연구의 선진화와 생산성 향상에
필요한 최적의 소프트웨어와 컨설팅을
공급하도록 노력하겠습니다.

Ship Detection

  • Ship detection in Synthetic Aperture Radar (SAR) is an important component of Maritime domain Awareness (MDA). This is because SARs, such as RADARSAT-1 and 2, offer a wide area surveillance capability (see for example J.K.E. Tunaley, "Algorithms for Ship Detection and Tracking using Satellite Imagery", IEEE, Proc. IGARSS '04, pp 1804-1807, 2004 and J.K.E. Tunaley and A.J. Higginson, "OceanSuite: A RADARSAT-2 Exploitation Tool", ASAR Conference, Canadian Space Agency, 2007 ). A common approach is based on a Constant False Alarm Rate (CFAR). Ships appear in SAR imagery as bright blobs against the background of sea clutter. A threshold for detection can be established based on the sea clutter statistics; statistical parameters are established by analyzing the image data itself.
    The process by which radar signals are scattered from the ocean surface can be modeled by the Bragg scatter mechanism plus some contribution from discrete scatterers. This suggests that the basic statistics for the complex radar return in single look imagery is normally distributed. This conclusion relies on the presence of many independent scattering centers within a radar resolution cell and is a consequence of the Central Limit Theorem. If the mean scattering level varies from cell to cell according to the family of gamma distributions, the resulting distribution of the signal intensity within a resolution cell is the K-distribution.
    The K-distribution involves modified Bessel functions of the second kind as well as gamma functions and can be calculated using algorithms that can be found in Numerical Recipes, for example. However, the code is not trivial and typically requires many iterations so that the calculation is quite computationally intensive. Of course, some of the problem can be circumvented by the use of look-up tables.

OceanSuite

History
  • In 2002, while employed at DRDC Ottawa, Dr. Tunaley began work on the ship detection software that eventually became OceanSuite. At that time, the goal was to detect ships in RADARSAT-1 Synthetic Aperture Radar (SAR) imagery of ocean scenes with a probability of detection of 95% and a probability of false alarm of about 2x10-9 per pixel. The principal algorithms were developed and by 2004 a working prototype had been created. Polar Epsilon was interested and Tunaley moved to NDHQ. Mr. Higginson was hired as an engineer/computer-scientist to assist in the conversion of the research prototype into viable operational software. They both developed the software: Tunaley concentrated mainly on improving detection performance and Higginson focused mainly on improving the user interface and coded some modules, such as the modules for reading RADARSAT-2 images, the land-mask and the OTH Gold parser (though there was often much overlap).
    As Project Polar Epsilon gained momentum, Tunaley became the technical lead for the Project dealing not only with OceanSuite but also other Project technical issues. In early 2008, when Tunaley left DND, the emphasis was on completing the test modules, polishing the code, and satisfying other requirements demanded by the IM Group. There were also unresolved problems with how to send the data to the Regional joint Operations Centres (RJOCs). Higginson has dealt with the remaining issues including the implementation of the MSSR modes.
    At one point, OceanSuite comprised three programs . One of these was for ship detection, another was for ocean wind extraction and a third was for oil spill detection. The ship detection program also contained a module for integrating the ship detections with concurrent AIS messages. It is understood that, for the present, these additional features have been removed.
    In the following, the OceanSuite software is described as it was in early 2008 or how it might have progressed since. More information can be found in A. Samoluk, DND Project Polar Epsilon, OceanSAR 2006 and P.W. Vachon and R.J. Quinn, Operational Ship Detection in Canada Using RADARSAT: Present and Future, June 2012.
Algorithms
  • The ship detection software is designed to be automatic and to avoid as much operator involvement as possible. The application is typically run on an image by image basis and after each image has been analyzed for the occurrence of ships, each of the detections may be viewed and assessed by the operator prior to sending the result to the RJOCs. The entire software package was originally coded under Microsoft?s Visual C++ Version 6 but was upgraded to Visual Studio 2003 and has undoubtedly been upgraded further to take advantage of additional features. No licences are required by the Government of Canada for operation of the basic system. The code can be compiled in 32 bit and 64 bit executables. The latter is preferred so that large blocks of contiguous memory can be allocated to accommodate large rectangular images with a size exceeding 1.5 GB. Because time latency is important for Polar Epsilon, the processing of a normal square image, which used to take up to about a minute, has been reduced to a small fraction of this. To avoid overloading the detection process by false alarms from land targets, a land mask must be employed and if a world mask is not used, an appropriate mask must be loaded that covers the area under consideration. (A world mask is used in newer applications of this type in which the software has been improved so that large masks can be loaded rapidly.) Because of inaccuracies in the masks and azimuthal shifts in apparent ship position, the masks can be extended out into the ocean by a pre-specified distance, which can be varied by the operator; normally the default distance would be accepted. The land mask information was in an ESRI polygon format but we expect that the latest versions of the application could accept vector maps of other formats. The detection process is based on the assumption that the sea clutter, which forms the background against which bright returns from a ship must be detected, is compound K-distributed. This distribution is characterized by three parameters, the mean, the shape parameter and the number of independent looks. The three parameters are assumed to be non-integer and the calculation involves evaluation of modified Bessel functions of the second kind. The number of looks is determined by the SAR image processor and the number of independent looks is derived from this by assuming an appropriate correlation between looks. The sea clutter parameters are estimated across the ocean scene by dividing it into blocks, estimating statistical parameters and smoothing the results. The default algorithm for extracting the shape parameter is the Maximum Mean Likelihood (MML) method, since this is believed to be more accurate than the Mean Variance (MV) method, though the operator can choose the latter if desired. If a ship is present within a clutter block, the clutter statistics will be distorted and the threshold for detection, which depends on the Probability of False Alarm (PFA), will be too high. The algorithm takes care of this by a process of sampling and comparing. The statistical estimates provide a detection threshold for image pixels that may represent a ship. However, simple approaches tend to suffer from false alarms that arise from artifacts in the image (e.g. range and azimuth ambiguities), features in the sea clutter that tend to occur at high angles of incidence, and failures to estimate the clutter distribution accurately. Most azimuthal ambiguities can be identified and removed by estimating their position from the radar PRF. To overcome the remaining problems, the returns are compared with models based on ship statistics. These include segmentation of the ship return in the clutter and a shape analysis as well as the effects of radar polarization. Decisions are based on fuzzy logic. To avoid repeated detections of small islands, oil drilling platforms, etc., it is possible to suppress these through an ?ignore list?. This can be updated during image processing, since latitude and longitude are available. By clicking on a target, its position can be added to the list. After ships have been detected, the operator can review the detections sequentially and accept or reject them. The detections that have been accepted are added to an OTH Gold message, or possibly a text message in another format, and can be sent over a VPN to the RJOCs. Prior to 2008, email was another option. Comprehensive help files are included to assist the operator. The expected future of DND?s maritime surveillance effort is expected to include space-based AIS.

