LIU Changjun, ZHOU Jian, WEN Lei, MA Qiang, GUO Liang, DING Liuqian, SUN Dongya

According to the problems faced in the flood prediction in poor-gauged small-sized catchment, such as the complicated runoff generation mechanism and the difficulty of obtaining model parameters, a spatiotemporally-mixed runoff model and a CART machine learning algorithm are proposed in this paper. Based on the division of geomorphologic and hydrological response units in small watershed and the unsaturated infiltration GARTO calculation model, a spatiotemporally-mixed runoff model is constructed from three aspects of horizontal mixing, vertical mixing and temporal mixing of infiltration excess and saturation excess runoff mechanism, and the model parameter regionalization is studied by using machine learning CART method. The applicability of the model and the regionalization method of parameters were verified by selecting 15 catchments with different morphological properties and 19 small basins in Henan Province. The results show that the average Nash coefficient of spatiotemporally-mixed runoff model is 0.78, 20% higher than that of other models. The average Nash coefficient machine-learning regionalization simulation of Henan 19 basins is 0.70, 35% higher than random transplantation. The model and parameter regionalization method presented in this paper show higher applicability in flood simulation of middle and small-sized mountainous catchments.

Keywords: poor-gauged small-sized catchments,spatio temporally-mixed runoff model,parameter regionalization,machine-learning algorithms