Deep learning for image-based weed detection in turfgrass

Abstract

Abstract Precision spraying of herbicides can significantly reduce herbicide use. The detection system is the critical component within smart sprayers that is used to detect target weeds and make spraying decisions. In this work, we report several deep convolutional neural network (DCNN) models that are exceptionally accurate at detecting weeds in bermudagrass [Cynodon dactylon (L.) Pers.]. VGGNet achieved high F1 score values (>0.95) and out-performed GoogLeNet for detection of dollar weed (Hydrocotyle spp.), old world diamond-flower (Hedyotis cormybosa L. Lam.), and Florida pusley (Richardia scabra L.) in actively growing bermudagrass. A single VGGNet model reliably detected these summer annual broadleaf weeds in bermudagrass across different mowing heights and surface conditions. DetectNet was the most successful DCNN architecture for detection of annual bluegrass (Poa annua L.) or Poa annua growing with various broadleaf weeds in dormant bermudagrass. DetectNet exhibited an excellent performance for detection of weeds while growing in dormant bermudagrass, with F1 scores >0.99. Based on the high level of performance, we conclude that DCNN-based weed recognition can be an effective decision system in the machine vision subsystem of a precision herbicide applicator for weed control in bermudagrass turfgrasses.