Method for multi-feature identification of radar sea clutter environment

Abstract

The present application relates to a kind of radar sea clutter environment multi-feature identification method, belong to radar signal processing technical field.The present application is first to each data sequence is extracted, and normalization is completed.Four kinds of feature values of each resolution cell are sent into random forest classifier to complete determination, and the resolution cell determined as sea clutter is set as 0;Resolution cell determined as ground clutter is set as 1.The 0 / 1 matrix obtained by determination result is converted into gray scale chart, and is used as primary identification result.To improve the identification result, in turn using closed operation and hole filling is handled to primary identification result, improves the identification effect, obtains secondary processing result.The present application has higher robustness and accuracy, and the extraction of potential target area is helpful to improve target detection efficiency.

Description

Technical Field

[0001] The invention relates to a radar ground-sea clutter environment multi-feature identification method, belonging to the technical field of radar signal processing. Background Art

[0002] Detecting nearshore targets presents multiple interference issues, including ground and sea clutter. Ground clutter generated by islands and landmasses can result in patches or discrete strong clutter points, severely impacting subsequent detection. Therefore, separating ground and sea clutter in the early stages of target detection is crucial. After land and sea identification, ground clutter interference is reduced, allowing detection to focus on the identified sea area. This facilitates more efficient target detection and clutter analysis, while also yielding more accurate results.

[0003] Currently, many land and sea clutter identification schemes have been proposed based on radar echo data. Xia Xiaoyun et al. proposed a phase-based land and sea clutter identification method for shore-based radars. This method characterizes the differences between land and sea clutter using the phase change between adjacent pulses within the same unit. Deng Jiaqi et al. used a heuristic identification algorithm for the mean clutter amplitude statistic in radar echo data to globally separate stationary echo regions within each region in the range dimension and distinguish between land and sea clutter regions based on the proportion of the DC component. While both methods achieve good identification results, as single-metric identification algorithms, they both have potential characteristic limitations and are easily affected by various factors, such as sea conditions, platform conditions, and radar systems. In 2002, Bandt et al. proposed permutation entropy, which is used to describe the complexity of time series or chaotic dynamical systems. Permutation entropy exhibits high sensitivity to time, strong noise immunity, and the ability to detect sudden changes in complex nonlinear systems. Therefore, permutation entropy has a certain ability to characterize the differences in the nonstationary characteristics of land and sea clutter.

[0004] In summary, as data volumes and resolution increase, higher requirements are placed on the speed of land-sea separation algorithms. At the same time, it is necessary to combine different features to provide information at different levels and angles to cope with more complex measurement environments. Therefore, more accurate and faster land-sea clutter identification methods are needed. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and to propose a multi-feature identification method for radar land and sea clutter environments. The method has higher robustness and accuracy, and the extraction of potential target areas helps to improve target detection efficiency.

[0006] The multi-feature identification method of radar land and sea clutter environment is special in that it includes the following steps:

[0007] Step 1: Extract the echo data of 13 adjacent pulses in the same distance unit of the resolution unit to be detected to form a data sequence. The DC component ratio feature of the sequence is extracted through time series analysis, and the spectrum kurtosis feature is extracted through spectrum analysis. The permutation entropy feature is calculated after the sequence is modulo-calculated. The phase information of the sequence is extracted and the phase quadratic fitting feature is calculated to complete the feature extraction of the data to be detected.

[0008] Step 2: Construct a dataset based on the features extracted in step 1 and normalize them separately. Then, the four eigenvalues ​​of each resolution unit are combined into a feature vector and sent to the random forest classifier to complete the judgment. The 0 / 1 matrix obtained from the judgment result is transformed into a grayscale image and used as the primary recognition result. The primary recognition result is processed in sequence using isolated point removal, closing operation and hole filling to improve the recognition effect and obtain the secondary processing result.

[0009] Step 3: Use the two empirical methods of the number of condensed points and the area width greater than N meters to complete the classification of large and small areas in the secondary processing results obtained in step 2. Extract the amplitude mean of the radar echo after modulo along the azimuth dimension in each small area, compress the two-dimensional area into a one-dimensional vector. The vector length is the same as the number of distance units occupied by the area. By performing a spectral peak search on the one-dimensional vector, the number and position of peaks are obtained. Based on the number of peaks, the type of small area is identified and it is judged as a target or a small island.

