Computers & Geosciences Vol. 21, No. IO, pp. 1201-1203, 1995 Copyright 0 1995 Elsevier Science Ltd Printed in Great Britain. All rights reserved OO!S300495)00051-8 009%3004/95 $9.50 + 0.00
SHORT NOTE DETECTION
OF SHIP WAKES IN SAR IMAGES MORPHOLOGICAL OPERATORS
USING
A. GARZELLI Dipartimento
Universita di Firenze, Firenze, Italy (e-mail:
[email protected])
di Ingegneria
Elettronica,
(Received 21 June 1994; revised 9 February 1995)
INTRODUCTION A moving ship produces a set of waves which can be detected in ocean imagery produced by Synthetic Aperture Radar (SAR) sensors operating at L band. Natural ocean phenomena, ships, and ship wakes are represented in high-resolution SAR images of the ocean surface. The detection of ship wakes from SAR images is an important task because it can provide useful information about size, direction, and speed of ships. Moreover, as a moving-ship image is displaced in azimuth from the wake, possibly into a high-clutter area, it could be barely detected. Conversely, some wake components may extend for 5-15 km behind the ship and may be represented clearly in a spaceborne SAR image (Skoevl, Wahl, and Eriksen, 1988). Therefore, wake detection is an essential assignment of satellite ship-surveillance systems. Significant researches on this topic have been carried out at the Norwegian Defence Research Establishment (NDRE) (Skoevl, Wahl, and Eriksen, 1988; Eldhuset, 1988) and at the Defence Research Establishment in Ottawa (DREO) (Rey and others, 1990). In the paper of Eldhuset (1988), ships are detected first and then a wake is searched around each ship, thus reducing considerably the computing time with respect to an extended search approach. However, as outlined previously, a reliable ship-surveillance system should detect ships and wakes independently. In Rey and others (1990) an automatic detection algorithm based on the Radon transform is developed and applied to the Seasat SAR imagery. Pre-processing and post-processing techniques also are employed to reduce the probability of false alarm; however the false alarm rate is unacceptably high for satellite ship-surveillance applications. This paper presents a new algorithm using morphological operators (Serra and Vincent, 1992) to detect automatically ship wakes from SAR images. The algorithm has been mapped onto a parallel architecture of INMOS T800 transputers to achieve real-time
performance. The proposed scheme of wake detection belongs to a complete system for ship surveillance which detects possible targets by searching in parallel for elongated wakes and ships. As this system performs a cross-validation of wakes against ships, it provides an estimate of the reliability associated with each detected target. Effective ship detectors can be realized simply, because ships are characterized by a high reflectance that produces generally well-defined bright spikes in SAR images. Therefore, only the phase of wake detection will be described in this paper. SHIP-WAKE DETECTION The detection of ship wakes has been structured in seven steps which follow a prefiltering operation, carried out by a speckle reducing filter, such as the Geometric filter (Crimmins, 1985) or an Alternating Sequential Filter (Serra and Vincent, 1992). (I) Binarization
of the prejiltered
image
This operation employs a p-tile threshold that selects the corresponding left tail of the image histogram (about 20%). Therefore, potential ship wakes are searched for in the dark regions of the image as suggested by the decrease in radar cross section associated with the turbulent wake astern of a ship (for L-band imaging SAR such as Seasat). Skoevl, Wahl, and Eriksen (1988) noted that the wake feature usually observed in more than 200 ship wake appearances in L-band Seasat SAR images was a dark band behind the ship, termed the turbulent wake (Fig. 1A). The mean thickness of this dark band and the bright band present on the wind-facing side of the wake also were measured (3 pixels and 1 pixel, respectively). For these reasons, the wake searching has been restricted to the dark regions of the image. The dimensions of a wake with respect to the area represented in a 512 x 512 pixel image (about 170 km*) ensure that a quartile threshold (p = 25%) does not eliminate any potential wake (Fig. 1B).