The K-Distribution

  • In the limit when the modulating or randomizing distribution is very compact or "sharp", the K-distribution goes over to the exponential distribution for signal intensity or equivalently to the Rayleigh distribution for signal amplitude. When the imagery is multi-look, a resolution cell contains power from several independent looks so that the basic intensity is no longer exponentially distributed but becomes gamma distributed. Again this can be randomized by another gamma density to represent variations from cell to cell and this results in the compound K-distribution.
  • In single look images, the sea clutter is K-distributed and, once the Bessel function has been evaluated, both the probability density and the distribution can be found easily; the distribution can be expressed in closed form. On the other hand, for multi-look images only the probability density can be expressed in closed form so that the distribution, which is needed to establish detection thresholds, must be determined by numerical integration. CFARs tend to be very small to avoid false alarms, which are likely to be costly. Since images can contain of the order of 100,000,000 resolution cells, false alarm rates of 10-9/cell are appropriate. (It is worth noting that the threshold must be calculated by integrating the density down from large values rather than up from zero; otherwise rounding errors will dominate the result even with double precision.)
  • The time required for the calculation of detection thresholds may be quite important as imagery must be processed within seconds of receipt of an image file. This is because any contribution to the Recognized Maritime Picture (RMP) should be as fresh as possible; data becomes stale and its utility decreases within minutes.
  • London Research and Development Corporation has developed a new algorithm that provides an alternative to the K-distribution detection threshold. It yields detection thresholds that are close to those from the K-distribution and runs orders of magnitude faster. Look-up tables can be avoided. An evaluation Dynamic Linked Library (DLL) based on Microsoft's Common Language Runtime (CLR) is available. This can be called from C++, C-sharp and Visual Basic. The definition of the function is as follows: static double Threshold(double cfar, double nu, double L) where:
    • cfar = the false alarm rate
    • nu = the shape parameter (same as K-distribution)
    • L = the number of independent looks
    The details of the theory are provided in the paper J.K.E. Tunaley, "K-Distribution Algorithm", LRDC Report, August, 2010
  • The Cramer Rao lower bound can be used to determine the best unbiased estimators for the two significant parameters of the K-distribution, namely the mean and the order parameter. The Central Limit Theorem can be invoked for a large number of samples and the standard deviation of parameters can be determined from the Fisher Information matrix. When this is done for order parameters in the usual range for sea clutter, it is found that the covariance of the two parameters is quite small so that the mean and the order parameter can be established almost independently. Also the difference between the optimal determination of the mean and the usual average is negligible. Therefore there is no point in pursuing two-dimensional approaches. As has been shown by Blacknell, the Methods of Moments in which the ratio of the standard deviation to the mean is used leaves much to be desired. Though the logarithm of the sample data gives an improvement, this also seems to be less than satisfactory. The reason that the logarithm is an improvement is because, for small order parameters, the Fisher information tends to be concentrated near the origin of the probability density. The Method of Moments entirely ignores this region. The uncertainty in the parameters for finite numbers of samples may create excessive bias in ship detection thresholds unless the number of independent samples exceeds 1000. The effect is described in the paper: J.K.E. Tunaley, Ship Detection in SAR Imagery, LRDC Technical Report, December 2010. The effect is likely to be serious especially in spiky clutter.
  • A presentation dealing with ship detection and ship wakes in SAR images was delivered at the TEXAS IV meeting in Washington, September 28th-30th, 2010; see J.K.E. Tunaley, Progress in SAR Ship Detection and Wake Analysis, TEXAS IV, Washington, September 2010. Alternatively workshop presentations can be found at TEXAS IV 2010, Briefs.
  • In the Maritime Surveillance section at the ASAR-2011 conference at the Canadian Space Agency, a presentation was given on the contributions of SAR: J.K.E. Tunaley, SAR Contributions to Ship Detection and Characterization, June 2011. This included material on both surface and internal wave wakes and space-based AIS performance.