Identification of tree species using a machine learning algorithm based on leaf shape and venation pattern
2019
Minowa, Y. (Kyoto Prefectural University, Kyoto (Japan). Graduate School of Life and Environmental Sciences) | Takami, R. | Suguri, M. | Ashida, M. | Yoshimura, Y.
The aim of this study is to identify tree species based on the leaf shape and venation pattern using a machine-learning algorithm. A total of 380 leaves (38 species) collected at the Kyoto University Campus served as samples in this study. The following leaf shape parameters were measured: Circularity, ratio of minor axis to major axis, and ratio of perimeter to optimal elliptical orbit. The venation patterns were evaluated by histograms of oriented gradients (HOG). The fractal dimensions for the entire leaf and parts of the leaf were also calculated. Two decision-tree algorithms (J48, RandomForest) and a neural network (MultilayerPerceptron) were used for machine-learning classification. A performance evaluation of the proposed model was performed with the correct ratio, the ratio of correct predictions to the total number of predictions. The classification accuracy for unknown data was verified by the 10-fold cross-validation method or test data. Verification for under/over fitting was proposed the ensemble-learning algorithm. The obtained results of the classification accuracy for the training data indicated that only the fractal dimension model showed a low correct ratio. In contrast, the leaf shape model showed the highest correct ratio; moreover, its classification performance could be improved by adding the fractal dimension. Overall, the venation pattern model showed the highest correct ratio. Regarding classification accuracy for unknown data, the correct ratio of the leaf shape with fractal dimension model ranged from 65.3% (J48) to 78.8% (RandomForest), and that of the venation pattern model ranged from 12.5% (J48) to 43.9% (MultilayerPerceptron, learning iterations were 500).
اظهر المزيد [+] اقل [-]