As explained by Kourkoumelis et al. [28] the interpretation of Raman spectra is often hindered by a broad background signal, which is mostly due to fluorescence from organic molecules and contaminants. To try to remove fluorescent background noise, a low-order polynomial approximation of the spectrum can be subtracted from the original (Figs. 3–6). Two algorithms were explored:
least squares approximation and Chebyshev approximation [29]. The least squares approximation minimises the average error in the approximation whereas the Chebyshev approximation minimizes the maximum error and reduces the risk of significant localised
error. Both performed well for polynomials up to order 5, but as it avoids the potentially ill-conditioned numerical matrix inversion used for least squares approximation, the use of the orthogonal Chebyshev polynomials [30] means that any order of approximation can be robustly computed.