九州大学大学院生物資源環境科学府環境農学専攻生産環境科学教育コース生産環境情報学分野
Laboratory of Bioproduction and Environment Information Sciences, Course of Bioproduction Environmental Sciences, Department of Agro-environmental Sciences, Graduate School of Bioresource and Bioenvironmental Sciences, Kyushu University
九州大学大学院農学研究院環境農学部門生産環境科学講座生物生産工学研究分野
Laboratory of Bioproduction Engineering, Division of Bioproduction Environmental Sciences, Department of Agroenvironmental Sciences, Faculty of Agriculture, Kyushu University
九州大学大学院農学研究院環境農学部門生産環境科学講座生物生産工学研究分野
Laboratory of Bioproduction Engineering, Division of Bioproduction Environmental Sciences, Department of Agroenvironmental Sciences, Faculty of Agriculture, Kyushu University
九州大学大学院農学研究院環境農学部門生産環境科学講座生物生産工学研究分野
Laboratory of Bioproduction Engineering, Division of Bioproduction Environmental Sciences, Department of Agroenvironmental Sciences, Faculty of Agriculture, Kyushu University
九州大学大学院農学研究院環境農学部門生産環境科学講座生物生産工学研究分野
Laboratory of Bioproduction Engineering, Division of Bioproduction Environmental Sciences, Department of Agroenvironmental Sciences, Faculty of Agriculture, Kyushu University
We evaluated the support vector machine (SVM), a supervised learning method used for pattern recognition, as a fundamental analysis method for building prediction models of rice yield and quality. First, prediction models of patterns regarding the yield and protein content of brown rice were built using a directed acyclic graph SVM (DAGSVM), which is a multiclass classifier. These models predict patterns of yield and protein content on the basis of seven variables, including the state of rice plant after heading (nitrogen nutrition and yield capacity) and meteorological environment (air temperature and solar radiation) after the late spikelet initiation stage (i.e., 14 days before heading). Data used for building the models were obtained from surveys of 47 paddy fields conducted in 2009 in the village of Hoshino, in 2009 and 2010 at the experimental farm of Kyushu University, and in 2010 in the cities of Itoshima and Fukuoka. The Hinohikari cultivar was grown in all the surveyed fields. Next, the validity of the SVM as an analytical method for building prediction models was evaluated in terms of the classification rate and an adjustment function of generalization. The classification rates of yield and protein content were found to be relatively high, i.e., 85.1% and 76.6%, respectively. Further, it was confirmed that an adjustment function of generalization with soft margin was effective in training the data that were not linearly separated. The results indicated that SVM, with high accuracy and high generalization performance, is an effective method for building prediction models of rice yield and quality that are affected by various factors.