[0010] Preferably, the specific steps of extracting the spectrum kurtosis feature in step 1 are:

[0011] The echo data of 13 adjacent pulses in the same range unit of the resolution unit to be detected are extracted to form a data sequence. A 256-point FFT is performed on each data vector. The spectrum kurtosis of each resolution unit is calculated based on the frequency domain data after FFT. The spectrum kurtosis of all resolution units is obtained in one scan data of the radar:

[0012]

[0013] Where K is the kurtosis of the data sequence after FFT, x i is the resolution unit to be detected, n is the number of scans to x i The total number of beams, is the average value of the unit to be tested and the adjacent units. The kurtosis of the standard normal distribution is 3. When the peak value of the time series curve is higher than that of the normal distribution, the kurtosis is greater than 3, that is, it presents a peak distribution. When it is lower than that of the normal distribution, the kurtosis is less than 3, that is, the peak value is relatively flat. The peak value has no dimension and its value range is between 1 and the number of sample data.

[0014] Preferably, the specific steps of extracting the DC component ratio feature in step 1 are:

[0015] The DC power and average power of the data sequence are expressed as formula (2) and (3), respectively denoted as s and p:

[0016]

[0017]

[0018] Where N is the number of pulses, x i is the i-th radar echo data. By calculating the ratio of the two, the DC component ratio can be obtained, which is recorded as:

[0019]

[0020] The DC component ratio ranges from [0,1], the DC component ratio of ground clutter is close to 1, and the DC component ratio of sea clutter is close to 0.

[0021] Preferably, the specific steps of extracting the phase quadratic fitting feature in step 1 are:

[0022] Select 13 adjacent echo data under the same distance unit of the detection unit, extract the phase value of each data to form a new sequence, and perform phase unwrapping processing. After phase unwrapping, use the least squares method to perform quadratic fitting on the new data sequence, and complete feature extraction by calculating the fitting error. Use the least squares method to perform quadratic fitting on the new data sequence, and compare each original phase data with the corresponding fitting value to calculate the residual, that is, the residual is equal to the original phase value minus the fitting phase value. Then, sum the squares of all residuals to obtain the total fitting error, which is the phase quadratic fitting feature.

[0023] Preferably, the specific steps of extracting the permutation entropy feature in step 1 are:

[0024] Select N adjacent pulse data under the same distance unit of the detection unit and take the modulus to form a time series {X(1),X(2),...,X(N)}. Use the phase space reconstruction method to obtain the reconstruction matrix:

[0025]

[0026] The time series N is 13, m and t represent the embedding dimension and delay time respectively. The embedding dimension m is 4 and the delay time t is 1. Each row in the reconstruction matrix is ​​a reconstruction component, with a total of K. The reconstruction components are rearranged in ascending order to obtain a new sequence. The column index of each element position in the new sequence constitutes a set of symbol sequences:

[0027] S(l)={j1,j2,…,j m}, l=1,2,…,k, and k≤m! (6)

[0028] There are m! kinds of symbol sequences in the m-dimensional phase space mapping. The probability of each symbol sequence appearing is calculated by dividing the number of times each symbol sequence appears by the total number of times m! different symbol sequences appear, that is, P1, P2, ..., P k , the permutation entropy is obtained based on the probability of occurrence of various symbol sequences:

[0029]

[0030] When P j =1 / m!, H pe (m) obtains the maximum value ln(m!), and normalizes the permutation entropy, that is: 0≤H pe =H pe (m) / ln(m!)≤1.

[0031] Preferably, the specific steps of step 2 are:

[0032] Step 2-1: Construct a data set based on the four features extracted in step 1 and normalize them separately to obtain an N*4 data matrix, where N is the number of samples to be detected. The four eigenvalues ​​of each resolution unit form a feature vector and are sent to the random forest classifier for judgment. The resolution unit judged as sea clutter is set to 0; the resolution unit judged as ground clutter is set to 1. The result obtained in this step is the random forest judgment result.

[0033] Step 2-2: A 0 / 1 matrix is ​​formed based on the random forest judgment results obtained in step 2-1, and the matrix is ​​transformed into a grayscale image. This matrix is ​​used as the primary identification result. The primary identification result is processed in sequence using isolated point removal, closing operation, and hole filling to obtain the optimized post-processing result. The isolated point removal is to delete all connected components with less than P pixels from the primary identification result, where the value of P can be determined according to the specific situation of the data to be detected. The closing operation is to design a certain structural element X. The result after removing the isolated points is first expanded by X and then eroded by X. X also needs to be determined according to the situation of the data to be detected. The result of the above operations is to disconnect the small connections in each area as much as possible to facilitate the implementation of the subsequent step 3. The hole filling operation has no parameter setting and is directly applied to the result after the closing operation.

