Suspended sediment load is expensive to measure and sparsely gauged, while streamflow and gridded climate records are abundant. Whether machine learning can bridge that gap — and which forcing data it actually needs — was not fully established.
Boosted decision-tree models (XGBoost, LightGBM, CatBoost) estimate daily suspended sediment load from hydro-climate forcings: streamflow, precipitation, and maximum and minimum temperature. They are trained and tested on NOAA's nClimGrid-daily and GHCNd records at 47 gauges across the United States, spanning 1950 to 2021.
Abstract as published, quoted from the authors' companion Zenodo record; the publisher exposes no machine-readable abstract.
The potential of machine learning (ML) models to estimate and predict sediment transport issues utilizing hydro-climate variables has not been fully investigated. In this study, we examine the potential of ML models for modeling sediment transport by utilizing hydro-climate variables as forcing data. To accomplish this, we used various boosted decision tree models such as Xgboost, LightGBM, and Catboost to estimate suspended sediment load (SSL) using hydro-climate variables such as streamflow, precipitation, maximum and minimum temperature. The models are trained and tested using dataset of the National Centers for Environmental Information (NCEI)'s NOAA Daily U.S. Climate Gridded Dataset (nClimGrid-daily) from the Global Historical Climatology Network-daily (GHCNd), respectively, at 47 gauges across the United States (US) between 1950 and 2021. The findings indicate that the proposed methodology has the potential to predict SSL at the daily scale in various climate zones with reasonable precision, as demonstrated by the average Nash-Sutcliffe Efficiency (NSE) of >0.66, Kling-Gupta Efficiency (KGE) of >0.67, and of Willmott Index (WI) of >0.83 during the unseen (test) phase.