Oil Slicks Detection From Polarimetric Data Using Stochastic Distances Between Complex Wishart Distributions

Polarimetric synthetic aperture radars (PolSAR) have been used to detect oil slicks at the sea surface. Different techniques to extract information from polarimetric data, using an adequate statistical distribution are currently available. A region-based classifier for PolSAR data - named PolClass - uses a supervised approach to compare stochastic distances between scaled complex Wishart distributions and hypothesis tests to associate confidence levels into the classification results. In this paper, the integrated use of these distances together with the uncertainty maps is applied for the first time to detect oil slicks. A quad-pol Radarsat-2 data, acquired during one open-water controlled exercise, was used to perform this test. The PolClass achieved similar overall accuracies for four stochastic distances, reaching 96.54% of global accuracy, the best result obtained by the Hellinger distance. A comparison between the full- and dual-pol matrices indicated that the results obtained with the VV-HH-HV, HH-HV, and VV-HV polarizations are statistically equivalent, but different from that obtained using the HH-VV. Therefore, the exclusion of the HV channel affected the detection of only mineral oils. The classifier demonstrated the potential to detect the three types of oils released, being more effective in detecting biogenic oils rather than mineral oils. The uncertainty levels increase from the center to the border of the mineral oil slicks, indicating the presence of transition regions, possibly related to different weathering mechanisms. The proposed approach will contribute to the understanding of where different physical and chemical processes may be acting, associating confidence levels to the classification results.