关键词: 水稻, 始穗期, 主成分分析, 随机森林算法, 误差反向传播神经网络算法

Abstract:

The initial panicle date of rice is a critical period during which disease−causing bacteria, such as rice smut, rice blast disease and sheath blight, invade organs like grains, leading to reduce the rice quality and lower yield In order to improve the ability of rice disease control and better meet the requirements of "prevention first, comprehensive control, green damage control, drug reduction and efficiency", taking Jiangsu as an example, using historical meteorological data and rice growth period observation data, based on the analysis of the characteristics and key influencing factors of the beginning panicle period, four groups of simulation schemes were set up by using principal component analysis, error back propagation neural network algorithm and random forest algorithm to establish the prediction model of the beginning panicle period of rice. In addition, the coefficient of determination and root mean square error were taken as the evaluation indices to analyze and evaluate the accuracy and compatibility of the model. The results showed that: the initial panicle stages of rice in northern Jiangsu, central Jiangsu and southern Jiangsu were concentrated from August 4 to 31, August 9 to September 18 and August 16 to September 20. The average standard deviation of each region was 4d, 6d and 5d respectively. The key factors affecting the initial panicle stage of rice in Jiangsu were basically the same. The days sequence of the three growth stages before the initial panicle stage of rice was the most critical. The temperature factors was significantly more important than the precipitation and sunshine hours factors during three growth stages, from sowing to tillering, from tillering to jointing and from jointing to booting. Compared with the model based on RF algorithm, the model based on BP algorithm had higher simulation accuracy, and had better "acceptability" for the predictors with weakened correlations after PCA treatment. In addition, based on the proposed method, the simulation prediction error of rice initial panicle stage of rice epidemic in all regions of Jiangsu was within 2d, with a prediction advance of about 10d, which can provide a reference basis for accurately capturing the key period of rice disease control. 

Key words: Rice, Initial panicle stage, Principal component analysis, Random forest algorithm, Neural network algorithm