Textural segmentation of natural water scenes for surface pollution detection

Abstract

Water-borne pollutants are currently monitored through spot sampling. This gives an occasional, localised and therefore unreliable picture of the level of contamination. This research is part of a larger EU funded project, ‘Blue Water’ the aim of which is to develop a system capable of continuous and automatic monitoring of water-borne pollution, through the use of remotely sensed visible band camera images. Water-borne pollution generates surface slicks which have a different texture to normal turbulent waves. In this thesis we develop a pixel by pixel segmentation algorithm which classifies the image into slick and non-slick textured regions. We test the algorithm on a set of grey scale lake images. The main stages of the algorithm are preprocessing of the images, feature extraction, classification, and finally postprocessing of the segmentation results. The segmentation process is based on a novelty detection approach. We build histogram and multivariate Gaussian density models of slick feature vectors which then represent ‘normality’. Receiver operating characteristic curves are used to set the decision boundaries of normality for these models, using expertly labelled slick and non-slick data. Each unseen pixel is then classified according to this model as either normal or novel, i.e having slick or non-slick like texture. A range of feature extraction techniques have been investigated namely, statistical moments,principal components analysis and finally one and two dimensional fast Fourier transforms.