Comparative Analysis of Feature Extraction Algorithms with Different Rules for Hyperspectral Anomaly Detection

Anomaly detection is one of the most important applications for hyperspectral images. Conventional algorithm such as Reed-Xiaoli (RX) detector fails to be applied to hyperspectral images, which have high spectral dimensionality and complicated correlation between spectral bands. Therefore, effective feature extraction methods and selection rules are necessary. In this paper, comparative analyses of feature extraction methods with different rules are conducted, illustrating their effects on application of anomaly detection in hyperspectral images. The algorithms for feature extraction include principal component analysis (PCA), minimum noise fraction (MNF) transform and their kernel versions. The rules for feature selection are energy and signal-to-noise (SNR). Furthermore, a local singularity (LS) measure is introduced to select the most singular component transformed for anomaly detection, based on local high-order statistics, i.e., skewness and kurtosis. Numerical experiments are performed on real hyperspectral images. The results show that using KPCA with LS measure greatly improves detection performance of the conventional RX algorithm and achieves satisfying effect, and the LS measure is more effective than other rules.