Semi-automatic core characterisation based on geochemical logging data

Abstract

Textural variations in sediment are considered to be one of the primary controls on the geochemical composition. This relation has been widely exploited for the climatological interpretation of unconsolidated Quartenary sediment cores. Despite the common objectives in respectively unconsolidated and consolidated sedimentological studies, for the latter the use of novel techniques to acquire geochemical data, such as non-destructive XRF on cores, has not been applied yet. Because there is an ever increasing need for detailed (textural) sedimentological data, this study investigated the possibilities of in-situ XRF for semi-automatic textural characterization. An improvement to the current work flow of XRF core scanning data acquisition was suggested by a pre-calibration filtering step, in order to increase the signal to noise ratio. Apart from that, a sample selection routine was developed which, given the underlying calibration strategy, was expected to yield improved calibration results. Embedding the two algorithms into the calibration work flow proved that the sample selection algorithm as well as the pre-calibration filtering step result in improved calibration output. They are also found to be effective for real data; when applied to two datasets of unconsolidated cores, similar results were obtained. The second part of this study included the design of a sedimentary basin infill model ”LINMIX”, based on the mixing of different sediment sources with a grain size dependent chemical composition. By doing so, the observed geochemical signal is decomposed into a portion that is the result of textural variations and a portion that reflects differences in provenance. When applying LINMIX to a record of Quaternary sediment (offshore Senegal) as a proof-of-concept, the model was able to reconstruct the geochemical record satisfyingly by linear mixing of 3 endmembers with 3 unique Grain Size Distributions (GSD’s) and 2 unique functions characterising the compositional change in the grain size spectrum (Transfer function or ’TF’). This result implies that the chemical variation induced by the relative mixing of two endmembers was only the result of grain size variation and not of source material variation. The two endmembers that were designated a common TF, have in a previous study been interpreted as both reflecting material from the same sediment source. This implies that the LINMIX model has potential for semi-automatic provenance interpretation of sediment cores. Additionally the endmember compositions correspond fairly well with the present-day composition of the Senegal river and African eolian dust. Finally it was investigated how geochemical data can support semi-automatic grain size prediction of consolidated sediment, for which holds that there is no sediment source variation. This was done using a dataset of Carboniferous material, which is highly variable in terms of lithofacies. Around 11 meter of core has been geochemically logged with an XRF core scanner and successfully calibrated using 40 calibration samples (20 unique sample depths). The quality of the core scanning data was found to be high; the main rock-forming elements were calibrated with a signal to noise ratio larger than 4. Subsequently the textural information content of the geochemical data was investigated in two different setups. Initially the data was used as a quantitative tool to fill in the gaps between the grain size derived from plug data using Multi Variate Regression (MVR). Apart from that it was used as input for an unsupervised Bayesian classification scheme in terms of grain size classes. Whereas the former yielded a residual variance on the input data smaller than 1.5 f -units in 66 % of the cases, the latter identified two core sections that, given RGB and geochemical data, should be classified differently. In both cases, an additional validation step should give more insight in the performance of the scanner as a grain size predictor. Recommendations w.r.t the scanner include (1) embedding the suggested algorithms in the data acquisition work flow and (2) installing a higher resolution camera to extract textural proxies from images.