A two-stage deep learning strategy for weed identification in grassfields

Abstract

Machine vision strategies for weed identification, whether in industrial crops or grassfields, are fundamental to the development of automated removal systems necessary to increase agricultural yield and field maintenance efficiency. Identifying plant species considered invasive on grassfields is particularly challenging due to reduced color and morphological contrast, as well as phenotypic variability. This work presents a two-stage weed identification strategy using visible spectrum images. The first stage employs a convolutional siamese neural network to identify candidate regions that may contain weeds of irregular or regular morphology. The second stage employs a convolutional neural network to confirm the presence of irregular morphology weeds. The results of each stage are combined to produce an output containing a per-pixel probability of irregular weed and bounding boxes for the morphologically regular weed. The two-stage strategy has an accuracy score of 97.16% and a balanced accuracy score of 89.94% and macro F1 score of 81.14%. In addition to the good performance scores obtained with the proposed approach, it is to be noted that the convolutional Siamese network allows achieving a good performance with a relatively small dataset compared to other strategies that employ data-intensive training phases for optimizing the convolutional neural networks. The results were obtained with a dataset of weeds that has been made publicly available, as well as the neural network models and associated computer code. The dataset contains samples Trifolium repens and Lectuca virosa on grass obtained with two different cameras under varying illumination conditions and different geographic locations. The lightweight nature of the proposed strategy provides a solution amenable to implementation using currently existing embedded computer technology for real-time weed detection.