A High-Frequency Ground Wave Radar Target Detection Method Based on Decision Tree Classification

Abstract

The present invention provides a high-frequency ground wave radar target detection method based on decision tree classification (DTC). First, the longitude, latitude, azimuth, and speed information of AIS vessels are mapped onto the radar range-Doppler map through coordinate transformation. Then, peak and constant false alarm rate detections are performed on the radar range-Doppler map, and the detected target signals are matched with AIS targets to obtain positive and negative sample sets for decision tree classification. Next, the feature vectors of the samples in three dimensions, namely the range dimension, Doppler dimension, and two-dimensional dimension, are obtained, and the signal-to-noise ratio features are calculated for each dimension through strategies such as cell averaging, ordered statistics, deleted mean, subtracted mean, maximum, minimum, and first-order difference minimum. The measured results show that the detection method based on decision tree classification can integrate the advantages of different calculation strategies, realize the distinction between "true" and "false" targets in the feature space, and can achieve the false alarm adjustment function by changing the number of training samples. Compared with the traditional sliding window CFAR method, the detection ability of vessel targets is improved.

Description

Technical Field

[0001] The present invention relates to the technology of detecting moving ship targets on the sea surface by high-frequency ground-wave radar, and particularly to a method for detecting targets by high-frequency ground-wave radar based on decision tree classification. Background Art

[0002] High-frequency ground-wave radar can achieve over-the-horizon detection of "soft targets" and "hard targets" on the sea surface by utilizing the characteristics of the diffraction of vertically polarized electromagnetic waves with frequencies of 3 - 30 MHz along the sea surface, and has the ability of all-weather and real-time operation. According to the antenna form, high-frequency ground-wave radar can be divided into array type and compact type. The compact high-frequency ground-wave radar has the advantages of small floor area, easy maintenance and installation, and low cost, and has been emphasized and widely used by coastal countries. However, affected by factors such as sea clutter, noise, and external interference, the echo signals of ship targets on the sea surface are complex and variable, showing low observable characteristics, resulting in a low detection probability.

[0003] The inventors of the present invention found through experimental research that: after pulse compression and coherent accumulation of ship targets, they show a two-dimensional Sinc function in range-Doppler, presenting a spike shape similar to a cone. The AIS ship information can be mapped into the range-Doppler map through coordinate transformation, but it does not necessarily fall on the position of the radar echo spike. In order to increase the credibility of positive samples, a method of matching the target sets detected by peak detection and CFAR with AIS targets is used to obtain the sample set. In addition, the noise floor calculation method of the sliding window type CFAR technology is single, and the form of the reference window is fixed and cannot be changed once set, making it difficult to intelligently fuse multiple strategies for target decision-making, resulting in poor target detection performance of high-frequency ground-wave radar. Therefore, the present invention proposes a method for detecting targets by high-frequency ground-wave radar based on decision tree classification by means of multi-feature fusion. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting targets by high-frequency ground-wave radar based on decision tree classification, which is used to solve the technical problem of low target detection probability of high-frequency ground-wave radar in the existing technical methods.

[0005] In order to achieve the above-mentioned invention purpose, the present invention adopts the following technical solutions. A method for detecting targets by high-frequency ground-wave radar based on decision tree classification includes:

[0006] S1: Calculate the distance between the ship and the radar station and the azimuth angle relative to the radar station according to the longitude and latitude information of the ship; then calculate the radial velocity of the ship relative to the radar by using the heading, speed, and azimuth angle information of the AIS ship relative to the radar, and then convert the distance and radial velocity into specific range-Doppler units of the radar according to the range and Doppler resolutions of the radar;

[0007] S2: Obtain radar target signals by means of peak detection + range - dimension constant false - alarm rate (CFAR) detection + Doppler - dimension CFAR detection, so that the radar - detected target signals match the AIS targets. The successfully - matched targets are used as the positive - sample set for decision - tree training, with the label of "+1", and the label of negative samples is "-1". The ratio of positive to negative samples is 1:1;

