Towards the Spectral Restoration of Shadowed Areas in Hyperspectral Images Based on Nonlinear Unmixing

This work proposes a new shadow restoration method for hyperspectral images based on nonlinear unmixing. A physical model is introduced to estimate the shadowed spectrum from a spectrum of the same material exposed to direct sunlight. By defining pure spectra receiving direct and indirect illumination as sunlit and shadowed endmembers, respectively, the proposed method estimates the abundance maps for both sunlit and shadowed endmembers pixelwise, taking into account nonlinear effects up to the second order, which are of particular importance in shadow areas. Subsequently, the spectrum of a pixel in a scene is restored by a linear combination of sunlit and shadowed endmembers. Experimental results show that shadowed spectra can be successfully recovered and their true reflectance better estimated. In addition, the proposed method solves shadow detection and restoration simultaneously, so that it does not need a shadows mask as an additional input.