A Novel Smoothing-Based Deep Learning Time-Series Approach for Daily Suspended Sediment Load Prediction

• Published in: Water resources management Vol. 37; no. 11; pp. 4271 - 4292
• Main Authors: Sahoo, Bibhuti Bhusan, Sankalp, Sovan, Kisi, Ozgur
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01-09-2023; Springer Nature B.V
• Subjects: Atmospheric Sciences; Civil Engineering; Daily; Deep learning; Earth and Environmental Science; Earth Sciences; Environment; Errors; Fourier series; Fourier transforms; Geotechnical Engineering & Applied Earth Sciences; Hydraulic engineering; Hydraulic loading; Hydrogeology; Hydrology; Hydrology/Water Resources; Long short-term memory; Memory cells; Missing data; Neural networks; Recurrent neural networks; Rivers; Sediment; Sediment load; Statistical methods; Suspended load; Suspended sediments; Time series; Fourier transformation; Recurrent neural network; Smoothed long short-term memory; Suspended sediment load
• ISSN: 0920-4741; 1573-1650
• DOI: 10.1007/s11269-023-03552-7

Precise assessment of suspended sediment load (SSL) is vital for many applications in hydrological modeling and hydraulic engineering. In this study, a smoothed long short-term memory (SM-LSTM) model was used to predict day-to-day SSL at two stations over two rivers namely Thebes station on the Mississippi River and Omaha station on the Missouri River. The model first removes the interference factors in the SSL time series by Fourier Transformation (FT) de-noising and then feeds into a long short-term memory (LSTM) network to forecast the SSL. Before de-noising, missing data in the time series is computed using the Monte Carlo multiple imputation technique. LSTM networks are a type of recurrent neural network (RNN) that incorporates memory cells, which makes them well-suited for learning temporal associations over the previous time steps. The model was built using daily observed time series of SSL in the Mississippi and Missouri rivers in the United States. The developed model was then assessed and compared to LSTM and RNN. These models were trained using 4 different time lags of the SSL time series as inputs. The SM-LSTM model with 12 lagged inputs outperformed the other models with the lowest root mean square errors (RMSE) = 32254 ton and mean absolute errors (MAE) = 19517 ton, and the highest Nash–Sutcliffe efficiency (NSE) = 0.99 for the Thebes Station while the model with 3 lagged inputs acted as the best with the lowest RMSE = 2244 ton and MAE = 1370 ton, and the highest NSE = 0.989 for the Omaha Station. The comparison of prediction accuracies showed that the SM-LSTM model can more satisfactorily predict daily SSL time series compared to LSTM and RNN.