Study on the issue of noise estimation in dimension reduction of hyperspectral images

The study on the influence of noise is never discontinuous in hyperspectral image processing. This paper studies this key element in dimension reduction methods based on orthogonal transformation of hyperspectral images. Firstly, distribution features of noise in spectral and spatial dimension are analyzed. Then several traditional dimension reduction methods are discussed. And, noise estimation methods based on spectral and spatial correlation are applied on Maximum Noise Fraction (MNF) transform respectively. From the experimental analysis, it is found that spectral and spatial de-correlation algorithm with image regular partitioning (e.g. rectangle) is more suitable for noise matrix estimation in MNF. Finally, these dimension reduction methods are contrastively used for extracting information from hyperspectral images. From the comparison of results, the optimized MNF considering characteristics of noise can extract more efficient features than others.