Oil spill segmentation of SAR images via graph cuts

Segmentation of dark patches in SAR images is an important step in any oil spill detection system. Segmentation methods used so far include 'Adaptive Image Thresholding', 'Hysteresis Thresholding', 'Edge Detection' (see [1] and references therein) and entropy methods like the 'Maximum Descriptive Length' technique [2]. This paper extends and generalizes a previously proposed Bayesian semi-supervised segmentation algorithm [3] oriented to oil spill detection using SAR images. In the base algorithm on which we build on, the data term is modeled by a finite mixture of Gamma distributions, with a given predefined number of components, for modeling each one of two classes (oil and water). To estimate the parameters of the class conditional densities, an expectation maximization (EM) algorithm was developed. The prior is an M-level logistic (MLL) Markov Random Field enforcing local continuity in a statistical sense. The methodology proposed in [3] assumes two classes and known smoothness parameter. The present work removes these restrictions. The smoothness parameter controlling the degree of homogeneity imposed on the scene is automatically estimated and the number of used classes is optional. To extend the algorithm to an optional number of classes, the so-called a-expansion algorithm [4] has been implemented. This algorithm is a graph-cut based technique that finds efficiently (polynomial complexity) the local minimum of the energy, (i.e, a labeling) within a known factor of the global minimum. In order to estimate the smoothness parameter of the MLL prior, two different techniques have been tested, namely the Least Squares (LS) Fit and the Coding Method (CD) [5]. Semi-automatic estimation of the class parameters is also implemented. This represents an improvement over the base algorithm [3], where parameter estimation is performed on a supervised way by requesting user defined regions of interest representing the water and the oil. The effectiveness of the proposed approach is illustrated with simulated SAR images and real ERS and ENVISAT images.