Dark spot detection for characterization of oil spills using PolSAR remote sensing

Remote sensing using Synthetic Aperture Radar(SAR) is one of the most commonly used methods for detecting and characterizing oil spills in seas and oceans. However, distinguishing true oil spills from false look-alike features like biogenic oil is a major challenge. The aim of this study is to use polarimetric decompositions like H/A/alpha decomposition to separate oil spill areas from look-alike features using Maximum Likelihood Classification (MLC) and Markov Random Field (MRF) classification methods. The classification algorithm is applied to a quadpolarized Ground Range Detected (GRD) dataset of Uninhabited Ariel Vehicle Synthetic Aperture Radar (UAVSAR) sensor of an experimental oil spill in Norway. The dataset contains a strip of plant oil and three strips of different types of crude oil-water emulsions with varying concentrations of oil. Polarimetric features are extracted using covariance matrix, coherency matrix and H/A/alpha polarimetric decomposition. An MRF classifier is trained and its parameters are optimized over the image. The kappa value in the case of MLC was 0.295 and 0.266 in case of MRF. Using the MRF classifier on features results does not significantly improve classification of different type of oil spills. The confusion between classes increases as the level of dilution in the oil-water emulsion increases. The procedure works well in separating oil spills from water. It also helps to better characterize oil spills which have higher concentration of oil. The incidence angle effect also contributes to the misclassification of oil spills in the area. This can be rectified by applying incidence angle correction.