[0008] S3: Select reference - cell clutter in the range dimension, Doppler dimension, and two - dimensional dimension;

[0009] S4: Extract the signal - to - noise ratio (SNR) eigenvalues of the target signals in the range dimension, Doppler dimension, and two - dimensional rectangular window by means of cell - averaging, ordered - statistics, deleted - mean, subtracted - mean, maximum, minimum, and first - order - difference minimum methods. Before training, normalize the sample data of different feature dimensions to the range of 0 to 1;

[0010] S5: When generating the decision tree, divide the sample attributes of the decision - tree nodes to generate a decision tree with strong generalization ability;

[0011] S6: Use the generated decision - tree model to classify new input data. If the predicted label is "+1", it is a "true" target; if the predicted label is "-1", it is a false target.

[0012] Preferably, in S2, the method for obtaining radar target signals is as follows:

[0013] S2.1: First, perform peak detection on a set of radar data;

[0014] S2.2: Detect radar target signals in the range dimension according to the set false - alarm rate;

[0015] S2.3: Detect radar target signals in the Doppler dimension according to the set false - alarm rate;

[0016] S2.4: The intersection of the three - time detected target sets is the finally - detected radar - target - signal set;

[0017] When the radar - detected target signals match the AIS targets, the range error does not exceed one range - resolution cell, and the velocity error does not exceed three Doppler - resolution cells.

[0018] Preferably, in S4, the calculation method of the eigenvalues is as follows:

[0019] (1) The cell - averaging method selects the average power of all reference cells as the estimate of the clutter - power level and calculates the target SNR:

[0020]

[0021] (2) The ordered - statistics method sorts all reference cells according to their power magnitudes and calculates the target SNR:

[0022]

[0023] (3) The mean deletion method sorts all reference units according to their power levels, then deletes the r larger reference values starting from the maximum value, and takes the average of the remaining reference units as the estimated clutter power level to calculate the target signal-to-noise ratio:

[0024]

[0025] (4) The trimmed mean method sorts all reference units according to their power levels, cuts off the r1 smaller reference units starting from the minimum value and the r2 larger reference units starting from the maximum value, and takes the average of the remaining reference units as the estimated clutter power level to calculate the target signal-to-noise ratio:

[0026]

[0027] (5) The maximum value method sorts all reference units according to their power levels, takes the maximum value as the estimated clutter power level to calculate the target signal-to-noise ratio:

[0028]

[0029] (6) The minimum value method sorts all reference units according to their power levels, takes the minimum value as the estimated clutter power level to calculate the target signal-to-noise ratio:

[0030]

[0031] (7) The first-order difference method sorts all reference units according to their power levels to obtain the first-order difference values, and then takes the reference unit corresponding to the position of the smallest first-order difference value as the estimated clutter power level to calculate the target signal-to-noise ratio:

[0032]

[0033] In the above formula, S represents the power of the target signal, R represents the range dimension, Doppler dimension, and two-dimensional sliding window reference unit sample power, where R i , i = 1, …, 2n, diff(R) represents taking the first-order difference of the reference samples, and index() represents the index of the minimum value.

[0034] Preferably, in S5:

[0035] The information gain method is used to perform attribute partitioning on the decision tree nodes, and through recursion and attribute partitioning, a decision tree with strong generalization ability is generated:

[0036] Assume that the proportion of samples of the kth class in the current training sample set D is p k(k = 1, 2, …, M), the information entropy of D is defined as:

[0037]

[0038] The smaller the value of Ent(D), the higher the purity of D. According to the different numbers of samples contained in different branch nodes, the weight |D v | / |D| is assigned to the branch node, and the "information gain" obtained by dividing the sample set D by the characteristic attribute a can be calculated

[0039]

[0040] The greater the information gain, the greater the "purity improvement" obtained by dividing using the attribute a. Therefore, the attribute selection of the decision tree is carried out using the information gain, that is

[0041]

[0042] a * represents the optimal division attribute of the current node, and A is the set of training sample attributes.

