Marine transportation is the most important mode of international conveyance, accounting for about two-thirds of international trade and 90% of China's total import and export freight. Due to the relatively low cost of shipping and growth in demand, the total dead weight tonnage (dwt) of the global merchant ships in 2016 was more than 1.7 billion dwt (Clarkson Research Services, 2016). Such a large number of ships has had a large impact on the marine environment. According to the United States National Research Council 2002 report on marine environmental pollution, 35% of marine pollutants come from ships. Ships carry many types of marine pollutants, including oil, domestic sewage, and garbage, of which oil pollution is the most serious. This includes the oil spill discharge of oil tankers, residue, washing oil and water mixtures, and mechanical oil. Ship oil pollution is a major source of marine pollution, and poses a major threat to the marine environment. The maritime sector has strict rules regarding ship discharge, but it remains challenging to effectively control illegal discharge. To enforce regulations on illegal discharge from ships, law enforcement officers rely on boarding inspections and fixed processes to determine whether illegal discharge has occurred (Li, 2014). With the large number of ships in operation, the workload is large. As a potential alternative, satellite remote sensing, with its capability to monitor large areas, has become an important mean of monitoring illegal sewage and collecting evidence of such a discharge.
Several studies have proposed satellite remote sensing monitoring methods of illegal oil spills. Busler et al. (2015) used a National Oceanic and Atmospheric Administration model to combine the wind field and flow field to predict the occurrence of an oil spill, and then matched the ship position provided by the automatic identification system (AIS) and other means to find the potential oil spill. Lupidi et al. (2017) demonstrated the functionality of a new maritime control system based on COSMO-SkyMed synthetic aperture radar (SAR) image processing, which is designed to quickly detect the illegal discharge of suspected ships. The results showed that the system was effective. These studies aimed at monitoring large areas of oil on the surface of the sea and show that SAR remote sensing technology has important potential value in marine oil spill monitoring. However, illegal oil spill discharge by ships is often intermittent and occurs in small quantities, and large oil patches are not produced. Nevertheless, small spills leave an oil wake in the ship's track, usually presenting as an elongated shape, because the oil film follows the track of the ship. Therefore, it is possible to find a ship that may be illegally polluted by detecting the wake of the ship with oil spills. So, what kinds of ship wakes in SAR imagery are related to oil spills? Ship wakes can be imaged by satellite SAR (Vesecky and Stewart, 1982; Swanson, 1984; Lyden et al., 1988). Lyden et al. (1988) systematically summarized images of a ship wake using Seasat SAR data. Hennings et al. (1999) conducted further systematic analysis of ERS-1/2 SAR and Radarsat SAR ship wake images. Based on systematic analysis of these studies, ship wakes observed by SAR images follow a common classification, and can be divided into four categories: turbulent wakes, Kelvin wakes, narrow-V wakes, and internal wave wakes. We analyzed the formation mechanism of each type of wake and found that the turbulent wake is the only type of wake that may be associated with oil spills.
Some research has been done on the detection of turbulent wakes. Several researchers have proposed automatic detection algorithms based on observations of ship wakes in SAR images. For example, Hendry et al. (1988) conducted a wake test of the Seasat trawler fleet and Eldhuset (1996) reported on the use of SAR images for vessel and wake detection. Based on these studies, many improved algorithms have been developed, such as ship wake detection algorithms based on the recursive modified Hough transform domain (Wang and Chen, 2009) and based on image segmentation and normalized gray-scale Hough transform (Ai et al., 2010). In general, wake detection from SAR imagery mainly uses a linear feature detection method, such as a Radon transform, Hough transform, or the circumferential scanning method. However, if the wake of the ship is distorted or the oil film diffuses and drifts, the above method is limited, especially in cases where the dark wake and the dark area of the sea are mixed.
This paper presents a new method of combining oil spill and ship wake detection as a rapid detection method of small oil spills based on SAR imagery. This method can detect potential long ship wakes maybe from oil spills and provides a basis for determining the occurrence of illegal oil spills. This article describes the accurate identification of the detection area, rapid detection of oil spill traces, and experimental verification of the proposed method.2 LONG SHIP WAKE
What is a long ship wake? Long ship wake is a turbulent wake and turbulent wake is the most common wake type on SAR images. Through statistical analysis of 71 ENVISAT Advanced Synthetic Aperture Radar (ASAR) images, we obtained probability tables (Table 1) for the four types of observed wakes (Table 1). From these results, the ship wake in an ASAR image is dominated by the turbulent wake and Kelvin wake, where the turbulent wake is 84% and the Kelvin wake is 14.6%. Narrow-V wakes and internal wave wakes have little probability of occurrence.
