Skip to main content

Partial least squares regression calibration model to predict biogenic silica and organic carbon percentages in High-Arctic lake sediment core samples

Presented by:
Vivienne Maxwell & Gregory de Wet (Smith College); Sara Stoudt (Bucknell University)
Abstract:

Partial least squares (PLS) regression is employed when working with highly correlated variables. In this case, multicollinear absorbance spectra are used to create a calibration model capable of predicting percentages of biogenic silica (BSi) and total organic carbon (TOC), which are proxies for temperature in High-Arctic settings. This is useful for paleoclimatologists who use lake sediment core samples to reconstruct past environments. Recently paleoclimatologists have begun to use Fourier Transform Infrared (FTIR) spectroscopy to collect information on BSi and TOC. However, the offloaded relative absorbance data make comparison with other proxies difficult. The goal of this project is to develop a universal calibration model using PLS to convert absorbance spectra into percentages of BSi and TOC so as to provide paleoclimatologists with a more universal way of analyzing and understanding their results. The model was developed in R using an existing pls Package and factors such as data preprocessing, number of PLS components, and k-fold cross-validation are used to determine what yields the most accurate calibration model.