A Spatiotemporally-Mixed-Runoff-Model-Based Artificial Intelligence Parameter Regionalization Application in Henan Province of China

Abstract

The small-sized catchment in China often characterized with complex topography and geomorphology. And the mountainous small-sized catchments has distributed widely, which often suffered by serious flash flood disasters. Moreover, most of those catchment are mainly ungauged that brings big challenges in flood modelling and forecast. According the problem of low modelling accuracy, this paper presents a new operational approach which integrated the spatiotemporally-mixed runoff model and machine learning algorithms for the parameter regionalization application and further verified with the data collected from 19 small-sized catchments in Henan province. The results shows that by using more physically-based model in combination with principal component analysis could effectively reduce the noise in the machine learning algorithms caused by over fitting problem. The cross-validation method has been applied in this study to evaluate the effects of different parameter regionalization methods such as machine learning algorithms, shortest distance method and behavior similarity method. The machine learning algorithm (CART model) has been proved as the best approach for the parameter regionalization application. The approach presented in this paper shows higher level promising applicability for solving the problem of low modeling accuracy in small-sized ungauged catchments in China.