[0034] Preferably, the specific steps of step 3 are:

[0035] First, use the two empirical methods of condensed points and area width greater than N meters to complete the classification of large and small areas in the secondary processing results obtained in step 2. The empirical method of condensed points refers to extracting the connected areas determined as ground clutter in the secondary processing results one by one, checking the number of sampling points occupied by the connected areas, and treating the areas with a number of sampling points greater than P as land areas. Here, the P value is an empirical value and needs to be determined in combination with the actual area size represented by a single sampling point in the echo data to be measured. It is required that P*single sampling point area>conventional small islands. The empirical method of area width greater than N meters is to calculate the center of mass of each connected area for the remaining connected areas after condensed points, and use the horizontal coordinate K of the center of mass as the reference to check how many azimuth codes the connected area occupies in the vertical coordinate. At this time, let the number of occupied azimuth codes be J, and the angle interval occupied by each azimuth code be x, then the area width L is:

[0036] L=2Ktan(Jx / 2) (8)

[0037] Thus, the areas with a width greater than N meters are selected as land areas, where the width N is an empirical value;

[0038] For the remaining small areas after screening in the above steps, the amplitude mean of the radar echo after modulo is extracted along the azimuth dimension, and the two-dimensional area is compressed into a one-dimensional vector with the same length as the number of distance units occupied by the small area. By performing a spectrum peak search on the one-dimensional vector, the number and position of the peaks are obtained.

[0039] The present invention's multi-feature identification method for radar land and sea clutter environments selects spectral kurtosis, DC component proportion, phase quadratic fit, and permutation entropy to form a feature vector. A random forest learning algorithm is used to classify sea and land clutter data to form a primary identification image. Morphological filtering is then used to eliminate isolated points and fill holes in the primary identification results, thereby obtaining a secondary identification image. Because anchored vessels and small islands in radar echo data (after pulse compression) have similar feature expressions, the patented algorithm distinguishes potential target areas from island areas in the identification results and extracts the location information of potential targets within the potential target areas. Experimental results demonstrate that the algorithm can accurately identify land and sea clutter and provide more accurate regional guidance for sea surface target detection. By analyzing land and sea clutter characteristics, the present invention selects four features for random forest classification. The classification results are then subjected to morphological filtering to ensure regional connectivity. Potential target areas and islands are distinguished by compressing two-dimensional regions into one-dimensional vectors and distinguishing between them based on the difference in the number of spectral peaks. Verification of measured data demonstrates that this method effectively identifies land and sea areas, potential target areas, and island regions, providing potential target location information. Compared to other identification methods, this method offers greater robustness and accuracy, and its extraction of potential target areas helps improve target detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 :It is the algorithm flow chart of the present invention;

[0041] Figure 2 : Diagram of fitting error extraction process;

[0042] Figure 3 : is the distribution result diagram of the permutation entropy value; in the figure (a) is the distribution result diagram of the permutation entropy value when the embedding dimension is 3 under the same region; (b) is the distribution result diagram of the permutation entropy value when the embedding dimension is 4 under the same region; (c) is the distribution result diagram of the permutation entropy value when the embedding dimension is 5 under the same region;

[0043] Figure 4 : Result diagrams of each optimization step of the random forest recognition result; (a) shows the recognition result obtained after the four eigenvalues ​​are fed into the random forest; (b) shows the image after removing isolated points by performing a closing operation on the random forest recognition result; (c) shows the image after removing noise by performing a closing operation; (d) shows the image after hole filling after the closing operation.

