Hyperspectral classification optimization using multiobjective evolutionary algorithm

The development of hyperspectral imagers provides an avenue to exploit it as a bio-optical taxonomic identification tool. In order to classify the species based on their spectral characteristic, spectral similarity measures are usually applied. Five spectral similarity measures are examined: Spectral Correlation Angle (SCA) for calculating the angle of spectral cross-correlation vectors, Spectral Angle Measure (SAM) for the angle between two spectral vectors, Spectral Information Divergence (SID)for measuring the difference in the probability distribution of two spectra, and hybrid measures: SID-SAM and Normalized Spectral Similarity Measure (nSSM). However, currently there is no specific threshold value for each spectral similarity measure that defines positive classification of one specific species. The conventional method by applying either a fixed or an adaptive threshold value is found to be unreliable. The focus of this thesis is to explore the characteristics of different spectral similarity measures and to utilize the MOEA to find the best value for the threshold value. In addition, the research also put forward the parameter test to find the optimum parameter. A machine learning algorithm, SVM is used to compare the performance of MOEA. Non-Dominated Sorting Genetic Algorithm (NSGA-II) is a variant of MOEA that is recognized as robust variant, and is therefore selected for this research. In the comparative study of 6 type of seagrass and 4 terrestrial plants, the performance of the discriminating threshold is found to be statistically superior to those from adaptive threshold method. The parameters of NSGA-II were quantified and found that crossover, mutation rate, and different initial chromosome seeds to be significant. The regression models were obtained using this information. The best parameters' values were then used to optimize NSGA-II. The optimized NSGA-II performs on-par with the SVM. The obtained results combined with Probability of Spectral Discrimination (PSD) and Power of Spectral Discrimination (PWSD) were used to suggest preferred similarity measures for specific target seagrass.