A Wavelet Based Data Mining Technique for Suspended Sediment Load Modeling

Abstract

The suspended sediment load (SSL) modeling generated within a catchment is a significant issue in the environmental and water resources planning and management of watersheds. The estimation methods of SSL are limited by the important parameters and boundary conditions (which are based on the flow and sediment properties). In this situation, soft computing approaches have proven to be an efficient tool in modelling the sediment load of rivers. In this study, the hybrid Wavelet-M5 model was introduced to model SSL of two different rivers (Lighvanchai and Upper Rio Grande) at both daily and monthly scales. In this way, first, the runoff and suspended sediment load time series were decomposed using the wavelet transform to several sub-time series to handle the non-stationary of the runoff and sediment time series. Then, the obtained sub-series were applied to M5 model tree as inputs. The obtained results for the Upper Rio Grande River at daily time scale, showed the better performance of Wavelet-M5 model in comparison with individual Artificial Neural Network (ANN) and M5 models so that the obtained Nash-Sutcliffe efficiency (NSE) was 0.94 by the hybrid Wavelet-M5 model while it was calculated as 0.89 and 0.77 by Wavelet-ANN (WANN) and M5 tree models, respectively. Also, the obtained NSE for the Lighvanchai River at monthly time scale was 0.90 by the hybrid Wavelet-M5 model while it was calculated as 0.78 and 0.69 by Wavelet-ANN (WANN) and M5 tree models in the verification step, respectively.