Modeling winter wheat response to water in North China with feed-forward neural networks

The model of crop response to water describes the quantitative relationship between crop yield and water input, and is essential for the rational regulation of field water regime and the improvement of water use efficiency. A multi-layer feed-forward neural network (MFNN) was used to simulate the crop response to water for winter wheat in Xiaohe irrigation district in North China. The MFNN was trained with field experiment results using the Hybrid Algorithm of genetic algorithms and back-propagation algorithm. It was found that the MFNN is capable to describe wheat yield response to water well when using suitable parameters and training algorithms, while the over-fitting of the MFNN can be improved by decreasing the number of hidden nodes and introducing calibration samples. The simulation results indicate that the yield of winter wheat is sensitive to water stress during three mid-growing stages. Moderate water stress in these three stages has little influence on the yield, and thresholds of moderate water stress for these three stages can be used in irrigation scheduling.