Deep Learning-Based Modeling of Daily Suspended Sediment Concentration and Discharge in Esopus

Accurately estimating suspended sediment concentration (SSC) in hydrological processes is challenging due to nonlinear nature of sediment dynamics, which are influenced by basin characteristics and temporal climatic patterns. To address this, deep neural networks (DNNs) have emerged as cuttingedge modeling methods capable of capturing the inherent nonlinearity in hydrological systems. Building on previous studies where LSTM emerged as the best-performing time series prediction model, this research, uses the long short-term memory method (LSTM) on discharge, SSC, and turbidity to simulate the discharge-suspended sediment relationship in Esopus Creek, NY, USA. We explore various feature selection scenarios, including combinations of discharge, turbidity, and SSC. Additionally, we analyzed the variations in SSC and discharge between the three USGS gages along Esopus Creek, specifically the upper stream, lower stream, and downstream areas adjacent to the pourpoint. The effectiveness of the DNN-based models is evaluated, highlighting the importance of feature importance in accurately simulating daily SSC and discharge. Our findings provide new insights into the application of machine learning for water quality assessment, emphasizing the critical role of feature selection.