Ship Detection with RADARSAT-2 Quad-pol SAR Data Using Improved Notch Filter

based on Freeman Decomposition

Jianhong Chen*1, Yongjun Zhao2, Wei Liu3

1
Institute of Zheng Zhou Information Science and Technology ZhengZhou, China. e-mail: [email protected]
2
Institute of Zheng Zhou Information Science and Technology ZhengZhou, China. e-mail: [email protected]
3
Institute of Zheng Zhou Information Science and Technology ZhengZhou, China. e-mail: [email protected]

Abstract
Ploarimetric Synthetic Aperture Radar (POLSAR) provide more information of targets, advantages of which lays
in its capability to acquire useful images with any-weather conditions and at night time. POLSAR image has been used
to reconnaissance Ships and vessels. Specifically, this paper proposes an improved notch detector algorithm of ship
with full polarimetric image.
The methodology adopted the result of freeman decomposition instead of coherency matrix for constructing
vector space of polarimetric targets. Furthermore, a factor related power is added to the detector in order to separate
ships dominated by surface scattering mechanism form the sea. The proposed detector was tested with Radarsat-2 full
polarimetric data and show significant agreement with the available sea truth. Comparing with SPAN, Polarimetric
whitening filter and Polarimetric notch filter method, the proposed method enhance ship-sea contrast ratio and show
higher detection ability.

1. Introduction
The aim of this study described in this paper is ship detection based on Polarimetric Synthetic Aperture
Radar(PolSAR). Ship detection is a key topic for the surveillance of maritime areas largely due to the capability to
acquire valuable images independent of solar illumination and (to some extent) weather conditions. The studies on
POLSAR target detection mainly exploite the polarimetric statistical and scattering information. Some of the excellent
works include the polarimetric whitening filter detector[1], ship detection using polarization cross-entropy[2], and
polarimetric target detector using the Huynen fork[3].
It is worth mentioning polarimetric target detector. The algorithm is based on a physical rather than a statistical
technique. In the new procedure, ships are detected by exploiting the difference between the polarimetric characteristics
of sea clutter and the ships of interest. Actually the name Polarimetric Notch Filter was already introduced in the past by
at least two more authors[4][5][6]. In this regard, the latest results come from Geometrical Perturbation Polarimetric Notch
Filter(GP-PNF)[7] and improved notch filter[8]. The key of the algorithm is building target scattering model and setting
reduction ratio reasonably. Despite reduction ratio was discussed in Literature[7]-[8] in detail, certain results has been
achieved. However above algorithm vectorized straightforward coherent matrix as target scattering model, which is
difficult to accurately express target scattering mechanism. Thus result in that partial weak target appears lost.
We know that composition of various types of scattering mechanisms could be effectively obtained by model-based
polarization decomposition. Aimed at the problem, a new target space scattering vector was constructed with the results
of freeman decomposition instead of coherent matrix in the paper and then a new polarization notch filter was formed.
Experimental results show that the proposed algorithm effectively enhances the filtering effect and improve the ship
detection performance.

2. Freeman Decomposition
The scattering power decomposition method based on physical scattering models was first developed by
Freeman and Durden[8]. The three-component decomposition scheme divides the polarimetric data of the imaging pixel
area into surface scattering, double-bounce scattering, volume scattering. These scattering powers are calculated very
easily and are used to compose full color images with red, green, and blue (RGB) color coding: red for the double
bounce power, green for the volume scattering power, and blue for the surface scattering power, for which each color
brightness is corresponding to the magnitude. They have been successfully applied to POLSAR image analysis since
color-coded images are easier to understand and since each color represents a specific scattering mechanism.
Coherency matrix can be expanded into three submatrices which correspond to surface scattering, double-
bounce scattering, volume scattering, and helix scattering mechanisms:
[T ] = f s [T ] surface + f d [T ] double + f v [T ] vol (1)

1
978-1-4673-5225-3/14/$31.00 ©2014 IEEE
 For the three scattering, the expansion matrix is given in reference[8]. Where fs, fd and fv are the expansion
coefficients to be determined.The corresponding scattering powers (the surface scattering power Ps, the double-bounce
scattering power Pd, the volume scattering power Pv) are directly obtained from the expansion coefficients of these
matrices when applied to the decomposition.
⎧ ps = (1 + β 2 ) f s
⎪
⎪ 2
⎨ pd = (1 + α ) f s (2)
⎪p = f
⎪⎩ v v

Where α and β are unknowns to be determined.

