Diffuse reflectance spectra were measured after core splitting aboard the JOIDES Resolution (JR) using the Minolta CM-2002 handheld spectrometer. The technical characteristics of Minolta spectrophotometer and the methods for acquiring the reflectance spectra are those presented by Balsam et al. (1997). The measurements were taken at intervals varying from 2 to 5 cm (Keigwin, Rio, Acton, et al., 1998). A total of 4141 carbonate measurements, including those measured on board (Keigwin, Rio, Acton, et al., 1998), were used for calibration (Table T1; S.-O. Franz and R. Tiedemann, unpubl. data; M.-S. Poli, unpubl. data; and S. Hagen et al., unpubl. data). Both raw spectra (measured from 400 to 700 nm wavelength at 10-nm intervals) as well as pretreated reflectance spectra (see below), located within 3 cm on each side of the position of the samples measured for carbonate content, were used as input for a stepwise multiple linear regression. The transformations applied to the raw spectra as pretreatments consisted of squaring and first-order derivation. Squaring reduces the nonlinearity in the relationship between the reflectance and carbonate (Harris et al., 1997). The first derivative of the raw reflectance spectrum (computed as a first-difference derivative between each pair of successive reflectance values and plotted at the midpoint of the 10-nm interval between the two values; Balsam et al. (1997) exhibits an increased variability over the initial smooth spectral curves of deep-sea sediments and therefore helps discriminate mineral-specific spectral features (Balsam and Deaton, 1996). Attempts to include magnetic susceptibility in the regression as an additional inversely correlated proxy for carbonate content did not improve the results.
Because only reflectance data were used, all regression equations were forced through the origin because no minerals can be identified when the reflectance across the spectrum falls to zero (Balsam and Deaton, 1996). Following the strategy outlined by Mix et al. (1995), the carbonate measurements made on discrete samples were initially split into calibration and verification data sets. After verification, the regression-derived calibration equations were developed using all data available.