Benchmarking one-stage object detectors for field-based detection of Flavescence dorée vectors

Deep learning-based object detection has emerged as a promising paradigm for automating high-throughput insect monitoring in agricultural environments. In this context, we propose a computer vision framework for automated detection of Flavescence dor´ee (FD) vector insects from yellow sticky trap (YST) imagery, addressing the scalability limitations of manual entomological inspection. Flavescence dor´ee is a severe phytosanitary threat to viticulture in Europe, primarily transmitted by the leafhopper Scaphoideus titanus, with Orientus ishidae increasingly recognized as a secondary vector due to its role in phytoplasma transmission dynamics. Current monitoring strategies rely on field-deployed YSTs followed by manual expert analysis, which is labor-intensive, time-consuming, and difficult to scale to large vineyard networks. To address these limitations, a custom dataset of images acquired from field-deployed YSTs using a fixed- distance laboratory imaging setup was developed and used to fine-tune and benchmark multiple modern one- stage object detection architectures. Repeated training runs showed generally stable performance, with average mAP50 values above 0.93 across all configurations. For the best-performing configuration, the highest performing run achieved mAP50 =0.965, mAP50-95 =0.690, precision =0.949, and recall =0.914. Edge benchmarking further quantified deployment feasibility, with YOLO26n reaching 11.41 FPS on Raspberry Pi 5. Class-wise evaluation confirmed accurate detection of both S. titanus and O. ishidae, despite heterogeneous trap backgrounds and illumination conditions. These image-level results support further evaluation of automated FD- vector monitoring, while trap-level independence, whole-trap counting, and fully operational field acquisition remain to be validated. The annotated dataset and benchmark provide a basis for further research on data- efficient agricultural computer vision systems.