Suspended sediment simulation using machine learning algorithms and CHIRPS satellite precipitation data with emphasis on data clustering and gamma test, case study: Ramyan Watershed, Golestan Province

The cycle of soil erosion (including removal, transport and deposition) that controls the sedimentation of watersheds, includes a set of complex and highly nonlinear processes. On the other hand, the factors influencing sedimentation in watersheds are very diverse, and according to the specific conditions of climate, soil, vegetation, geology, topography, etc., in each basin, the weight and role of each of the mentioned factors in sediment production is very different. Accurately determining and measuring these factors and making mathematical relationships between them are often difficult, expensive, time-consuming and error-prone, and this is the case with the use of models based on computational intelligence and the use of a limited number of basin dynamic variables, it is possible to simulate the behavior of the watershed in sediment production. Regardless of the type of intelligent models, in most of the conducted research (especially in internal research), the simulation of suspended sediment is mainly based on the discharge variable and the role of variables such as precipitation (especially precipitation obtained from satellite images), which are effective in the sedimentation of basins, have received less attention. In addition to precipitation, the skewness of sediment measurement data is also one of the issues that lack of recognition and attention will reduce the efficiency of estimator models. In the present study, the role of variable daily rainfall (taken from CHIRPS satellite) in the simulation of suspended sediment of Qarachai River has been investigated.

Multi-layer perceptron artificial neural network was used in order to simulate the daily suspended sediment concentration of Qarachai River (at the Ramian hydrometer station in Golestan province). In this regard, the variables of discharge and previous discarge (in instantaneous and daily scales) as well as the average daily and previous rainfall of the basin (taken from CHIRPS satellite) for a statistical period of 37 years (1980-2017) as variables model input was used. In order to increase the generalization power of the models, self-organized mapping neural network (for data clustering) and gamma test was used to find the best combination of input variables. In order to improve the efficiency of network training, a variety of activation and loss functions as well as the overfitting prevention algorithm were used. In order to investigate the effect of using activation and loss functions in suspended sediment estimation, different scenarios were considered, which led to the construction of 9 models. After that, using validation indicators, the effectiveness of the models in simulating suspended sediment was investigated and compared, and then the best model was selected.

The results obtained from the present research showed that among the different models, the neural network model with Huber's activation function and ReLU loss function, having the average absolute value of the error equal to 368 mg/l, the root mean square error equal to 597 mg per liter, the Nash-Sutcliffe coefficient of 0.87 and the percent bias -2.2% were selected as the best model. The results also showed that the use of the rainfall variable (as one of the important factors in causing erosion and sediment transfer in the basin) has improved the efficiency of the models, therefore, considering the ease of using CHIRPS satellite rainfall data, it is suggested in order to simulate the suspended sediment of rivers, this data is also used along with other predictive variables.

In the simulation of suspended sediment, discharge variable is often used as the only predicting variable of suspended sediment, while in basins with rainy, or rainy-snow regimes, the role of precipitation in the production of surface runoff and soil erosion is very important and plays an important role in the production and transport of sediment in the basin. In this regard, although the use of rainfall data obtained from ground rain gauge stations has played an effective role in increasing the efficiency of data-based models in estimating suspended sediment, however, the preparation of hundreds of spatial distribution layers of daily rainfall from the data point data of ground stations, the use of this variable in the simulation of the suspended sediment of the basin has been faced with many problems (such as the lack or inappropriateness of the spatial distribution of rain gauge stations, statistical deficiencies, the use of inappropriate interpolation methods and time-consuming calculations). Therefore, in practice, the variable of river flow is often used as a predictor of sediment, and precipitation is used less often. One of the solutions to the problem mentioned in the present study is the use of CHIRPS satellite data, which was investigated for the first time in this study. These data, available since 1981, can easily be used to simulate suspended sediment or other applications related to watersheds. Another important point that needs to be taken into account in the simulation of suspended sediment is the presence of high skewness in sediment measurement data (both suspended sediment and flow rate), which lack of attention in the process of training (or recalibration) and testing the models leads to It will lead to the construction of weak models in terms of efficiency and the existence of uncertainty in the accuracy of their results. In this regard, it is necessary to use logarithmic transformations or suitable functions of activation and loss in the training process, which in this research, two functions, ReLU and Huber, were proposed respectively. Another important point is to pay attention to the generalization power of data-based models, which is largely dependent on the data used in their calibration or training process. These data should be selected in such a way that while they are representative of the data in the entire statistical period, they are similar and have the same distribution with other data sets (such as cross-validation or test sets). According to the results obtained from the present research and in order to increase the efficiency of artificial neural network models in estimating the suspended sediment of watershed hydrometric stations, it is suggested to use the experiences obtained in this research in other sediment measuring stations of the country.