Padrões Bioacústicos como identificadores precisos da presença de abelha rainha em colmeias de abelhas com e sem ferrão

• Ano de defesa: 2023
• Autor(a) principal: Rodrigues, Ícaro de Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Processamento de áudio; Aprendizado de máquina; Abelhas com e sem ferrão; Ciência de dados; Audio processing; Machine learning; Stingless and honey bees; Data science; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/76442

It is estimated that 35% of agriculture crops and almost 90% of wild flower plants depend on pollinating agents. Among the pollinating insects bees are the main and most important. However about 40% of the world’s bee species are dying. To monitor their colonies and avoid risks to their well-being, beekeepers do periodic checks traditionally done manually and invasively by opening the hive. A risk that can lead to the collapse and even the death of the colony is the loss of the queen bee. A non-invasive way of remote, real-time monitoring is through microphones and audio sensors that detect the colony’s bioacoustic patterns (e.g. wing flapping, vibrations). With that in mind, in this work we performed experiments with the audio produced by two different bee species: africanized honey bees (sting) and jataí bees (stingless) to identify the presence/absence of a queen. Both species with widely presence in Brazil. The first experiment has a real-time monitoring approach and datastreams and therefore incremental classifiers were applied for performance evaluation and validation of this approach. The incremental classifiers used were Naive Bayes, Hoeffding Tree and Adaptive Random Forest. In the second experiment, in batch classifiers were used to classify among multiple hives to determine which ones have queen presence and which ones have queen absence. The classifiers in batch used were Multilayer Perceptron, Extreme Learning Machine and AdaBoost. in both experiments, the classifiers underwent performance evaluation comparing the metrics of execution time and accuracy. As result, the Naive Bayes classifier obtained the best performance for the case of incremental classifiers and datastreams. Furthermore we conclude that using 10 windows of 1 second for daily audio sampling is enough to detect the queen presence and is much less expensive computationally. For the case of in batch classifiers, MLP obtained a better performance. Finally, the colony real-time monitoring was also more accurate and precise than a multiple cross-colony approach using data batches.