Information theoretic assessment of correlated noise in hyperspectral signal unmixing

Hyperspectral imaging sensors simultaneously acquire data in hundreds of spectral bands, facilitating detailed study of a scanned object. Unmixing the hyperspectral data as well as estimating the intrinsic dimension of hypercube requires an accurate evaluation of the noise structure. Existing methods mostly simplify the evaluation by considering a white Gaussian noise. However, due to the nature of the hyperspectral sensors,the noise is highly correlated in spectral dimension leading to an inaccurate estimation for white noise assumption. In this paper, we firstly evaluate the strength of the correlation in adjacent spectral bands. Evaluation results prove that only adjacent bands exhibit a significant correlation. Based on the results, we have proposed an explanatory model for the noise structure to extract the correlation coefficients and second order statistics of noise in spectral bands. Simulation results show that our proposed Hyperspectral Correlation Extractor (HYCE) method is accurately estimating the noise structure and is robust to the variation of noise statistics. Our method that is specifically proposed for hyperspectral imaging applications shows unmixing results with an accurate estimation of the pure materials (endmembers) and the related mapping.