NIRS quantification of lake sediment composition by multiple regression using end-member spectra

Here we develop a novel method for quantifying sediment components, e.g. biogenic silica, organic or mineral matter, from near infrared (NIR) spectra based on fitting by multiple regression of measured spectra for end-member materials. We show that with suitable end-members our new open-source multiple regression routine gives excellent simultaneous quantification of the major components of a sediment, the concentrations comparing well with independent methods of quantification. Widely used partial least squares regression approaches rely on large environmental training data sets; our method produces comparable results, but with the advantages of negating the need for a training dataset and with greater simplicity and theoretical robustness. We demonstrate that component NIR spectra are additive, a prerequisite for use of multiple regression to un-mix the compound spectra, and show that a number of environmental materials make suitable end-members for this analysis. We show that spectral mixing is not conservative with respect to mass proportion, but rather to the relative chromatic intensity of contributing sediment components. Concentrations can be calculated using the measured spectra by correction using a chromatic intensity factor, the value of which can be measured independently. We have applied our approach to a postglacial sediment sequence from Loch Grannoch (SW Scotland) and reveal a down-core pattern of varying dominance by biogenic silica, organic and mineral content from the late glacial to present. With isolation and measurement of appropriate end-members this multivariate regression approach to interrogating NIR spectra has utility across a wide range of sedimentary environments and potentially for other spectral analytical methods.