Therefore, turbulent wakes are the most common wake type in SAR images. Turbulent wakes usually appear as a dark narrow line, but sometimes appear as a bright narrow line, or both. Observed turbulent wakes generally include one of three characteristics: one dark narrow line, one bright narrow line, or an oil spill dark line. Figure 1 presents the images of three typical turbulent wakes.
The formation mechanism of a dark turbulent wake includes two types of sea surface flows resulting from the ship's movement: hull vortex caused by the flow on both sides of the route of the current, and backward flow. These two streams suppress sea surface Bragg waves, thereby weakening the backscattering energy in this region, making it appear as a dark wake in SAR images (Lyden et al., 1988).
Table 2 presents a description of typical turbulent wakes. In theory, due to the impact of waves, dark turbulence wakes exist for only a short period of time and are thus relatively shorter in length. However, when the vessel discharges, a large amount of oil film gathers on the wake. Because of its strong viscosity, the oil film inhibits Bragg waves, weakening the backscattered energy and resulting in a dark line. In contrast, oil film effects last longer, and the wakes are usually longer. After analyzing these wakes, the main differences are that long ship wakes for oil spills are thin and long, with clear edges, and markedly lower backscattering than the surrounding sea. The average length of an oil spill long wake is more than 10 km, whereas that of a common turbulent wake is about 5 km.
Therefore, a ship with potential illegal discharge may have such a wake feature. Rapid and efficient long wake detection is a key technique for extracting candidate ships from a large number of SAR images.3 METHOD FOR THE RAPID DETECTION OF THE LONG SHIP WAKE
In order to overcome the deficiencies of traditional algorithms, we propose a rapid long ship wake detection method. The method mainly includes three steps: first, the detection area of the ship is determined, then the detection area is binarized, and finally, the linear fitting parameter is solved. It can then be determined whether there is a long ship wake in the area from the fitting parameter and the wake's length.
The individual steps are described in detail below.3.1 Determination of the detection area
To detect ships, we used an operational ship detection system, which for the sake of brevity, is not described in detail here which has been described in detail elsewhere (Chen et al., 2017). The system can give a ship's center and axial position. From this, we needed to determine how to reduce the long ship wake detection area.
During SAR imaging, Doppler displacement of the moving vessel occurs. The range of the displacement depends on the speed of the ship and the angle between the direction of navigation and the SAR range direction. Doppler displacement, D, is determined as follows (Chen, 2004):
where H is the relative height of the satellite to the earth surface, θ is the angle of incidence, ϕ is the angle between the direction of motion and the range direction, u is the velocity of the ship, and VSAR is the azimuth velocity of the satellite.
The Doppler displacement value reaches a maximum when the vessel is navigating along the range direction at the limit speed (set to 40 knots). For example, the ENVISAT satellite the track height is about 800 km, the speed is 7.45 km/s, the incident angle is about 23°, the maximum speed of ship is 40 knots, and the maximum Doppler displacement reaches 937 m. For the Radarsat-2 satellite, the orbit height is about 798 km, the flight speed is about 7.5 km/s, the average incident angle is 35°, and the maximum Doppler displacement reaches 1 500 m. For the GF-3 satellite, the orbit height is about 770 km, the flight speed is about 7.5 km/s, the average angle of incidence is about 42°, and the maximum Doppler shift reaches 1 425 m. For these three satellites, taking 1 500 m as the maximum value of Doppler displacement is appropriate.
According to Eq.1, the maximum displacement also changes as the ship's sailing direction changes, and Fig. 2 shows the variation curve. The Doppler displacement value reaches the positive maximum when the ship is moving along the range direction, is zero when the ship is moving along the azimuth direction, and reaches the negative maximum when the ship is moving along the opposite range direction. In this way, the Doppler displacement can be increased or decreased depending on the axial direction of the ship, thereby reducing the long ship wake detection area.
