Feature selection algorithms for anomaly detection in hyperspectral data

We address the problem of feature selection and anomaly detection in hyperspectral (HS) imaging data. We consider feature selection because it leads to savings in the cost of the sensor system and speed in real-time applications. We propose three new feature selection algorithms: the adaptive branch and bound (ABB) algorithm, the improved forward floating selection (IFFS) algorithm, and the fast ratio feature selection algorithm. We use the new ABB algorithm to select optimal subsets of features. Our ABB algorithm is an improved version of the BB algorithm and is much faster than an exhaustive search and other prior versions of the BB algorithm. However, when the number of original features is very large, optimal solutions are not possible because the required computational load is very excessive. We thus introduce our new IFFS algorithm that provides quasi-optimal or near-optimal solutions. The IFFS algorithm is an improvement on the sequential forward floating selection (SFFS) algorithm and is much faster than optimal techniques such as the BB algorithm. It is shown to outperform other state-of-the-art quasi-optimal feature selection algorithms. We also consider a new fast ratio feature selection algorithm to select sets of ratio features (the ratio of the responses at two different spectral bands) for classification. Our three new algorithms are shown to be of use on HS imaging data for two product inspection problems. The two case studies include the detection of chicken skin tumors and chicken contaminants on poultry carcasses. The detection results on the two applications demonstrate the excellent performance of our new algorithms.