An evaluation of class knowledge transfer from synthetic to real hyperspectral imagery

Hyperspectral imagery provides ample signal content to identify and distinguish between spectrally similar, but compositionally unique, materials, but representative training samples for all materials in a given scene are often unavailable in remote sensing applications. We propose a technique which leverages training spectra of materials from one hyperspectral image (the source), for classifying another (target) hyperspectral image, allowing robust target detection of source classes in the target image. By locating spectra of known material species in the source image which are also known in the target image, we derive a transformation that compensates for systematic spectral differences between the images, including atmospheric or seasonal effects, calibration differences, and noise. With this transformation applied as a similarity measure we can adapt a classifier trained on spectra from the source image, to classify the target image. We evaluate our technique between a pair of synthetic hyperspectral images with systematic differences in the spectral signatures of corresponding materials. Then we assess the feasibility of using synthetic imagery for training a classifier to identify spectral species in similar, real images, and show knowledge transfer results between a synthetic (HYDICE-type) source image and a real low-altitude AVIRIS target image of a complex urban area.