Uso de aprendizado de máquina na borda para um sistema não invasivo de monitoramento de colmeias

Abstract

Climate change poses significant threats to bee populations, making continuous monitoring of hive activity essential for both apiculture management and ecological research. This work presents an innovative and non-invasive bee monitoring system that combines machine learning and computer vision techniques for automated bee counting at hive entrances. The system employs the YOLOv8n object detection model, implemented on an embedded Raspberry Pi Zero 2W platform, enabling local processing without the need for cloud transmission. The performance evaluation demonstrated a Mean Average Precision (mAP50) of 0.987, with precision of 0.91, recall of 0.96, and an F1-score of 0.94 in bee detection tasks. Validation against the gold standard, through Pearson correlation analysis, revealed a strong agreement (𝑟 = 0.93) between automated and manual counts. The Bland-Altman analysis showed a small mean bias of 3.4 bees, with 95% limits of agreement within ±1.96 SD. The automated approach eliminates the need for manual observation, reduces labor costs and colony disturbance, while maintaining high measurement accuracy, offering significant advantages for precision apiculture and ecological monitoring initiatives.