Modelos de classificação para predição do bem estar de colônias da abelha Apis mellifera

• Ano de defesa: 2020
• Autor(a) principal: Braga, Antonio Rafael
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Teleinformática; Abelha - Criação; Abelha - Classificação; Apis mellifera; Data mining; Clustering
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/51581

Bees are the main pollinators of most species of wild plants pollinated by insects and are essential for the maintenance of plant ecosystems and food production. However, in the past three decades, they have suffered numerous health challenges, including changes in habitat, pollutants, toxins, pests, diseases, and competition for resources. An attempt to mitigate this problem is to estimate the health status of the colonies and indicate a state of imminent collapse for beekeepers. To estimate the health status of bee colonies, we propose three methods of data analysis that calibrate classification and regression algorithms based on supervised and unsupervised machine learning approaches. To validate the first proposed method, a real dataset from two hives obtained from the HiveTool.net portal was used with internal temperature, relative humidity and weight of Apis mellifera beehives. From Calinski-Harabasz index and the k-means algorithm, 6 colony health patterns related to transitions between seasons were found. From the found patterns, three classification algorithms were trained, validated and tested. To validate the second method, a data-set obtained from 6 apiaries was used. In this data-set, 27 Apis mellifera beehives were monitored over three years. Three classification algorithms were trained, validated and tested. In terms of attributes, the internal temperature and the weight of the hive were used, in addition to climatic data (external temperature, dew point, wind direction, wind speed, precipitation, and daylight). Also, 703 in loco apiary inspections carried out weekly were also used to put labels in sensors data. Finally, to validate the third method, the Long Short-Term Memory (LSTM) algorithm was applied to a real data-set obtained through the Arnia remote monitoring system. The data-set has data on brood temperature (internal temperature), internal humidity, average ventilation, average flight noise, hive weight and external temperature collected throughout the European autumn in 2017. The results obtained with the application of the methods suggest that the classification and regression algorithms are efficient to obtain high precision models for predicting colony health levels.