Abstract Real‐time unmanned aerial vehicle (UAV)‐based detection plays a critical role in precision agriculture. Current convolutional neural network‐based models and related model compression and deployment techniques cannot satisfy the real‐time early diagnosis of plant diseases with high accuracy. This paper proposes FeatherTeaNet, an ultra‐lightweight object detection framework designed for edge devices. By combining channel width compression with a novel Wing Shuffle module, the model reduces parameters by 88.6% (7.02 M→0.80 M) while preserving multi‐scale feature interactions. The inner intersection over union mechanism enhances sub‐pixel localization under occlusion, and hierarchical pruning further compresses the model to 0.14 M parameters with minimal accuracy loss. Deployed on an NVIDIA Jetson Orin NX, FeatherTeaNet achieves 55.9 frames per second with a 67.1% reduction in floating point operations (15.8G→5.2G), demonstrating strong applicability to real‐time, large‐scale crop disease monitoring in resource‐constrained UAV systems.
Yao et al. (Sun,) studied this question.