A new scattering similarity based metric for ship detection in polarimetric synthetic aperture radar image

A new scattering similarity based metric for ship detection in polarimetric synthetic aperture radar image

PDF

Haitao Lang¹^,²^,^*, Yunhong Tao¹, Lihui Niu¹, Hongji Shi¹

Acta Oceanologica Sinica | 2020, 39(5) : 145 - 150

Less

Acta Oceanologica Sinica | 2020, 39(5): 145-150

• Marine Technology •

A new scattering similarity based metric for ship detection in polarimetric synthetic aperture radar image

Full

Haitao Lang¹^,²^,^*, Yunhong Tao¹, Lihui Niu¹, Hongji Shi¹

Affiliations

¹ College of Mathematics and Physics, Beijing University of Chemical Technology, Beijing 100029, China

² Beijing Key Laboratory of Environmentally Harmful Chemicals Analysis, Beijing University of Chemical Technology, Beijing 100029, China

Published: 2020-05-25 doi: 10.1007/s13131-020-1563-7

Outline

Abstract

Less

A new paradigm for ship detection in polarimetric synthetic aperture radar (Pol-SAR) image is presented. We firstly utilize the scattering similarity parameters to investigate the differences of scattering mechanism between ships and sea clutter. Based on these differences, we propose a novel ship detection metric, denoted as the scattering similarity based metric (SSM), to conduct ship detection task. The distribution model of SSM metric is investigated and modeled by kernel density estimation (KDE). Based on the statistical distribution, an adaptive constant false alarm rate (CFAR) detection scheme is implemented. We compare the proposed SSM with two classic polarimetric metrics, i.e., the polarimetric cross-entropy (PCE) and the reflection symmetry metric (RSM). The experimental results conducted on C-band RADARSAT-2 Pol-SAR data demonstrate the feasibility and advantage of the proposed SSM metric both in sea clutter modeling and in ship detection.

Key words

Pol-SAR / scattering similarity / KDE / CFAR / ship detection

Cite this Article

Haitao Lang, Yunhong Tao, Lihui Niu, Hongji Shi. A new scattering similarity based metric for ship detection in polarimetric synthetic aperture radar image[J]. Acta Oceanologica Sinica, 2020 , 39 (5) : 145 -150 . DOI: 10.1007/s13131-020-1563-7

Full Text

Less

1 Introduction

Less

Ship detection plays an important role in the field of marine surveillance applications, such as fisheries management, search and rescue of vessels, and pollution monitoring (Lang et al., 2014). The synthetic aperture radar (SAR), due to the ability to image through cloud cover and independence of illumination, together with a fine spatial resolution, has become the key remote sensing tool to observe metallic targets at sea. With the development of sensor techniques, advanced polarimetric SAR (Pol-SAR) systems, which contain multiple channels, provide a more complete description of target scattering behavior than traditional single-channel SAR systems. Recent years, ship detection using Pol-SAR data has drawn much attention (Wang et al., 2015).

The literature focusing on ship detection techniques based on Pol-SAR data is extensive. A recent survey of ship detection techniques can be found in (Ouchi, 2016). Fundamentally, polarimetric analysis techniques can be divided into two categories: one exploits different polarimetric channels as independent measurements for the same scene (Wang et al., 2015); the other constructs a specific measurement metric that can jointly exploit information from polarimetric data and then be linked to a decision criterion (Liu et al., 2005). The latter is rather popular. The key issue is the construction of a reasonable and feasible metric based on the scattering differences of the physical characteristics between sea clutter and ships.

The polarimetric scattering similarity parameter proposed by Yang et al. (2001) is an effective and convenient method to understand the scatterer. By checking the similarity between a given scatterer and some other canonical scatterers, a direct characterization of the scatterer can be obtained. However, Yang’s parameter can only measure the similarity between two single scatterers. To overcome this shortcoming, a random similarity parameter is proposed by Li et al., which can measure the similarity of any two scatterers (Li and Zhang, 2015). In this paper, we utilize Li’s parameter to analyze the scattering mechanism of ships and sea clutter. The experimental results reveal the differences between them. Based on these differences, a novel metric denoted as the scattering similarity based metric (SSM) is proposed for ship detection. Then the probability density function (PDF) of SSM is modeled by the method of the kernel density estimation (KDE), which yields much smaller bias than some widely used fitting methods. Based on the statistical model, the constant false alarm rate (CFAR) detection is conducted. Experiments performed on the RADARSAT-2 Pol-SAR data demonstrate the effectiveness of the new metric and the usefulness of the model for practical ship detection applications.

