Automated Detection and Chemical Characterization of Anomalous Grains in Scientific Ocean Drilling Legacy Cores

Abstract. The dynamic behavior of glaciers under varying climatic conditions plays a crucial role in Earth's climate systems, necessitating reliable records of glacial extent throughout the Cenozoic era. Ice-rafted debris (IRD) provides insights into past iceberg activity, glacial erosion rates, sediment transport, and meltwater delivery. Polar IRD reconstructions can enhance our understanding of ice sheet responses to climate shifts, however, traditional methods for detecting IRD are often destructive and labor-intensive. This study proposes a new methodological framework for detecting IRD in marine sediments as a proxy for reconstructing polar paleoclimate variability and glacial dynamics. Our new multi-proxy approach utilizes micro-X-ray fluorescence (μXRF) 2D imaging and computed tomography (CT) 3D scanning on five Deep Sea Drilling Project (DSDP) Site 266 archive half-sections. Our findings reveal that non-destructive scanning techniques have significant potential to effectively identify and quantify IRD, with μXRF providing high-resolution geochemical mapping and CT imaging enabling three-dimensional visualization of sediment structures. Machine learning classification of CT data using the Waikato Environment for Knowledge Analysis (WEKA) in the Fiji software package is a powerful tool for fast detection of anomalous components in IODP cores. However, we established that the finding of such coupled machine learning-CT approaches must be ground-truthed to ensure accuracy: our application of automated (k-means cluster analyses) chemical fingerprinting based on the μXRF data revealed many of the grains identified in the CT classification were false positives due to contamination from drill pipe flakes. We confirmed this contamination using scanning electron microscopy energy dispersive spectroscopy (SEM-EDS) on targeted samples. Fundamentally, our multi-proxy approach underscores that this approach alone cannot provide the geochemical fingerprinting necessary for unequivocal IRD identification. We therefore recommend a combined approach applying μXRF for geochemical fingerprinting and CT for structural detection to ensure a comprehensive, non-destructive method for IRD verification and preliminary provenance analysis. Ultimately, our work underscores the importance of applying stringent ground truthing to machine learning and high-resolution image datasets to ensure the data quality necessary for robust palaeoclimate reconstructions.