[0043] The inventive concept of the present invention is as follows:

[0044] The high-frequency ground wave radar target detection method based on decision tree classification first maps the AIS vessel information onto the radar range-Doppler map through coordinate transformation, then obtains the radar target signal by using the intersection of the peak detection + range dimension constant false alarm detection + Doppler dimension constant false alarm detection set, obtains the positive sample set of decision tree classification by matching with the AIS target, and randomly selects negative samples in the non-AIS target area, with the positive and negative sample ratio being 1:1; then obtains eigenvalue data in the range dimension, Doppler dimension, and two-dimensional dimension by using methods such as cell averaging, ordered statistics, deleted mean, subtracted mean, maximum value taking the minimum value, and first-order difference minimum value, and traverses each radar data to obtain the training sample set; finally, uses the trained decision tree prediction model for the classification of new radar echo data.

[0045] In the present invention, the target samples and sample labels of decision tree classification are obtained by matching with the AIS target through the method of peak detection + range dimension constant false alarm detection + Doppler dimension constant false alarm detection. In feature extraction, the advantages of methods such as cell averaging, ordered statistics, deleted mean, subtracted mean, taking the maximum value, taking the minimum value, and taking the first-order difference minimum value are integrated for clutter power level estimation, thereby overcoming the disadvantage of the fixed parameter setting of a single sliding window type CFAR method and improving the high-frequency ground wave radar target detection performance.

[0046] Compared with the prior art, the technical advantages of the present invention are:

[0047] 1. The parameter settings are flexible, which can be conveniently embedded into other supervised machine learning methods for research.

[0048] 2. Through feature extraction, it can integrate the advantages of various clutter power level estimation methods and has high robustness.

[0049] 3. By changing the number of training samples, the number of prediction targets can be adjusted to achieve a function similar to false alarm adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 It is the AIS target mapping diagram in the embodiment of the present invention.

[0052] Figure 2 It is the range-Doppler power spectrum of the AIS target and the CFAR target in the embodiment of the present invention.

[0053] Figure 3 It is the positive and negative sample acquisition area in the embodiment of the present invention.

[0054] Figure 4 It is the feature distribution diagram in the embodiment of the present invention; among them, (a) range dimension and Doppler dimension reference windows; (b) two-dimensional reference window.

[0055] Figure 5 It is the eigenvalue distribution under different numbers of training samples in the embodiment of the present invention.

[0056] Figure 6 It is the high-frequency ground wave radar target detection decision tree in the embodiment of the present invention.

[0057] Figure 7 It is the detection result diagram of the DTC method and the OS-CFAR method (the DTC method is the method of the present invention, and the OS-CFAR method is the comparison method).

[0058] Figure 8 It is the matching rate comparison curve in the embodiment of the present invention, (the DTC method is the method of the present invention, and other methods are the comparison methods);

[0059] Figure 9 It is the radar target detection process in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0061] Embodiment 1

[0062] This embodiment provides a high-frequency ground wave radar target detection method based on decision tree classification, and its specific steps are as follows:

[0063] S1: During a radar echo data time period, align the start time and end time of AIS vessel information with the radar data time. Then, use the linear fitting method to obtain the longitude, latitude, heading, and speed change curves of the AIS vessel during this time period. Then, take the values at the middle moment to represent the longitude, latitude, heading, and speed of the vessel for coordinate conversion ( Figure 1 ). When performing linear fitting, the threshold of the number of AIS information of the vessel is greater than or equal to 3, and those with less than 3 are deleted;