2 Data and methods

Less

2.1 Scattering mechanism analysis

As for a given mixed scatterer T and a canonical mixed scatterer T_c, the similarity parameters modified by Li and Zhang (2015) can be defined as

(1)

$r\left({T,{T_c}} \right) = \frac{{{\rm {Trace}}\left({TT_c^\rm H} \right)}}{{{\rm {Trace}}\left(T \right){\rm {Trace}}\left({\left[ {{T_c}} \right]} \right)}},$

where Trace (·) denotes the operation of matrix trace, superscript H denotes the operation of conjugate transpose. According to Eq. (1) and the coherence matrix of canonical scatterer, the similarity parameters corresponding to the different canonical targets can be derived and listed in Table 1.

We utilize two real SAR images as shown in Fig. 1 to analyze the SSM between ships and sea clutter. A Pol-SAR data set acquired by RADARSAT-2 is processed to analyze the scattering mechanism of ships and sea clutter and validate the proposed method in this paper. The detailed information of Data#1 and Data#2 is listed in Table 2.

In order to statistically analyze the scattering mechanism of ships and sea clutter, we first manually sampled the pixel of ships and sea in two SAR images. After the sampling is completed, according to the similarity parameters corresponding to the different canonical targets as shown in Table 1, we calculate the similarity parameters of each pixel corresponding to the odd-bounce scatterers, the even-bounce scatterers, the volume scatterers, and the extension volume scatterers, respectively. Then we calculate the average values of the similarity parameters of the pixels belonging to ships and sea clutter separately. The results are shown in Fig. 2. In the process of sampling, we acquire more than 10⁵ sampling points of ships and more than 10⁶ sampling pixels of sea clutter on each SAR image, to ensure the validity of the statistical results.

According to Fig. 2, as the extension of r_v, the parameter r_v1, r_v2 and r_v3 is competent to model the highly random volume scattering resulted from the even-bounce and odd-bounce structures. They are not competitive but cooperative. The mean of these parameters belonging to ships and sea clutter is calculated respectively. As shown in Fig. 2, for r_o, r_e and r_v, ships are dominated by r_o and r_e, while sea clutter is dominated by r_o, which is much larger than that of ships. Although both ships and sea are not dominated by r_v, r_v for ships are slightly larger than that for sea clutter; for r_v1, r_v2 and r_v3, sea clutter is dominated by r_v2 and r_v3, ships are dominated by them all at the same time, and r_v1 of ships is larger than that of sea clutter obviously. The reason behind the phenomenon above is that the even-bounce scattering of ships may mainly attributed to the dihedral reflectors, such as the deck-hatch, deck-cargo handling gears structure (Xi et al., 2017), while the volume scattering is mainly caused by some complex structures on board; the odd bounce, for ships, may mainly be contributed by the trihedral and the flat structure, the odd bounce for the rather rough sea surface is mainly caused by the Bragg scattering

2.2 Detection metric

As analyzed above, the scattering similarity parameters reveal the differences between ships and sea background, which motivates us to construct a metric for ship detection. The r_e, r_v and r_v1 of ships are larger than those of sea clutter, while the r_o, r_v2 and r_v3 of sea clutter are larger than those of ships, therefore a new ship metric can be constructed as Eq. (2). Obviously, the SSM of ships is much larger than that of sea clutter, which can improve the ship-sea contrast effectively. In Eq. (2), λ₃ is the third eigenvalue of the coherency matrix which can discriminate the system noise (Sun et al., 2012).

(2)

${\rm {SSM}} = \frac{{\left({{r_e} + {r_v}} \right)}}{{{r_o}}} \cdot \frac{{{r_{v1}}}}{{\left({{r_{v2}} + {r_{v3}}} \right)}} {\lambda _3}.$

As we can see, the sum of the polarimetric scattering similarity parameters corresponding to odd-bounce, even-bounce and volume scattering is equal to one. Theoretically, this indicates the combination of r_o, r_e and r_v can achieve a good cover of the scattering mechanism; thus good characterization is guaranteed. In other words, the scattering mechanism of each scatterer can be fully represented by them. Another similar case is the polarimetric scattering similarity parameters r_v1, r_v2 and r_v3. On the other hand, the metric constructed by these two sets of the similarity parameters meets the law of conservation of energy.

