RDD-SPA: An efficient visual recognition algorithm for robot-assisted rapeseed pest control systems

Terraced rapeseed fields present persistent challenges for pest control owing to complex terrain, low mechanization, and the limited computing capability of mobile spraying robots. To address these constraints, this study proposes RDD-SPA, a lightweight perception–actuation framework that integrates a symptom-oriented object detector (RDD-YOLO) with a Spraying Point Adaptor (SPA) for real-time precision spraying in terraced environments. A field dataset comprising 3309 images collected at the seedling and bud stages was constructed, containing 13,530 annotated instances of rapeseed plants, leaf holes, and leaf discoloration, and reflecting dense canopies, irregular leaf morphology, and frequent occlusions. RDD-YOLO enhances YOLO11 through the introduction of Lite-PSConv, C3k2-IDC modules, and the SIoU loss, achieving an mAP 50 : 95 of 64.0% while reducing parameters and computational cost by 12.02% and 3.02%, respectively. After pruning and knowledge distillation, the model is compressed to 1.12 million parameters and maintains stable real-time inference at 60 FPS on the AX650N embedded platform under a practical robot speed of 0.18 m/s. Based on binocular depth estimation and geometric calibration, SPA translates detection results into adaptive spraying commands, dynamically adjusting nozzle position, angle, and spray volume. Field experiments on 40 rapeseed seedlings demonstrate that the proposed system increases leaf liquid retention by 33.3% and reduces chemical drift by 27.1% compared with manual spraying. These results indicate that RDD-SPA provides an effective and deployable solution for precision pesticide application in resource-constrained terraced agriculture. • A lightweight perception–actuation framework (RDD-SPA) is proposed for precision spraying in terraced rapeseed fields. • A customized YOLO-based detector (RDD-YOLO) improves small and rotated leaf symptom detection under real-time constraints. • Model compression enables stable ≥ 60 FPS inference on an embedded AX650N platform at a practical robot speed of 0.18 m/s. • A spraying point adaptor converts visual detections into spray coordinates, angles, and dosage in real time. • Controlled experiments demonstrate improved leaf retention and reduced chemical drift compared with manual spraying.