Sediment texture and geochemistry as predictors of sub-depositional environment in a modern estuary using machine learning: A framework for investigating clay-coated sand grains

Sedimentary cores from the Ravenglass Estuary lack some of the sedimentary structures which can be seen in other estuarine sands due to their unconsolidated nature, making it difficult to meaningfully interpret depositional environments using standard sedimentological facies analysis. Here we explore how sediment texture, obtained from laser particle size analysis, and geochemistry, obtained from portable X-ray fluorescence, can be used independently, or in combination, to automatically classify sub-depositional environment and estuarine zone in a modern estuary. We have adapted an established Extreme Gradient Boosting workflow to select the most informative geochemical elements to be included in a training set to automatically classify sub-depositional environment at the surface of the Ravenglass Estuary, NW England, UK. Models that are trained exclusively on textural data significantly outperform those that use geochemical data when classifying sub-depositional environment but are comparable when classifying estuarine zone. However, the combination of textural and geochemical data in training sets improves model performance in all but one class when compared to separate textural and geochemical models. We have applied surface-calibrated combined textural and geochemical models to classify palaeo sub-depositional environment in three cores obtained from a tidal flat in the Ravenglass Estuary which are interpreted to record initial outer estuary deposition which transitioned to an inner estuary setting dominated by deposition from suspension. The subsurface classifications provide a framework to investigate the occurrence of reservoir quality-preserving detrital grain coats which vary in abundance as a function of sub-depositional environment at the surface. A review of literature suggests that the mixed flat sub-depositional environment is an ideal target for optimum clay grain coat occurrence, and we show that this sub-depositional environment is reliably identified by all models. We also discuss the implications of the modelling, how it compares to other machine learning approaches to understand reservoir quality in ancient systems, and how the workflow may be adapted for application to reservoir core. This study demonstrates value in utilising textural and geochemical data in conjunction with machine learning methods to help reveal the environmental evolution of marginal marine sands.