In the heart of North Dakota, where sunflowers stretch towards the sky, a groundbreaking study is reshaping how we assess and maximize crop yields. Maria Villamil-Mahecha, a researcher from the Department of Agricultural and Biosystems Engineering and the School of Natural Resource Sciences at North Dakota State University, has been at the forefront of this agricultural revolution. Her recent work, published in the journal *Inteligencia Artificial en la Agricultura* (Smart Agricultural Technology), delves into the world of deep learning and its potential to transform sunflower mapping and stand counting.
Sunflowers are not just a picturesque part of the landscape; they are a strategically important oilseed crop. Accurate mapping of single and double plants, along with precise stand counting, is crucial for yield assessment. Traditionally, these tasks have been labor-intensive and time-consuming. However, Villamil-Mahecha and her team have leveraged the power of deep learning to automate these processes, potentially saving time and resources for farmers and energy companies alike.
The study compares two approaches: using built-in deep learning models within geographic information system (GIS) platforms like ArcGIS Pro or QGIS, and an external, GUI-based pipeline developed in Python using the Plotly Dash framework. The team used high-resolution imagery captured by unmanned aerial systems (UAS) to train and deploy models based on SSD and YOLOv3 architectures within ArcGIS Pro, and YOLOv11 outside of it.
One of the standout findings was the performance of the YOLOv11m model. “The YOLOv11m model achieved the best balance between precision and recall, with a mean average precision (mAP) of 0.8 and an F1 score of 0.9 in mapping singles and doubles,” Villamil-Mahecha explained. “The root mean squared error (RMSE) for stand count was 26.7 sunflowers per tile, confirming the model’s accuracy.”
In contrast, while SSD produced full-field mapping, YOLOv3 exhibited scalability limitations when handling large orthomosaics. This suggests that although closed-loop commercial software like ArcGIS Pro provides deep learning features for model training, it still remains limited in adapting to custom, high-resolution, agriculture-centered applications.
The implications of this research are significant for the energy sector, particularly for companies involved in biofuel production. Accurate yield assessment can lead to better planning and resource allocation, ultimately improving the efficiency and profitability of biofuel production. As Villamil-Mahecha noted, “Large-scale imagery processing tasks, particularly those involving mapping or counting, are better suited to open-access environments, offering improved scalability, flexibility, and overall performance.”
This study not only highlights the potential of deep learning in agriculture but also underscores the need for flexible, customizable solutions. As we look to the future, the integration of advanced technologies like deep learning and UAS imagery could revolutionize the way we approach crop assessment and management, paving the way for more sustainable and efficient agricultural practices.