Infra-sample, inter-sample and down-core microvariation in sea-surface temperature estimates obtained from planktonic foraminifera in the NE Atlantic

Quantitative investigation of populations of planktonic foraminifera in core-top and down-core (top 10 cm) samples from 14 sites in the NE Atlantic, between 19° to 52°N and 11° to 30°W, have revealed no intra-sample, inter-sample or down-core variability beyond that to be expected on counts of 300 specimens (2 standard deviations, 95% confidence interval). Application of F20 transfer function analysis shows that estimates of sea-surface temperature, based on these counts, fall within a range of ±1.9°C (precision). However, when these estimates are compared with ‘actual’ temperatures within the 1° latitude/longitude squares containing the sites, the estimate errors range as high as ±3.6°C (accuracy). These results indicate there is a continuing need to improve methods for estimating proxy temperatures from planktonic foraminifera, in order to match the requirements of numerical modellers of past climates.


INTRODUCTION
Calibration (or training) sets of core-top assemblage data are crucial for the inference of accurate micropalaeontological proxy-values for past ocean temperatures, salinities, productivities, etc. The first such sets for planktonic foraminifera were assembled in connection with the CLIMAP programme development of Transfer Function Analysis (TFA) (Imbrie & Kipp, 1971;Imbrie et al., 1973;Kipp, 1976), and additional compilations were prepared in connection with the development of the Modern Analog Technique (MAT) (Prell, 1985). The need for further expansion of such calibration sets has recently been recognized by the establishment of SCOR Working Group 100, and the IMAGES programme of the PAGES core project of IGBP (SCOR = Scientific Committee on Oceanic Research, IMAGES = International Marine Global Environmental Changes, PAGES =Past Global Environmental Changes, IGBP = International Geosphere Biosphere Programme). Meantime several centres, in addition to ourselves, and including Barash (Moscow) and Pllaumann (Gel), have been independently adding to knowledge of ocean core-top distribution of planktonic foraminiferal assemblages. In 1992 we embarked on a study of the core-top distribution of planktonic foraminifera in a selected area of the NE Atlantic, mainly utilizing the archive of cores at the NERC's Institute of Oceanographic Sciences Deacon Laboratory (Wormley).
As part of our investigation we looked again at the errors inherent in sea-surface temperature (SST) estimates made using a specific TFA, that are due to: (a) intra-sample, (b) intersample, and (c) down-core variations in the percentage abundances of planktonic foraminifera1 species. For this purpose we examined planktonic foraminifera from 14 sites, ranging between 19O and 52ON and between 1l0 and 3OoW ( Fig. 1). For ease of reference in this paper we have labelled these sites V1-6 (intra-sample and down-core investigation), M 1-5 (inter-sample investigation) and TI-3 (trans-turbidite investigation); the original core station numbers, locations and other details are given in Table 1.
For our intra-sample investigation we have taken two subsamples from the same core-top intervals, and subjected them to Location of sites ( 0 ) studied in this investigation, and location of core-top sites (+) used to generate the calibration data for the F20 Transfer Function (Molfino et ul., 1982).
the same preparation and species-counting procedures. Since CLIMAP it has been customary to count (identify) approximately 300 specimens to estimate the percentage abundance of species in the total population. The associated statistical error (2 standard deviations, 95% confidence interval) on a count of 300  Funnel1 & Swallow specimens varies from f5.8% at 50% abundance to f0.7% at 0.3% abundance. In practical terms this means that replicate counts of 300 specimens randomly selected from a natural population will yield percentages that fall within the stated statistical error limits 95 times out of 100. A count of 150/300 (50%) could represent an actual percentage in the total population of between 44.2 and 55.8%; a count of 5/300 (1.7%) an actual percentage of between 0.1 and 2.9%; and a count of 1/300 (0.3%) an actual percentage of between 0.0 and 1.0%0).] The results of our counting of duplicate core-top samples from cores V1 to V6 are given in Table 2.
For our inter-sample investigation we have taken samples from Multi-corer core-tops. Multi-corers take 12 separate cores within a 1 m radius. Therefore they sample any microvariation in the lateral distribution of microfossils at and near the sea-bed surface at a < 2 m scale. Gooday & Lambshead (1989) have reported on results for benthic foraminifera. As far as we know ours is the first investigation of the microdistribution of planktonic foraminifera in sea-bed sediments. The same preparation and species-counting procedures were used as for the intra-sample investigation. The results of our counting of multiple core-top samples from Multi-corer deployments M 1 to M5 are given in Table 3.
For our down-core investigation we have taken samples from successive 1 or 2cm intervals down to lOcm depth, and subjected them to the same preparation and counting procedures as the intra-and inter-sample samples. In most deep-sea cores this IOcm zone is likely to be one of active, present-day bioturbation, and the contained microfossils therefore represent the climatic equivalent of the latest Holocene, rather than solely the present-day. Consequently little change in the planktonic foraminiferal assemblages over this interval is to be expected. On the other hand confirmation of this expectation would indicate that samples taken from cores whose top few centimetres are unsampleable, or have been lost on collection, may still be useful contributors to a calibration set. Some 'coretop' samples used in the CLIMAP calibration set were obtained from lOcm depth. The results of our counting of down-core samples are given in Table 4.
In addition to our study of intra-sample, inter-sample and down-core variation we also made trans-turbidite comparisons on three cores where core-top samples were immediately underlain by turbidites, and could be directly compared with the sub-turbidite assemblages. (The supra-turbidite samples in these cases would not have been able to develop mixing by bioturbation with the underlying earlier Holocene pelagic sediment.)

