Predictive models for yield and protein content of brown rice using support vector machine

Rice production in Japan is facing problems of yield and quality instability owing to recent climate changes, aging of farmers, and a decrease in the farmer population. Thus, it is becoming important to develop an improved rice production technology that utilizes collected data about rice production rather than relying on the conventional technology that is based on the experience and knowledge of individual farmers. We developed predictive models for yield and protein content of brown rice that can provide useful knowledge to support farmer's management decision-making, utilizing data sets from 47 paddy fields where rice was produced under various environments and management styles. Support vector machines (SVMs) were applied to build the predictive models based on explanatory variables representing the growth and nutrition conditions after the heading stage and the meteorological environment after the late spikelet initiation stage. The models achieved quantitative accuracy that was within approximately 1 t ha−1 in yield for 85.1% of the total data sets and within 0.8% in protein content for 76.6% of the total data sets, respectively. Further, patterns of explanatory variables classified in three classes of yield and protein content, which were visualized by the predictive models, were reasonable in terms of knowledge of crop science. We found that the predictive models using SVMs had the potential to describe a relation between yield or protein content and multiple explanatory variables that reflected diverse rice production in actual fields, and could provide useful knowledge for decision-making of topdressing and basal fertilization.