Locating Open-field Broccoli Plants with Unmanned Aerial Vehicle Photogrammetry and Object Detection Algorithm: A Practical Prediction Approach

Abstract

We developed a practical approach to locate individual open-field broccoli plants with a position error of less than 5 cm, using the georeferenced high-resolution orthomosaic imagery generated through the unmanned aerial vehicle-based photogrammetry and the YOLOv5 object detection model. The feasibility of our method was evaluated on the basis of two angles: the cost of preparing training data and the accuracy of object detection. The orthomosaic imagery was generated for two plots: Plot A, which experienced large variations in plant growth due to drought-induced mortality and replanting, and Plot B, which showed small variations under normal growing conditions. On the basis of the result of analysis under our recommended settings for the training data, we found that (1) the detection accuracy improved with an increase in the amount of training data in both Plots A and B; (2) in Plot A, 95% of a total of 21277 plants were detected using training data for approximately 630 plants selected to represent the individual differences in growth; and (3) in Plot B, 98% of all 7836 plants were detected using training data for approximately 126 plants selected randomly. Our findings can guide the optimal balance between the cost of training data preparation and the desired accuracy level of object detection in precision crop management, particularly for broccoli production.