Reliably estimating the noise in AVIRIS hyperspectral images

A new method is presented for computing the noise affecting each band of an AVIRIS hyperspectral image. Between-band (spectral) and within-band (spatial) correlations are used to decorrelate the image data via linear regression. Each band of the image is divided into small blocks, each of which is independently decorrelated. The decorrelation leaves noise-like residuals whose variance estimates the noise. A homogeneous set of these variances is selected and their values are combined to provide the best estimate of that band's noise. This method provides consistent noise estimates from images with very different land cover types. Its performance is validated by comparing its noise estimates with noise measures provided with two AVIRIS images. The method works well with inhomogeneous images (e.g., of a vegetated area such as Jasper Ridge) unlike a method described recently by Gao. The method is automatic and does not require the intervention of a human operator. Noise estimates are presented for 10 images recorded in the years 1989, 1990 and 1992. We show that the method can be used for both radiance and reflectance (atmospherically corrected) images.