Transformation of hyperspectral data to improve classification by mitigating nonlinear effects

Non-linear effects in hyperspectral data are caused by varying illumination conditions, different viewing angles or multiple scattering of the incident light. These effects interfere with commonly used data analysis procedures. Manifold learning procedures are slow and require certain assumptions about the data structure that do not necessarily hold in real hyperspectral data. In this paper, a transformation is proposed that uses neighborhood distances to track the nonlinear structures of multiple classes simultaneously. The transformation is evaluated using a hyperspectral data set containing nonlinearities. A classification is performed and the results on the original and the transformed data are compared.