Reducing noise in hyperspectal data — A nonlinear data series analysis approach

Abstract

Hyperspectral data are subject to a variety of noise sources associated with the physical processes involved during data acquisition, which distort signal statistical properties and limit the applications of hyperspectral data for information extraction. Noise reduction is, therefore, a prerequisite for many hyperspectral data applications based on classification, target identification, and spectral unmixing. Studies have found that hyperspectral data are more complicated than realizations of linear stochastic processes, upon which many hyperspectral noise reduction algorithms are based. The noise in hyperspectral data may be non-Gaussian and signal dependent. Moreover, as demoustrated in our previous work, hyperspectral data exhibit apparent nonlinear characteristics, which suggests that the noise may exist in broad-band in the frequency domain. An algorithm is introduced in this paper with the intention to improve the noise reduction for hyperspectral data. The effectiveness of the algorithm is evaluated using multiple metrics focusing on both noise reduction and spectral shape preservation.