Deep learning for sunflower (Helianthus annuus L.) mapping and stand counting: Trade-offs between closed vs. open-access methods

Sunflower (Helianthus annuus L.) is a strategically important oilseed crop in North Dakota (ND), where accurate mapping of singles and doubles along with stand count is critical for yield assessment. In recent years, researchers have leveraged built-in deep learning (DL) models available within the geographic information system (GIS) platforms, such as ArcGIS Pro or QGIS, to automate mapping and counting tasks. However, these built-in DL models often lack full customization capabilities to effectively adapt to diverse datasets. Therefore, this study presents a comparative evaluation of ArcGIS Pro integrated deep learning tools and an external, GUI-based pipeline developed in Python using the Plotly Dash framework. The use of highresolution unmanned aerial system (UAS) imagery was assessed by initially training and deploying models (SSD and YOLOv3 architectures) within ArcGIS Pro, and outside using You Only Look Once (YOLOv11). In parallel, a custom web-based tool was developed using advanced YOLOv11 medium variants, enhanced by Sliced Aided Hyper Inference (SAHI) for high-resolution image processing. Among these, the YOLOv11m model achieved the best balance between precision (98.9%), recall (88.1%), and inference time, resulting in a mAP@[0.5:0.95] of 0.8 and an F1 score of 0.9 in mapping singles and doubles. The root mean squared error (RMSE) for stand count was 26.7 sunflowers per tile further confirmed the model’s accuracy. In contrast to the models deployed within the GIS environment, while SSD produced full-field mapping, YOLOv3 exhibited scalability limitations when handling large orthomosaics. These results suggests that although closed-loop commercial software such as ArcGIS Pro provides DL features for model training, it still remains limited in adapting to custom, high-resolution, agriculture-centered applications. Furthermore, 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 [ 17 ]