Cassava detection under real field conditions using YOLOv5

Abstract

Plant detection is a critical step in many farm-management tasks. Many object detection models lack robustness for plant detection in real field conditions. Two versions of the deep learningbased YOLOv5 models – YOLOv5n and YOLOv5s – were evaluated and compared for cassava crop detection. Images for model training and validation were collected using an unmanned aerial vehicle (UAV) under varying real-life conditions. The effect of varying input image resolutions on the model performance was also evaluated. YOLOv5s recorded mean average precision (mAP) of up to 0.965 better than 0.947 for YOLOv5n; but at the cost of up to 19.2% increase in training time and over 20% reduction in detection speed. For both models, higher image resolutions yielded better mAP, while the detection speed decreased with increase in image resolution.