METHODS
The Box-corer subcores and Multi-corer cores used in the intrasample and down-core investigations, ( V 4 of this paper), were obtained during various cruises of RRS Discovery. Full details of the Cruise and Core numbers and Station data are given in Table 1. The Box-corer subcores had been previously sectioned and used for pore-water analysis, and the samples taken for the present investigation were obtained from the remaining 'squeezed' cakes. (The samples obtained from this source proved to contain relatively low percentages of fragmented planktonic foraminifera, indicating that the prior processing of the samples had not damaged the planktonic foraminifera.) The Multi-corer cores had also been previously sectioned, but the resultant core slices were preserved in buffered (neutral) formaline or alcohol.
The Multi-corer cores used in the inter-sample investigation (MlLM5 of this paper), were obtained by RRS Challenger (Table 1).
The Box-corer cores, used for the trans-turbidite investigation (Tl-3 of this paper), were obtained by RRS Discovery (Table 1).
All samples from both Box-corer and Multi-corer sources were oven-dried to disaggregate the sediment, then washed through a 63 pm sieve. The > 63 pm fraction was dry-sieved to separate the > 150 pm and the 1 5 M 3 pm fractions.
150pm fractions were sub-divided, using a random sediment splitter to ensure equal division of larger and smaller planktonic foraminifera, in order to obtain sub-samples containing approximately 300 whole specimens. The planktonic foraminifera were identified (Plate 1) in accordance with the standard taxonomy used in connection with the F20 transfer function (Molfino et al., 1982), and the planktonic/benthic ratio determined at the same time. The number of fragments of planktonic foraminifera encountered while counting c.300 whole planktonic foraminifera was also recorded. All results are set out in Tables 2 4 . The

