Context and objectives
In remote sensing applications, hyperspectral imagery captures the reflectance spectra of the different materials that are present on the surface, with a high spectral resolution in a large wavelength range spanning visible and infrared light. Spectral mixing plays an important role in these airborne and space borne sensors, as the field of view of each pixel is relatively large. The spectrum in each pixel is a mixture of all components inside, and often even outside, the field of view. Spectral unmixing aims to decompose each pixel again into its constituent spectra, their abundances, and possible additional details such as geometrical or atmospheric information.
Several popular unmixing strategies exist, based on linear or bilinear models, and radiative transfer approaches. These techniques suffer from several drawbacks, and are typically incapable of dealing with higher-order optical interactions, atmospheric interference, or geometrical issues such as lighting received from objects in adjacent pixels or shadowing. Furthermore, most approaches assume that a single pure spectrum exists for each spectral component, unable to accommodate for naturally occurring spectral variability between specimens of the same class.
In this project, we will tackle these problems by developing an unmixing methodology with capabilities that greatly surpass existing methods. In several work packages, we will design methods for handling spatial interference, shadowing, and spectral variability, based on concepts that we recently developed. The proposed approaches rely on geometrical and statistical modeling of the optical interactions, constrained optimization techniques for solving the resulting minimization problems, and exploiting insights into equivalent geometrical problems for speeding up search strategies, while not losing sight of the physical interpretation, feasibility and applicability in practical scenarios.
Expected scientific results
As the proposed methodology will be able to cope with many aspects that are now often missing from current unmixing methods, we expect this method to have a much better unmixing performance than alternative methods. Furthermore, by construction, the model will generate secondary information which can be of great interest to researchers and users, such as the level of nonlinearity, shadow, adjacency, pixel bleeding, and three-dimensional structure. In total, such an unmixing methodology would greatly surpass any current approach, both in quality of results, and in the amount of additional information generated. We expect this to generate a lot of interest from the remote sensing research community, and to quickly find real-world applications in spectral unmixing, replacing older, less flexible methods.
|Project leader(s):||UA - Vision Lab - Visielab|