2.3 Kernel density estimation

To implement the CFAR technique according to the SSM, a crucial and fundamental issue is to develop a precise and effective distribution to model the SSM. Kernel density estimation (KDE) is a nonparametric procedure for probability density modeling, which has been widely and successfully used in various fields (Bors and Nasios, 2009). In KDE, a kernel function is assigned to each data sample

$\left\{ {{X_i},} \right.\left. {i = 1, 2,\cdots,N} \right\}$

, and the probability density function (PDF) that represents these data is approximated by

(3)

$\psi \left(X \right) = \frac{1}{{Nh}}\mathop \sum \limits_{i = 1}^N \kappa \left({\frac{{X - {X_i}}}{h}} \right),$

where κ is the kernel function, and h is the kernel bandwidth. The kernel function and bandwidth are two fundamental parameters that control the smoothness and modeling ability of the KDE. We follow the existing studies and choose the Gaussian as the kernel function due to its properties of approximation and for having the derivatives of all orders defined over the entire space (Jones et al., 1996). A classical bandwidth estimation method derives from minimizing the error between the underlying distribution and its estimation (Sheather, 2004), and the optimal bandwidth can be obtained

(4)

${h^{\rm {opt}}} = \frac{{1.06}}{{{N^{\frac{1}{5}}}}}\sigma,$

where σ is the standard deviation of the given data

$\left\{ {{X_i},} \right.i\left. { = 1, 2,\cdots,{N}} \right\}$

3 Experiments

Less

3.1 Experiments data

We utilize two real SAR images as shown in Fig. 3 to test the fitting performance and detection performance of the method we proposed. The detailed information of two SAR images is listed in the Table 3. Data#3 is a multi-look PolSAR data with 3-look in azimuth direction and 1-look in range direction while Data#4 is a multi-look PolSAR data with 2-look in azimuth direction and 1-look in range direction. We sampled two sub-regions consisting of 356×386 pixels from the Data#3 and 1 061×1 161 pixels from the Data#4, respectively to conduct our experiments.

3.2 Fitting results

The proposed KDE methodology is applied on the test data. For comparison, some widely used fitting models such as the K distribution, Weibull distribution (WBL), logarithmic normal distribution (LGN), are also used to fit the test data. Furthermore, to provide a quantitative comparison, the commonly used symmetric Kullback-Leibler (KL) distance (Kullback and Leibler, 1951), which measures global similarity between estimated and theoretical PDFs, is employed as the goodness-of-fit (GoF) measure in our experiments. Figure 4 and Table 4 provides the fitting results and the KL values of different distributions for the test data. For Data#3, the KL value of proposed KDE is just 0.002 3, which is almost 85% smaller than the WBL, whose KL value 0.015 3 is second best in all compared methods. For Data#4, the KL value of proposed KDE is just 0.001 8, which is almost 97% smaller than the LGN, whose KL value 0.065 0 is the second best in all compared methods. The visual inspection also clearly shows that the proposed distribution almost perfectly matches the real SSM statistics histograms.

3.3 Ship detection

After the effectiveness of the proposed modeling method for the SSM is validated, the next goal is to investigate its suitability and capability in practical ship detection applications. To complete this task, the experiment is performed on two sub-regions as shown in Fig. 3. Because of the lack of available automatic identification system (AIS) data for the region, the ships are identified by experts in SAR image analysis, as shown in Fig. 5a and Fig. 6a. For Data#3, to present a detailed understanding of the proposed metric, we conduct a series of experiments by varying probability of false alarm (PFA) from 0.000 07 to 0.001 at an interval of 0.000 005. For Data#4, we conduct a series of experiments by varying probability of false alarm (PFA) from 0.000 1 to 0.01 at an interval of 0.000 05. For each experiment, we employ pixel-level figure of merit (FoM) to evaluate the ship detection performance, which can be defined as

(5)

${\rm {FoM}} = \frac{{{N_d}}}{{{N_t} + {N_f}}},$

where N_d denotes the number of correctly detected pixels, N_f is the number of false alarm pixels, and N_t is the ground truth number of ship target pixels.

