Non-Linear Spectral Unmixing: A Case Study On Mangalore Aviris-Ng Hyperspectral Data

Abstract

Due to the low spatial resolution of the sensor, multiple scattering and intimate mixing at the ground, the majority of the pixels in the hyperspectral image are of mixed type. In this case, spectral unmixing is used to decompose this mixing effect. From the literature, it is clear that non-linear unmixing is more accurate and robust compared to linear unmixing. In this paper, we take a Mangalore dataset captured using Airborne Visible/ Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) camera to compare various non-linear unmixing. The paper presents and extensive comparison of various endmember extraction algorithms and abundance estimation algorithms. The performance of algorithms was assessed using two quality metrics (Spectral Angle Mapper and Reconstruction Error). Three types of experiments were carried out; endmember extraction accuracy assessment, testing of abundance estimation efficacy and comparison of linear and non-linear models. The simulation results conclude that Energy-based Convex Set (ECS) and Polynomial PostNonlinear Model (PPNM) give accurate results on the considered study site for endmember extraction and abundance estimation respectively.