HRRP multi‐target recognition in a beam using prior‐independent DBSCAN clustering algorithm
2019; Institution of Engineering and Technology; Volume: 13; Issue: 8 Linguagem: Inglês
10.1049/iet-rsn.2018.5598
ISSN1751-8792
AutoresPengcheng Guo, Zheng Liu, Jingjing Wang,
Tópico(s)Seismic Imaging and Inversion Techniques
ResumoIET Radar, Sonar & NavigationVolume 13, Issue 8 p. 1366-1372 Research ArticleFree Access HRRP multi-target recognition in a beam using prior-independent DBSCAN clustering algorithm Peng-cheng Guo, Peng-cheng Guo National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of China Xi'an Electronic Engineering Research Institute, Xi'an, People's Republic of ChinaSearch for more papers by this authorZheng Liu, Corresponding Author Zheng Liu lz@xidian.edu.cn National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of ChinaSearch for more papers by this authorJing-jing Wang, Jing-jing Wang National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of ChinaSearch for more papers by this author Peng-cheng Guo, Peng-cheng Guo National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of China Xi'an Electronic Engineering Research Institute, Xi'an, People's Republic of ChinaSearch for more papers by this authorZheng Liu, Corresponding Author Zheng Liu lz@xidian.edu.cn National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of ChinaSearch for more papers by this authorJing-jing Wang, Jing-jing Wang National Laboratory of Radar Signal Processing, Xidian University, Xi'an, People's Republic of ChinaSearch for more papers by this author First published: 14 June 2019 https://doi.org/10.1049/iet-rsn.2018.5598Citations: 2AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract During the operation of monopulse radar, the high-resolution range profiles (HRRPs) of multiple targets may be overlapped, which will reduce the target recognition performance. In this study, a novel multi-target recognition method is proposed based on prior-independent density-based spatial clustering of applications with noise (PI-DBSCAN) algorithm. In the training phase, various features of training samples are extracted, among which the distribution of strong range cells is utilised to obtain the parameters of PI-DBSCAN algorithm while the others are used to train the support vector machine (SVM) classifier. In the test phase, PI-DBSCAN algorithm is exploited at the radial distance-azimuth plane to segment the test multi-target HRRP. Afterward, the features of the segmented HRRPs are extracted and fed to the SVM classifier to be recognised. The proposed method has no constraint on the target motion and does not need to set any parameters manually, which benefits its application. The experiment results of real measured data show that the proposed method is robust against noise and the increasing target HRRP overlap ratio comparing with traditional methods. 1 Introduction In recent years, radar automatic target recognition (RATR) has been regarded as one of the main areas of radar research in military applications. The high-resolution range profile (HRRP) is the coherent summation of the complex time returns from target scatterers projected on the radar line-of-sight (LOS), which reflects the target structure signature. Compared with SAR and ISAR images, it is easier to get and process. Therefore, RATR based on HRRP has attracted great attention from many researchers [1-3]. In the multi-target situation, such as group targets or formation targets, HRRPs from multiple targets may be overlapped rather than separated from each other because of their close radial distances in a beam. In this case, the overlapped HRRPs collected from different targets may be considered as HRRP from the same target, causing reduced recognition performance. Hence, the study of multi-target recognition algorithm is very important in the real application. The monopulse radar is widely used due to its high measurement accuracy of target angle, strong anti-jamming ability and high data rate [4, 5]. The multi-target HRRP recognition method with monopulse radar has attracted great attention from many researchers and the strategies can be mainly divided into two categories. The first strategy segments the multi-target echo first to obtain separated echo for each target, and then classifies the separated target echo, respectively. The key of this method is the multi-target segmentation algorithm which falls into four classes. The first one is the segmentation algorithm based on the ISAR image. In [6], the multi-target ISAR image is segmented using K-means algorithm and then the segmented ISAR image is inversely transformed back into the time domain. However, this method requires the target rotation relative to the radar, which restricts its utilisation. The second method segments the multi-target echo in the time-frequency domain [7]. Targets with different velocities will lead to the differences in the time-frequency curves, which can be used to segment multi-target