RESULTS
Twenty-eight species of planktonic foraminifera, including the Neogloboquadrina pachyderrna (dextra1)ldutertrei intergrade form (Kipp, 1976), were identified and percentage abundances calculated. Coiling direction, colour and form of species were also distinguished wherever appropriate. Percentage benthic foraminifera, and percentage planktonic foraminiferal fragments were separately determined relative to the total count (c.300) of whole planktonic foraminifera. All results are set out in Tables 2 4 . All the intra-sample and inter-sample sample sets, and almost all the down-core and trans-turbidite sample sets show no variation between counts obtained on individual cores exceeding that which would be expected from statistical counting error (at the 2 standard deviations, 95% confidence interval level). This result is at the same time both encouraging (because it implies minimal variability arising from intra-sample, inter-sample, down-core and trans-turbidite variations in the original populations), and disappointing (because it reveals no such variations and leaves statistical counting errors as probably the principal source of lack of precision in TFA estimates of SST).
In the following account we use the terms precision to refer to the consistency of TFA estimates of SST obtained from single sample sets (i.e. from one site), and accuracy to refer to the consistency between TFA-estimated SST values and 'actual' present day SST in the same l o latitude/longitude squares. The 'actual' present day SST values we have used are from Levitus (1982).
To evaluate the effects of statistical counting errors on TFA estimates of SST, we have applied the F20 transfer function to all our data. The F20 transfer function (Molfino et al., 1982) is a late version of the type of transfer function originally developed and used for the CLIMAP programme. Imbrie et al. (1973) had noted that the estimate precision of the earlier F3 transfer function varied according to geographical       location as the transfer function was applied to faunas with different community structures. With the F3 transfer function mid-latitude samples tended to have lower precision, whereas single assemblage samples, e.g. polar, had higher precision. Imbrie et al. (1973) measured sample precision using five replicate samples. Their highest precision was ~tO.40-0.54"C (range 0.8&1.07"C) and lowest f2.00°C (range 4.00"C). Molfino et al. (1982) assessed the accuracy limits of the F20 transfer function as h1.2"C (range 2.4"C) for SST-,,, (i.e. summer) estimates.
The likely reliability of TFA estimates of SST can also be evaluated by calculating a communality value for each sample. This measures the similarity of the planktonic foraminiferal assemblages under investigation to those used in the original calibration set used to generate the transfer function (Imbrie & Kipp, 1971). A communality value of 1 indicates that the calibration and investigated assemblages are identical. Imbrie et al. (1973) considered that values >0.8 indicated that the assemblages under investigation fell significantly within the range of the assemblages used to calibrate the transfer function. In our results (Tables 5B-8) Kipp (1976) from the same general area.

DISCUSSION
Our intra-sample variability results, expressed in terms of F20 SST,,, (=summer) and SSTcold (=winter) estimates, are summarized in Table 5A, and the precision and accuracy of these estimates are compared in Table 5B. The minimum range of the SST, , , estimates is f0.O0C and the maximum f0.51"C; for SSTWld the comparable figures are +o.Ol"c and f0.70°C, respectively. However, when we come to compare these estimates with the 'actual' present-day sea-surface temperatures in the relevant 1" latitude/longitude square, we find much larger discrepancies. SST,,, estimates range from an underestimate of -335°C to an overestimate of +3.5"C, whereas SSTcold estimates range from an underestimate of -2.92"C to an overestimate of 0.79"C (based on the mean, not the extreme, estimate values). The range of variation in the accuracy of these results is considerably larger than the variation in the precision.
Our inter-sample variability results, are summarized in Table  6A, and the precision and accuracy of these estimates are    Table 6B. Comparison of multi-core F20 SST estimates with 'actual' SST at sites M1-5.  Table 6B. Here the minimum range of the SST,,, estimates is *0.14OC and the maximum f0.95'C; for SSTcold the comparable figures are fO.2O0C and 1.08"C respectively. Relative to present-day sea-surface temperatures the SST,,, estimates range from an underestimate of -2.46"C to an underestimate of -1.13OC, and the ssTcOld estimates from an underestimate of -3.14OC to an underestimate of -0.18OC.
Here again the variation in accuracy is larger than the variation in precision, and all SSTs are underestimated. It should be noted that the communalities for these samples are somewhat less than 0.8 (see above), but high communalities in the intra-sample set are associated with even higher inaccuracies. Once again only mean SST estimates are compared with the present-day seasurface temperatures.