Spectral band selection for ensemble classification of hyperspectral images with applications to agriculture and food safety

• Main Author: Samiappan, Sathishkumar
• Other Authors: Lori M. Bruce
• Format: Others
• Language: en
• Published: MSSTATE 2014
• Subjects: Electrical and Computer Engineering
• Online Access: http://sun.library.msstate.edu/ETD-db/theses/available/etd-06262014-114816/

<p>In this dissertation, an ensemble non-uniform spectral feature selection and a kernel density decision fusion framework are proposed for the classification of hyperspectral data using a support vector machine classifier. Hyperspectral data has more number of bands and they are always highly correlated. To utilize the complete potential, a feature selection step is necessary. In an ensemble situation, there are mainly two challenges: (1) Creating diverse set of classifiers in order to achieve a higher classification accuracy when compared to a single classifier. This can either be achieved by having different classifiers or by having different subsets of features for each classifier in the ensemble. (2) Designing a robust decision fusion stage to fully utilize the decision produced by individual classifiers. </p> <p> This dissertation tests the efficacy of the proposed approach to classify hyperspectral data from different applications. Since these datasets have a small number of training samples with larger number of highly correlated features, conventional feature selection approaches such as random feature selection cannot utilize the variability in the correlation level between bands to achieve diverse subsets for classification. In contrast, the approach proposed in this dissertation utilizes the variability in the correlation between bands by dividing the spectrum into groups and selecting bands from each group according to its size. The intelligent decision fusion proposed in this approach uses the probability density of training classes to produce a final class label. The experimental results demonstrate the validity of the proposed framework that results in improvements in the overall, user, and producer accuracies compared to other state-of-the-art techniques. The experiments demonstrate the ability of the proposed approach to produce more diverse feature selection over conventional approaches. </p>