echoes. However, this method may fail when the velocities of targets are the same. The next approach exploits the synthesis model of multi-target echo [8-13] to estimate the parameter of an individual target. For example, [8-10] estimate the individual target angle relative to the radar using the instantaneous matching moment method, model simplified likelihood solution, and maximum likelihood estimation, respectively. In [11, 12], the target angles are obtained for two targets based on four-channel monopulse radar and super-resolution algorithm. Using maximum likelihood estimation method and the minimum description length criteria in [13], the parameters, including distance-azimuth angle, and elevation angle can be estimated for as many as five targets simultaneously. However, these methods cannot obtain the echo of the target and the estimated parameters may be insufficient to recognise targets. The last method segments the target echoes in the time domain using the independent component analysis principle [14]. This method is applicable to the situation where the echoes of multiple targets are separated. The strong correlation of overlapped echoes will cause a decrease in its performance. The other approach directly recognise the overlapped echoes without segmenting the multi-target HRRP. As discussed in [15], the scattering centres (SCs) of multiple targets are extracted using the ESPRIT method and the template database containing multi-target SCs are built. This method does not require to segment target echoes, which can avoid the error caused by segmentation results. However, due to the aspect-sensitivity of HRRP, the template database is large and can bring the heavy computational load to the real RATR systems. In this paper, a novel HRRP multi-target recognition method is proposed using prior-independent density-based spatial clustering of applications with noise (PI-DBSCAN) algorithm. In the training phase, the HRRPs separately collected from individual targets are used as the training samples. Based on the extracted feature of HRRPs, the preset parameters of the PI-DBSCAN algorithm can be obtained with the number and distribution of the strong range cells (SRCs). The parameters of support vector machine (SVM) classifier [16, 17], i.e. the support vector and the penalty factor, can be obtained as well. In the test phase, the PI-DBSCAN algorithm is exploited to segment the test HRRP on the radial distance-azimuth plane of SRCs. Then, the features of the segmented HRRPs are extracted for recognition using the SVM classifier. The advantages of the proposed method can be summarised as follows. First, there is no constraint on the states of target rotation or velocities, which widens its application. Second, the range cells which are heavily corrupted by noise or interfered by other targets can be eliminated effectively. Thus, the accurately segmented HRRP can be obtained by the clustering result. Finally, the preset parameters of the clustering algorithm are learned automatically during the training phase, which can eliminate the error caused by manually setting parameters and improve the segmentation performance of HRRP. The remaining part of this paper is organised as follows. Section 2 analyses the separability of the SRCs of the multi-target HRRP in the range-azimuth angle plane. The multi-target recognition method based on PI-DBSCAN clustering algorithm is introduced in Section 3. The detailed experiment results using measured multi-target HRRP data are provided in Section 4. Section 5 presents the conclusion. 2 Analysis of range-azimuth angle distribution of the SRCs In this section, the property of multi-target HRRP is analysed by using the measured data collected from two targets, i.e. car and truck. Fig. 1 shows the HRRPs of the two targets at different relative positions in a beam. We can see that the HRRPs of the two targets are overlapped as the distance of the radar from the car is close to that from the truck. Fig. 1Open in figure viewerPowerPoint Measured HRRPs containing two targets, i.e. car and truck, which are illuminated by the transmitted signal simultaneously. The truck is stationary, whereas the car is moving away from the radar as shown in Fig. 6. Note that the HRRPs of the car vary due to its changing aspects relative to the radar (a) Separate HRRP of the car and truck when the radial distance between them is 3 m, (b) Completely overlapped HRRP of the car and truck, (c) Partial overlapped HRRP of the car and truck, (d) Separate HRRP of the car and truck when the radial distance between them is 1 m The overlapped HRRP can be segmented into the separated HRRPs of each target with a clustering algorithm operating on its range cells. However, the appropriate window length of the clustering algorithm is hard to determine. A large clustering window length can cause a smaller number of segmented targets than the real value, resulting in the missed segmentation of some targets. In contrast, the