Using non-negative matrix factorization in the "unmixing" of diffuse reflectance spectra

Abstract

Diffuse reflectance spectrophotometry (DRS), often termed simply color scanning, is a technique applied commonly to marine sediments to provide records of compositional variations. Measured DRS spectra, however, represent the bulk response of a sediment's constituents and are therefore difficult to interpret in their raw form. A quantification technique will be proposed and discussed which approaches the analysis of DRS data sets as a linear mixing problem and applies a non-negative matrix factorization (NMF) algorithm in their decomposition. The presented methodology allows the spectra of the end-member sediment constituents and their fractional abundances to be determined using only the measured data set. Unlike other DRS data processing techniques, NMF only allows additive combinations of end-members to explain the data set and will therefore return only positive abundances. Analysis of sediments from the eastern Mediterranean Sea demonstrates the applicability of the NMF approach and the new method of Orbital Cycle Stacking establishes that the "unmixed" data is consistent with expected environmental change.