[0064] S2: Match the radar target signals detected by peak detection + range dimension constant false alarm detection + Doppler dimension constant false alarm detection with the AIS targets. The successfully matched targets are used as positive samples, and the negative samples are randomly selected in the non-AIS target area ( Figure 3 ), and the ratio of positive and negative samples is 1:1; among them, when matching, the distance error does not exceed the next distance cell, that is, plus or minus 2.5 km, and the speed error does not exceed three Doppler resolution units, that is, plus or minus 0.25 m / s ( Figure 2 ); specifically, the steps to obtain the radar target signals are as follows:

[0065] S2.1: First, perform peak detection on a set of radar data;

[0066] S2.2: Detect radar target signals in the range dimension according to the set false alarm rate;

[0067] S2.3: Detect radar target signals in the Doppler dimension according to the set false alarm rate;

[0068] S2.4: The intersection of the three detected target sets is the final detected radar target signal set;

[0069] S3: When calculating the signal-to-noise ratio feature in the range dimension, take 5 reference cells outside 1 guard cell above and below the target cell; when calculating the signal-to-noise ratio feature in the Doppler dimension, take 6 reference cells outside 1 guard cell to the left and right of the target cell; when calculating the signal-to-noise ratio feature in two dimensions, select a 5×5 data centered on the target unit, with 8 cells in the middle layer as guard cells and 16 cells in the outer layer as reference cells ( Figure 4 );

[0070] S4 extracts the target signal SNR eigenvalue in the range dimension, Doppler dimension, and two-dimensional rectangular window by means of cell averaging, ordered statistics, mean deletion, clutter averaging reduction, maximum value, minimum value, and minimum of the first-order difference ( Figure 5 ), where Figure 5 (a) shows the distribution of feature 7 and feature 9 when the number of training samples is equal to 3500. The eigenvalue of positive and negative samples hardly overlaps, and they can be basically separated, with good feature separation; Figure 5 (b) shows the distribution of feature 7 and feature 9 when the number of training samples is equal to 7000. Most of the eigenvalue of negative samples overlaps with that of positive samples, with poor feature separation. Therefore, by adjusting the number of training samples, the false alarm rate of the prediction target can be adjusted. By traversing the three receiving channels of the monopole crossed-loop antenna, 63 SNR eigenvalues of the target signal can be cumulatively extracted. Before training, the sample data of different feature dimensions are normalized to between 0 and 1;

[0071] Among them,

[0072] (1) The cell averaging method selects the average power of all reference cells as the clutter power level estimate and calculates the target SNR:

[0073]

[0074] (2) The ordered statistics method sorts all reference cells according to the power magnitude, calculates the target SNR, and selects the 9th, 7th, and 11th in the Doppler dimension, range dimension, and two dimensions respectively as the clutter power level estimate:

[0075]

[0076] (3) The mean deletion method sorts all reference cells according to the power magnitude, then deletes r larger reference values starting from the maximum value, where r is 2, and takes the average value of the remaining reference cells as the clutter power level estimate to calculate the target SNR:

[0077]

[0078] (4) The clutter averaging reduction method sorts all reference cells according to the power magnitude, cuts off r1 smaller reference cells starting from the minimum value and r2 larger reference cells starting from the maximum value, where r1 and r2 are 1, and takes the average value of the remaining reference cells as the clutter power level estimate to calculate the target SNR:

[0079]

[0080] (5) The maximum value method sorts all reference cells according to the power magnitude, takes the maximum value as the clutter power level estimate, and calculates the target SNR:

[0081]

[0082] (6) Minimum value method: Sort all reference units according to power magnitude, take the minimum value as the clutter power level estimate, and calculate the target signal-to-noise ratio:

[0083]

[0084] (7) First-order difference method: Sort all reference units according to power magnitude to obtain the first-order difference values, and then take the reference unit corresponding to the position of the smallest first-order difference value as the clutter power level estimate, and calculate the target signal-to-noise ratio:

[0085]

[0086] Traverse the three channels of the monopole / crossed loop to obtain the eigenvector containing 63 eigenvalues.

