In the quest for precision agriculture, one of the most pressing challenges is automating the counting of arable crop seedlings. Accurate plant counts are crucial for optimizing resource use, improving yields, and reducing environmental impact. However, the question of how much annotated data is needed to train effective object detection models for this task has remained largely unanswered—until now.
A recent study published in *Remote Sensing* (translated from Italian as “Remote Sensing”) and led by Samuele Bumbaca from the Department of Agricultural, Forest and Food Sciences at the University of Turin, Italy, sheds light on this critical issue. The research systematically evaluated the minimum dataset requirements for fine-tuning object detectors to count maize seedlings in orthomosaic imagery captured by drones.
The study compared traditional deep learning models—such as YOLOv5, YOLOv8, and YOLO11—with newer approaches like CD-ViTO (which requires fewer training examples) and OWLv2 (which requires zero labeled examples). The researchers also included a handcrafted computer graphics algorithm as a baseline. They tested these models with varying training sources, dataset sizes (ranging from 10 to 150 images), and annotation quality levels (10–100%).
The findings were clear: models trained on out-of-distribution data—meaning data that didn’t match the conditions of the target environment—failed to achieve acceptable performance, regardless of dataset size. However, models trained on in-domain data—data that closely matched the conditions of the target environment—reached the benchmark accuracy (R² = 0.85) with as few as 60 to 130 annotated images, depending on the model architecture.
One of the most intriguing results was that transformer-based models, such as RT-DETR, required significantly fewer samples (just 60) compared to CNN-based models (110–130). “Transformer-based models showed a remarkable efficiency in learning from smaller datasets, which could be a game-changer for precision agriculture applications,” Bumbaca explained. However, these models also exhibited different tolerances to reduced annotation quality, maintaining acceptable performance even when only 65–90% of the original annotation quality was retained.
Despite recent advances in few-shot and zero-shot learning, neither approach met the minimum performance requirements for practical deployment in precision agriculture. This underscores the importance of high-quality, in-domain training data for reliable results.
The implications of this research are significant for the energy sector, particularly in the context of bioenergy crops. Accurate plant counting can optimize land use, reduce costs, and improve the sustainability of bioenergy production. As the demand for renewable energy grows, so too will the need for efficient and scalable agricultural technologies. This study provides a roadmap for developing robust maize seedling detection systems, offering practical guidance for researchers and industry professionals alike.
As the field of precision agriculture continues to evolve, these findings highlight the need for tailored solutions that balance model complexity, dataset size, and annotation quality. The future of automated plant counting lies in leveraging the right combination of these factors to achieve reliable, scalable, and cost-effective results.