Noise modelling and estimation of hyperspectral data from airborne imaging spectrometers

摘要

The definition of noise models suitable for hyperspectral data is slightly different depending on whether whiskbroomor push-broom are dealt with. Focussing on the latter type (e.g., VIRS-200) the noise is intrinsically non-stationaryin the raw digital counts. After calibration, i.e. removing the variability effects due to different gains and offsetsof detectors, the noise will exhibit stationary statistics, at least spatially. Hence, separable 3D processes correlatedacross track (x), along track (y) and in the wavelength (λ), modelled as auto-regressive with GG statistics havebeen found to be adequate. Estimation of model parameters from the true data is accomplished through robust techniquesrelying on linear regressions calculated on scatter-plots of local statistics. An original procedure was devisedto detect areas within the scatter-plot corresponding to statistically homogeneous pixels. Results on VIRS-200 datashow that the noise is heavy-tailed (tails longer than those of a Gaussian PDF) and somewhat correlated along andacross track by slightly different extents. Spectral correlation has been investigated as well and found to depend bothon the sparseness (spectral sampling) and on the wavelength values of the bands that have been selected.