Hyperspectral Vnir - Swir Image Registration: Do Not Throw Away Those Overlapping Low Snr Bands

Abstract

We study the influence from a series of factors on the quality of the registration of VNIR and SWIR hyperspectral images. We specifically consider the registration of VNIR and SWIR images of different spatial resolutions acquired in controlled laboratory conditions on different historical artefacts. This registration problem is defined by a large scaling difference and small translation and rotation. We compare four methods and demonstrate that the largest effects on the quality metrics are due to the image contents and to the deployment of specific ranges of spectral bands. In particular, we demonstrate that using bands of similar nominal wavelengths gives significant advantage. While those bands have typically low values of Signal-to-Noise ratio and are frequently discarded, for this application, they can be treated as a valuable asset.