应用神经网络和统计模型预测大豆生长发育阶段

Prediction of Soybean Growth and Development Stages Using Artificial Neural Network and Statistical Models

预测大豆的物候期对指导大豆生产、安排农事活动等具有重大意义。本研究构建了一种简单有效的预测模型,使大豆种植者能够较准确地预测大豆各生育阶段的具体日期。试验地位于美国密西西比Delta研究推广中心(经度:90°55'W,纬度:33°25'N)。试验进行了5年,前4年(1998—2001)数据构建模型,第5年(2002)的数据验证模型。为简化模型,杂草、病虫害、干旱等干扰因素被优化排除。采用逐步回归(SR)、神经网络(ANN)以及内插法构建模型。营养(V-stage)和生殖(R-stage)生长阶段分别建模。结果表明,通过播种日期(PD)和从播种到某阶段的相对平均天数能很准确地预测营养生长各阶段的具体日期;可通过播种日期和成熟期组数值(MG)准确预测生殖生长各阶段的具体日期。3种方法中,神经网络所构建的模型准确度最高,具有较好的推广应用价值。

The prediction of soybean phenology is important in many aspects of soybean production. The study objective was to develop predictive models, using a simple and effective modeling technique, which can allow producers to predict soybean growth and development stages in their fields. The experiments were conducted at the Delta Research and Extension Center in Stoneville, Mississippi （latitude： 33025＇ N, longitude： 90055＇ W） under irrigated conditions. The models were constructed using four-year field data （1998 to 2001）, and validated with the fifth-year data （2002）. Potential factors affecting stages of soybean growth and development were considered for developing the models. Affecting factors, such as weeds, insects, diseases, and drought stress, were controlled optimally to simplify the modeling procedures. In addition, stepwise regression （SR） analysis, artificial neural networks （ANN）, and interpolation approaches were used to construct the models. The modeling of soybean growth and develop- ment processes was separated into two distinct periods： vegetative growth stage （V-stage） and reproductive growth stage （R-stage） The models included ten V-stages （up to V8） and eight R-stages. In the V-stage models, PD （planting date） and mean relative time-span from planting to a particular stage were the only significant parameters, whereas in R-stage models, PD and MG （maturity group） were significant. The models obtained accurate predictions when only using PD, MG. and mean relative time-span from planting to a particular stage. The ANN method provided the greatest accuracy in predicting phenological events, indicating that the ANN method can be effectively applied in crop modeling.