A combined methodology to produce highly accurate classification for AVIRIS hyperspectral data

Abstract

The primary objective of this paper was to produce highly accurate supervised classification for airborne visible and infrared imaging spectrometer (AVIRIS) hyperspectral data for Chequamegon National Forest vegetation species in northern Wisconsin. This was done using a combined methodology of an atmospheric correction scheme and neural network together with data reduction and an increase in feature separation through three types of transformations, namely the fast Fourier transform (FFT), principal component (PC), and canonical transformation (CT). The atmospheric correction scheme was performed by removing the contaminating effects of aerosol particle scattering that are contained within the hyperspectral dataset due to the significant variation of the solar zenith angle and aerosol optical properties corresponding to cloud-free Earth targets. An algorithm in C++ has been developed to correct for the influence of solar zenith angle and the atmospheric components on scattering. The accuracy assessment criteria including overall accuracy, Ǩ (Kappa or KHAT statistics), and the statistical separability between classes were used. Using the correction of aerosol scattering, transformations, and the neural network classification, the classification accuracy of the vegetation species was shown by the overall accuracy of 97.09%, Ǩ = 0.96, and the average statistical separability of 1.99. The classification accuracy of the vegetation species for the uncorrected image was shown by the overall accuracy of 84.40%, Ǩ = 0.80, and the average statistical separability of 1.91. Therefore, the combined methodology including the atmospheric correction scheme, transformations, and the neural network classification yielded relatively high classification accuracy.