Abstract
Recently, identification (discriminant) models based on neural network are being introduced in several agricultural researches. In this study, we discussed the method to select the supervisor data set from whole data to perceptron neural network for the evaluation of soybean plant shape. Because a perceptron neural network is trained to respond to only a supervisor data set, the general identification (discriminant) efficiency of the neural network for non-supervisor data strongly depends on the method to select the supervisor data. Though the method to select the supervisor data takes one of the most important roles to develop neural network models, it has not been fully established especially in such a case as soybean plant shape evaluation where the features (variables) to identify the shape are not clear and the distributions of those features are hardly known. In this study, we applied neural network model to evaluation of soybean plant shape for the substitution of human visual judgments and examined several supervisor data sets to find the method to select the most effective supervisor data set. The results of the examinations indicate two strategies to select the supervisor data set. The first that the distribution of target output in the supervisor data set is not biased and the second is that the supervisor data set contains the largest and smallest values for each component of input vectors. The neural network trained to the supervisor data set selected based on those strategies showed the identification (discriminant) efficiency 20% higher than that given by the neural network trained to ordinarily selected data sets.
