Benefits of signal-dependent noise reduction for spectral analysis of data from advanced imaging spectrometers

Abstract

This paper shows that the signal dependent nature of the noise introduced by up to date imaging spectrometers is crucial for the spectral analysis carried out by the maximum noise fraction (MNF) transformation, which requires a preliminary estimation, either supervised or not, of the covariance matrix of the noise. Once the parametric noise model of the instrument has been estimated with the aid of calibration panels placed within the imaged scene, the mixed noise, i.e. photonic + electronic, can be removed. Noise filtering provides negligible improvements in the signal to noise ratio (SNR), at least whenever SNR is sufficiently high, but allows a correct spectral analysis to be accomplished via the MNF transformation, also in the absence of calibration panels. Conversely, the unsupervised estimation of the covariance matrix of the signal dependent noise may introduce unpredictable gross errors in the calculation of MNF transformation, thereby leading to transformed components that do not adequately capture the energy of the hyperspectral data.