Walnut-YOLO: a lightweight and occlusion-robust object detection model for UAV-based walnut image-level counting

Abstract Accurate walnut object detection is the key to achieving automated yield estimation and harvesting in orchards. However, the high-density clustering and severe occlusion of walnuts in natural environments cause traditional lightweight models to be highly prone to missed detections, while large networks are difficult to deploy on edge computing devices with limited computational power. To this end, this study proposes Walnut- you only look once (YOLO), an efficient and lightweight detection model with strong occlusion robustness. Using YOLO26n as the baseline, the model first introduces the state space model (Mamba) to acquire global contextual information, compensating for local feature loss caused by occlusion with an acceptable computational overhead. Second, it integrates the bidirectional feature pyramid network to enhance multi-scale feature interaction, effectively distinguishing the targets from severe foreground–background color confusion and low-contrast green backgrounds. Finally, it employs the dynamic focusing mechanism of the wise-intersection over union (WIoU) loss function to optimize localization capabilities for severely occluded samples. Validated on a secondarily refined annotated unmanned aerial vehicle remote sensing walnut dataset, Walnut-YOLO achieves a mean average precision (mAP@0.5) of 0.787 while maintaining low parameters (2.86 M) and computational overhead (5.7 GFLOPs), significantly outperforming the baseline model by 2.4 percentage points. Particularly in the severely occluded (complex) category, the model successfully reduces the missed detection rate (i.e. false negative rate) to 17%. Furthermore, in the walnut counting regression analysis for image-level target counting demonstration, the model achieves a coefficient of determination ( R 2 ) of 0.882 and a root mean square error of 7.49. This study effectively overcomes the target detection challenges in complex orchard scenarios, providing a high-precision, deployable visual perception foundation for automated image-level counting in smart agriculture.