Feature extraction using partitioning of feature space for hyperspectral images classification

Hyperspectral images provide valuable sources of information for discriminant of different classes in land covers. Because of limitation of available training samples, feature extraction is an important preprocessing step before classification for avoiding Hughes phenomenon. The huge volume of continues bands in hyperspectral data has high correlation and thus produces redundancy. We propose partitioning of spectral signature of pixels to some disjoint parts using a proper approach so that each part containes bands which are correlated or similar together and are different from bands involved in other parts. Then we obtain the position and shape of each part using calculation mean and variance of that part. We represent some approaches for partitioning of feature space such as uniform based partitioning, correlation based partitioning and k-means clustering based partitioning. We compared these different approaches with the most commonly used unsupervised feature extraction method, principal component analysis (PCA). The experiments were performed using Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) hyperspectral image data and the results show the goodness of proposed method using k-means partitioning approach.