VBP-YOLO-prune: Robust apple detection under variable weather via feature-adaptive fusion and efficient YOLO pruning

Apple-picking robots are increasingly applied in smart agriculture, but their performance is limited by complex orchard conditions such as unstable lighting, occlusion, and weather variations. This study proposes an optimized lightweight detection model, VBP-YOLO-prune, based on YOLOv8n, to enhance detection accuracy and deployment efficiency on edge devices. The model incorporates a V7 downsampling module, BiFPN feature fusion, and an improved PIOUv2 loss function, aiming to improve multi-scale representation and bounding box regression. A custom apple dataset was augmented with diverse lighting and weather conditions to improve generalization. Experimental results using 10-fold cross-validation show that VBP-YOLO-prune achieves 89.0 % mAP50 and 66.26 % mAP50–95, with Precision of 84.01 % and Recall of 80.52 %. Additionally, it reduces parameters by 79.7 %, FLOPs by 60.9 %, and increases FPS by 29.2 % compared to YOLOv8n.The final model contains only 0.61 M parameters, 3.2 GFLOPs, and runs at 102.6 FPS on NVIDIA Jetson Orin Nano. These results demonstrate that VBP-YOLO-prune provides a practical and efficient solution for real-time fruit detection in complex environments. Future research may extend this approach to other crop types and explore full integration into autonomous harvesting systems.