The way the images were collected was not designed to provide an estimate for analytical precision of the digital-image-based color measurement method. Even if a digital camera remains stable over a longer period of time, reproducibility of RGB values of a specific object is not in itself a measure for precision. The accuracy and precision of the final color data is mostly dependent on the processing steps (i.e., conversion to calibrated color values and correction for uneven light distribution within single images). We cannot as yet estimate the error involved in the data processing, but inspection of the generated color time series yields some rough measure of what the analytical accuracy of the method can be.
The fact that processed color values for Holes 1019C and 1019E are very similar, even though the calibration settings of the camera were changed considerably, shows that the imaging method can yield accurate color data (Fig. 1). The color-conversion equation developed here provides a method that is robust as far as changing the setting of a given camera is concerned. Further testing should show if consistent results can be obtained also when different camera systems and (or) different light sources are used to image one and the same object.
Visual examination of a* values, which, of the three color variables, has the lowest amplitude signal, shows that the noise level is low enough to recognize even subtle color trends. Figure 3 shows examples of parallel holes in two sites. Site 1010 has a low-amplitude signal in a* but shows a consistent trend in all three holes, which presumably registers a Milankovitch signal. Site 1022 shows a much weaker and less regular signal, which is less easily correlated between parallel holes. Presumably, a large part of the variation at that site is due to noise and may provide a first estimate of analytical precision. Wavelengths of 1 m and more were filtered out of the data in this interval at Site 1022, and the standard deviation was calculated over the remaining high-frequency data. Standard deviations are between 0.5 and 0.8 for all three color variables, depending on which hole is used. This provides a rough maximum estimate for the analytical error, which is sure to be an overestimation of the real error, as it includes "random" but real color variation as well as error attributable to the method. The main source of analytical error at present is in the way the correction for light is carried out. Here, an average regression curve for a large number of images is used to estimate a light correction curve, which means that individual images with flatter or steeper light deviations are not corrected adequately. On the other hand, given that real variation in color is present at decimeter to meter scale, using only a few images to estimate a correction curve could too easily result in a correction that filters out part of the real color variation. For now, the data contain a systematic error, which cannot be filtered out, with a deviation from the real color as a function of the position in the image relative to the light source. Future work will be done to test how an overlap of several centimeters between images can be used to refine the light-correction procedures, and thereby reduce the analytical error.
The reprocessed digital image color data set included on the volume CD-ROM (back pocket, this volume) covers all sites drilled during Leg 167, except Site 1015. Part of the images for Site 1015 could not be retrieved from their storage medium for reprocessing, and, in combination with relatively low core recovery, the resulting time series is too incomplete to warrant inclusion. For all other sections, except Site 1022, composite depth was converted into an age scale by linear interpolation between shipboard age data. A zero age is assumed for the top of each site; ages for sediments below the oldest age-control point were assigned an age by extrapolating accumulation rates. Site 1022 has too few age control points to provide a meaningful age interpolation. Figure 4, Figure 5, and Figure 6 present filtered composite records plotted against age, consisting of the shipboard splice, below which data from the deepest hole at each site are added.
A preliminary indication of how sediment composition determines color can be obtained from a comparison of color data with shipboard chemical data. Organic carbon (TOC) and carbonate content measurements and the semiquantitative estimates from smear slides are matched to color data points in the nearest centimeter in the section, and correlation coefficients are calculated between the compositional data and the three color variables (Table 1). For smear-slide data, only data from major lithologies were used; the sum of all siliceous microfossils is used as an estimate of biogenic silica (BSi) content. Clay content is included as a test for consistency of the semiquantitative data. Smear-slide and chemical samples cannot be matched directly, because they are not taken in the same interval; however, positive correlations between L* and carbonate and negative ones between clay and L* (Table 1) are as expected and suggest that smear-slide data are internally consistent. As such, BSi estimates can provide a useful first comparison with color data.
Figure 4 shows the color record for the three deep-water sites, Sites 1010, 1016, and 1021. At these sites, long-term variation at Ma-scale occurs in carbonate content (Lyle, Koizumi, Richter, et al., 1997). TOC is generally low. BSi is high in sediments older than 7 Ma in Sites 1010 and 1021 and is variable but high on average throughout the section at Site 1016.
L* correlates strongly with carbonate content; b* follows L* at Sites 1016 and 1021 but not at Site 1010. The color variables show the long-term, high-amplitude variation that correlates with the trends in carbonate content. Overprinted on this are Milankovitch-scale cyclicities that have much lower amplitude. In these sections, carbonate appears to be the dominant control on color composition. For Sites 1010 and 1021, with high BSi deep in the section and low values throughout the younger parts, a correlation between BSi and color is absent. At Site 1016, with lithologies consisting of diatoms and calcareous nannofossils mixed in variable proportions with each other and with clay, BSi content has a negative correlation with L* (i.e., BSi-rich intervals are darker than average). In carbonate-free sediments, high silica content usually corresponds to lighter than average sediments. Presumably, at Site 1016, lightness is more strongly influenced by carbonate in the three-component system of carbonate, clay, and BSi.
Figure 5 and Figure 6 show the color record for the past 5 Ma in all sections for which an age interpolation was calculated. The northern shallow- and intermediate-water depth sites have generally subtle variation in composition, with low carbonate content and variable amounts of biogenic silica. In the color variables, this is reflected in low-amplitude time series and generally weak correlations with compositional data. Geochemical data will be collected at finer resolution in specific intervals to unravel what controls the subtle color variations in these sections. Still, even though the trends are low amplitude, the data show the presence of variation at the scale of orbital cycles. For example, L* in Sites 1018 and 1020 show the presence of cycles at the scale of 100 ka.
The highest amplitude high-frequency signal is found in the southern shallow- and intermediate-water depth section. Sites 1011 to 1014 have pronounced variation in TOC and carbonate content, which produces clear cycles in L* and b*. Overall, carbonate-rich sediments are lighter and more yellow, whereas organic-rich sediments are darker and more gray. Sites 1013 and 1014, which have the highest TOC content on average of all sites, are the only two sites that show a relatively high-amplitude signal in a*, with values periodically reaching moderately high positive values and with a positive correlation with TOC (i.e., organic-rich sediments contain a red component).
The color data presented here are one of the data sets that can be used for high-resolution correlation of sites with each other as well as fine tuning of the age-depth scale. At the resolution used for this study (1 cm), the color time series extracted from digital images are comparable with data that can be obtained with the Minolta spectrophotometer that is now part of the standard shipboard analytical tools, but was not yet used during Leg 167. The Oregon State University spectral reflectometer has similar spatial resolution, but can measure color in a wide range of wavelengths, which can be translated into a good proxy for chemical sediment composition (Mix et al., 1995). In contrast, the video system as well as the Minolta system measure light only in three broad bands (red, green, and blue for the video system). Choice of method is mostly dependent on the actual application. Video images or still camera images can provide a real resolution of up to ~100 measurements per centimeter of sediment (or 0.1 mm), which allows registration of very fine-scale color variation, for example in laminated sediments (Merrill and Beck, 1995; Schaaf and Thurow, 1995, 1998). Methods presented here illustrate that after refinement of the light correction procedure, digital images can deliver good quality calibrated color data also at lower resolutions. A prerequisite is that the core surface is scraped carefully with a tool without sharp edges (e.g., a dough scraper) to remove scratches and other surface irregularities that would produce shadows or reflections causing artifacts in the image data.