[0087] S5: When generating the decision tree, use the information gain method to divide the sample attributes of the decision tree nodes, and generate a decision tree with strong generalization ability through recursion and attribute partitioning ( Figure 6 ), where the maximum number of splits of the decision tree is set to 100;

[0088] S6: Use the generated decision tree model to classify the new input data. If the predicted label is "+1", it is a "true" target; if the predicted label is "-1", it is a false target. Figure 7 It is the result of predicting the single-site radar data using the generated model. The DTC method can not only detect most of the OS-CFAR targets, but also detect other suspected targets additionally, and there are more matching pairs with AIS targets, such as Figure 7 shown by the black dashed box in. Changing the number of training samples can adjust the number of predicted "true" targets ( Figure 8 ), to achieve a function similar to false alarm adjustment, where the number of detected targets is 9000 - 60000 and the corresponding number of training targets is 3000 - 6000. The prediction results show that the overall detection performance of the DTC method is stable, and the matching rate is 1.1% - 5.9% higher than that of the other 4 CFAR methods, proving the superiority of the detection performance of the DTC method.

[0089] Figure 9 It is the principle block diagram of the high-frequency ground wave radar target detection method based on decision tree classification of the present invention, including two parts: model training and prediction. The prediction part includes sample set construction, data preprocessing, feature extraction, model design, and model training. The prediction part uses the trained model to classify the new radar data, where "H1" represents a "true" target and "H0" represents a "false" target.

[0090] The specific examples of implementation described in the present invention are only illustrative of the methods and steps of the present invention. Those skilled in the technical field of the present invention can make corresponding modifications, supplements or variations to the described specific implementation steps (i.e., adopt similar alternative methods), but will not deviate from the principles and essence of the present invention or exceed the scope defined by the appended claims. The scope of the present invention is only defined by the appended claims.

Claims

1. A high-frequency ground wave radar target detection method based on decision tree classification, characterized in that It includes the following steps: S1: Calculate the distance between the vessel and the radar site and the azimuth angle relative to the radar site based on the longitude and latitude information of the AIS vessel; then calculate the radial velocity of the vessel relative to the radar using the course, speed, and azimuth angle information of the AIS vessel relative to the radar. After that, convert the distance and radial velocity to the specific range-Doppler cells of the radar according to the range and Doppler resolutions of the radar. S2: Use a three-level detection method of peak detection + constant false alarm rate detection in the range dimension + constant false alarm rate detection in the Doppler dimension to obtain the radar target signal, and make the radar-detected target signal match the AIS target. The successfully matched targets are used as the positive sample set for decision tree training, with the label "+1", and the negative samples are randomly selected in the non-AIS target area, with the label of the negative samples being "-1". The ratio of positive to negative samples is 1:

1. S3: Select the reference cell clutter in the range dimension, Doppler dimension, and two-dimensional dimension. S4: Use methods such as cell averaging, ordered statistics, mean deletion, clutter averaging reduction, maximum value, minimum value, and minimum of the first-order difference to extract the signal-to-noise ratio eigenvalue of the target signal in the range dimension, Doppler dimension, and two-dimensional rectangular window. Before training, normalize the sample data of different feature dimensions to between 0 and 1. S5: When generating the decision tree, divide the sample attributes of the decision tree nodes to generate the decision tree. S6: Use the generated decision tree model to classify new input data. If the predicted label is "+1", it is a "true" target; if the predicted label is "-1", it is a false target. In S4, the cell averaging method selects the average power of all reference cells as the clutter power level estimate. The ordered statistics method sorts all reference cells according to the power magnitude, and respectively selects the 9th, 7th, and 11th in the Doppler dimension, range dimension, and two-dimensional as the clutter power level estimate. The mean deletion method sorts all reference cells according to the power magnitude, then deletes r larger reference values starting from the maximum value, and takes the average value of the remaining reference cells as the clutter power level estimate. The clutter averaging reduction method sorts all reference cells according to the power magnitude, cuts off r1 smaller reference cells starting from the minimum value and r2 larger reference cells starting from the maximum value, and takes the average value of the remaining reference cells as the clutter power level estimate. The maximum value method sorts all reference cells according to the power magnitude and takes the maximum value as the clutter power level estimate. The minimum value method sorts all reference cells according to the power magnitude and takes the minimum value as the clutter power level estimate. The first-order difference method sorts all reference cells according to the power magnitude to obtain the first-order difference value, and then takes the reference cell corresponding to the position of the minimum first-order difference value as the clutter power level estimate.

