Precision Agriculture Сельское хозяйство и биологические науки (общие)
A UAV-based real-time detection system of flowering nectar plants using a lightweight deep learning model
Аннотация
By integrating UAV-based imaging, accelerated deep learning, and edge computing, this study developed a real-time UAV-edge prototype for nectar flower detection in Rhus chinensis , a nectar plant, under complex mountainous field conditions. The proposed system enables efficient and timely monitoring of nectar plant flowering, providing a foundation for the development of practical decision-support tools for precision apiculture. By facilitating rapid flowering assessment during short blooming periods and supporting timely management decisions, this approach contributes to improved spatial-temporal management and more efficient resource allocation in precision agriculture. First UAV-based real-time counting system for mountainous nectar flowers. Propose an improved high-precision and lightweight DL model for UAV edge computing platform. Achieved 89.5% mAP in complex environment, a 3.4% improvement over original YOLOv5s. Provides a potential prototype for flowering assessment to support timely decision-making in smart apiculture. Purpose Timely knowledge of flowering status and peak blooming periods in transferred areas is crucial for migratory beekeepers to make informed management decisions. Traditional manual field surveys are labor-intensive, time-consuming, and unsuitable for frequent monitoring. Current monitoring approaches using Unmanned Aerial Vehicles (UAV) rely on offline processing and require complex data analysis by specialists, making them impractical for beekeepers to implement in apicultural settings. Given the short time window for capturing flowering events, especially for species with short florescence periods, an accurate, fast monitoring solution is needed. We developed a UAV-based real-time investigation prototype leveraging edge-intelligence computing for nectar flower detection of Rhus chinensis in complex mountainous field conditions. Methods An improved, high-precision, and lightweight deep learning model based on YOLOv5s was proposed. The model was optimized with acceleration strategies during deployment to reduce inference time. Results Experimental results showed that the proposed model for flower detection achieved a mean Average Precision (mAP) of 89.5 and an inference speed of 56 frames per second (FPS). After applying acceleration strategies during model deployment on the edge computing device, the processing throughput reached 192 FPS with a mAP of 89.2; a 3.1% increase in mAP and approximately a 5.6 times improvement in FPS compared to the original YOLOv5s. Conclusions These results demonstrate that the proposed lightweight model and acceleration strategies effectively balance high accuracy with high-throughput onboard image processing. The developed UAV-edge computing system provides a promising prototype for future efficient nectar flowering status monitoring, with potential applications in smart apiculture.