clustering method with a small window length tends to divide the HRRP into more target segments than the true target number and get the fragmentary HRRPs. In this paper, the range-angle two-dimensional (2D) information of HRRPs is utilised for HRRP segmentation and recognition. Although the distance of multiple targets in a beam may be close, leading to the overlapped HRRPs in time domain, their azimuth or pitch angles relative to the radar may be different, which can be used to segment the HRRPs from multiple targets. In the context of the missile-borne monopulse radar, the differences between the pitch angles of multiple targets are small, which has little contribution to the HRRP segmentation. Therefore, the azimuth angle and radial distance of SRCs are exploited to segment the overlapped HRRP in the proposed method. Fig. 2 shows the overlapped HRRPs of two targets, as well as the corresponding range and azimuth angle of the SRCs. Clearly, it is difficult to distinguish multiple targets based only on HRRP 1D-information, but multiple targets have good separability in the 2D plane of radial distance-azimuth error. It is worth noting that a few SRCs are isolated along the azimuth angle dimension which may be caused by noise, clutter or great error of the measured angle. These outliers will adversely affect the HRRP segmentation and should be eliminated. Fig. 2Open in figure viewerPowerPoint Multi-target HRRPs and the corresponding range and azimuth angle of the SRCs (a) Overlapped HRRP of car and truck, (b) Radial range and azimuth angle of the RCSs shown in a, (c) Separate HRRP of car and truck, (d) Radial range and azimuth angle of the RCSs shown in c 3 Multi-target recognition method based on prior-independent DBSCAN clustering algorithm Section 2 reveals the shortcomings of the clustering algorithm with the undesirable window length and the separability of the SRCs of the multi-target HRRP in the range-azimuth angle plane. Based on this, an HRRP multi-target recognition method using PI-DBSCAN clustering algorithm is proposed. In this section, we briefly introduce the procedure of the proposed method, followed by the detail of PI-DBSCAN clustering algorithm. 3.1 Main procedure of the proposed method The proposed method consists of two phases, i.e. the training phase and the classification phase, as shown in Fig. 3. In the training phase, the target HRRP template database is composed of the HRRPs collected from individual targets. First, the six target features, i.e. size, moment, distance between the maximum peak and the target edge, the amplitude ratio of the peak with maximum amplitude to that with second largest amplitude, number of SRCs, position of SRCs, are extracted with the training samples. Then, the number and the position of SRCs are exploited to learn the parameters of PI-DBSCAN clustering algorithm. At the same time, the four target features, consisting of size, moment, distance between the maximum peak and the target edge, the amplitude ratio of the peak with maximum amplitude to that with second largest amplitude, are used to train the SVM classifier and get the parameters of SVM, such as the support vectors and the penalty factor. Fig. 3Open in figure viewerPowerPoint Flowchart of the proposed multi-target recognition method In the test phase, the SRCs of the test HRRP are first obtained by setting the amplitude threshold. Second, the range and azimuth distance of all the SRCs are determined by the monopulse. Then, the prior-independent DBSCAN clustering algorithm is exploited to cluster the SRCs from different targets. Next, the multi-target HRRP is segmented according to the clustering result of SRCs. Finally, the features of the segmented HRRPs are extracted and classified by the SVM. 3.2 Prior-independent DBSCAN clustering algorithm As introduced by [18], DBSCAN clustering algorithm is a density-based method aiming at classifying elements into several categories. The motivation for adopting it includes the following factors. First, different from K-means clustering algorithm in [19] which is only suitable for the convex cluster, it can be applied to the cluster with arbitrary shape. Secondly, it can eliminate the isolated noise point and is insensitive to corruption. Finally, the number of clusters can be determined automatically. Despite its advantages, two parameters of the DBSCAN algorithm, i.e. the neighbourhood radius and the minimum number of neighbours included in the -neighbourhood of a point should be preset manually. The inappropriate parameters will decrease the clustering performance. To solve this, this paper proposes a PI-DBSCAN clustering algorithm in which the preset parameters are automatically obtained by learning the training HRRP data. In order to better illustrate the parameter estimation method proposed in this paper, the most important concept in DBSCAN algorithm, namely core object, is described here. For an object p in the object set D, it can be viewed as the core object