On The Identification of Some Species of the Cassieae Tribe Harvested in Cameroon Using Three Machine Learning Technics

Species from the Cassieae tribe are widely used as ornamental, medicinal and food plants despite their apparent similarities. In this paper, we study identification of these species by means of the description of their characteristics and by using three machine learning methods (Decision Tree, k-Nearest Neighbors and Support Vector Machine). For that, we collect, in the cities of Douala and Yaoundé in Cameroon, a set of 390 specimens (13 species and 30 per specie) and we describe each of them based on 24 variables (23 features variables and one target variable given the name of the specie). These algorithms are implemented on the obtained database by simple cross validation and 10-folds cross-validation, the performance of each of them was evaluated by means of four indicators: the error rate/accuracy of the model, the sensitivity, the specificity and the Area under the ROC curve (AUC). The minimum accuracy is 95.4% obtained with 10-folds cross-validation. These algorithms perform better on the balanced dataset than on the unbalanced dataset except for SVM which performs better on the unbalanced dataset than on the balanced dataset in 10-folds cross-validation (99.74% vs 99.48%).

Keywords:Species Identification; Decision tree learning; k-Nearest Neighbors; Support Vector Machine; Multiclass Classification.