1201
Short Note
Figure 1. A 256 x 256 portion of Seasat-SAR image representing two ships and two wakes: A, original; B, after p-tile filtering (20%); C, D, E, F, after steps 2, 3, 4, and 5, respectively, along direction NE/SW; G, after step 5. along actual direction of longest wake; H, final result, synthetic lines, whose extremes are computed by step 7, are superimposed on original image. (2) Morphological element
closing with a linear structuring
Morphological closing is used to suppress the residual effects of speckle noise on ship wakes. The closing operation using a 4 pixel structuring segment joins those pixels with high gray levels which are no more than a 4 pixel distance along the segment direction (Fig. IC). Starting from step 2, the method of ship-wake detection is subdivided into twelve identical but independent parts, one for each direction of ship-wake searching. This ensures that the algorithm can be implemented easily on a parallel architecture. The resulting accuracy of direction identification is 0.5 x 180/12 = 7.5.‘. (3) Morphological opening (first phase) This step eliminates the bright pixels of the image that are not aligned with at least 30-40 other bright pixels. Such a value represents the linear dimension of the structuring element used by the morphological opening which corresponds to about 1 km at the scale of the real scene (Skoevl, Wahl, and Eriksen, 1988) (Fig. ID). (4) Morphological opening (second phase) A lower limit to the wake thickness is imposed. For each given direction d an opening procedure is performed. It employs a linear structuring element, 3 pixels long and perpendicular to d (Fig. IE).
(6) Morphological subtraction
opening
(third
phase)
and
An upper limit to the wake thickness is imposed. The output image of step 5, namely i5, is opened with respect to a structuring segment, whose length is 20-30 pixels, perpendicular to the direction of interest. The resulting image, i6 then is subtracted from i5, producing an image which is free of linear segments that are too thick. This step is useful particularly in the situation of sea-coast scenes, where different areas of land can appear as dark and connected regions that can be eliminated only by imposing a thickness constraint. (7) Ship-wakes localization Starting from 12 sets of segments, all the ship wakes are localized finally: from a set of intersecting segments, the longest segment is selected (Fig. 1F-H). Information relating to ship bodies and to wakes has been integrated in two phases. First, a coupling algorithm which associates ships to wakes has been implemented. It minimizes the quantity cd,., where d,,, is the distance between ship i and wake j and the sum is computed on all possible ship-wake couples. Note that the number of ships can be different from that of wakes. Second, for each detected ship-wake pair a coupling coefficient is computed, on the basis of the relative position of the two constituent objects independently detected as ship and wake.
(5) Morphological reconstruction This operation (Serra and Vincent, 1992) allows recovering of the shapes of the connected structures after step 3, which are not filtered out by step 4 (Fig. 1F).
EXPERIMENTAL
RESULTS AND CONCLUSIONS
The processing chain has been tested on 42 Seasat SAR 4-look images, representing either the open sea or a region of sea and coast. The elongated and
Short
“narrow V” wakes have been identified and correctly localized with probability P = 97.5%. All the false alarms (PFA = 11%) have been rejected by the phase of association between possible ships and wakes, thus providing a correct interpretation of the scene. The method worked on sea-coast images correctly, by the thickness control on the detected objects. Potential ships and elongated ship wakes are searched for in parallel to obtain a higher detection reliability. The proposed algorithm presents some important advantages with respect to classical wakedetection methods using Hough or Radon transform: (I) edge extraction, critical for noisy images, is not required; (2) accurate localization of wakes is performed without a further investigation on the original image; (3) low-processing time is accomplished by using binary images, morphological operators, and parallel computing.
1203
Note
Acknowledgments-This Alenia S.p.A.
work
has been supported
by
REFERENCES Crimmins, T. R., 1985, Geometric filter for speckle reduction: Applied Optics, v. 24, no. 10, p. 1438-1443. Eldhuset, K., 1988, Automatic ship and ship wake detection in spaceborne SAR images in coastal regions: Proc. of IGARSS ‘88 Symp., Edinburgh (13-16 September 1988). p. 152991533. Rey, M. T., Tunaley, J. K., Folinsbee, J. T., Jahans. P. A., Dixon, J. A. and Vant, M. R., 1990, Application of Radon transform techniques to wake detection in Seasat-A SAR images: IEEE Trans. Geoscience and Remote Sensing, v. 28, no. 4, p. 553-560. Serra, J., and Vincent, L., 1992, An overview of morphological filtering, in Astola, J., and Neuvo, Y., eds., Circuit Systems and Signal Processing, v. 11, no. 1, Birkauser. Boston, p. 47-108. Skoevl, A., Wahl, T., and Eriksen, S., 1988, Simulation of SAR imaging ofship wakes: Proc. of IGARSS ‘88 Symp., Edinburgh (13-16 September, 1988), p. 152.5-1528.