For Data#3, with the increment of PFA, the FoM of SSM detector increases gradually, and the maximum FoM 0.850 8 is acquired when PFA is 0.000 3, the corresponding detection result is shown in Fig. 5d. where twenty-two targets are detected correctly and there is no false alarm target. The performance of the proposed method is compared with two classic polarimetric approaches, i.e., the polarimetric cross-entropy (PCE) (Chen et al., 2009) and reflection symmetry metric (RSM) (Nunziata et al., 2012). PCE output, as shown in Fig. 5b, witnesses that in this case the filter is not able to work properly. Although the targets have been correctly detected, many false alarms are present in the sea clutter area and part of targets are incomplete, leading to the FoM of PCE only 0.355 0. The output of the RSM is shown in Fig. 5c. Although the targets have been correctly detected and no false positive is present, the FoM of RSM is only 0.762 1, many pixels belonging to ships are missing.

For Data#4, with the increment of PFA, the FoM of SSM detector increases gradually, and the maximum FoM 0.712 8 is acquired when PFA is 0.002 9, the corresponding detection result is shown in Fig. 6d. where eighteen targets are detected correctly and there is no false alarm target. The performance of the proposed method is compared with two classic polarimetric approach, i.e., the PCE and RSM. PCE output, as shown in Fig. 6b, witnesses that in this case the filter is not able to work properly. Only the sidelobes around the ship targets are highlighted and many false alarms are present in the sea clutter area, leading to the FoM of PCE just 0.088 9. This odd result is inherently due to the norm that PCE filter uses to distinguish targets from clutter, i.e., the polarimetric entropy. The output of the RSM is shown in Fig. 6c. Although the targets have been correctly detected and no false positive is present, the FoM of RSM is only 0.608 1, many pixels belonging to ships are missing.

4 Conclusions

Less

In this paper, we have investigated the scattering mechanism of ships and sea clutter using the scattering similarity parameters. Based on the differences between them, we propose a new detection metric SSM, and the third eigenvalue of the coherency matrix is utilized to suppress the disturbances. At the same time, we develop a statistical model of the metric with KDE method and based on which a CFAR detection scheme is achieved. The extensive experiments verify the good performance of the developed statistical model and demonstrates the advantage of the CFAR detection based on the proposed SSM metric.

Funding

Less

The National Natural Science Foundation of China under contract No. 61471024; the National Marine Technology Program for Public Welfare under contract No. 201505002.

References

Less

Bors A G, Nasios N. 2009. Kernel bandwidth estimation for nonparametric modeling. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6): 1543–1555, doi: 10.1109/TSMCB.2009.2020688

Chen Jiong, Chen Yilun, Yang Jian. 2009. Ship detection using polarization cross-entropy. IEEE Geoscience and Remote Sensing Letters, 6(4): 723–727, doi: 10.1109/LGRS.2009.2024224

Jones M C, Marron J S, Sheather S J. 1996. A brief survey of bandwidth selection for density estimation. Journal of the American Statistical Association, 91(433): 401–407, doi: 10.1080/01621459.1996.10476701

Kullback S, Leibler R A. 1951. On information and sufficiency. The Annals of Mathematical Statistics, 22(1): 79–86, doi: 10.1214/aoms/1177729694

Lang Haitao, Zhang Jie, Zhang Ting, et al. 2014. Hierarchical ship detection and recognition with high-resolution polarimetric synthetic aperture radar imagery. Journal of Applied Remote Sensing, 8(1): 083623, doi: 10.1117/1.JRS.8.083623

Li Dong, Zhang Yunhua. 2015. Random similarity between two mixed scatterers. IEEE Geoscience and Remote Sensing Letters, 12(12): 2468–2472, doi: 10.1109/LGRS.2015.2484383

Liu C, Vachon P W, Geling G W. 2005. Improved ship detection with airborne polarimetric SAR data. Canadian Journal of Remote Sensing, 31(1): 122–131, doi: 10.5589/m04-056

Nunziata F, Migliaccio M, Brown C E. 2012. Reflection symmetry for polarimetric observation of man-made metallic targets at sea. IEEE Journal of Oceanic Engineering, 37(3): 384–394, doi: 10.1109/JOE.2012.2198931

Ouchi K. 2016. Current status on vessel detection and classification by synthetic aperture radar for maritime security and safety. In: Proceedings of the 38th Symposium on Remote Sensing for Environmental Sciences. Japan: Aichi.

