Using spatial correspondences for hyperspectral knowledge transfer: Evaluation on synthetic data

Abstract

We describe a proof of concept for class knowledge transfer from a labeled hyperspectral image to an unlabeled image, captured with a different (hyper-/multi-spectral) sensor, when the spatial extents of the images partially overlap. By defining a set of spatio-spectral correspondences between the labeled source image and the unlabeled target image, we create a mapping between the images we can use to propagate labels from the source to the target image. This mapping allows us to classify the target image using the source labels without manually defining training labels in the target image. We evaluate the technique using state of the art synthetic hyperspectral imagery.