Classificação de pássaros por meio de cantos e chamados: uma abordagem integrada com espectrogramas em conjunto com descritores handcrafted e transfer learning

Abstract

The classification of bird species is an important task in environmental monitoring and conservation. This study aimed to develop a machine-learning model to classify bird species based on their songs and calls. Initially, spectrograms were generated for each recording available in the database. Subsequently, a strategy of dividing these spectrograms into smaller segments, called patches, was adopted to investigate the influence of this approach on the model’s performance. Various feature extraction techniques were explored, including handcrafted descriptors such as LBP and Gabor Filters, and descriptors derived from Transfer Learning models, such as VGG16, ResNet50, DenseNet121, and MobileNet. This process used sets with two, three, five, and ten classes to analyze the outcome behavior as this subset gradually increased. Additionally, experiments using the SVM classifier were conducted to evaluate the accuracy and effectiveness of the model. The results revealed that the descriptors based on Transfer Learning showed superior performance, with particular emphasis on ResNet50, which achieved an F1 score of 0.8994 in classifying two birds with the spectrogram divided into three patches. Other models, such as DenseNet121 and MobileNet, also showed competitive results in the same configuration with two birds: DenseNet121 achieved an F1 score of 0.8992 using just one patch. In contrast, MobileNet scored 0.8989 with three patches. These findings suggest that dividing the spectrograms into patches, combined with advanced feature extraction techniques, can effectively classify bird species based on their songs and calls. The results also show that classification performance drops sharply as the number of species increases.