[0044] Figure 5 : The classification results of potential target areas; (a) is the large area screening map; (b) is the extraction area map with a width of >350 meters; (c) is the extraction potential target area map. DETAILED DESCRIPTION

[0045] For better understanding and implementation, the following specific embodiments are given to explain in detail a radar ground-sea clutter environment multi-feature identification method of the present invention. The following is a detailed description of the processing flow of the present invention:

[0046] The algorithm flow chart of this embodiment is as follows Figure 1The radar echo data after pulse compression is analyzed, and the features of each data sequence are extracted and normalized. Then, the four feature values of each resolution cell are combined into a feature vector, which is input into the random forest classifier to complete the judgment. The resolution cells judged as sea clutter are set to 0, and the resolution cells judged as ground clutter are set to 1. The 0 / 1 matrix obtained by the judgment result is converted into a gray image, which is used as the primary recognition result. In order to improve the recognition result, the primary recognition result is processed by using the closing operation and hole filling in turn to improve the recognition effect and obtain the secondary processing result. In order to solve the judgment problem of the target area and small islands which are greatly affected by the pulse compression distance sidelobes, the present patent first uses two empirical standards of the number of condensation points and the area width greater than N meters to screen large areas, and then extracts a plurality of small areas left after screening. The amplitude mean value of the radar echo after taking the modulus along the azimuth dimension in each small area is extracted, and the two-dimensional area is compressed into a one-dimensional vector. The vector length is the same as the number of distance cells occupied by the area. The peak value number and the peak value position are obtained by performing spectral peak search on the one-dimensional vector. According to the peak value number, a series of small areas are identified to determine whether they are targets or small islands. The following is a step-by-step explanation:

[0047] Step 1 (ground and sea clutter feature extraction):

[0048] (1) Spectral kurtosis feature extraction

[0049] The radar beam presents a petal shape. In actual work, the data of a certain resolution cell is often collected at the maximum gain point. However, due to the influence of the main lobe beam width, the adjacent resolution cell information is also carried at the non-maximum gain point of the main lobe. That is, the same resolution cell can be scanned by multiple adjacent beams. Therefore, in combination with the measured data used in the present patent, the present patent selects to extract the echo data of 13 adjacent pulses of the resolution cell to be detected at the same distance unit to form a data sequence. The 256-point FFT is performed on each data vector, and the spectral kurtosis of each resolution cell is calculated according to the frequency domain data after FFT. Through this extraction method, the spectral kurtosis of all resolution cells can be obtained in one scan data of the radar.

[0050]

[0051] In the formula, K is the kurtosis of the data sequence after FFT, x i is the resolution cell to be detected, n is the total number of beams scanning x i , is the average value of the unit to be measured and the adjacent unit, and the kurtosis of the standard normal distribution is 3. When the curve peak value of the time sequence is higher than the normal distribution, the kurtosis is greater than 3, that is, it presents a sharp peak distribution. When it is lower than the normal distribution, the kurtosis is less than 3, that is, it presents a relatively flat peak value. The peak value has no dimension and the value range is between 1 and the number of sample data.

[0052] (2) Feature extraction of DC component ratio

[0053] The proportion of the DC component of the signal can be obtained by the ratio of DC power to average power. The DC power and average power expressions of the data sequence are shown in equations (2) and (3), denoted as s and p respectively:

[0054]

[0055]

[0056] Where N is the number of pulses, x i is the i-th radar echo data. By calculating the ratio of the two, the proportion of the DC component can be obtained, which is recorded as:

[0057]

[0058] The DC component ratio ranges from [0,1], the DC component ratio of ground clutter is close to 1, and the DC component ratio of sea clutter is close to 0.

[0059] (3) Phase quadratic fitting feature extraction

[0060] Select 13 adjacent echo data points within the same range bin of the detection unit and extract the phase values ​​from each data point to form a new sequence. Note that phase wrapping must be unwrapped. Phase wrapping means that the calculated phase value is not the true phase. Instead, the true phase falls on the range [-π, π] after a 2π cycle. Therefore, the calculated value needs to be unwrapped.

[0061] After phase unwrapping, the least squares method can be used to perform quadratic fitting on the new data sequence, and feature extraction can be completed by calculating the fitting error. The extraction process is as follows: Figure 2 shown.

[0062] (4) Permutation entropy feature extraction

[0063] Select N adjacent pulse data in the same distance unit of the detection unit and take the modulus to form a time series {X(1), X(2), ..., X(N)}. Use the phase space reconstruction method to obtain the reconstruction matrix.

