Deep learning-based grain-size decomposition model: A feasible solution for dealing with methodological uncertainty

Abstract

Terrigenous clastic sediments cover a large area of the Earth's surface and provide valuable insights into the Earth's evolution and environmental change. Sediment grain-size decomposition has been widely used as an effective approach to inferring changes in sediment sources, transport processes and depositional environments. Several algorithms, such as single sample unmixing, end-member modelling analysis and the universal decomposition model, have been developed for grain-size decomposition. The performance of these algorithms is highly dependent on parameter selections, introducing subjective uncertainty. This uncertainty could undermine the reliability of decomposition results, limit the application of grain-size decomposition techniques and reduce comparability across different studies. To mitigate the methodological uncertainty, a novel deep learning-based framework for grain-size decomposition of terrigenous clastic sediments is proposed. First, an improved universal decomposition model is used to analyse the collected grain-size data, in order to provide training sets for the end-to-end decomposers. To meet the data size requirements of supervised learning, generative adversarial networks are also trained for data augmentation. The performance of the new framework is then evaluated using a small-scale dataset (73 393 samples from 18 sites) of three sedimentary types (loess, fluvial and lake delta deposits). The decomposed grain-size results demonstrate high feasibility and great potential of the framework in constructing a robust grain-size decomposition model. Finally, it is proposed that future grain-size research should aim to establish guidelines for grain-size data sharing and produce a big grain-size database for deep learning.