Classification of Marine Oil Spills and Look-alikes in Sentinel-1 TOPSAR and Radarsat-2 ScanSAR Images

The main focus of this thesis is to investigate the potential for discrimination between marine oil spills and look-alikes in Synthetic Aperture Radar (SAR) imagery, using log-cumulants and other parameters describing the characteristics of a dark feature. Look-alikes denotes other phenomena known to cause similar SAR signatures as mineral oil. It is vital for the companies managing oil detection services, e.g. Kongsberg Satellite Services (KSAT), to minimize the number of false alarms caused by look-alikes. A selection of parameters are investigated in this thesis, including descriptors currently used by KSAT, and parameters from literature quantitatively describing the same characteristics. Especially the statistical descriptors known as log-cumulants are thoroughly explored and analysed, which are currently not embedded in KSATs oil spill detection service. The potential for using the parameters to discriminate between mineral oil and look-alikes are evaluated using a linear Support Vector Machine (SVM). The analysis is performed on SAR data from KSATs detection service, i.e. large scale, low resolution and single/dual polarization SAR scenes, for which in-situ information is available. The separability obtained between mineral oil and look-alikes using log-cumulants are found to be highly sensor specific. No significant separation are found for dark features acquired by RADARSAT-2 (RS2) in ScanSAR mode. An improved separation are found for the data acquired by Sentinel-1 (S1). Especially the separation between mineral oil and look-alikes assumingly created by atmospheric/oceanographic phenomena proves to be promising. The discrimination obtained in the log-cumulant domain appears to increase with resolution. No clear separation is identified for mineral oil and the class of look-alikes composed of other matter on the surface. The S1 log-cumulant analysis indicates that mineral oil tends to generate dark features with both a greater damping effect and texture than both classes of look-alikes. The SVM is found to discriminate well between mineral oil and look-alikes assumingly created by atmospheric/oceanographic phenomena using S1 data. The first Hu-moment, compactness, coefficient of variation, along with the normalized first order log-cumulant are identified as the most promising parameters. These are objective quantitative parameters, measuring the same characteristics as some of the descriptors currently used in KSATs oil detection service, which are mainly set qualitatively by visual inspection. In general, the quantitative measures performs better and more consistently compared to the descriptors used by KSAT. Resolution is found to be vital to obtain separation between the classes. The work presented in this thesis adds to the research already conducted on the discrimination of mineral oil and look-alikes in SAR imagery. To the authors knowledge, this is the first log-cumulant analysis conducted on wide swath SAR data, i.e. the data type currently being used operationally by the oil detection services. The classification results obtained indicates that the parameters used today can advantageously be quantified, which can potentially contribute towards further automation of the oil detection services