[0064]

[0065] Here, the time series N is 13, and m and t represent the embedding dimension and delay time, respectively. Based on the experimental results above, the embedding dimension m is set to 4. Since the number of pulses is small, there is no need to downsample the signal, so the delay time t is set to 1. Each row in the reconstruction matrix represents a reconstruction component, for a total of K. The reconstruction components are rearranged in ascending order to obtain a new sequence. The column index of each element position in the new sequence forms a set of symbol sequences:

[0066] S(l)={j1,j2,…,j m}, l=1,2,…,k, and k≤m! (6)

[0067] There are m! kinds of symbol sequences in the m-dimensional phase space mapping. The probability of each symbol sequence appearing is calculated by dividing the number of times each symbol sequence appears by the total number of times m! different symbol sequences appear, that is, P1, P2, ..., P k The permutation entropy is obtained based on the probability of occurrence of various symbol sequences:

[0068]

[0069] When P j =1 / m!, H pe (m) obtains the maximum value ln(m!), so the permutation entropy can be normalized, that is: 0≤H pe =H pe (m) / ln(m!)≤1.

[0070] The extracted features are used to construct a data set and normalized. The four eigenvalues ​​of each resolution unit are then combined into a feature vector and fed into a random forest classifier to complete the judgment. The resulting 0 / 1 matrix is ​​transformed into a grayscale image and used as the primary recognition result. The primary recognition result is then processed in sequence using isolated point removal, closing operations, and hole filling to improve the recognition effect and obtain the secondary processing result, as shown below.

[0071] Construct a dataset using the four features extracted in step 1 and normalize each one to create an N*4 data matrix (N is the number of samples to be tested). The four eigenvalues ​​of each resolution unit form a feature vector and are fed into a random forest classifier for classification. Recognize resolution units identified as sea clutter as 0, and those identified as ground clutter as 1. The result of this step is the random forest classification result.

[0072] The above is the identification result of the trained model directly. If you need to analyze based on a self-constructed dataset, you need to retrain the random forest model and adjust the model parameters based on the relationship between generalization error and model complexity. This patent uses learning curves and grid search methods to continuously adjust the model complexity to get as close as possible to the lowest point of generalization error (Note: There are no strict restrictions on the parameter adjustment method in the actual method, and third-party libraries can also be used for automatic parameter adjustment). The training parameters mainly focused on in this patent are the number of subtrees, the maximum growth depth of the tree, and the minimum number of leaf samples.

[0073] The random forest judgment results obtained according to the above steps are composed into a 0 / 1 matrix, which is then transformed into a grayscale image and used as the primary identification result. The primary identification results are processed in sequence using isolated point removal, closing operation, and hole filling to obtain the optimized post-processing result. Removing isolated points means deleting all connected components with less than P pixels from the primary identification result, where the value of P can be determined according to the specific situation of the data to be detected. The closing operation is to design a certain structural element X, and the result after removing isolated points is first expanded by X and then eroded by X. X also needs to be determined according to the situation of the data to be detected. The result of the above operations is to disconnect the small connections in each area as much as possible to facilitate the implementation of the subsequent step 3. The hole filling operation has no parameter settings and can be directly used on the result after the closing operation.

[0074] Using the two empirical methods of clustering points and region width greater than N meters, we classify the small and large regions in the secondary processing results obtained in the above steps. The amplitude mean of the radar echo modulo along the azimuth dimension is extracted within each small region. The two-dimensional region is compressed into a one-dimensional vector with a length equal to the number of distance units occupied by the region. A spectral peak search is performed on the one-dimensional vector to determine the number and location of peaks. Based on the number of peaks, the small region type is identified, determining whether it is a target or a small island. The details are as follows:

[0075] First, the two empirical methods of condensation points and area width greater than N meters are used to complete the classification of large and small areas in the secondary processing results obtained in the above steps. The empirical method of condensation points refers to extracting the connected areas determined as ground clutter in the secondary processing results one by one, checking the number of sampling points occupied by the connected areas, and treating the areas with more than P sampling points as land areas. The P value here is an empirical value, which needs to be determined in combination with the actual area size represented by a single sampling point in the echo data to be measured, requiring P*single sampling point area>conventional small islands. The empirical method of area width greater than N meters is to calculate the center of mass of each connected area for the remaining connected areas after the condensation points, and use the horizontal coordinate K of the center of mass (i.e., the distance unit in the B display image) as the reference to check how many azimuth codes the connected area occupies in the vertical coordinate. At this time, let the number of occupied azimuth codes be J, and the angle interval occupied by each azimuth code be x, then the area width L is:

[0076] L=2Ktan(Jx / 2) (8)

[0077] Thus, areas with a width greater than N meters are selected as land areas. The width N is an empirical value, usually no greater than the length of a large ship, but it also needs to be selected based on actual conditions.

