Benchmark studies on pixel-level spectral unmixing of multi-resolution hyperspectral imagery

Abstract

ABSTRACT Spectral unmixing-based estimation of material abundances in hyperspectral imagery has a variety of applications in mineralogy, environmental monitoring, agriculture, food processing, pharmacy, etc. A substantial body of literature is available on different inversion algorithms, optional pre-processing such as dimensionality reduction, and algorithms for endmembers extraction. The quality of abundance estimation depends on the number of materials, size, the geometrical orientation of materials, the source of endmembers, and the inversion algorithm used. However, there is a lack of studies on one-to-one assessment of the retrieval of abundances under various scenarios of spectral material distributions, the spatial resolution of the imagery, and the potential of in-situ reflectance measurements as candidate endmembers. The unavailability of comprehensive benchmark data coupled with pixel-to-pixel ground truth data has impeded comprehensive assessment of the first principles of spectral unmixing from a verifiable experimental perspective. The objective of this research is assessing the dynamics of material abundance as a function of the source of endmembers, spatial resolution, number of materials, and the size of materials. Linear and its sparse-based spectral unmixing algorithms were implemented on the datasets acquired for the estimation of abundances, considering the different scenarios of material distributions, spatial resolution, and the source of endmembers. We validated the results using pixel-to-pixel ground truth maps prepared for the different cases of spectral unmixing. The results provide answers to some critical open challenges in spectral unmixing, such as, (i) for an unambiguous detection, the fractional distribution of material has to be at least 1% of the pixel, (ii) endmembers from the in-situ spectra based on the external spectral library can offer reasonably good abundance estimates (an error of up to 20% compared to the image-based endmembers), and (iii) geometric orientations of materials in the ground sampling distance influence the abundance estimations. The benchmark dataset generated in this work is a valuable resource for addressing intriguing questions in spectral unmixing using hyperspectral imagery from a multi-resolution perspective.