SAR is ascending and right-looking (Fig. 3). If the ship is sailing toward the northeast, the D value is determined by the angle between the ship's axial angle and X-axis, and the wake must be in the left portion of the shadow area. If the ship is sailing southwest, the wake must be in the right portion of the shadow area. If the direction cannot be determined, only the left and right sides of the shadow area must be detected. As the ship's axial direction changes, the shadow region changes accordingly. Taking into account the axial error, ϕ is the buffer angle, and its value is set to 30°. In this paper, we discuss the ascending and right-looking case only; however, other cases can be treated similarly.
Figure 4 presents an ENVISAT ASAR image of a dark turbulent short wake (A) and a typical long ship wake (B). Figure 5 shows the detection area determined using the SAR imaging parameters and the ship's axis. Compared with the traditional method, the detection area is greatly reduced.3.2 Detection area binarization
This method uses the threshold T to binarize the detection area. T is determined by the following equation:
where μ is the mean gray value of the sea surface image and σ is the standard deviation of the sea surface gray value.
After binarization, isolated noise points must be removed. This method uses a morphological filtering algorithm to carry out isolated point elimination.
Jiang et al. (2000) used a morphological filter to eliminate speckle noise, which we used to develop this filter for dark isolated points. In this method, let the image be A, and the structural element be S. Sv is the reflection result of S, and the filter performs a corrosion operation:
Figure 7 shows the results of morphological filtering, which preserves the shape of the wake well.3.3 Linear fit
A long generally ship wake is linear, or partially curved. Therefore, by simple linear fitting to determine the correlation coefficient and length, we can determine whether the wake is a long ship wake. Figure 8 shows the results of the linear fit from the data in Fig. 7, with a correlation coefficient of 0.95 (high degree of correlation) and length of more than 10 km. These results indicate that it is a long ship wake.
Other dark turbulent wakes, long ship wakes, and non-wake images were used to test the method. To determine whether an image shows a long ship, we applied two criteria: 1) the correlation coefficient after linear fitting is more than 0.85, and 2) the wake length is greater than 10 km. If these two conditions are met, the dark strip is judged to be a long ship wake.
Table 3 presents the results of several typical examples of ENVISAT ASAR images. Image 2 is a typical long ship wake, but there is a large, dark area of interference at the front of the wake, resulting in a low correlation coefficient of fitting, and incorrect classification as a long ship wake. Image 6 is affected by the overall darkening of the background, and there is also a misjudgment. The other dark stripes in the list are accurately distinguished. Thus, the overall accuracy of the decision reaches 71.4%. It can be seen that the linear fitting method has a good effect and the calculation amount is much less than the traditional algorithm.4 VALIDATION AND RESULT
To test this method, we used data from China's newly launched GF-3 SAR satellite for testing. The GF-3 satellite is China's first C-band multi-polarized high-resolution microwave remote sensing satellite, and was launched successfully from the Taiyuan Satellite Launch Center on August 10, 2016. The GF-3 SAR satellite has a high resolution, a wide swath, and multi-imaging modes. The standard strip (SS) mode is suitable for ship wake detection. Figure 9 shows a GF-3 SAR image in SS mode. The image shows two clear long ship wakes that have been distorted by the impact of ocean currents. The yellow dashed box is the determined detection area. Table 4 shows the linear fitting results for the two wakes, both of them are judged long ship wakes. These results show the adaptability of the proposed method, particularly for long ship wakes that do not have clear linear features.5 CONCLUSION
In this paper, a new method for detecting long ship wakes is proposed. The method uses the SAR imaging mechanism and the axial direction of the ship to accurately determine the wake detection area. The dynamic threshold method is used to binarize the detection region. The binarized image is filtered using a morphological filter, and the dark points are fitted using a linear fitting method. In this method, the linear fitting correlation coefficient must be higher than 0.85 and the wake's length more than 10 km as the criteria to confirm the presence of a long ship wake. However, these two cut-off values are empirically based and were obtained from a statistical analysis, and may not be very precise. Experiments were conducted using ENVISAT ASAR and GF-3 SAR images. The results showed that the proposed method has a good ability to detect oil spill long ship wakes, with a comprehensive accuracy rate greater than 71%. The test results indicate that this method can be used for the detection of long ship wakes, and if subsequent AIS or other information sources can be used to identify the ship's identity, the method will provide a basis for illegal oil spill discharge monitoring. Compared with other methods, the proposed method offers two innovations, the ability to determine accurately the wake detection area, and the proposal of two criteria to identify long ship wakes. This method has several limitations. For example, if satellite imaging is begun just as the ship has begun to discharge oil, the wake would be shorter. Therefore, future research will examine more complex conditions.6 DADA AVAILABILITY STATEMENT
The authors declare that the data supporting the findings of this study are available within the article.7 ACKNOWLEDGMENT
CHEN Peng conceived the whole method, designed and performed the experiments, and wrote the paper; LI Xiunan and ZHENG Gang implemented the method.