3. Rewised Polarimetric Notch Filter

The ship detector presented in this paper shares the same general methodology of GP-PNF. More details regarding
the mathematical and physical justification of the algorithm can be found in [8].
The first step is to construct a vector containing the second order statistics of the observed target. A feature partial
scattering vector is introduced:
t = [ ps , pd , pv ]
T
(3)
Where t lies in a subspace C3 representing all the physically feasible partial targets. The normalized version of t
can be considered: tˆ = t / t . After a series of mathematical manipulations, the final expression of the RPNF is:
1
γd = > Tn
⎛ ⎞
b 2
tˆ ⋅ tˆ
*T
(4)
1+ 2 ⋅⎜ − 1⎟
a span 2 ⎜ ˆt *T tˆ T
2
⎟
⎝ ⎠
where tˆ T = [ 0,1,1] represents the signature of the target to be detected (and can be any unitary vector in the space of

the physically feasible targets), tˆ is the partial vector extracted from the scene (i.e., observables), Tn is the threshold and

b 2 / a 2 is a detector parameter that can be set using a rationale based on the SCR.

4. Validation with RADARSAT-2 Polsar Data

A. Presentation of the Datasets
To demonstrate and validate the proposed target detection algorithm, a RADARSAT-2 Fine Quad-Pol mode scenes
were acquired off the east-south port in Tokyo Bay at the 28th December 2012. To facilitate the analysis, Google Earth
optical image was given for comparison. Fig 1.a depicts the Google Earth image of the area. The size of this image patch
is 500Pix × 500Pix with a resolution of 8 meters. Fig 1.b illustrates the Pauli RGB image. The size of experimental
image slice in red box is 270Pix × 270Pix, which contains a total of 12 ships. Fig 1.c illustrates the corresponding Pauli
RGB image.

(a) (b) (c)

Fig. 1. (a) Google Earth image (b)whole Pauli RGB image
(c) Pauli RGB image of experimental area

2
 B. Validation Results and Analysis
The Selected comparison methods conclude classical Span detector, PWF detector, PCE entropy detector,
generalized optimal polarization(GOPCE) detector and improved polarization Notch Filter(IPNF) detector.
For characteristics image of Span, PWF, PCE and GOPCE, the results of detection ship were obtained by dual-
parameter CFAR. Here probability of false alarm was set 10-5. The maximum size of ship is 400m*200m and the
minimum size of 15m*10m. Fig 2.a-2.d illustrate corresponding test results. The value of threshold Tn in IPNF and the
proposed method is 0.5. Reduction Ratio b2/a2 is 0.95. Fig 2.e-2.f illustrate corresponding test results.
For the quantitative evaluation of the performance of the above algorithm, correct detection rate(Cdr), false alarm
rate(Far), leakage alarm rate(Lar) and quality factor(Fq) indicators were used. Here quality factoras follows:

The number of correct detection (5)

Fq =
The number of false alarms + The actual number

4
8

1 9
5
23

7 10
6

(a) SPAN (b) PWF (c) PCE

43 3 4

8 6 7
9 8
1 1 9
1 10 7
4 5
5 6 9
2 12 2 3 11

5 6
7 10
11 8
2

(d) GOPCE (e) IPNF (f) the proposed method

Fig. 2. Test results of various methods

When the number of false alarms detected or undetected too much, the quality factor will decrease. Only when all
the real targets correctly were detected and false alarm and missing alarm does not exist, its value was 1. the smaller its
value is the worse detection results.
From the test results, the SPAN detector does not detect any target, the ship is fully submerged in strong noise. The
proposed algorithm detected 11 correct targets and the detection rate of 100%. Only a ship is missed. Table 1 shows the
comparison result of the five detection algorithms apart from SPAN detector.

