A novel machine learning fingerprinting method using sparse representation for provenance detection in delta sediments

Quantitative research on the sediment source to sink process is critical to revealing the evolution of the Earth's surface. The fingerprinting technique is useful for sediment source detection, and machine learning methods are being used with this technique. However, paleo-provenance research with limited measured data from sediment cores is not adequate for big data analysis which demands a large amount of data. Hence, determining how to assess authentic source contributions by using fingerprinting techniques based on very limited measurement data is a key scientific task. Sparse representation (SR), a machine learning method with advantages in feature extraction that treats the source as a linear combination of multiple tracers, has the potential to provide satisfactory answers. In this paper, a sparse representation-based sediment fingerprinting (SRSF) method is introduced. Based on size-specific magnetic and geochemical characterizations, we applied this method to quantize the provenance of three sediment cores from the Yangtze River Delta (YRD), where Yellow River sediment has contributed to delta formation over the past 400–600 years. Our results showed that the sediments of the Yellow River contributed 19.3 ± 5.8 % of the clay and fine silt (<16 μm) and 23.5 ± 9.2 % of the medium silt (16–32 μm) in the delta, which was consistent with previously published results. This study improves the current fingerprinting technique in the feature extraction and selection stage by leveraging SR and provides reference data for the development of the modern YZD delta. The SRSF method has many potential applications, especially in paleo-provenance detection with highly limited measurement data, which could help expand our understanding of paleo-Earth surface evolution.