Suspended sediment load modeling by machine learning algorithms in low and high discharge periods (Case study: Kashkan watershed)

Introduction
Sediment that moves with water is called suspended sediment, and the amount of suspended sediment material that passes through a river section in a certain period of time is called suspended load. The suspended sediment load (SSL) of a watershed, which passes through a certain section of the river, depends mainly on the climatic characteristics, the characteristics of the watershed and the capacity of carrying sedimentary materials. Actully Suspended sediment transport in the river is a function of meteorological and hydrological parameters as a complicated process. The input suspended load is one of the important and influencing factors on the amount of sediment input to reservoirs of dams and lakes. Determining the amount of sediment carried by rivers is important in many aspects. The calculation of suspended load is very important because of various reasons, one of the most important reasons is the role of suspended sediment load in the quantitative and qualitative management of surface water resources. Therefore, the distribution and transportation of suspended sediment load (SSL) in rivers have a significant effect on the water resource management, design of hydraulic structures, river morphology, water quality, and aquatic ecosystems. In fact, accurate and reliable modeling of suspended sediment load (SSL) is very important for planning, managing, and designing of river systems and water resource structures.  In addition, the determination of dry and wet periods is very important in studies related to water resources management, especially in arid and semi-arid regions.
 
Materials and Methods
To campare the result of the proposed models’ performance, the Cham Anjir, Bahram Jo, Kaka Reza and Sarab Syed Ali hydrometry stations in Khorramabad, Biranshahr and Alashtar sub-watersheds (a part of Kashkan watershed) in western of Iran, is used as a case study area. The geographic coordinates of the Cham Anjir, Bahram Jo, Kaka Reza and Sarab Syed Ali are 48° 15 '34" E 33° 26' 55" N, 48° 17' 45"E 33° 34' 8" N, 48° 13' 51" E 33° 43' 39" N and 48° 12' 14" E 33° 44' 55" N, respectively. The studied area has a semiarid climate with a mean annual rainfall Less than 500 mm. The studied area has a maximum elevation of 3578 m in Alashtar watershed and the minimum elevation of 1158 m in khorramAbad watershed. Most parts of the studied sub-watersheds are rangeland, while forest, dry farming, and irrigation lands are in considerable quantities. The surface lithology in the KhorramAbad, Alashtar and Biranshahr watersheds are covered by the Eocene, Quaternary, Cretaceous, Miocene, Oligocene, Paleocene, and Pliocene geologic formations.
Predicting suspended sediment load (SSL) in water resource management requires efficient and reliable predicted models. The present study was carried out for the modeling of Suspended sediment load by learning algorithms in low and high discharge periods. In this study, five soft computing techniques, GP-PUK, GP-RBF, M5P, REEP Tree and RF were used to predict the SSL in Cham Anjir, Bahram Jo, Kaka Reza and Sarab Syed Ali hydrometry stations in Khorramabad, Biranshahr and Alashtar sub-watersheds. Total data set consists of rain, discharge and suspended sediment load (in a period of 18 years from 2000 to 2018). of three sub-watersheds out of which 70% data used to train the model and 30% data were used to test the model. Finally, the models’ accuracy was assessed using three performance evaluation parameters, which were Correlation Coefficient (C.C.), Root Mean Square Error (RMSE) and Maximum Absolute Error (MAE). Finally, a sensitivity investigation was executed to catch the best noteworthy input parameter during the modeling process. This process was carried out by eliminating the one input parameter and noted the output in terms of RMSE and C.C.
 
Results and Discussion
The obtained results suggest that the Gaussian Process (GP) model with two PUK and RBF kernels is more accurate to estimate the suspended sediment load (SSL) compared to the M5P, ReepTree and Random Forest (RF) models for the given study area. According to the results of the test part of the GP-PUK model, it has given us the best result, which are the correlation coefficient, the root mean square error and the mean absolute error in Bahram Jo station (0.55, 0.42, and 0.27), Cham Anjir station (0.74, 018, and 0.08), Sarab Seyed Ali station (0.71, 016, and 0.07) and Kakareza station (0.71, 0.24, and 0.15), respectively. In general, the Gaussian Process-PUK model, is the powerful model for the prediction of suspended sediment load (SSL) in low and high discharge periods. Therefore, according to the obtained results from this research, these optimal models can be used to costly and time-consuming tasks of the estimation of suspended sediment load from river. Also, these models can be used to estimate the suspended sediments load of nearby rivers by/without hydrometry station for the management of the quantity and quality of surface water. Also, sensitivity analysis suggests that  in Cham Anjir, Bahram Jo and Sarab Syed Ali hydrometry stations and rain in Kaka Reza hydrometry station, are the most significant parameters in estimation/prediction of SSL.
 
Conclusion
The present study focused on the development of a GP-PUK, GP-RBF, M5P, REEP Tree and Random forest (RF) models to estimate the suspended sediment load. For this purpose, the hydrometry and hydroclimatology data of the Bahram Jo, Cham Anjir, Sarab Syed Ali and Kaka Reza stations in Khorramabad, Alashtar and Biranshahr sub-watersheds composed of Suspended Sedimend Load (SSL), discharge, and rainfall data were used. In general, The major conclusions of the study are as follows: Among those models with the highest performance, the GP-PUK has the highest performance in both testing and training phases. -The GP-PUK predicted data are closer to observational data compared with the other model’s output data. Besides, the GP-PUK is the nearest predicted model with observational data.
The GP-PUK model is one of the most extensively used data driven models in the erosion and sediment literature, while the usages of other data-driven models are comparatively lesser. Also, the structure of the GP-PUK is very simple and very less time consumable. Thus, the GP-PUK model can be useful in the Suspended Sediment Load (SSL) modeling not only foraccuracy but also for its time-saving nature and simple structure compared with other models.