基于机器学习的水稻发育期预测模型构建  

A predicting model based on machine learning for rice development

作　　者：乌玲瑛[1] 徐奂[1] 蔡喨喨[1] 严力蛟[1] 

机构地区：[1]浙江大学生命科学学院/生态规划与景观设计研究所,浙江杭州310058

出　　处：《扬州大学学报（农业与生命科学版）》2012年第3期44-50,共7页Journal of Yangzhou University：Agricultural and Life Science Edition

基　　金：国家自然科学基金资助项目(69673044);国家高技术研究发展计划项目(863-2007AA10Z220)

摘　　要：采用机器学习中的支持向量机(SVM)方法,建立以适应区域尺度生产指导为目的的水稻发育期预测模型。通过整合水稻发育期数据和气象数据,构建训练集与测试集,并应用SVM算法建立针对5个不同发育阶段,应用2种不同样本构建方法的10个发育期预测模型。对其逐一进行评估,最终挑选出具有最佳预测效果的模型作为研究成果。结果表明:采用第1类样本(提前150d的样本)生成策略的5个发育期模型,其预测精度均大于80%,甚至达到95%的水平;而采用第2类样本(提前30d的样本)生成策略的5个发育期模型,其精度普遍在80%左右。与此同时,对这2种样本构建方法分别进行了敏感性及假阳性比较。结果表明:虽前者敏感性高于后者,但其假阳性也高,预测误差在9d左右,而第2类样本的预测误差则能控制在4～5d内,更符合模型构建的要求。采用第2类样本生成策略进行发育期模型的研究可获得更准确的预测结果。This research adopted a machine learning method named support vector machine （SVM） in order to establish prediction model of rice-development, which would guide the rice production in regional-scale. Rice growth data and cli- mate data were used in this research. After dividing the data into training and testing sets in using k-fold cross validation method, SVM was introduced to construct models. We focused on five main growth-nodes in the rice growth period, and used trained SVM to forecast them respectively. In this research we compared two different sample strategies, 150 days in advance versus 30 days in advance, to choose the negative samples for training SVM. The results showed that the prediction accuracy of five developmental models which using the first sample generation strategy （150 days in advance） were all greater than 80%, the best even reached 95%. But the models which adopting the other strategy （30 days in advance）, the prediction accuracy were approximately 80%, and some of them were less than 80%. Comparing the sensitivity and false-positive of the two different kinds of strategies, the results showed that although the first sample models were more sensitive than the second one, the false positive were higher, the forecast error of the first one were 9 days, while the second sample models could control the forecast error within 4--5 days. So the second sample strategy was much more suitable for constructing models.

关 键 词：水稻 发育期模型 支持向量机(SVM) 

分 类 号：S126[农业科学—农业基础科学] S511