Transferencia de Aprendizaje Hacia el Modelo ResNet-50 para la Clasificación de Especies de Aves como Churrín Negruzco, Soterrey Cola Pálida y Soterrey Montés de Pecho Gris

The automatic recognition of bird vocalizations through deep learning is an important tool for monitoring and conservation of species, especially in fragile ecosystems where traditional methods, such as manual identification by experts, are expensive and not scalable. The objective of this work was to determinate the overall percentage of accuracy when applying transfer learning to the ResNet-50 model to classify the vocalizations of three Ecuadorian bird species: Blackish Churrin, Plain-tailed Wren, and Grey-breasted Wood Wren. Following the CRISP-ML(Q) methodology, three main phases were developed for the first objective. In the first phase, an interview with Machine Learning expert, Engineer Oscar Cumbicus was conducted as an initial context to guide the development of the project and define its technical feasibility. In the second phase, 1,164 recordings were collected from the Xeno-canto platform, which were preprocessed, normalized, and transformed into spectrograms. In addition, data augmentation techniques were applied to improve representativeness. In the third phase, the ResNet-50 model was tuned by transfer learning, optimizing hyperparameters and implementing callbacks to improve performance and prevent overfitting. For the second objective, the ML model evaluation phase was executed, in which the model achieved an accuracy of 91% in the training set and 89.33% in the test set. Its performance was validated with external audios through a web interface that allowed its use by non-specialized users, achieving an overall hit rate of 91.19%. These results not only demonstrate the robustness of the proposed approach, but also establish a benchmark in bioacustic classification, offering new perspectives for the development of advanced tools in the study and conservation of biodiversity.

• Autor principal: Cajamarca Escaleras, Jimmy Alexander (author)
• Formato: bachelorThesis
• Idioma: spa
• Publicado em: 2025
• Assuntos: TRANSFERENCIA DE APRENDIZAJE; APRENDIZAJE AUTOMATICO
• Acesso em linha: https://dspace.unl.edu.ec/jspui/handle/123456789/32063