[0078] The remaining small area after screening by the above steps is extracted along the azimuth dimension to obtain the amplitude mean value of the radar echo after mode extraction. The two-dimensional area is compressed into a one-dimensional vector, and the vector length is the same as the number of distance units occupied by the small area. The number of peaks and the peak position are obtained by performing spectral peak search on the one-dimensional vector. Since the potential target to be detected in this paper is more prominent in the time series amplitude value, the spectral peak search algorithm in this step has no strict requirements, but the minimum peak convex height needs to be considered to prevent the algorithm performance from being reduced due to too many peak values. In the ideal state, the peak number histogram of the island area and the potential target area will have a relatively obvious double peak, and the smaller cluster of peak values is the potential target area. Thus, the small area type recognition is completed according to the peak value, and the potential target area is basically extracted.

[0079] Since the embedding dimension is an important parameter for calculating permutation entropy, it affects the final calculation result. To achieve better recognition effect, further experimental analysis is needed to analyze the influence of embedding dimension on permutation entropy value, and to determine when the embedding dimension is taken as a value, the difference of sea clutter is most significant. Figure 3 (a), (b), (c) represent the distribution results of permutation entropy values when the embedding dimension is 3, 4 and 5 respectively.

[0080] As shown by the experimental results, Figure 3 (a) When the embedding dimension is 3, the entropy value curve of the sea clutter is relatively chaotic, and there are many overlapping parts, so it is difficult to find a threshold to distinguish the sea clutter and the ground clutter.(b) When the embedding dimension is 4, it can be seen that the sea clutter curve is significantly larger than the ground clutter curve. Compared with the embedding dimension of 3, the sea clutter entropy curve is more stable, and the overlapping points of the two clutter curves are significantly reduced.(c) When the embedding dimension is 5, the entropy value curves of the two types of clutter tend to be stable. However, compared with the embedding dimension of 4, the discrimination degree of the two clutters does not increase significantly, and the increase of the embedding dimension makes the calculation more complex, and the calculation amount also increases. According to the above research, considering the recognition effect and the calculation amount, the embedding dimension selected in the subsequent feature extraction is 4.

[0081] Step 2: Random forest classification result generation and classification result optimization:

[0082] The four feature values of each resolution unit are combined into a feature vector and input into the random forest classifier to complete the judgment. The resolution unit judged as sea clutter is set to 0, and the resolution unit judged as ground clutter is set to 1. The 0 / 1 matrix obtained by the judgment result is converted into a gray scale image and used as the primary recognition result. In order to improve the recognition result, the primary recognition result is processed in turn using the closing operation and hole filling to improve the recognition effect and obtain the optimized processing result.

[0083] Figure 4(a) shows the recognition result after feeding the four eigenvalues ​​into the random forest. (b) shows the image after removing isolated points by performing a closing operation on the random forest recognition result. (c) shows the image after removing noise by performing a closing operation. (d) shows the image after filling holes after the closing operation.

[0084] It can be seen from the experimental results that compared with the single-feature identification method, the multi-feature identification results have a certain degree of improvement on the sidelobe effect of the target area and the land area fragmentation problem mentioned above. It can be seen from the land and sea clutter boundary obtained after hole filling that the land and sea clutter boundary obtained by the method of this patent is closer to the coastline in the radar P display image. Morphological filtering is mainly achieved through set operations by structuring elements of different forms. The structuring elements are composed of 0 / 1 matrices, and the area with 1 defines the scope of the morphological operation. Under the premise of ensuring the basic outline of the land and sea clutter boundary, the structuring elements used in this patent when processing the random forest identification results are smaller than the structuring elements used by the above-mentioned single-feature method, which means that more boundary details are retained.

[0085] Therefore, the method of this embodiment comprehensively utilizes the complementarity and correlation between multiple land and sea clutter classification features to more comprehensively describe the difference characteristics of different types of clutter, thereby improving the accuracy and stability of classification.

[0086] Step 3: Extract potential target areas

[0087] Figure 5 The algorithm uses multi-feature recognition results to filter large areas, extracting regions with a width greater than N meters (the value of N takes into account the common sizes of vessels in the monitored sea area). It then classifies and determines the remaining small regions, providing the location information of suspected targets within those regions. In (a) and (b), dark blue represents land areas identified by the algorithm, while light blue and yellow in (c) represent suspected islands and suspected targets, respectively.

