An adaptive similarity measure for classification of hyperspectral signatures

Capturing both the shape of the spectral continuum and the positions/widths of absorption bands is essential to accurately measure similarity between hyperspectral signatures. Furthermore, the relative importance of these features is data dependent. In this letter, we present an adaptive version of our recently proposed continuum-intact (CI)/continuum-removed (CR) similarity measure which automatically determines a convex weighting between similarity measurements of CI and CR signatures according to input data. We describe an efficient technique to calculate an optimal weight for a linear combination of CI and CR similarity measurements. We evaluate the technique on the Airborne Visible/Infrared Imaging Spectrometer spectra sampled from a well-studied urban scene and show that our technique yields improved classification accuracy in comparison to CI or CR similarity measurements alone and performs comparably to calculating the weight via brute-force search, at a much reduced computational cost. A source code implementation of our algorithm is provided online.