Quantifying hourly suspended sediment load using data mining models: Case study of a glacierized Andean catchment in Chile

Suspended sediment has significant effects on reservoir storage capacity, the operation of hydraulic structures and river morphology. Hence, modeling suspended sediment loads (SSL) in rivers contributes to various water resource management and river engineering. An evaluation of stand-alone data mining models (i.e., reduced error pruning tree (REPT), M5P and instance-based learning (IBK)) and hybrid models, (i.e., bagging-M5P, random committee-REPT (RC-REPT) and random subspace-REPT (RS-REPT)) for predicting SSL resulting from glacial melting at an Andean catchment in Chile has been conducted in this study. The best input combinations are constructed based on Pearson correlation coefficient (PCC) of hourly SSL time series data with water discharge (Q), water temperature (T) and electrical conductivity (C) for different time lags. Seventy percent of the available data (one year of hourly data) is used to calibrate the models (dataset training) and the remaining 30% is used for model evaluation (dataset testing). The performances of the models are evaluated using several quantitative and graphical criteria, including coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency (NSE), percentage of bias (PBIAS), the ratio of RMSE to the standard deviation of observation (RSR), a Taylor diagram and a boxplot. All the models performed well in predicting SSL. However, the Friedman and Wilcoxon signed rank tests revealed that predicted SSL significantly differed for different models except between IBK (or M5P) and REPT. The hybrid models performed better than individual models. The bagging-M5P had the best predictive capability while the KEPT had the poorest.