Evaluating ability of three types of discrete wavelet transforms for improving performance of different ML models in estimation of daily-suspended sediment load

The main aims of this study are (a) evaluating effects of discrete wavelet transforms (DWTs) on the machine learning (ML) models, (b) evaluating ability of two types of ML models (regression and classification models) for estimating daily-suspended sediment load (SSL), and (c) selecting the best ML model or hybrid ML model with DWT by new methods such as the TOPSIS (technique for order of preference by similarity to ideal solution) method. The used DWTs are classical DWT, maximal overlap DWT (MODWT), and MODWT-multiresolution analysis (MODWT-MRA) with two types of mother wavelet functions db4 and sym7. The independent variables used as model inputs are daily flow discharge, stage, and precipitation covering period from 2000 to 2018 and collected in the Navrood Watershed (the north of Iran). The obtained results show that DWT can considerably improve accuracy of the ML models with regression nature such as multivariate adaptive regression splines (MARS). The DWT decreased root mean square error (RMSE) of the MARS model from 10,477 to 2038 kg/day. However, DWT has negligible effect on performance of the ML models with classification nature such as the random tree (RT), random forest (RF), M5 model tree (M5T), and classification and regression tree (CART). The TOPSIS method and Taylor diagram show that the RT-MODWT-db4 is the superior model for estimation of the daily SSL. The R and RMSE of this model are 1 and 122.18 kg/day, respectively. Using this model is recommended in the mountainous and forest watersheds for estimation of daily SSL.