A combination method of differential evolution algorithm and neural network for automatic identification oil spill at Vietnam East Sea

Abstract

This paper proposes an automatic detection of oil spills on Synthetic Aperture Radar (SAR) images using differential evolution (DE), Neutral network and Back Propagation algorithm (BP). Here, DE and BP are combined to train a multilayer perceptron (MLP) network for achieving the global extreme with a better convergence speed. The input data of neural networks are the geometrical characteristics of oil spills (e.g. area, perimeter, complexity) and the physical behavior of oil spills (e.g. mean or max backscatter value, standard deviation of the dark formation). The out data is oil spill or look-alike. We experiment ALOS/PALSAR and EnviSAT ASAR on East sea area of Viet Nam. The experimental results show that the combination algorithm converges faster and has significantly better capability of avoiding local optima.