if the number of samples in its radius is no less than . The core object is the basic element of the cluster. The sample within the radius is calculated by (1). Therefore, as long as is greater than or equal to , p is the core object (1) where is the distance between the points p and q. The smaller the radius , the better the resolution of the DBSCAN algorithm. However, too small radius will cause the truncation of the HRRP of one target to several parts. So the radius should be greater than the maximum distance between two adjacent SRCs of the training samples. Therefore, is calculated by the following equation: (2) where index the SRC of target HRRP and index the training sample in the HRRP template database. M is the total number of training samples and is the number of SRCs of the jth training sample. stands for the index of the ith SRC of the jth training sample, and is the range resolution which can be calculated by , where B denotes the bandwidth and c denotes the light speed. The large will enhance the robustness of the DBSCAN algorithm against noise. However, in order to prevent the target from being judged as noise, the value of should be less than the minimum number of SRCs among all the training samples. Thus is obtained with (3) where is the number of SRCs of the jth training sample. In addition, because the definition of does not include the object itself, we subtract 1 from . In the monopulse radar, radial distance and azimuth angle of the range cell can be obtained directly. However, they are of different scales, which adversely affect the clustering performance. Therefore, the azimuth error should be transformed into the cross distance by (4) where R is the radial distance between the radar and target, and is the azimuth error obtained by the monopulse radar. The prior-independent DBSCAN clustering algorithm consists of two phases which are conducted in the training phase and classification phase, respectively. The details are summarised in Algorithms 1 and 2 (see Fig. 4 and 5). Fig. 4Open in figure viewerPowerPoint Algorithm 1: First phase of prior-independent DBSCAN clustering algorithm Fig. 5Open in figure viewerPowerPoint Algorithm 2: Second phase of prior-independent DBSCAN clustering algorithm 4 Experiment results 4.1 Measured data description The performance of the proposed multi-target recognition method is tested on the real HRRP data collected from two targets, i.e. truck and car, as shown in Fig. 6. The size of the two targets is shown in Table 1. The training HRRP data of the two targets are collected separately within the azimuth angle from to . The two-target HRRPs collected in the scenario presented in Fig. 6 are viewed as test data. The truck is stationary. At first, the car is 10 m in front of the truck, 5 m to the left. Then, the car moves in steps of 0.1 m until it stops 5 m to the right of the truck. The two targets are always within one-fourth beam width around the beam centre. The radar depression angles of the training data and test data are both . The transmitted signal is a chirp pulsed waveform with a bandwidth of 1 GHz. The values of signal-to-noise ratio (SNR) of the HRRP from two targets are approximately equal to 32.06 and 25.76 dB, respectively. In our experiments, the number of training samples of the car and the truck is both 4500, and the test data consists of 520 multi-target HRRPs. Table 1. Information of training data and test data Target type Sample size Length, m Width, m training data truck 4500 6.8 2.7 car 4500 4.8 1.9 test data truck + car 520 — — Fig. 6Open in figure viewerPowerPoint Top view of radar-target geometry All the experiments are performed using the software Matlab on a PC with an 8 GB RAM and a 3.6 GHz Intel CPU. The rest of this section includes two parts. In the first part, the performance of PI-DBSCAN clustering algorithm is tested. In the second part, the recognition performance robustness of the proposed method is evaluated using the measured data. 4.2 Performance of PI-DBSCAN clustering algorithm In the following, we provide the SRC clustering result and HRRP segmentation result on the measured data. Fig. 7 gives the two examples of clustering and HRRP segmentation for multi-target HRRP data by using PI-DBSCAN clustering algorithm. The figures indicate that PI-DBSCAN clustering algorithm can effectively cluster the SRCs of HRRP from the same target and eliminate noise point effectively due to the appropriate parameters setting. Moreover, by clustering the SRCs, the multi-target HRRP can be segmented accurately. Fig. 7Open in figure viewerPowerPoint Clustering results and HRRP segmentation results (a) Clustering result of the SRCs from the overlapped HRRP with PI-DBSCAN clustering algorithm, (b) HRRP segmentation result by using the clustering result in a, (c) Clustering result of the SRCs from the separate HRRP with PI-DBSCAN clustering algorithm, (d) HRRP segmentation result by using the clustering result in c 4.3 Measured data description In this