RemoteVision: A Multi-Scale Graph Attention Network and Edge GPU Integrated UAV for Remote Aerial Plant Detection

This study introduces RemoteVision, an advanced real-time AI-driven plant recognition system operating on UAV without requiring post-processing on external computers. RemoteVision uses a novel object detection model called GraphFusionDet, which is built on EfficientDet. It includes a special network called Graph Attention Based Multi-Scale Fusion Network (GAMSFN) that uses a graph structure to improve detection accuracy. GAMSFN assimilates long-range dependencies and facilitates cross-scale feature fusion for the detection of objects at varying altitudes. Unlike traditional methods that analyze aerial films after the flight, RemoteVision uses real-time video streaming during the flight with an NVIDIA Jetson Nano. This setup allows for very low delays and efficient use of resources, making it ideal for precision agriculture. Field trials conducted in potato fields and attained a 98% detection accuracy at 30 FPS, surpassing overall benchmarks established by EfficientDet-D7, YOLOv8, and GraphSAGE in terms of detection accuracy, FPS and inference speed. The GraphFusionDet underwent pruning and quantization to enhance performance and reduce computational overhead while maintaining accuracy. These developments make RemoteVision highly scalable for extensive agricultural monitoring. This research establishes a novel benchmark in AI-driven precision agriculture by achieving real-time responsiveness on UAV-based agricultural platforms, which enhances productivity and sustainability in modern agricultural sectors.