Heavy-load agricultural UAVs operating at low altitude over farmland often encounter three major difficulties: unreliable recognition of distant small obstacles, unstable localization of elongated targets, and strong interference from cluttered backgrounds. To cope with these challenges, this work introduces a lightweight real-time obstacle detection framework by redesigning YOLOv8n for farmland scenes. In the backbone, SPD-Conv, retained high-resolution \(\:{P}_{2}\)features, and the DGA-C2f module are jointly used to preserve fine-grained cues for small and slender obstacles. In the feature aggregation stage, a lightweight scale-difference fusion network is constructed, where the LSDF module is embedded into a bidirectional interaction scheme to strengthen cross-level feature collaboration. In the prediction stage, a direction-aware decoupled head is adopted so that orientation modeling can assist localization and improve the regression quality of elongated targets. Experiments on a self-built farmland obstacle dataset show that the resulting model reaches 89.3% Precision, 88.0% Recall, 91.8% mAP@0.5, and 85.0% mAP@0.5:0.95. Compared with YOLOv8n, the proposed model improves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 by 1.5, 1.8, 2.4, and 2.3% points, respectively. It also maintains real-time inference performance with appropriate parameters, 8.9 GFLOPs, and 123 FPS, demonstrating a clear balance among detection accuracy, lightweight complexity, and deployment efficiency for farmland obstacle perception.