section, to quantitatively evaluate the proposed method, several state-of-the-art multi-target recognition algorithms shown in Table 2, are exploited as the comparative methods. The first method segments the multi-target HRRP by using the rule-based method operated on the radial distance of the SRCs in the HRRP. The other three segmentation methods exploit three clustering methods, i.e. K-means, DBSCAN and PI-DBSCAN, which are operated on SRCs in the radial distance-azimuth distance dimension. In the K-means method, we set the clustering number as the real target number of 2 and employ K-means++ algorithm to initialise other parameters [20]. The parameters of DBSCAN are calculated by the adaptive algorithm as introduced in [21]. In all the multi-target recognition methods, the target features are extracted from the segmented HRRPs and recognised by the SVM classifiers. Table 2. Multi-target recognition methods used in this paper clustering dimension segmentation method radial distance rule-based method radial distance-azimuth distance K-means radial distance-azimuth distance DBSCAN radial distance-azimuth distance PI-DBSCAN 4.3.1 Robustness against the overlap ratio The robustness to the overlap ratio of the HRRP from multiple targets is crucial for a multi-target recognition method, as the overlap ratio of the test multi-target HRRP is out of control. Therefore, the proposed method is tested against overlap ratio variance in this part. In this paper, the overlap ratio is defined as (5) where , and and are the lengths of two targets along the radar LOS, respectively. Fig. 8 shows the recognition accuracy versus the overlap ratio of multi-target HRRPs. It is apparent that the proposed method outperforms the other three multi-target recognition methods. When the overlap ratio equals to 0.6, the average recognition rate is still >0.9. The reason for that is two-fold. First, the overlapped SRCs belonging to different targets may be separated in the azimuth dimension, which can be used to segment the multi-target HRRP. Secondly, the parameters of PI-DBSCAN algorithm are determined by the training HRRP data which can reduce the impact of manual settings and eliminate the noise points effectively, so the overlapped HRRP can be segmented appropriately. Fig. 8Open in figure viewerPowerPoint Recognition performance of the proposed method and the comparative methods relative to the overlap ratio of multi-target HRRP The performance of the recognition method using the rule-based segmentation algorithm is much worse than the other methods. This is because that the rule-based method is operated only on the range distance dimension and that the rule is too coarse to segment the multi-target HRRP. The DBSCAN parameters are set according to the point distribution and can be easily disturbed by noise, so the error of parameters can be too large to achieve stable recognition performance. The K-means algorithm cannot eliminate the noise point effectively, which causes that the segmented HRRP contains some redundant range cells, as shown in Fig. 9. In this case, the target feature will be inaccurate, which degrades the recognition performance. Fig. 9Open in figure viewerPowerPoint Clustering result by K-means and the corresponding HRRP segmentation results (a) Clustering result of SRCs by K-means algorithm, (b) HRRP segmentation results by using the clustering result of a When the overlap ratio is larger than 80% the recognition performance of the proposed method declines obviously. This is mainly because the high overlap ratio leads to a large number of interferometric range cells. Since the error of the measured angle of the interferometric range cell may be large, the clustering method may fail easily as in Fig. 10. Even though the cluster of SRCs is effective, the segmented HRRP loses too much information of target because the interferometric range cells are eliminated by PI-DBSCAN algorithm. This case will also lead to bad recognition performance of the proposed method. Fig. 10Open in figure viewerPowerPoint Clustering failure case 4.3.2 Robustness against SNR Owing to the noise from the radar system and background, the test HRRP may be collected under variant levels of SNR. Therefore, it is necessary for the multi-target recognition method to have robust performance against SNR. The SNR of HRRP is defined as (6) where denotes the power of the ith SRC, N denotes the number of SRCs, and denotes the power of noise. The recognition performance with respect to the SNR of test samples is given in Fig. 11. We can see that the overall trends in recognition performance of all these three methods are declining with the decrease of SNR. However, the average recognition rates of the proposed method are the highest of the three methods. When , the average recognition rates of the proposed method are >80%, whereas that of the other two methods are always smaller than 80% when . Therefore, the recognition performance of the proposed method satisfies