Using BP Network to Estimate Initial Soil Water Storage in Xin’anjiang Model

Abstract

【Objective】 Xin’anjiang model is a hydrological model widely used for catchment modelling, but it application needs to know the initial soil moisture storage. Such initial soil moistures were traditionally estimated using the empirical reduction coefficient method which has some shortcomings, and this paper aims to present an alternative method to improve the estimate of this initial soil moisture storage when applying the model to humid and semi humid areas. 【Method】 The proposed method is based on the BP neural network and calculates the initial soil moisture storage using some easy-to-measure factors that are thought to affect moisture distribution in soil. 【Result】 Using three input factors, when the number of nodes in the hidden layer was more than 11, the accuracy index of the BP network model reached first-class level during the training stage. Of the nine samples used in the test of the model, six met the required criterion. It was also found that when the number of the nodes in the hidden layer varied between 4 and 21, the Nash-Sutcliffe efficiency coefficient in the model evaluation increased from 0.51 to 0.97, with the associated root mean square errors decreasing from 11.77 to 2.74. Compared with the traditional empirical reduction coefficient method, the BP neural network model is superior in resolving the constraints in the former, including that it needs a rainstorm or a long drought to start the calculation and that the calculation needs to be continuous in time. 【Conclusion】 The BP neural network model proposed in this paper is feasible to calculate the initial soil water storage when applying the Xin’anjiang model to humid and semi humid areas. It can accurately estimate the initial soil water storage in a catchment if the number of input factors and the nodes in the hidden layers are rationally selected.

Keywords

• initial soil water storage
• bp neural network
• xin’anjiang model
• runoff simulation