Synthetic training data from a virtual light trap enables insect detection for greenhouse pest management without manual annotation

Automated insect identification with deep learning is normally limited by the need to collect and manually annotate thousands of real images, which is slow, error-prone, and impractical for rare or seasonal species. This study indicates that the real-data collection step can be substantially reduced. We build a virtual light trap, a digital twin of a physical greenhouse trap implemented in Blender. From a handful of ordinary 2D reference photographs, it generates large synthetic training sets with bounding-box labels derived automatically from the rendered segmentation mask, with no manual annotation. Insect geometry is reconstructed from the 2D images via an AI-based image-to-3D model and rendered with multi-pass compositing for ground truth that is exact with respect to the rendered geometry and independent of scene lighting, and with targeted domain randomization of the trap's LED illumination. A lightweight YOLO26n detector trained exclusively on this synthetic data, with no real training images of the target domain, already detected real greenhouse insects across four species with a mean average precision (mAP50) of 78.6%, demonstrating genuine simulation-to-reality transfer. Under the stricter mAP50–95 criterion, however, the synthetic-only model reaches 0.367 against 0.814 for real-data training. A separately collected set of real trap images serves as the benchmark: a held-out portion forms the independent test set, and the remainder trains a real-data baseline and hybrid models that mix synthetic and real images, allowing us to assess whether such mixing is beneficial. Synthetic data from the virtual trap can therefore remove most of the real-data collection burden for edge-deployed insect detection, while a small real seed set remains valuable where tight localization or the separation of visually similar species matters.