

[0073]

[0074] In the formula: The phase of the same polarization ratio corresponding to the i-th pulse is denoted as .

[0075] Further, the range of the obtained same polarization ratio phase results is taken, and normalized using the mean to obtain the same polarization ratio phase range ξ3:

[0076]

[0077] Then, calculate the mean amplitude of the HH channel and the mean amplitude of the VV channel for the extracted pulse scattered echoes, and divide them to obtain the same polarization amplitude-average ratio ξ4:

[0078]

[0079] In the formula: S hh,i S vv,i , respectively, represent the scattering intensities of the HH and VV polarization channels of the i-th pulse.

[0080] Step 3: After processing in Step 2, the four feature quantities obtained are: mean phase difference, mean phase ratio of same polarization, phase range of same polarization, and mean amplitude ratio of same polarization. These are then used to construct feature vectors F = [ξ1, ξ2, ξ3, ξ4] for complex and simple targets. These vectors are then used as training sets and input into a pre-defined support vector machine for classifier training. The kernel function selected in the support vector machine is the Gaussian radial basis kernel function.

[0081] In the pre-defined support vector machine, the cost factor in the penalty matrix is ​​modified to improve the traditional support vector machine into one that allows setting the false alarm probability. The specific training process based on the improved support vector machine includes the following sub-steps:

[0082] Step 3.1: Set the target false alarm probability PF0, error ratio K, and penalty matrix. Weight cap C h and weight lower bound C lThe penalty matrix The mathematical expression is:

[0083]

[0084] In the formula: C ij C represents the cost factor for classifying a sample of class i as class j. 12 The false alarm cost factor, which corresponds to the cost of the classifier classifying sea clutter as target echo and the missed detection cost factor, C 21 The cost factor for missed detections is the cost factor corresponding to the classifier classifying the target echo as sea clutter; C 11 C 22 Typically, C is initialized to 0. 21 =1,C 12 =1; In this example, the error ratio K is taken as an empirical value of 0.1, and the upper limit of the weight C is... h and lower limit C l Take the empirical values ​​of 15 and 0.1 respectively.

[0085] Step 3.2: Train a support vector machine classifier based on the training feature set and calculate the false alarm probability PF1 at this time;

[0086] Step 3.3, verify whether the training false alarm probability PF1 is within the error range:

[0087] |PF0-PF1|≤K·PF0

[0088] If the above conditions are met, proceed to step 3.6; otherwise, proceed to the next step.

[0089] Step 3.4: Determine the relationship between the target false alarm probability PF0 and the training false alarm probability PF1, and use the bisection method to continuously search for the required false alarm cost between the upper and lower limits of the weights: if PF0 < PF1, let C l =C 12 C 12 =0.5(C h +C l Otherwise, let C h =C 12 C 12 =0.5(C h +C l );

[0090] Step 3.5, return to step 3.2, and based on the updated penalty matrix Retrain;

[0091] Step 3.6: Complete classifier training.

[0092] Step 4: Receive radar echo signals and obtain sample feature vectors after processing in Step 2. Use the trained classifier to classify the sample feature vectors and output classification labels to distinguish complex structural targets from sea clutter.

[0093] To verify the effectiveness and practicality of the weak target detection method using amplitude and phase feature fusion for sea-based polarimetric radar provided in this application, this embodiment further illustrates the method through an actual measurement data processing experiment using a polarimetric coherent radar:

[0094] Figure 2 An illustrative comparison of the original phase and the expanded phase obtained after processing is provided. Clearly, the original phase initially did not undergo any abrupt changes, and the expanded phase overlapped with it. Between pulses 8 and 9, and in subsequent pulses, the original phase underwent multiple abrupt changes, becoming continuous after expansion. Therefore, this method can avoid the impact of abrupt changes at π on phase analysis.

