Robust Supervised Method for Nonlinear Spectral Unmixing Accounting for Endmember Variability

Due to the complex interaction of light with mixed materials, reflectance spectra are highly nonlinearly related pure material endmember spectra, making it hard estimate fractional abundances materials. Changing illumination conditions and cross-sensor situations cause spectral variability, further complicating unmixing procedure. In this work, we propose a supervised approach unmix mineral powder mixtures, containing variability. First, estimated by calculating geodesic distances between mixtures endmembers. It is argued experimentally validated that abundances, although not correct, invariant external Then, applied learn mapping from obtained follow linear model. To mapping, groundtruth number training samples required. Although any nonlinear regression method can be used Gaussian process found suitable when limited available. The trained model applicable all manifolds contain similar behavior as manifold, e.g., same measured another sensor. Using output simple inversion reveals true abundances. Experiments conducted on simulated real mixtures. particular, developed data sets homogeneously acquired two different sensors, an Agrispec spectrometer snapscan shortwave infrared (SWIR) hyperspectral camera, under strictly controlled experimental settings. proposed compared other approaches mixture models.