2. The high-frequency ground wave radar target detection method based on decision tree classification according to claim 1, characterized in that, In S2, S2.1: First, perform peak detection on a set of radar data. S2.2: Detect the radar target signal in the range dimension according to the set false alarm rate. S2.3: Detect the radar target signal in the Doppler dimension according to the set false alarm rate. S2.4: The intersection of the three detected target sets is the final detected radar target signal set. When the radar-detected target signal matches the AIS target, the range error does not exceed one range resolution cell, and the velocity error does not exceed three Doppler resolution cells.

3. The high-frequency ground wave radar target detection method based on decision tree classification according to claim 1, characterized in that, In S4, the calculation method of the eigenvalue is as follows: (1) The cell averaging method selects the average power of all reference cells as the clutter power level estimate and calculates the target signal-to-noise ratio: (2) The ordered statistics method sorts all reference cells according to power magnitude and calculates the target signal-to-noise ratio. Among them, the 9th, 7th, and 11th are respectively selected as the clutter power level estimates in the Doppler dimension, range dimension, and two-dimensional: (3) The deleted mean method sorts all reference cells according to power magnitude, then deletes r larger reference values starting from the maximum value, and takes the average value of the remaining reference cells as the clutter power level estimate to calculate the target signal-to-noise ratio: (4) The trimmed mean method sorts all reference cells according to power magnitude, cuts off r1 smaller reference cells starting from the minimum value and r2 larger reference cells starting from the maximum value, and takes the average value of the remaining reference cells as the clutter power level estimate to calculate the target signal-to-noise ratio: (5) The maximum value method sorts all reference cells according to power magnitude and takes the maximum value as the clutter power level estimate to calculate the target signal-to-noise ratio: (6) The minimum value method sorts all reference cells according to power magnitude and takes the minimum value as the clutter power level estimate to calculate the target signal-to-noise ratio: (7) The first-order difference method sorts all reference cells according to power magnitude to obtain the first-order difference values, and then takes the reference cell corresponding to the position of the smallest first-order difference value as the clutter power level estimate to calculate the target signal-to-noise ratio: In the above formula, S represents the power of the target signal, and R represents the sample power of the reference unit in the range dimension, Doppler dimension, and two-dimensional rectangular window, where R i , i = 1, …, 2n, diff(R) represents the first-order difference of the reference samples, and index() represents the index of the minimum value.

4. The high-frequency ground wave radar target detection method based on decision tree classification according to claim 1, characterized in that, In S5, the information gain method is used to perform attribute partitioning on the decision tree nodes, and a decision tree is generated through recursion and attribute partitioning: Suppose the proportion of samples of the $k$-th class in the current training sample set $D$ is $p_k$ k (where $k = 1, 2, \ldots, M$), then the information entropy of $D$ is defined as: The smaller the value of Ent(D), the higher the purity of D; according to the different numbers of samples contained in different branch nodes, a weight |D v | / |D| is assigned to the branch node, and the "information gain" obtained by dividing the sample set D by the feature attribute a can be calculated. The larger the information gain, the greater the "purity improvement" obtained by using attribute a for partitioning. Therefore, the information gain is used for attribute selection of the decision tree, that is a * Denote the optimal splitting attribute of the current node, and A is the set of attributes of the training samples.

Images