[0095] Figure 3 This is a comparison chart of the mean probability distribution of the unfolded phase difference between the echo of a complex structure target and sea clutter, obtained from processing field measurement data in this embodiment. From... Figure 3 As can be seen, the average phase difference of sea clutter spreads between 0.8 and 2.9 rad, concentrated around 2.6 rad; the phase difference of complex target features spreads between 0.25 and 2.2 rad, concentrated around 0.6 rad. Generally, the inter-pulse phase change of sea clutter is greater than that of complex target echoes. This is because the target has a large mass and a relatively small radial velocity on the undulating sea surface, thus producing a smaller phase difference within the same pulse interval.

[0096] Figure 4 This is a comparison chart of the mean phase probability distribution of the same polarization ratio of the echo of a complex structure target and sea clutter, obtained from the processing of field measurement data in this embodiment. From... Figure 4 As can be seen, the average polarization ratio phase of complex structure targets is mainly distributed between 0.2 and 2.6 rad, with a peak value around 0.6 rad, and the distribution is relatively concentrated; while the characteristic value of sea clutter ranges between 0.5 and 2.8 rad, with a peak value at 1.6 rad, and its distribution range is relatively wide and flat.

[0097] Figure 5 This is a comparison chart of the polarization ratio phase difference probability distribution between the echo of a complex structure target and sea clutter, obtained from processing field measurement data in this embodiment. From... Figure 5It can be seen that the phase difference of the same polarization ratio for complex structural targets ranges from 0.8 to 6, with a concentration around 3; while sea clutter is mainly distributed between 0.2 and 3, concentrated at 1.7. From the mathematical definition of the same polarization ratio phase, this characteristic reflects, to some extent, the echo phase difference between the HH and VV polarization channels. Therefore, it can be seen from the above two characteristic distribution diagrams that the phase difference of sea clutter between the same polarization channels is greater than that of complex structural targets, but its dispersion is less than that of complex structural targets.

[0098] Figure 6 This is a comparison chart of the co-polarization amplitude-average ratio probability distribution of the echo of a complex structure target and sea clutter, obtained from the processing of field measurement data in this embodiment. From Figure 6 As can be seen, the distribution range of sea clutter is 0.5 to 4, and the distribution is discontinuous. There is a notch at around 1.5, and the overall distribution shows a bimodal distribution. The polarization amplitude ratio of complex structure targets is roughly between 0.5 and 2.5, and is concentrated at 1.5.

[0099] Figure 7 This figure compares the detection performance of the proposed weak target detection method for sea-polarized radar in this embodiment with that of some traditional target detection methods on a field measurement data test set. The experiment was conducted with false alarm probabilities ranging from 0.001 to 0.1 to test the detection performance of the proposed method. Performance was also compared with classical statistical theory detection methods such as the cell-averaging constant false alarm (CA-CFAR) detector before and after coherent accumulation, the ordered statistic (OS) CA-CFAR detector, and the adaptive normalized matched filter (ANMF). The results show that although coherent accumulation improves the signal-to-clutter ratio, the detection performance of both CA-CFAR and OS-CFAR incoherent CFAR detectors is improved, but the accuracy is still below 60%, which cannot meet the detection requirements. ANMF also exhibits poor detection performance due to strong clutter, further demonstrating that classical statistical theory detection methods are not suitable for scenarios with strong sea clutter interference. The proposed method achieves a detection accuracy of 79.22% with a false alarm probability of 0.001, 87.34% with a false alarm probability of 0.01, and 96.44% with a false alarm probability of 0.1. Clearly, the proposed method can effectively detect weak targets on the sea surface against a background of strong sea clutter.

Claims

1. A method for detecting weak targets using a sea-based polarimetric radar, characterized in that, Specifically, the steps include the following: Step 1: Obtain the radar echo signal training dataset; Step 2: Preprocess the training dataset to obtain the training feature set, including: The phase curve expansion method is used to preprocess the echo phase in the echo signal training data to obtain the expanded phase, and the mean value of the expanded phase difference is further calculated. Calculate the phase mean of the same polarization ratio, the phase range of the same polarization ratio, and the amplitude-average ratio of the same polarization as polarization characteristics; The mean phase difference and polarization characteristics of each echo signal data are used to form a feature vector, and then the training feature set is obtained. Step 3: Substitute the training feature set into the preset support vector machine to train a classifier that can distinguish weak target echoes from sea clutter. After each training round, adjust the false alarm cost factor in the corresponding penalty matrix of the support vector machine according to the difference and magnitude between the current training false alarm probability and the preset target false alarm probability. Then, substitute the training feature set again for the next round of training until the difference between the training false alarm probability and the set false alarm probability falls within the preset error range, and then the training ends. Step 4: Use the trained classifier to classify the sample features obtained from radar echo signal processing in order to achieve effective detection of complex structural targets.

