Comparison of two classifiers in the identification of tree species based in continuous characters

Abstract

The species name assignment to a vegetable branch is the process called botanical identification and consists in a classification problem. The inclusion of continuous variables in this problem is not new, but it’s a growing topic. The objective of this paper was evaluate two classifiers that assign a specific name to a set of continuous measures of a leaf. There were collected 352 leaves from 5 species of Myrtaceae botanical family. There were measured 5 variables manually: maximum blade width, petiole width, leaf, blade and petiole maximum length. There were utilized two classifiers linear discriminant analysis (LDA) and random forests (RF). The data set was divided into train (70%) and test (30%) and 2000 iterations were conducted to each of 31 possible combination of variables. The models and classifiers were compared by the mean of the successful classification rate in the test set obtained in the 2000 iterations. As a result, considering all variable combinations, the LDA accuracy was 98.2% while the RF classified 96.8% correctly. The best isolated variable was the petiole maximum length and the best combination of two variables was the petiole maximum length and blade maximum width. LDA had a better performance than RF in the greater part of the variables combinations. These findings show the potential that this approach has to contribute as an aid to the botanical identification process.