The emergence rate of seedlings is an early season measurement that plant breeders consider when selecting varieties. B. carinata and B. napus, important crops in North America, have low emergence rates, and so these crops are densely seeded, which makes measuring emergence challenging. Current methods of automating emergence counting require human annotations to locate individual plants in images. However, state-of-the-art object detection models still struggle to detect individual plants in densely planted clumps with significant overlap and occlusion between plants. One solution to this issue is to acquire yet larger amounts of training images, increasing the time and expense involved in data annotation. As an alternative, we propose a new approach for generating synthetic data that is representative of clumps of densely overlapping plants. Using microplot images of B. carinata, we select 330 isolated, individual plants to generate composite training images containing one to 15 overlapping plants. 75,000 synthetic images were used to extract morphological features used to train various regression models such as Linear Regression, Random Forests and a three-layer Convolutional Neural Network. We show that the models trained on synthetic images achieve higher correlation with ground-truth plant counts from the field compared to our baseline model trained with real image data alone. We compared the counting models and tested them on 680 real microplot images of B. carinata.