Sheather S J. 2004. Density estimation. Statistical Science, 19(4): 588–597, doi: 10.1214/088342304000000297

Sun Yuan, Zhang Bo, Wang Chao, et al. 2012. Ship detection based on eigenvalue-eigenvector decomposition and OS-CFAR detector. In: Proceedings of 2012 International Conference on Computer Vision in Remote Sensing. Xiamen, China: IEEE, 350–355

Wang Wenguang, Ji Yu, Lin Xiaoxia. 2015. A novel fusion-based ship detection method from pol-SAR images. Sensors, 15(10): 25072–25089, doi: 10.3390/s151025072

Xi Yuyang, Lang Haitao, Tao Yunhong, et al. 2017. Four-component model-based decomposition for ship targets using polSAR data. Remote Sensing, 9(6): 621, doi: 10.3390/rs9060621

Yang Jian, Peng Yingning, Lin Shiming. 2001. Similarity between two scattering matrices. Electronics Letters, 37(3): 193–194, doi: 10.1049/el:20010104

Appendix

Less

Year 2020 volume 39 Issue 5

PDF

Cite this Article

BibTeX

Article Info

doi: 10.1007/s13131-020-1563-7

Receive Date：2019-01-28
Online Date：2026-03-31
Published：2020-05-25

Article Data

Affiliations

History

Received：2019-01-28
Accepted：2019-05-13

Funding

The National Natural Science Foundation of China under contract No. 61471024; the National Marine Technology Program for Public Welfare under contract No. 201505002.

Affiliations

¹ College of Mathematics and Physics, Beijing University of Chemical Technology, Beijing 100029, China

² Beijing Key Laboratory of Environmentally Harmful Chemicals Analysis, Beijing University of Chemical Technology, Beijing 100029, China

Corresponding:

* E-mail: langht@mail.buct.edu.cn

References

Share

https://castjournals.cast.org.cn/joweb/aos/EN/10.1007/s13131-020-1563-7

Share to

Scan QR to access full text

Cite this article

BibTeX

Citations

表12种不同金属材料的力学参数

科 Family	属数 Number of genus	种数 Number of species	占总种数比例 Percentage of total species (%)	属 Genus	种数 Number of species	占总种数比例 Percentage of total species (%)
鹅膏菌科Amanitaceae	2	11	5.26	鹅膏菌属 Amanita	10	4.78
小菇科 Mycenaceae	2	12	5.74	丝盖伞属 Inocybe	5	2.39
多孔菌科 Polyporaceae	8	14	6.70	蜡蘑属 Laccaria	5	2.39
红菇科 Russulaceae	3	23	11.00	小皮伞属 Marasmius	6	2.87
				小菇属 Mycena	11	5.26
				光柄菇属 Pluteus	5	2.39
				红菇属 Russula	17	8.13
				栓菌属 Trametes	5	2.39

