Ensemble Wavelet-Support Vector Machine Approach for Prediction of Suspended Sediment Load Using Hydrometeorological Data

Explicit prediction of the suspended sediment loads in rivers or streams is very crucial for sustainable water resources and environmental systems. Suspended sediments are a governing factor for the design and operation of hydraulic structures, like canals, diversions and dams. In recent decades, to model hydrological phenomena which are complex in nature the machine learning models are used commonly. In the present study, support vector machine (SVM) with wavelet transform (WASVM) has been employed for prediction of daily suspended sediment load (SL) for two south Indian watersheds (Marol and Muneru) using hydrometeorological data. A 40-year daily observed data (1972–2011) have been used for the analysis, where past SL, streamflow ($Q$), and rainfall ($R$) data were used as the model inputs, and SL was the model output. Using conventional correlation coefficient analysis between input and output variables, the best input of WASVM model was identified. The reliability of SVM and WASVM models were evaluated on the basis of different performance criteria, i.e., coefficient of determination ($R^2$), root mean square error (RMSE), normalized mean square error (NMSE), and Nash-Sutcliffe coefficient (NS). Initially, 1-day ahead SL prediction was performed using the best WASVM model. The results showed that, 1-day predictions were very precise, showing a close agreement with the observed SL data ($R^2=0.94$, $\mathrm{NS}=0.94$ for the Marol watershed, and $R^2=0.77$, $\mathrm{NS}=0.77$ for the Muneru watershed) in the testing period. The same WASVM model was then used for the prediction of SL for the higher lead periods. The NMSE value for the Marol watershed was found as low as 0.06 for 1-day ahead prediction, and increases subsequently as 0.29, 0.46, and 0.70 for 3-, 6-, and 9-day higher leads, respectively. Likewise, for the Muneru watershed, the NMSE value was found as low as 0.21 for 1-day ahead prediction, and increases subsequently as 0.42, 0.53, and 0.68 for 3-, 6-, and 9-day higher leads, respectively. Further, the model was evaluated on the basis of its capability of predicting peak SL and cumulative SL for 1- to 6-day leads. The statistical analysis shows that the developed WASVM model can predict the target value successfully up to a 6-day lead and is not suitable for higher lead specifically in the selected watersheds having similar hydroclimatic conditions like the ones selected in this study. Predictions by the WASVM model were found significantly superior to the ones obtained by the conventional SVM model. The results revealed that the WASVM model provides a very good accuracy in predicting SL and can be used as an effective forecasting tool for hydrological applications.