the requirement of the real application. Fig. 11Open in figure viewerPowerPoint Average recognition rate at different SNR of the test data 5 Conclusion In this paper, a novel multi-target recognition method based on PI-DBSCAN clustering algorithm is presented. The preset parameters of DBSCAN are first learned based on the distribution of SRCs in HRRP. Then the SRCs are clustered in the dimensions of range-azimuth angle by the proposed PI-DBSCAN clustering algorithm. Next, the multi-target HRRP is segmented according to the clustering result of SRCs. Finally, the target features are extracted using the segmented HRRP and classified by the SVM. The robust performance of the proposed method is verified by the experiments conducted on the measured data. The proposed method does not have any restriction on the motion state of multiple targets in a beam, and can get the preset parameters automatically. Therefore, it can be used widely in practice. Note that the performance of all the recognition methods above will decrease dramatically with the increase of overlap ratio, so the multi-target recognition method with robust performance under high overlap ratio will be further studied in the future. 6 References 1Du, L., Wang, P., Liu, H. et al: 'Bayesian spatiotemporal multitask learning for radar HRRP target recognition', IEEE Trans. Signal Process., 2011, 59, (7), pp. 3182– 3196 2Liu, H., Chen, B., Feng, B. et al: 'Radar high-resolution range profiles target recognition based on stable dictionary learning', IET Radar Sonar Navig., 2016, 10, (2), pp. 228– 237 3Guo, Y., Xiao, H., Kan, Y. et al: 'Learning using privileged information for HRRP-based radar target recognition', IET Signal Process., 2018, 12, (2), pp. 188– 197 4Sherman, S. M.: ' Monopulse principles and techniques' ( Artech House, Dedham, MA, 1984) 5Nickel, U.: 'Overview of generalized monopulse estimation', IEEE Trans. Aerosp. Electron. Syst., 2006, 21, (6), pp. 27– 56 6Xiao, D., Su, F., Wu, J.: ' A method of ISAR imaging for multiple targets'. IEEE Int. Conf. on Signal Processing, Beijing, China, 2013, pp. 30– 33 7Li, Y., Fu, Y., Li, X. et al: ' An ISAR imaging method for multiple moving targets based on fractional Fourier transformation'. IEEE Int. Conf. on Radar, Pasadena, CA, USA, 2009, pp. 1– 6 8Blair, W.D., Brandt-Pearce, M.: 'Monopulse DOA estimation of two unresolved Rayleigh targets', IEEE Trans. Aerosp. Electron. Syst., 2001, 37, (2), pp. 452– 469 9Wang, Z., Sinha, A., Willett, P. et al: 'Angle estimation for two unresolved targets with monopulse radar', IEEE Trans. Aerosp. Electron. Syst., 2013, 40, (3), pp. 998– 1019 10Sinha, A., Kirubarajan, T., Bar-Shalom, Y.: 'Maximum likelihood angle extractor for two closely spaced targets', IEEE Trans. Aerosp. Electron. Syst., 2002, 38, (1), pp. 183– 203 11Zheng, Y., Tseng, S. M., Yu, K. B.: 'Closed-form four-channel monopulse two-target resolution', IEEE Trans. Aerosp. Electron. Syst., 2003, 39, (3), pp. 1083– 1089 12Crouse, D. F., Nickel, U., Willett, P.: 'Comments on 'closed-form four-channel monopulse two-target resolution'', IEEE Trans. Aerosp. Electron. Syst., 2012, 48, (1), pp. 913– 916 13Zhang, X., Willett, P. K., Bar-Shalom, Y.: 'Monopulse radar detection and localization of multiple unresolved targets via joint bin processing', IEEE Trans. Signal Process., 2005, 53, (4), pp. 1225– 1236 14Lee, S. H., Lee, S. J., Choi, I. O. et al: 'ICA-based phase-comparison monopulse technique for accurate angle estimation of multiple targets', IET Radar Sonar Navig., 2018, 12, (3), pp. 323– 331 15Jia, H., Li, J., Yang, T. et al: ' Based on improved ESPRIT algorithm radar multi-target recognition'. IEEE Int. Conf. on Dependable, Autonomic and Secure Computing, Chengdu, China, 2013, pp. 416– 421 16Yu, P. S., Yang, T. C., Chen, S. Y. et al: 'Comparison of random forests and support vector machine for real-time radar-derived rainfall forecasting', J. Hydrol., 2017, 552, pp. 92– 104 17Liu, J., Fang, N., Xie, Y. J. et al: 'Multi-scale feature-based fuzzy-support vector machine classification using radar range profiles', IET Radar Sonar Navig., 2016, 10, (2), pp. 370– 378 18Ester, M., Kriegel, H. P., Xu, X.: ' A density-based algorithm for discovering clusters a density-based algorithm for discovering clusters in large spatial databases with noise'. AAAI Int. Conf. on Knowledge Discovery and Data Mining, Portland, OR, USA, 1996, pp. 226– 231 19Kanungo, T., Mount, D.M., Netanyahu, N.S. et al: 'An efficient k-means clustering algorithm: analysis and implementation', IEEE Trans. Pattern Anal., 2002, 24, (7), pp. 881– 892 20Arthur, D., Vassilvitskii, S.: ' K-means++: the advantages of careful seeding'. 18th Symp. on Discrete Algorithms (SODA), Philadelphia, PA, USA, 2007, pp. 1027– 1035 21Daszykowski, M., Walczak, B., Massart, D. L.: 'Looking for natural patterns in data', Chemom. Intell. Lab. Syst., 2001, 56, (2), pp. 83– 92 Citing Literature Volume13, Issue8August 2019Pages 1366-1372 FiguresReferencesRelatedInformation
Referência(s)