2. The method according to claim 1, characterized in that, In step 2, the unfolded phase is obtained as follows: Step 2.1: Fix the phase of the first pulse echo sampling point within the same distance unit. The initial phase for unfolding the phase curve, i.e. ; Step 2.2, sampling the phase from the second pulse echo point within the same distance cell. Start calculating the phase difference with the previous sampling point ,when hour, and The phase values ​​of subsequent pulse echo sampling points are all subtracted. ;when hour, and The phase values ​​of subsequent pulse echo sampling points are all added ; Step 2.3: Repeat the above steps until all sampling point phases have been traversed to obtain the expanded phase sequence. .

3. The method according to claim 1, characterized in that, In step 2, the average of the expanded phase results of multiple pulses is randomly selected and defined as the average expanded phase difference. : ， In the formula: For the same distance unit The expanded phase corresponding to each pulse. The number of pulses selected.

4. The method according to claim 3, characterized in that, In step 2, the phase mean of the same polarization ratio is calculated as follows: : For multiple randomly selected pulses, calculate the same polarization ratio and phase of the scattered echo from a single pulse. as follows: ， In the formula: , The scattering intensities of the HH and VV polarization channels are respectively. , These are the real and imaginary parts of a complex number, respectively. The average value of the obtained results is used to obtain the phase mean of the same polarization ratio. as follows: In the formula: For the first The same polarization ratio phase corresponding to each pulse.

5. The method according to claim 4, characterized in that, In step 2, the range of the obtained same polarization ratio phase results is taken, and then normalized using the mean to obtain the same polarization ratio phase range. as follows: 。 6. The method according to claim 3, characterized in that, In step 2, the average amplitude of the HH channel and the average amplitude of the VV channel are calculated for the extracted pulses, and the difference is used to obtain the same polarization amplitude-average ratio. as follows: In the formula: , denoted as the scattering intensities of the HH and VV polarization channels of the i-th pulse, respectively.

7. The method according to claim 1, characterized in that, Step 3 involves the following steps: First, initialize the false alarm cost factor in the penalty matrix. and the cost factor of missed detection ; The classifier is trained based on the initial penalty matrix to obtain the false alarm probability. and the target false alarm probability Compare; if and The absolute value of the error between them is less than the error ratio. and If the product of these factors is used, the classifier training is complete; otherwise, the false alarm cost factor is modified. And then, further retraining.

8. The method according to claim 7, characterized in that, False alarm cost factor and the cost factor of missed detection Initialize to 1.

9. The method according to claim 7, characterized in that, In step 3, the classifier is trained using the following steps: Step 3.1, set the target false alarm probability Error ratio Punishment Matrix Weight cap and weight lower bound The penalty matrix The mathematical expression is: , In the formula: This represents the false alarm cost factor, which corresponds to the cost of the classifier classifying sea clutter as target echoes and the missed detection cost factor. This is the cost factor for missed detection, which corresponds to the cost factor for the classifier to classify the target echo as sea clutter; Step 3.2: Train a support vector machine classifier based on the training feature set, and calculate the false alarm probability at this point. ; Step 3.3, Verify the training false alarm probability Is it within the error range? If the above conditions are met, proceed to step 3.6; otherwise, proceed to the next step. Step 3.4, determine the false alarm probability of the target. and training false alarm probability The size relationship between them, if ,make , Otherwise , ; Step 3.5, return to step 3.2, and based on the updated penalty matrix Retrain; Step 3.6: Complete classifier training.

10. The method according to claim 8, characterized in that, Weight cap and lower limit Take 15 and 0.1 respectively.

Images