Noise reduction is a crucial step in hyperspectral imagery pre-processing. Based on sensor characteristics, the noise of
hyperspectral imagery represents in both spatial and spectral domain. However, most prevailing denosing techniques
process the imagery in only one specific domain, which have not utilized multi-domain nature of hyperspectral imagery.
In this paper, a new spatial-spectral noise reduction algorithm is proposed, which is based on wavelet analysis and least
squares filtering techniques. First, in the spatial domain, a new stationary wavelet shrinking algorithm with improved
threshold function is utilized to adjust the noise level band-by-band. This new algorithm uses BayesShrink for threshold
estimation, and amends the traditional soft-threshold function by adding shape tuning parameters. Comparing with soft
or hard threshold function, the improved one, which is first-order derivable and has a smooth transitional region between
noise and signal, could save more details of image edge and weaken Pseudo-Gibbs. Then, in the spectral domain, cubic
Savitzky-Golay filter based on least squares method is used to remove spectral noise and artificial noise that may have
been introduced in during the spatial denoising. Appropriately selecting the filter window width according to prior
knowledge, this algorithm has effective performance in smoothing the spectral curve.
The performance of the new algorithm is experimented on a set of Hyperion imageries acquired in 2007. The result
shows that the new spatial-spectral denoising algorithm provides more significant signal-to-noise-ratio improvement than
traditional spatial or spectral method, while saves the local spectral absorption features better.