Abstract:
This study concerns with classification techniques in high dimensional space such as
that of Hyperspectral Imaging (HSI) data sets, with objectives of understanding the
strength and weakness of various classifiers and at the same time to study how their
performances can be assessed particularly when there is an absence of ground truth
target map in the data set. The thesis summaries the work that carried out during the
course of this study and it encompasses a brief survey of machine learning and
classification theories, an outline of the HSI instrumentations, data sets that collected in
the study and classification analysis.
It is found that the supervised classifiers such as the Maximum Likelihood (QD) and the
Mahalanobis Distance (FD) classifiers, especially when they are coupled with
techniques like Regularised Discriminant Analysis (RDA) or leave-one-out covariance
estimations (LOOC), have demonstrated excellent performances comparable to that of
the more complicated and computational costly classifiers like the Support Vector
Machine (SVM). This work has also revealed that separability measures such as the
Total Transformed Divergence (TTD) and Total Jeffries-Matusita Distance (TJM) can be
an invaluable method for assessing the goodness of classification in principle. However,
the present methods for the evaluation of the separability measures are insufficient for
achieving this goal and further work in this area is needed. This study has also
confirmed the effectiveness for using RDA and LOOC techniques for a better estimation
of the covariance when the sample size is small, ie when the sample size per class to
band ratio is less than 100.
Through team work this study has contributed partially a number of publications in the
area of hyperspectral imaging and machine visions.