Signal-Dependent Noise Parameter Estimation of Hyperspectral Remote Sensing Images

In new-generation hyperspectral sensors, the electronic noise is not dominant and the photon noise has to be taken into account. Therefore, a parametric model that accounts for both signal-dependent and signal-independent noise on the useful signal is established. A novel algorithm to estimate the parameters of the model is proposed, which consists of two steps. First, the residual image is calculated by the multiple linear regression in spectral domain to decouple the strong spectral correlation. Then, local sample statistics of the hyperspectral image and its residual image are calculated, and the system of linear equations with respect to the signal-dependent and signal-independent noise variances is established. The least square solution of the equations is the estimation of the signal-dependent and signal-independent noise variances. Experiments on the simulated hyperspectral data analyze the accuracy of the method and experiments on the real-life data show its effectiveness.