Relationship between sample size and classification accuracy for tree species identification based on leaf shape

Abstract

Yasushi Minowa and Tsukumo Nakanishi: Relationship between sample size and classification accuracy for tree species identification based on leaf shape. Jpn. J. For. Plann. 55: 61~75, 2022　The aim of this study was to verify the relationship between sample size and classification accuracy for tree species identification based on leaf shape. In this study, 6,516 leaves （20 species, 250-350 leaves of each tree species） served as samples. The following leaf-shape parameters were measured: circularity, ratio of minor axis to major axis for fitting optimal elliptical, ratio of perimeter to optimal elliptical orbit, and convexity. Two decision-tree algorithms （J48, random forest） and a neural network （multilayer perceptron） were used for machine-learning classification. A performance evaluation of the proposed model was performed using the Matthews correlation coefficient （MCC）. The estimation method of the necessary sample size for measuring each leaf-shape parameter used the design method of sample size with power analysis. The MCC of the training data ranged from 0.755 to 1.000, and that of the test data ranged from 0.737 to 0.744. Every machine learning model resulted in high classification accuracy for training and test data, respectively. Almost all models showed a relatively high classification accuracy when the number of tree species was small; however, the classification accuracy tended to decrease with increasing number of tree species. When the sample sizes were small, the classification accuracy of each machine learning model was low. Even if the sample sizes were larger, the classification accuracy did not necessarily improve for any machine learning model. The results of the estimation for sample size for each leaf-shape parameter suggested that the sample size in this study was insufficient to identify tree species from leaf shape.