Synthetic aperture radar (SAR) has been widely used to detect oil-spill areas through the backscattering intensity difference between oil and background pixels. However, since the signal is similar to that produced by other phenomena, positive identification can be challenging. In this study we developed an algorithm to effectively analyze large-scale oil spill areas in SAR images by focusing on optimizing the input layer to artificial neural network (ANN) through removal the factor of lowering the accuracy. An ANN algorithm was used to generate probability maps of oil spills. Highly accurate pixel-based data processing was conducted through false or un-detection element reduction by normalizing the image or applying a non-local (NL) means filter and median filter to the input neurons for ANN. In addition, the standard deviation of co-polarized phase difference (CPD) was used to reduce false detection from the look-alike with weak damping effect. The algorithm was validated using TerraSAR-X images of an oil spill caused by stranded oil tanker Volganefti-139 in the Kerch Strait in 2007. According to the validation results of the receiver operating characteristic (ROC) curve, the oil spill was detected with an accuracy of about 95.19% and un-detection or false detection by look-alike and speckle noise was greatly reduced.
Keywords: artificial neural network (ANN); co-polarized phase difference (CPD); non-local means filter (NL-means filter); oil spill detection; probability map.