Vegetation detection and discrimination within vegetable plasticulture row-middles using a convolutional neural network

Weed control between plastic covered, raised beds in Florida vegetable crops relies predominantly on herbicides. Broadcast applications of post-emergence herbicides are unnecessary due to the general patchy distribution of weed populations. Development of precision herbicide sprayers to apply herbicides where weeds occur would result in input reductions. The objective of the study was to test a state-of-the-art object detection convolutional neural network, You Only Look Once 3 (YOLOV3), to detect vegetation both indiscriminately (1-class network) and to detect and discriminate three classes of vegetation commonly found within Florida vegetable plasticulture row-middles (3-class network). Vegetation was discriminated into three categories: broadleaves, sedges and grasses. The 3-class network (Fscore = 0.95) outperformed the 1-class network (Fscore = 0.93) in overall vegetation detection. The increase in target variability when combining classes increased and potentially negated benefits from pooling classes into a single target (and increasing the available data per class). The 3-class network Fscores for grasses, sedges and broadleaves were 0.96, 0.96 and 0.93 respectively. Recall was the limiting factor for all classes. With consideration to how much of the plant was identified (broadleaves and grasses), the 3-class network (Fscore = 0.93) outperformed the 1-class network (Fscore = 0.79). The 1-class network struggled to detect grassy weed species (recall = 0.59). Use of YOLOV3 as an object detector for discrimination of vegetation classes is a feasible option for incorporation into precision applicators.