Analysis on the percolation from root zone of winter wheat: Combination of a numerical model and BP Artificial Neural Network

摘要

This study attempts to analyze the percolation from root zone of winter wheat combining a numerical model and BP Artificial Neural Network(BP-ANN).Based on field experimental data during winter wheat seasons from 2007 to 2009 in Shijing Irrigation Scheme in Hebei province in China, a numerical model was developed to analyze the percolation, employing Hydrus-1d soft package. A BP Artificial Neural Network (BP-ANN) regression model for percolation was then designed on the basis of comparisons of different algorithms and a variety of hidden unit numbers. Sample data for BP-ANN establishment were obtained by simulations in the numerical model of eight hundred and sixty-four input scenarios which were formed by combining five percolation-influenced factors (initial 2m-soil storage saturation, the date and amount of 1st Spring-Irrigation, 2nd Spring-Irrigation volume and time intervals between two Spring-Irrigation activities). Still, the specified BP-ANN was also compared to a multiple liner regression model (MLRM). The results showed :( 1) induced by two Spring-Irrigation activities, a double humps-shape in percolation process was formed, which overlapped with each other. Cumulative percolation in the whole growing season of winter wheat was up to almost 30% of the sum of irrigation and precipitation due to flood irrigation, (2) Levenberg-Marquardt(LM) algorithm was better than both the traditional and improved BP ones in this study. The most appropriate hidden unit number is 10 and the best structure for BP-ANN was 5-10-1, which showed a higher computational accuracy than multiple liner regress model (MLRM).