A real-time kinematic global navigation satellite system (RTK-GNSS) mounted on a drone was used for the generation of a relatively accurate orthophoto and a digital elevation model (DEM) with a minimum number of ground control points (GCPs) in a large-scale onion field. Subsequently, deep learning was applied to detect onion bulbs in the orthophotos. Consequently, 16,812 onion bulbs were annotated using 250 images for object detection. After the machine learning process, the detection rate was 0.60, whereas the recall was 0.16. When the trained model was applied to the orthophotos, the number of onion bulbs ranged from 29,970 to 109,694, which was 5–19 % of the estimated number of onion plants. When the size of the annotation bounding box was expanded to include the surrounding image of the onion bulb, the detection rate was improved to 41–47 %. In the future, our objective will be to improve the trained model by reducing the false-negative value in the confusion matrix.
View full abstract