Recovery of ground-truth pixel information from airborne hyperspectral images

Airborne hyperspectral imaging sensors provide hundreds of contiguous measurements in the visible through infrared region. The availability of data in the spectral dimension provides an opportunity to extract information related to material surfaces corresponding to a sensor pixel. The measured spectral radiance signal of a hyperspectral sensor is dependent on the material surfaces present within the pixel and the illumination and atmospheric conditions. We present nonlinear algorithms to estimate the reflectance spectrum and the orientation of ground material corresponding to a single material pure pixel. The algorithms are based on a nonlinear physics-based image formation model. We model the reflectance spectra using a low-dimensional subspace model. We model the common dependence of the illumination spectra and the path-scattered radiance on the viewing geometry and atmospheric condition by using a low dimensional coupled subspace. The algorithms exploit the fact that reflectance and illumination spectra typically lie in distinct subspaces. The algorithms involve solving constrained nonlinear optimization problems to estimate the material surface reflectance spectrum and it orientation. We extend the nonlinear algorithms to extract subpixel information for pixels with more than one material. It involves solving a constrained nonlinear optimization problem to estimate the pixel area fractions of different materials present within the pixel. We evaluated the utility of our algorithms on a large set of synthetic and real data.