Table 1 Comparison test results

Total Detection Number of correct detection Number of errors detected Number of missing target
Detector
(Fq) (Cdr) (Far) (Lar)
PWF 12(0.71) 10(83.3%) 2(16.7%) 2(16.7%)
PCE 10(0.50) 8(66.7%) 4(40%) 3(25%)
GOPCE 12(0.71) 10(83.3%) 2(16.7%) 2(16.7%)
IPNF 9(0.75) 9(100%) 0(0%) 3(25%)
proposed method 11(0.92) 11(100%) 0(0%) 1(8%)

As can be seen from Table 1, false alarm rate of PWF, PCE and GOPCE are higher that notch-like algorithms. We
can find these false alarms are mainly caused by the target split from Fig 2.b-2.d.

3
 Fig 2.e depicts the IPNF mask exploiting quad polarimetric data. Fig 2.f depicts the proposed method mask
exploiting quad polarimetric data. In contrast, the proposed method reduces leakage alarm rate of ship targets. Are shown
in Fig 2.f weak small target 8 and target 2 were effectively strengthened and detected. Success is that results of based-
model polarimetric decomposition can be more accurate reflection of ocean scattering mechanisms. The traditional notch
method use directly odd scattering coherent matrix, surface scattering of pauli decomposition as bragg scattering to
characterize marine electromagnetic scattering model. In fact besides surface scattering components, there is a certain
double-bounce scattering component in sea clutter. It is difficult to accurately reflect the ocean surface scattering only
using pauli decomposition results. In contrast, the results of based-model polarimetric decomposition can be more
accurately reflect the actual marine feature.
To further analyze the performance of the detectors, statistics histogram was extracted from feature maps of IPNF
and the proposed method. The results shown in Fig. 3. It can be seen that the proposed method effectively enhance
contrast between ship and sea. In term of quality factor, the value of the proposed method was 0.92, which the largest
fully reflect the its superiority compared to the rest of the algorithm.
2000
4000
1800
3500
1600
3000
1400

1200 2500

1000 2000

800
1500
600
1000
400
500
200

0 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(a) (b)
Fig. 3. Statistics Histogram
(a) IPNF (b) the proposed method

5. Conclusion
A new kind of polarization notch filter detector was developed based-model polarization decomposition.
Detector combine creatively freeman decomposition and notch filter, which effectively distinguish ship and the sea
clutter. The algorithm has been tested using satellite data (RADARSAT-2 acquired over Tokyo Bay) showing the
capability of the proposed algorithm for vessel detection.

6. References
[1] L. M. Novak and M. C. Burl, “Studies of target detection algorithms that use polarimetric radar data,” IEEE Trans. Aerosp.
Electron. Syst., vol. 25, no. 2, pp. 150–165, Mar. 1989.
[2] J. Chen, Y. Chen, and J. Yang, “Ship detection using polarization crossentropy,” IEEE Geosci. Remote Sens. Lett., vol. 6, no. 4,
pp. 723–727, Oct. 2009.
[3] A. Marino, S. R. Cloude, and I. H. Woodhouse, "A Polarimetric Target Detector Using the Huynen Fork," IEEE Transaction on
Geosciences and Remote Sensing, vol. 48, pp. 2357-2366, May 2010.
[4] A. J. Poelman, “Virtual polarization adaptation.Amethod of increasing the detection capability of a radar system through
polarization vector processing,” Proc. IEE, vol. 128, no. 5, pp. 261–270, 1981.
[5] A. J. Poelman and K. J. Hilgers, The Effectiveness of Multi-Notch Logic Product Polarization Filters in Radar for Countering
Rain Clutter. Boston, MA: Kluwer Academic, 1992.
[6] K. Suwa, K. Yamamoto, C. Nonaka, A. Imamura, and T. Kirimoto, “A target detection algorithm using polarimetric notch
filter,” Electronics and Communications in Japan (Part I: Communications), vol. 88, no. 3, pp. 33–43, 2005.
[7] A. Marino, “A Notch Filter for Ship Detection With Polarimetric SAR Data”, IEEE JOURNAL OF SELECTED TOPICS IN
APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 3, JUNE 2013 1219-1232.
[8] Sun Yuan, Wang Chao, Zhang Hong, et, al. “Polarimetric SAR ship detection using improved notch filter”, JOURUAL OF
IMAGE AND GRAPHICS, 2013,18(10):1374-1381.
[9] A. Freeman and S. Durden, “A three-component scattering model for polarimetric SAR data,” IEEE Trans. Geosci. Remote
Sens., vol. 36, no. 3,pp. 963-973, May 1998.