Hyperspectral image enhancement with vector bilateral filtering

An approach is proposed to extend bilateral filtering to the vector case so as to simultaneously take spectral and spatial information into account by using spectral distances and multivariate Gaussian functions. To simplify the determination of the parameters of the corresponding covariance matrix, the data vectors are transformed to eigenspace through principal component analysis (PCA). By locally adapting to the spectral distribution in decorrelated PCA space, the proposed approach offers effective noise removal while keeping the spatial details in the band images. It also provides dynamic range enhancement of severely affected bands to make meaningful data extraction possible. Experimental results with the proposed approach using remote-sensed hyperspectral data demonstrate improved denoising and enhancement in comparison to existing methods.