A fusion approach to classify hyperspectral oil spill data

Abstract

Oil spills in the ocean are one of the major environmental concerns as they pose significant threat to the ecosystem. In the recent year’s hyperspectral sensors have been used to detect oil spills due to their ability to capture reflected solar signal with very narrow bandwidth, that enables to differentiate materials even with subtle signature differences. However high spectral dimensionality affects classification accuracy due to insufficient training samples. Conventional methods have dealt with this problem through band (or feature) selection or feature extraction (or transformation) approaches. In the case of an oil emulsion signal, as evident from the literature the oil characteristics are present only in certain bands of a spectral signature. Hence feature selection without focusing on right bands may lead to selecting features of less significance for target identification. Moreover, in feature extraction approaches during transformation from high dimensional space to low dimension space some of the important features could be lost. Hence, the proposed research takes advantage of different but complimentary benefits of both feature selection and feature extraction methods to obtain final features effectively. In the proposed research, features are selected by modelling the oil emulsion signal using derivative spectrum and calculation of partial sum of energies. Derivative spectrum represents variation in signal energies and calculation of partial sum of signal energies for each band facilitates the application of filters bank to capture oil sensitive signal characteristics. Feature extraction is done through wavelet transform by gradual multi-scale zooming of signals through partial analysis of time (space) frequencies. The obtained features are fused together and fed to Gaussian Mixture Model (GMM) classifier to classify oil emulsions. The proposed approach gives 5% to–10% higher classification accuracy as compared to some of the conventional techniques.