Classificação hierárquica de espécies de pássaros utilizando seus cantos

Abstract

This work aims to present an analytical comparison between the hierarchical classification approach and the plan classification approach, in the automatic bird species classification task. Furthermore, this work analyzed the effect of a segmentation pre-processing technique in the pursuit of the proposed task for both approaches. To conduct the experiments, it was used a database from the CLEF 2014 challenge, with 501 distinct species distributed in 9688 audio files with one meta-data XML file per audio containing information. This database was cloned and rearranged respecting a hierarchy of Family, Genus and Species by using MatLab® scripts. In the experiments it was used four classifiers (J48, KNN, Naïve Bayes and SVM) distributed in three comparative scenarios with one more general comparison. The results suggest that the hierarchical classification approach has better results than the plan classification approach only for the SVM classifier. Moreover, experiments combining the hierarchical classification approach with the segmentation preprocessing technique obtained a superior performance; however, when compared with flat classification using the segmentation preprocessing technique the results only were positive in one of the four classifiers.