Removing the Vegetation Effect in Mineral Maps Produced by Hyperion

Abstract

Estimation of mineral quantity with the presence of dry and green vegetation at pixel level is a challenging task in hyperspectral image analysis. A multiple linear regression model was trained to remove the effect of vegetation at pixel level in Hyperion image. The main diagnostic absorption features of the minerals in the study area are in 900, 2160, 2200, 2235 nm of the VNIR spectrum. In this study, the trained models did not show high sensitivity to the presence of noise in the spectrum and the vegetation cover changes on the ground. The maximum observed amount of green vegetation, dry vegetation, and mixed green and dry vegetation were respectively 41%–60%, 59%–64%, and 61%–76% in the study area. Field spectroscopy measurements were carried out by ASD-Fieldspec-3. Vegetation Corrected Continuum Depth (VCCD) model were developed for obtaining Continuum Removed Band Depth (CRBD) of mineral diagnostic absorption features. The model efficiency was evaluated in the presence of random noise and via introducing vegetation type changes. The accuracy of results was evaluated by the petrography, XRD and XRF of collected rock samples. Our results revealed that the presence of dry vegetation leads to underestimation of the frequency of all investigated minerals; similarly, green vegetation leads to underestimation of the frequency of the minerals, with the exception of hematite and goethite. Our results also showed that linear unmixing at pixel level will increase the accuracy of the algorithm to predict kaolinite, alunite, epidote, chlorite, pyrophyllite, goethite, muscovite and hematite minerals, respectively. In this work the maximum and minimum improvements of the accuracy were obtained as 0.40 and 0.09 for epidote and hematite, respectively.