Runoff simulation of small watershed in loess hilly region using dynamic parameter SCS-RF model

Rainfall and runoff events are two important parameters in the natural hydrological cycle. The rainfall also dominates the formation of runoff in many influence factors. In recent years, various human activities, such as the construction of projects for water conservancy, as well as conservation projects for soil sources, have posed a great impact on the soil infiltration and topography of basin, which further affected the evolution of surface runoff. Heavy rain and runoff can cause a series of natural disasters, such as water erosion and flood damage. Sediment loss from construction sites and soil erosion has become a serious source of water pollution in the loess hilly and gully area. The rapid situation can deeply deteriorate the water and soil resources, further to threaten the safety of flood control. Fortunately, Soil Conservation Service (SCS) model can be used to evaluate the impact of rainfall on runoff yield. The improved model was established to consider the impact of other rainfall characteristics in the supposed conditions and internal structure with the parameter calibration. However, the accuracy of runoff simulation depends mainly on the selection of model parameters, particularly on the regional characteristics of parameters. In this study, a dynamic parameter SCS-Random Forest (RF) model was constructed, according to the dataset from 307 times rainfall runoff in Wangjiagou basin and its sub basins in flood season. The rainfall characteristics were taken as splitting attributes of a decision tree, while the RF was used to determine the Curve Number (CN) and initial abstraction ratio in the model parameters. Various parameters were be calibrated, according to the rainfall characteristics of a same rainfall, and then compared with the SCS model without parameter improvement. The results showed that the Root Mean Square Error (RMSE) of SCS-RF model and SCS model were 1.06 and 6.64, while the Nash-Sutcliffe Efficiency (NSE) were 0.84 and -8.65, respectively. Moreover, the SCS-RF model achieved an excellent performance in each basin, where the simulation effect of SCS-RF model was better than that of SCS model. The SCS-RF model also considered the influence of rainfall characteristics on runoff yield. The process of parameter calibration was simplified, thereby to enhance the universality of the model. The different treatment in the basin made the distribution of parameters significantly different from others. There was small fluctuation for the CN in Yangdaogou basin, and the initial abstraction ratio in the Chacaizhugou basin. The initial abstraction ratios were less than the standard value of 0.2, after the control of soil erosion. The distributions of CN were approximate, comparing with the two basins that were comprehensively treated but with different areas. Specifically, there was a large area fluctuation for the CN and initial abstraction ratio in the Wangjiagou basin, where the CN in Wangjiagou basin was larger than that of Chacaizhugou basin. The compared area was similar, where the Yangdaogou basin did not be treated, while the Chacaizhugou basin was treated. The initial abstraction ratio was small in the basin of Chacaizhugou, but fluctuated greatly in Yangdaogou basin, whereas, the CN fluctuation was small in Yangdaogou basin. It infers that the distribution of initial abstraction ratio was similar. There was a negative correlation between CN and rainfall P in each basin. The data distribution of parameters and rainfall characteristics (P, rainfall intensity in 30 min I30) had a relatively obvious concentration area. The rainfall-runoff simulation can provide a theoretical basis for the conservation planning of soil and water, as well as the management of water resources.