Prediction of Sediment Concentration in Rivers by Recursive Least-Squares and Linear Minimum Variance Estimators

Peaks of suspended sediment concentration in rivers appear before peak discharge appears. The observation of discharge is usually done in many rivers for river management, but the suspended sediment concentration is not always observed. The suspended sediment concentration is expressed by a sediment-rating curve over a long period of time. Coefficients in the sediment-rating curve are time varying. The time variation of the suspended sediment concentration is predicted by the sediment-rating curve with the recursive least-squares and the linear minimum variance estimators. Because the sediment-rating curve is linearized by the logarithmic transformation, we can apply the simple linear models such as recursive least-squares and linear minimum variance estimators for the prediction of the suspended sediment concentration. These methods indicate that the coefficients in the sediment-rating curve are time varying. These methods can estimate sediment discharge by identifying the coefficients in the sediment-rating curve. Nash-Sutcliffe model efficiency coefficients indicate that the three-day ahead prediction is possible with accuracy. The recursive linear minimum variance estimator is compared with the recursive least-squares estimator. As a result, the recursive linear minimum variance estimator is not always better than the recursive least-squares estimator, although the former is theoretically superior to the latter as long as the assumption of the linear minimum variance estimator holds. The application of these methods to the prediction of the suspended sediment concentration is found to be possible because discharges are assumed to be observed daily in many rivers.