Deep Learning Detection of Weeds in Vegetable Fields

  Objective  Deep learning to accurately identify weeds for effective weeding in vegetable fields was investigated.

  Method   Image of a vegetable field was cropped into grid cells as sub-images of vegetables, weeds, and bare ground. Deep learning networks using the ShuffleNet, DenseNet, and ResNet models were applied to distinguish the target sub-images, particularly the areas required weeding. Precision, recall rate, F₁ score, and overall and average accuracy in identifying weeds of the models were evaluated.

  Result  Although all applied models satisfactorily distinguished weeds from vegetables, ShuffleNet could simultaneously deliver a 95.5% precision with 97% recall and a highest detection speed of 68.37 fps suitable for real-time field operations.

  Conclusion   The newly developed method using the ShuffleNet model was feasible for precision weed control in vegetable fields.