Ai J Q, Qi X Y, Yu W D, Liu F. 2010. A New ship wake CFAR detection algorithm in SAR images based on image segmentation and normalized Hough transform. Journal of Electronics & Information Technology, 32(11): 2 668-2 673. DOI:10.3724/SP.J.1146.2009.01527
Busler J, Wehn H, Woodhouse L. 2015. Tracking vessels to illegal pollutant discharges using multisource vessel information. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol(XL-7/W3), p.927-932, https://doi.org/10.5194/isprsarchives-XL-7-W3-927-2015.
Chen P, Yang J S, Wang J. 2017. Marine targets detection using GF-3 SAR data. In: Proceedings of 2017 IEEE International Geoscience and Remote Sensing Symposium. IEEE, Fort Worth. p.1 978-1 980, https://doi.org/10.1109/IGARSS.2017.8127368.
Chen P. 2004. The study of ship detection in spaceborne Synthetic Aperture Radar (SAR) image. Second Institute of Oceanography, SOA, China, Hangzhou, 82p.
Clarkson Research Services. 2016. Shipping review and outlook. Clarkson Research Services. London, p.118-119, 138-139, 150-159.
Eldhuset K. 1996. An automatic ship and ship wake detection system for spaceborne SAR images in coastal regions. IEEE Transactions on Geoscience and Remote Sensing, 34(4): 1 010-1 019. DOI:10.1109/36.508418
Hendry A, Skingley J, Rye A J. 1988. Automated linear feature detection and its application to curve location in synthetic aperture radar imagery. In: Proceedings of International Geoscience and Remote Sensing Symposium, 'Remote Sensing: Moving Toward the 21st Century'. IEEE, Edinburgh. p.1 521-1 524, https://doi.org/10.1109/IGARSS.1988.569509.
Hennings I, Romeiser R, Alpers W, Viola A. 1999. Radar imaging of Kelvin arms of ship wakes. International Journal of Remote Sensing, 20(13): 2 519-2 543. DOI:10.1080/014311699211912
Jiang Q, Aitnouri E, Wang S. 2000. Automatic detection for ship target in SAR imagery using PNN-model. Canadian Journal of Remote Sensing, 26(4): 297-305. DOI:10.1080/07038992.2000.10874780
Li H. 2014. Analysis and control of illegal sewage of ship.Dalian Maritime University, Dalian, China, p.2-12.
Lupidi A, Staglianò D, Martorella M, Berizzi F. 2017. Fast detection of oil spills and ships using SAR images. Remote Sensing, 9(3): 230. DOI:10.3390/rs9030230
Lyden J D, Hammond R R, Lyzenga D R, Shuchman R A. 1988. Synthetic aperture radar imaging of surface ship wakes. Journal of Geophysical Research:Oceans, 93(C10): 12 293-12 303. DOI:10.1029/JC093iC10p12293
Swanson C V. 1984. Radar observability of ship wakes. Report of Cortana Corporation, Falls Church, VA, USA.
Vesecky J F, Stewart R H. 1982. The observation of ocean surface phenomena using imagery from the Seasat synthetic aperture radar:an assessment. Journal of Geophysical Research:Oceans, 87(C5): 3 397-3 430. DOI:10.1029/JC087iC05p03397
Wang L L, Chen H X. 2009. Multi-ship wakes detection method based on recursive modified Hough transform field. Systems Engineering and Electronics, 31(4): 834-837. (in Chinese with English abstract)