Abstract
<title>Abstract</title> <p>Uneven occurrence of potato leaf diseases necessitates site-specific treatment rather than conventional uniform spraying. This study presents a vision-guided autonomous variable-rate spraying system integrating real-time deep learning-based disease detection, temporal disease severity estimation, and PWM-controlled precision spray actuation for targeted crop protection. A composite dataset comprising 1,561 field images containing 14,166 annotated leaf instances was developed under dense canopy conditions. Comparative evaluation of object detection architectures showed that YOLOv26 achieved the highest detection performance with a peak mAP@0.5 of 0.961, outperforming YOLOv8 (0.924) by 3.7% and YOLOv12 (0.938) by 2.3%. The lightweight YOLOv26n variant was selected for embedded deployment. When deployed on a Raspberry Pi 5, the optimized ONNX model achieved an inference speed of approximately 10 FPS, corresponding to a spatial sampling interval of 0.018 m at an operating speed of 0.18 m/s. A 2 s temporal sliding window generated a stable canopy-level disease severity index, which was mapped to PWM duty cycles for variable-rate pesticide application. Field experiments involving 598 spatial spray decision instances validated the proposed perception-driven spray control framework. Compared with conventional constant-rate spraying, the system reduced agrochemical consumption by 32.75% (from 458 to 308 L/ha). These results demonstrate the feasibility of integrating embedded deep learning and adaptive spray control into a field-deployable autonomous spraying platform for precision potato disease management.</p>