Integrating statistical and machine learning approaches for sediment transport prediction in a typical coarse sandy region of the Yellow River Basin

Inner Mongolia Autonomous Region (China). This study investigated the multiscale correlations among runoff, precipitation, potential evapotranspiration (PET), and normalized difference vegetation index (NDVI) with sediment load in the Ten Tributaries region from 2007 to 2021. Furthermore, sediment transport was predicted using statistical models and machine learning (ML) techniques to enhance understanding of sediment dynamics under varying environmental conditions. This work provided novel insights on the quantification of the scale-specific controls of sediment load in the coarse sandy region of the Yellow River. Multivariate empirical mode decomposition (MEMD) was employed to decompose the original time series of sediment load and its associated variables into five or six intrinsic mode functions (IMFs) and one residual component. Time-dependent intrinsic correlation (TDIC) analysis revealed that the relationships between sediment load and environmental factors exhibit dynamic, multi-scale properties. Runoff was the key factor affecting sediment load in Maobula and Xiliugou watershed. In Hantaichuan watershed, runoff dominated sediment load dynamics from IMF1 to IMF5, whereas PET governed the sediment transport process at IMF6. Three machine learning models, multilayer perceptron (MLP), convolutional neural networks (CNN) and particle swarm optimization-support vector regression (PSO-SVR) were applied to forecast sediment load in different basins. Integrating MEMD with ML significantly enhanced sediment load prediction accuracy. • Multi-scale relationships between sediment load and environmental factors were identified. • Runoff was the dominant factor controlling sediment load in the Ten Tributaries region. • MEMD combined with machine learning provided better sediment load prediction accuracy.