As the thermal environment experienced by the MESSENGER mission at Mercury has resulted to be warmer than was expected, the IR detectors on the Visible and Infrared Spectrograph (VIRS) onboard the spacecraft have recorded extremely large amounts of dark noise. In cases where data records include multiple consecutive dark observations, such as when the instrument is viewing the unilluminated surface or rides off the planet's limb, dark measurements exhibit potentially regular oscillations with time. In this paper, we propose to analyze spectrally those data in order to provide efficient reconstruction and characerization of the spurious records. Specifically, taking advantage of spectral unmixing methods, we aim to describe the dataset as a data cloud in a multidimensional space and to outline each spectrum as a nonlinear combination of noise patterns. Experimental results show how the proposed characterization is able to detail the dark spectra as a proper nonlinear combination of spectral signatures, s.t. the introduced spectral analysis methods can be used to characterize signals acquired over lit surface as well.