[0088] Depend on Figure 5 (b) It can be seen that after filtering the large areas and extracting the areas with width > 350 meters, there are still many small areas in the image. Figure 5 (c) It can be seen that after the potential target area detection, most of the islands and land parts of interest in this data can be identified. In addition, the areas that were misjudged as islands (such as Figure 5(c) The red box is the sidelobe area that is separated from the actual target after image processing. Such areas will not affect subsequent target detection. For island areas that are mistakenly judged as targets, the detection algorithm can be used to further detect the small area of ​​suspected targets. Comprehensively extracting potential target areas and analyzing their classification results, the land and sea clutter identification algorithm basically meets the requirements of practical applications, and possible false alarms can also be further eliminated through subsequent processing. In addition, compared with the full-area detection in previous algorithms, the method of target detection in the potential target area greatly saves calculation amount and detection time. In practical applications, the use of this patented method can complete land and sea identification and potential target area identification at one time, which helps to optimize the marine target detection process and effectively save computing resources.

[0089] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. Radar land and sea clutter environment multi-feature identification method, characterized by The following steps are involved: Step 1: Extract the echo data of 13 adjacent pulses in the same distance unit of the resolution unit to be detected to form a data sequence. The DC component ratio feature of the sequence is extracted through time series analysis, and the spectrum kurtosis feature is extracted through spectrum analysis. The permutation entropy feature is calculated after the sequence is modulo-calculated. The phase information of the sequence is extracted and the phase quadratic fitting feature is calculated to complete the feature extraction of the data to be detected. Step 2: Construct a dataset based on the features extracted in step 1 and normalize them separately. Then, the four eigenvalues ​​of each resolution unit are combined into a feature vector and sent to the random forest classifier to complete the judgment. The 0 / 1 matrix obtained from the judgment result is transformed into a grayscale image and used as the primary recognition result. The primary recognition result is processed in sequence using isolated point removal, closing operation and hole filling to improve the recognition effect and obtain the secondary processing result. Step 3: Use the two empirical methods of the number of condensed points and the area width greater than N meters to complete the classification of large and small areas in the secondary processing results obtained in step 2. Extract the amplitude mean of the radar echo after modulo along the azimuth dimension in each small area, compress the two-dimensional area into a one-dimensional vector. The vector length is the same as the number of distance units occupied by the area. By performing a spectral peak search on the one-dimensional vector, the number and position of peaks are obtained. Based on the number of peaks, the type of small area is identified and it is judged as a target or a small island.

2. The radar ground-sea clutter environment multi-feature identification method according to claim 1, characterized in that The specific steps of extracting the spectrum kurtosis feature in step 1 are: The echo data of 13 adjacent pulses in the same range unit of the resolution unit to be detected are extracted to form a data sequence. A 256-point FFT is performed on each data vector. The spectrum kurtosis of each resolution unit is calculated based on the frequency domain data after FFT. The spectrum kurtosis of all resolution units is obtained in one scan data of the radar: Where K is the kurtosis of the data sequence after FFT, x i is the resolution unit to be detected, n is the number of scans to x i The total number of beams, is the average value of the unit to be tested and the adjacent units. The kurtosis of the standard normal distribution is 3. When the peak value of the time series curve is higher than that of the normal distribution, the kurtosis is greater than 3, that is, it presents a peak distribution. When it is lower than that of the normal distribution, the kurtosis is less than 3, that is, the peak value is relatively flat. The peak value has no dimension and its value range is between 1 and the number of sample data.

3. The radar ground-sea clutter environment multi-feature identification method according to claim 1, characterized in that The specific steps of DC component ratio feature extraction in step 1 are: The DC power and average power of the data sequence are expressed as formula (2) and (3), respectively denoted as s and p: Where N is the number of pulses, x i is the i-th radar echo data. By calculating the ratio of the two, the DC component ratio can be obtained, which is recorded as: The DC component ratio ranges from [0,1], the DC component ratio of ground clutter is close to 1, and the DC component ratio of sea clutter is close to 0.

4. The radar ground-sea clutter environment multi-feature identification method according to claim 1, characterized in that The specific steps of phase quadratic fitting feature extraction in step 1 are: Thirteen adjacent echo data under the same distance unit of the detection unit are selected, and the phase values ​​of each data are extracted to form a new sequence. The phase is unwrapped. After the phase is unwrapped, the new data sequence is quadratically fitted using the least squares method. The feature extraction is completed by calculating the fitting error. The new data sequence is quadratically fitted using the least squares method. By comparing each original phase data with the corresponding fitting value, the residual is calculated, that is, the residual is equal to the original phase value minus the fitting phase value. Then, the sum of the squares of all residuals is summed to obtain the total fitting error, which is the phase quadratic fitting feature.

