IODP Proceedings    Volume contents     Search


Calibration of XRF data to CaCO3 measurements

The XRF normalized median-scaled (NMS) CaCO3 data from Lyle et al. (2012) was used for calibration and verification. Although the scaling and normalization steps make the NMS data more linear with respect to the calibration, they are not a calibration. The NMS data were calibrated with the CaCO3 measurements from Stockholm University listed in Table T1 to become true estimates of CaCO3 wt%.

We used 255 CaCO3 measurements as the calibration data set for the Site U1338 splice and reserved the remaining CaCO3 analyses as the check data set. The number in the calibration data set was determined by preliminary calibration trials. We made certain there was a range of CaCO3 wt% at all depths. Where there appeared to be systematic offsets, we added more calibration data to help the calibration better conform to the data set. The final calibration is from the linear regression shown in Figure F3 with an R2 of 0.87:

XRF CaCO3 = NMSCaCO3 × 1.0241 – 7.904,

where NMSCaCO3 is the XRF NMS CaCO3 component as reported in Lyle et al. (2012) and XRF CaCO3 is the XRF estimate of carbonate content. XRF CaCO3 estimates are listed in Table T2.

The calibration line is offset from the 1:1 line in Figure F3 because the NMS model used to precondition the Site U1338 XRF data is biased slightly high with respect to the actual CaCO3 measurements. Nevertheless, the uncalibrated data do a reasonable job of estimating differences downcore, as the difference between the NMS model CaCO3 and measured CaCO3 stays reasonably constant. The slope of the calibration line, in other words, is close to 1.

The remaining 850 CaCO3 analyses were used to check the quality of the XRF CaCO3 estimate. The locations of both check data and calibration data are shown with a profile of CaCO3 estimated by XRF in Figure F4, and differences between paired CaCO3 measurements and their XRF estimates are shown in Figure F5. Assuming a normal distribution, the standard deviation is ±4.98%. However, the distribution is more peaked than a Gaussian distribution. The data are marked by a few large mismatches, typically where there are large changes in CaCO3 wt% over short depth intervals.

The difference data have high kurtosis (more near-correct values than expected by a Gaussian distribution). Although 1 standard deviation is ±4.98%, 76% of the data is between ±5% of the correct value, rather than the expected 67%. We find that the highest differences between measured CaCO3 and the XRF estimate occur where there are rapid changes in CaCO3 content, suggesting that the greater differences between the two do not entirely represent error but also include actual differences in depth-matched samples when rates of change are high. Such differences might occur either where there are slight mismatches in depth or where there are significant differences between the sample composition near the center (and surface) of the split core where the XRF measurements were taken, versus near the edge (and below the surface) of the core where the discrete samples were taken. Alternately, the large tails may represent systematic mismatches at the extrema of the calibration. Examination of the data shows that the largest values clearly are “fliers” (i.e., rare cases where one data point was significantly mislabeled in depth or for some reason failed to be measured accurately).

We also find a slight but systematic positive trend in the differences between measured and estimated CaCO3 versus depth, suggesting that a small uncorrected porosity effect is still contained in the XRF estimate or that the slope of the calibration should be somewhat greater. Figure F6 shows the difference between measured and estimated CaCO3 versus depth in the core. For 0–100 compressed meters composite depth below seafloor (cmcd; using the CCSF-B method; see the “Methods” chapter [Expedition 320/321 Scientists, 2010a]), there is a systematic bias in the XRF estimate so that the XRF estimate averages 4% higher than the measured values with large scatter. Below 300 cmcd, the XRF estimate is systematically about 3%–4% lower than the measured values. This bias must be kept in mind but does not affect relative differences between samples on the 100 m scale.

Differences between measured and estimated CaCO3 are also somewhat smaller below 300 cmcd, but it is unclear whether this reflects lower errors because the CaCO3 content is less variable and the likelihood of calibration sample mismatches is less or whether the higher variability near the surface results from poorer ability to correct for porosity bias in the XRF data in the high porosity zone. Further work will help to better understand the porosity effect on the calibration.