Stochastic modeling of sediment yield using random forest and quantile regression

Background and objectives: Assessment of suspended sediment load is very important. Water quality and environmental is under impression of sediment load. As well as the design of hydraulic structures and other water supply facilities, watershed management, soil conservation programs and another major problem caused by sediment yield is dependent on the accurate estimation of sediment load. As a direct estimation of sediment load is very difficult and time consuming, so this led the researchers to estimate sediment load as indirect that it is possible to resort to various methods. One easy way to indirectly estimate the sediment load is sediment rating curve. It can only represent invariable amount of sediment in flow and due to various factors in nature may be there is several sediment load for a known flow rate. On the basis of this study quantile regression and random forest methods was used that can estimate sediment load for a given flow rate in the various probability. The use of these two methods can be analyzed sediment load in great flood and special events.
Materials and methods: In this study, sediment rating curve models, quantile regression and random forest was used to estimate sediment load in four stations Gorganrood River Jangaldeh, Nodeh, Arazkoose and Ghazaghli in Golestan province. For this purpose, flow and sediment data was collected at four studied stations and separated into two parts, 75% for training and 25% for testing. The rating curve was obtained using fitted power equation between discharge and sediment load. Quantile regression and random forest algorithms were implemented using R statistical software. The optimal values of the variable parameters of the two methods were determined using trial and error method. By running the model, the amounts of sediment associated with specified flow were calculated in different probability level (1% to 99%).
Results: Using these two methods, sediment load was estimated in quantiles 2.5, 50 and 97.5%, respectively and range of uncertainty was determined in each station. In Jangaldeh and Nodeh stations random forest were selected as best method with RMSE criterion 96 and 210 tons per day and quantile regression were selected as best method with RMSE criterion 6453 and 24886 tons per day in Arazkoose and Ghazaghli stations. Classic rating curve method estimate sediment load in Jangaldeh, Nodeh, Arazkoose and Ghazaghli stations with RMSE 199, 288, 7505 and 25811 tons per day respectively.
Conclusion: The results showed that classic sediment rating curve not only unable to estimate the sediment load in the range of uncertainties in specified flow rate but also estimates sediment load with more error. Quantile regression and random forest methods can be estimate sediment load in various probabilities for a specified flow and this has contributed greatly to accurate and comprehensive planning for the construction of hydraulic structures and in this way, the dangers of the destruction of the facility reduction due to the great flood.