The purpose of this article is to attempt classification of leaf images using a self-organizing map (SOM) and tree-based model. The number of samples was 420 (84 species), which were collected at the Kyoto Prefectural University campus. Data input into the models were used as 10 dimensions: circularity, ratio of minor axis to major axis, four capacity dimensions and four information dimensions. Both the capacity and information dimensions were calculated from the distance feature, which was calculated as the distance from the center of gravity in the figure to the circumference and was shown as the function of an angle, using the fractal dimensions of states with ε-entropy. As a classification method, SOM illustrated the 10 dimensions data on a two-dimensional plane as a nonlinear map. Moreover, the classification accuracy of decision trees derived from tree-based models was examined. Additionally, the samples were divided into five groups based on differences in leaf shape, as follows: simple leaves with leaf teeth, simple leaves without leaf teeth, lobed leaves, needle leaves and compound leaves. It was found that: (1) The fractal dimension showed different values, and this dimension was found to be effective as a factor for estimating and classifying; (2) when the number of leaning times and map sizes of SOM increases, all tree species were clearly classified on SOM map; and (3) as for classification by tree-based models, the correct ratios of each model varied widely, ranging from 42.1% (REPTree) to 100.0% (RandomTree) without cross-validation, and ensemble learning can improve the estimation accuracy of the models.
抄録全体を表示