关闭全屏

BibTeX
EndNote
RefWorks
TxT

Table 1. Modified scattering similarity parameters

Scattering model	Canonical scatterer (T_c)	Similarity parameter (r)
Note: T_mn represents the element value of the m-th row and the n-th column of the matrix T.
Odd-bounce scatterers	${\boldsymbol T_o} = \left[ {\begin{array}{*{20}{c}}1&0&0\\0&0&0\\0&0&0\end{array}} \right]$	${r_o} = \dfrac{{{T_{11}}}}{{{T_{11}} + {T_{22}} + {T_{33}}}}$
Even-bounce scatterers	${\boldsymbol T_e} = \left[ {\begin{array}{*{20}{c}}0&0&0\\0&1&0\\0&0&0\end{array}} \right]$	${r_e} = \dfrac{{{T_{22}}}}{{{T_{11}} + {T_{22}} + {T_{33}}}}$
Volume scatterers	${\boldsymbol T_v} = \left[ {\begin{array}{*{20}{c}}0&0&0\\0&0&0\\0&0&1\end{array}} \right]$	${r_v} = \dfrac{{{T_{33}}}}{{{T_{11}} + {T_{22}} + {T_{33}}}}$
Extension volume scatterers 1	${T_{v1}} = \dfrac{1}{{15}}\left[ {\begin{array}{*{20}{c}}0&0&0\\0&8&0\\0&0&7\end{array}} \right]$	${r_{v1}} = \dfrac{{8{T_{22}} + 7{T_{33}}}}{{15\left({{T_{11}} + {T_{22}} + {T_{33}}} \right)}}$
Extension volume scatterers 2	${T_{v2}} = \dfrac{1}{{30}}\left[ {\begin{array}{*{20}{c}}{15}&5&0\\5&7&0\\0&0&8\end{array}} \right]$	${r_{v2}} = \dfrac{{15{T_{11}} + 7{T_{22}} + 8{T_{33}} + 10Re\left({{T_{12}}} \right)}}{{30\left({{T_{11}} + {T_{22}} + {T_{33}}} \right)}}$
Extension volume scatterers 3	${T_{v3}} = \dfrac{1}{{30}}\left[ {\begin{array}{*{20}{c}}{15}&{ - 5}&0\\{ - 5}&7&0\\0&0&8\end{array}} \right]$	${r_{v3}} = \dfrac{{15{T_{11}} + 7{T_{22}} + 8{T_{33}} - 10Re\left({{T_{12}}} \right)}}{{30\left({{T_{11}} + {T_{22}} + {T_{33}}} \right)}}$

Table 2. Data used for SSM analysis.

	Sensor	Acquired time	Incidence/(°)	Pass	Pixel spacing/m	Mode
Data#1	Radarsat-2	2010-07-14 05:39:22	48.9	descending	5.2×4.7	Fine Quad
Data#2	Radarsat-2	2014-09-26 06:07:15	40.2	ascending	4.7×4.8	Fine Quad

Table 3. Data used for testing the modeling and detection performance of SSM

	Sensor	Acquired time	Incidence/(°)	Pass	Pixel spacing/m	Mode
Data#3	Radarsat-2	2015-11-21 19:14:14	20.6	ascending	4.8×4.7	Fine Quad
Data#4	Radarsat-2	2010-07-12 09:50:22	33.4	ascending	4.7×4.5	Fine Quad

Table 4. The KL distance between the distributions and image histogram

	LGN	WBL	NKG	K-root	G⁰	Our method
Note: The black bold denotes the best performance.
Data#3	0.060 2	0.015 3	0.018 0	0.039 3	0.083 0	0.002 3
Data#4	0.065 0	0.215 5	0.104 3	0.079 4	0.255 9	0.001 8

Fig. 1. Data used for SSM analysis

Fig. 2. Main scattering mechanism analysis of the Data#1 (a) and Data#2 (b).

Fig. 3. RADARSAT-2 data. The region of interest (black rectangle) is used in the detection experiment.

Fig. 4. Sea clutter modeling with different comparative methods.

Fig. 5. The detection results for Data#3. a. Visual interpretation image, b. the PCE detection result corresponding to FoM 0.355 0, c. the RSM detection result corresponding to FoM 0.762 1, and d. the SSM detection result corresponding to FoM 0.850 8. The red circle indicates the false alarm target, the yellow circle indicates that the target shape is incomplete, and the green circle indicates that the target area is significantly larger.

Fig. 6. The detection results for Data#4. a. Visual interpretation image, b. the PCE detection result corresponding to FoM 0.088 9, c. the RSM detection result corresponding to FoM 0.6081, and d. the SSM detection result corresponding to FoM 0.712 8.

Articles: Latest Articles; Most Read; Collections

Updates: Events; News; Multimedia

About: About Us

Contact

No. 86 Xueyuan South Road, Haidian District, Beijing

100081

010-62199257

qkjq@cast.org.cn

Copyright © 2025 China Association for Science and Technology. All rights reserved. For all open access content, the relevant licensing terms apply.
Sponsored by the Office of the Leading Group for Cybersecurity and Informatization of CAST, and supported by Science and Technology Review Publishing House