Deconvolution of SWIR reflectance spectra for automatic mineral identification in hyperspectral imaging

Hyperspectral sensors generally acquire images in the spectral range in more than one hundred contiguous narrow channels with a (deca)metric spatial resolution. Each pixel of the image is thus associated with a continuous spectrum which can be used to identify or map surface minerals. The most powerful algorithms (e.g., USGS (United States Geological Survey) Tetracorder) run with a standardized spectral library, are often supervised and require some expert knowledge. In this paper, we present an original method for mineral identification and mapping. Its originality lies in its fully automatic functioning for the full spectral range, from initialization using spectral derivatives, to spectral deconvolution and mineral identification, with a global approach. The modelling combines exponential Gaussians, a continuum including the fundamental water absorption at and deals with overfitting to keep only the relevant Gaussians. We tested the method in the SWIR (Short-Wave InfraRed, ) and for 14 minerals representative of industrial environments (e.g., quarries, mines, industries). More than 98% of the simulated spectra were correctly identified. When applied to two AVIRIS (Airborne Visible/InfraRed Imaging Spectrometer) images, results were consistent with ground truth data. The method could be improved by extending it to the VNIR (Visible and Near-InfraRed, ) spectral range to include iron oxides and by managing spectral mixtures.