Caracterização da diversidade de abelhas vibradoras a partir de propriedades acústicas de voo e visitas florais, usando inteligência artificial

Abstract

The correct classification of species is fundamental in ecological and evolutionary studies, especially in tropical regions with high diversity. This study investigated the classification of 11 bee species using machine learning techniques applied to an acoustic database with flight and buzz (floral visitation) vibrations of bees. The algorithms used were Random Forest, Support Vector Machine, and Perceptron, with Random Forest standing out by achieving an accuracy of 78.95%. These results demonstrate the robustness of the model, even in scenarios with increasing diversity. The fundamental frequency and the MFCC coefficients were essential for classification, along with the type of vibration performed by the bees. The confusion matrix of the Random Forest model showed that classes of similar sizes were more often confused with each other, and the model’s accuracy curve demonstrated that, although there was a decrease in test accuracy as the diversity index of the communities increased, this decrease did not exceed 10%. The model's achieved performance presents high potential for the development of automatic tools capable of classifying species in diverse communities, contributing to conservation and monitoring in regions with great diversity, where precise classification is essential.