Year: 2025
Type: preprint
Abstract: Riverine suspended sediment flux (SSF) is an important measure of water quality; however, limited in situ observations hinder our ability to comprehensively assess SSF at large scales. This study employed a deep learning approach (i.e., the Long Short-Term Memory network, or LSTM) to quantify SSF at a global scale. The LSTM models were trained in two ways: (1) A global LSTM that was trained and tested using 100,942 in-situ SSF measurements from 824 sites between the years 1984-2019 and (2) a set of LSTM models that were developed by grouping the 824 gauges into different clusters. Four different clustering techniques, based on basin area, stream order, land use, and climate, were assessed. For small to mid-sized watersheds (drainage area ≤ 105 km2), the area-based clustering approach yielded the best SSF estimates, with median relative error (RE), Nash-Sutcliffe Efficiency (NSE), and Kling-Gupta Efficiency (KGE) of 83%, 0.20, and 0.31, respectively. For large watersheds (drainage area > 105 km2), the global approach was most effective with median RE, NSE and KGE of 90%, 0.12 and 0.16. Furthermore, the LSTM performed significantly better than the Morgan-Morgan Finney (MMF) model at daily and monthly scales. LSTM found recent (~10 day) modelled river discharge most influential for SSF estimation, underscoring the need for better discharge estimates. Despite the small training dataset, the LSTM provides a valuable means for addressing the gaps in SSF measurements. Importantly, this study shows that training a deep learning model on all the watersheds may not be the best strategy.
Institutions: University of Cincinnati, University of Pittsburgh
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: gold
Award: G7401153355
OpenAlex ID: W4414593225
Cites:
Cited by:
Related to: