Oil spill candidate detection from SAR imagery using a thresholding-guided stochastic fully-connected conditional random field model

Abstract

The detection of marine oil spill candidate from synthetic aperture radar (SAR) images is largely hampered by SAR speckle noise and the complex marine environment. In this paper, we develop a thresholding-guided stochastic fully-connected conditional random field (TGSFCRF) model for inferring the binary label from SAR imagery. First, an intensity thresholding approach is used to estimate the initial labels of oil spill candidates and the background. Second, a Gaussian mixture model (GMM) is trained using all the pixels based on the initial labels. Last, based on the GMM model, a graph-cut optimization approach is used for inferring the final labels. By using a threholding-guided approach, TGSFCRF can exploit the statistical characteristics of the two classes for better label inference. Moreover, by using a stochastic clique approach, TGSFCRF efficiently addresses the global-scale spatial correlation effect, and thereby can better resist the influence of SAR speckle noise and background heterogeneity. Experimental results on RADARSAT-1 ScanSAR imagery demonstrate that TGSFCRF can accurately delineate oil spill candidates without committing too much false alarms.