5. The radar ground-sea clutter environment multi-feature identification method according to claim 1, characterized in that The specific steps of permutation entropy feature extraction in step 1 are: Select N adjacent pulse data under the same distance unit of the detection unit and take the modulus to form a time series {X(1),X(2),…,X(N)}. Use the phase space reconstruction method to obtain the reconstruction matrix: The time series N is 13, m and t represent the embedding dimension and delay time respectively. The embedding dimension m is 4 and the delay time t is 1. Each row in the reconstruction matrix is ​​a reconstruction component, with a total of K. The reconstruction components are rearranged in ascending order to obtain a new sequence. The column index of each element position in the new sequence constitutes a set of symbol sequences: S(l)={j1,j2,…,j m }, l=1,2,…,k, and k≤m! (6) There are m! kinds of symbol sequences in the m-dimensional phase space mapping. The probability of each symbol sequence appearing is calculated by dividing the number of times each symbol sequence appears by the total number of times m! different symbol sequences appear, that is, P1, P2, ..., P k , the permutation entropy is obtained based on the probability of occurrence of various symbol sequences: When P j =1 / m!, H pe (m) obtains the maximum value ln(m!), and normalizes the permutation entropy, that is: 0≤H pe =H pe (m) / ln(m!)≤1.

6. The radar ground-sea clutter environment multi-feature identification method according to claim 1, characterized in that The specific steps of step 2 are: Step 2-1: Construct a data set based on the four features extracted in step 1 and normalize them separately to obtain an N*4 data matrix, where N is the number of samples to be detected. The four eigenvalues ​​of each resolution unit form a feature vector and are sent to the random forest classifier for judgment. The resolution unit judged as sea clutter is set to 0; the resolution unit judged as ground clutter is set to 1. The result obtained in this step is the random forest judgment result. Step 2-2: A 0 / 1 matrix is ​​formed based on the random forest judgment results obtained in step 2-1, and the matrix is ​​transformed into a grayscale image. This matrix is ​​used as the primary identification result. The primary identification result is processed in sequence using isolated point removal, closing operation, and hole filling to obtain the optimized post-processing result. The isolated point removal is to delete all connected components with less than P pixels from the primary identification result, where the value of P can be determined according to the specific situation of the data to be detected. The closing operation is to design a certain structural element X. The result after removing the isolated points is first expanded by X and then eroded by X. X also needs to be determined according to the situation of the data to be detected. The result of the above operations is to disconnect the small connections in each area as much as possible to facilitate the implementation of the subsequent step 3. The hole filling operation has no parameter setting and is directly applied to the result after the closing operation.

7. The radar ground-sea clutter environment multi-feature identification method according to claim 6, characterized in that The specific steps of step 3 are: First, use the two empirical methods of condensed points and area width greater than N meters to complete the classification of large and small areas in the secondary processing results obtained in step 2. The empirical method of condensed points refers to extracting the connected areas determined as ground clutter in the secondary processing results one by one, checking the number of sampling points occupied by the connected areas, and treating the areas with a number of sampling points greater than P as land areas. Here, the P value is an empirical value and needs to be determined in combination with the actual area size represented by a single sampling point in the echo data to be measured. It is required that P*single sampling point area>conventional small islands. The empirical method of area width greater than N meters is to calculate the center of mass of each connected area for the remaining connected areas after condensed points, and use the horizontal coordinate K of the center of mass as the reference to check how many azimuth codes the connected area occupies in the vertical coordinate. At this time, let the number of occupied azimuth codes be J, and the angle interval occupied by each azimuth code be x, then the area width L is: L=2Ktan(Jx / 2) (8) Thus, the areas with a width greater than N meters are selected as land areas, where the width N is an empirical value; For the remaining small areas after screening in the above steps, the amplitude mean of the radar echo after modulo is extracted along the azimuth dimension, and the two-dimensional area is compressed into a one-dimensional vector with the same length as the number of distance units occupied by the small area. By performing a spectrum peak search on the one-dimensional vector, the number and position of the peaks are obtained.

Images