Mineral identification in LWIR hyperspectral imagery applying sparse-based clustering

An assessment of mineral identification applying hyperspectral infrared imagery in laboratory conditions is presented here and strives to identify nine different minerals (biotite, diopside, epidote, goethite, kyanite, scheelite, smithsonite, tourmaline, quartz). A hyperspectral camera in Long-Wave Infrared (LWIR, 7.7-11.8 m) with a LW-macro lens, an infragold plate, and a heating source are instruments used in the experiment. For automated identification, a Sparse Principal Component Analysis (Sparse PCA)-based K-means clustering is employed to categorise all pixel-spectra in different groups. Then the best representatives of each cluster (using spectral averaging) are chosen to compare these spectra to ASTER spectral library of JPL/NASA through spectral comparison techniques. Spectral angle mapper (SAM) and Normalized Cross Correlation (NCC) are two of such techniques, which are used herein to measure the spectral difference. In order to evaluate robustness of clustering results among the minerals spectra, we have added three levels of Gaussian and salt&pepper noise, , and , to input spectra which dropped the accuracy percentage from more than 84.73, for 0 added noise, to 44.57, for 2 average of both additive noise, and 22.21, for 4 additive noise. The results conclusively indicate a promising performance but noise sensitive behaviour of the proposed approach.