Reporting on Three Years of Activity on Oil Spill Classification from Optical Satellite Images

The presence of oil spills at sea represents a relevant issue for government institutions since these are causing serious damages to the marine and coastal ecosystem. The exploitation of satellite sensors for oil spill detection allows for large areas monitoring and remote zone control, thus offering many advantages in economical and time saving terms. In particular, the use of multi-spectral satellite data for the development of techniques for oil spill detection and classification could provide a valuable support to SAR-based solutions, in order to meet the need of environmental protection authorities for efficient and cost effective monitoring tools. This work describes the potential of oil spill detection from optical satellite images, as investigated during three years of activity. The work is focused on the classification of possible region of interests (ROIs), which have been extracted from multi-spectral satellite images, in two classes, namely oil spills and look-alikes. A dataset of more than 300 ROIs has been built by analyzing a number of MODIS-TERRA and MODIS-AQUA images, acquired during the years 2008 and 2009, over the entire area of the Mediterranean Sea. In particular, bands B1 and B2 [visible (VIS) at 0.65 μm, and near-infrared (NIR) at 0.85 μm], which are the only available bands from MODIS at the highest spatial resolution (250 m), have been used. In order to characterize the ROIs, a set of geometrical and grey level features from SAR literature have been exploited. These features have been used to feed different classical machine learning classifiers (multi-layer perceptron, radial basis function, neuro-fuzzy classifiers), achieving promising results. Moreover, recent results from an on-line cost oriented classification approach, based on an ensemble of support vector machines, are reported.