Estimation of Signal to Noise Ratio for Unsupervised Hyperspectral Images

Abstract

Hyperspectral sensors have become a standard technology used in the techniques of observation by satellite and aerial platform for observing the terrestrial ecosystem with particular interest in the detection and identification of minerals, vegetation, materials and artificial environments. The detection of real materials depends on the coverage spectral resolution and signal to noise ratio of the spectrometer itself, as well as the density of the material and the absorption characteristics for the material in the region of wavelength measured. The signal to noise ratio in particular is one of the parameters that need to be estimated to establish the quality of images acquired by these systems. In this contribution a method to estimate the Signal to Noise Ratio (SNR) for unsupervised hyperspectral images has been investigated. The method uses the computation of local means and local standard deviations of small homogeneous blocks in order to define respectively the average signal and the mean noise of the images. If the noise may be considered mainly addictive the local standard deviation may be considered as the mean noise of image. This method uses all the spatial information contained in the image scene giving a representative SNR of entire image. The technique has been engineered in IDL environment and applied to hyperspectral data of HYPER-SIMGA sensor, developed in the frame of AIRFIRE Project for wildfire detection by airborne remote sensing data. The SNR results point out that HYPER-SIMGA SWIR images are quite noisy and the spectral range that has to be taken into account for data analysis is from 1000 to 1700 nm.