A COMPARISON OF DIFFERENT UNMIXING TECHNIQUES AND BAND CLASSIFICATION IN HYPERSPECTRAL IMAGING

Abstract

In this thesis, three unmixing algorithms were employed on a mixture of washing powder HSI data, and to increase the computational efficiency of the unmixing, band classification was utilised. First, a mixture of three endmembers examined by the three algorithms NCLS, AMF and ACE. In case unconstrained linear model subpixel unmixing was used the solved abundance fractions prone to have some results with negative values which might poorly represent the scene. For that, NCLS algorithm was used to solve the nonnegative constrained least square problem. In the ACE, a background covariance matrix based algorithm is used to estimate the abundance fractions which explain its superiority on the unmixing when compared with AMF, as AMF technique posed the hypnosis and compute the image while assigning all the other materials as a background and focuses on one material at a time. The performance of these algorithms was compared to coloured hand-drawn ground truth maps of the powder mixture and a pre-mixed volume. From the analysis, NCLS outperformed all the other algorithms with the highest accuracy and lowest average error. For that, NCLS method was tested on additional sets of the powder. It succeeds to estimates the abundance fractions of a mixture having five endmembers; indeed, the accuracy of the unmixing was better than the mixture of three endmembers. However, when a mixture of seven endmembers was tested, the accuracy of the unmixing dropped by more the 25% owing to different particles sizes. Bands reduction has improved the performance of the unmixing by reducing the computational cost. The MI and Manual selection of the bands recorded the highest accuracy rate in band classification, but in industry-wise projects, manual selection appears not to be a practical solution. In the same manner, the random selection results was fluctuating, and the results was trivial. MI appears to have an issue with the coding that keeps using the same range of clusters which can be improved if there were enough time for the experiment.