Convolutional Neural Networks for Real-Time Weed Identification in Wild Blueberry Production

Abstract

Yield-limiting weeds in wild blueberry fields, including hair fescue and sheep sorrel, are traditionally managed with uniform applications of herbicides. Spot applications of herbicides reduce the volume required for management. Convolutional Neural Networks (CNNs) were trained to identify hair fescue and sheep sorrel in images of wild blueberry fields. Six CNNs identified targets with a minimum F1-score of 0.95 for hair fescue and 0.89 for sheep sorrel. Two CNNs were selected as viable for controlling applications from an eight-camera smart sprayer based on processing speeds above 9 frames per second and memory use below 6.4 GB. A graphical user interface was developed for monitoring CNNs and controlling hardware in real-time based on identification of target weeds. The results of this study indicate that CNNs are suitable for identifying hair fescue and sheep sorrel. Future research will involve using the output of the CNNs to automate spray applications, limiting herbicide use.