Training neural networks for plant phenotyping requires substantial labeled data, often a limiting factor due to time-consuming manual annotation. This study demonstrates that an instance segmentation neural network for barley seed morphology phenotyping can be effectively trained using a purely synthetic dataset generated through domain randomization. The trained model achieved 96% recall and 95% average precision on a real-world test dataset and showed effectiveness across various crops (rice, lettuce, oat, wheat). This synthetic data approach significantly reduces human labor costs in deep learning-based agricultural analysis.
