Unsupervised Nonlinear Feature Extraction Method And Its Effects On Target Detection In High-Dimensional Data
The principal component analysis (PCA) is one of the most effective unsupervised techniques for feature
extraction. To extract higher order properties of data, researchers extended PCA to kernel PCA (KPCA) by means of kernel
machines. In this paper, KPCA is applied as a feature extraction procedure to dimension reduction for target detection as a
preprocessing on hyperspectral images. Then the detection was done with a support vector data description (SVDD) algorithm
which is another type of one-class support vector machines (SVMs). The SVDD constructs a minimum hypersphere enclosing
the target objects as much as possible. For the supervised learning, SVDD has been trained with a training set which has been
chosen from target class. Balanced classification rate (BCR) and F-measure have been used to evaluate the performance of
proposed technique against full-band target detection. The experimental results on hyperspectral data from the HYDICE
sensors show that the KPCA based dimension reduction offers high performance for target detection applications by SVDD.
Index Terms—Feature extraction, hyperspectral imagery, kernel principal component analysis, support vector data
description, target detection.