Filtering based Spatial-Spectral Feature Extraction for Hyperspectral Image Classification

The advanced Hyperspectral (HS) Sensors developed recently, makes it is possible to capture rich information per pixel, thereby enabling proper characterization of objects in an image. But due to very large dimensionality, classification of Hyperspectral Image (HSI) becomes very challenging. In order to improve the classifier performance there is a need for accurate feature extraction (FE) techniques. Several FE algorithms like: Morphological and Attribute Filters, Gabor, Local Binary Pattern (LBP), Wavelet Features etc. are available in literature, which are generally operated on reduced HSI. These FE approaches are computationally expensive and more complex. In the proposed framework, a well known and simple bilateral-filter(BF) based spatial-spectral FE is incorporated. Compared to simple Gaussian and other filters the BF can preserve the edges with a maximum margin. Instead of operating on reduced HSI, the BF is applied to each band in the original HSI. So it turns the spectral information to a spatial-spectral relationship. BF improved the classification results (≈ 10%), but after combining it with local covariance matrix (LCMR) features still better results were obtained. These results are validated on the widely used Indian Pines and Salinas-A hyperspectral dataset with the help of K-Nearest Neighbor (KNN) classifier. The performance of the proposed approach are compared with other state-of-art methods and competent results were observed.