Improving the Classification in Shadowed Areas using Nonlinear Spectral Unmixing