Viability of using Convolutional Neural Networks for Real-Time Fescue and Sheep Sorrel Detection in Wild Blueberry Fields

Wild blueberry (Vaccinium angustifolium Ait.) production is hindered by weeds such as fescue (Festuca filiformis Pourr.) and sheep sorrel (Rumex acetosella L.). Commercial sprayers provide a uniform application of herbicide regardless of the incidence of weed coverage. Traditional methods of spot-applying herbicide to target weed locations relied on colour co-occurrence matrices that were limited by long processing times in addition to green colour segmentation lacking the ability to discriminate between weeds and crop canopy of similar colour. Deep learning convolutional neural networks (CNNs) are a modern processing technique which often uses powerful graphics processing units (GPUs) to classify images or detect objects within images. This novel research study featured two object-detection CNNs, YOLOv3 and YOLOv3-Tiny, trained to detect fescue using images of wild blueberry fields captured during application timing intervals in the 2019 growing season. A custom-built desktop computer using the Ubuntu 16.04 operating system and an NVIDIA GeForce RTX 2080 Ti GPU was used to test the CNNs using 1280x720 resolution images. YOLOv3 classified fescue images with an F1-score of 0.95 and YOLOv3-Tiny classified fescue images with an F1-score of 0.97. YOLOv3 and YOLOv3-Tiny classified sheep sorrel images with F1scores of 0.93 and 0.89 respectively. A laptop running Windows 10 Pro with an NVIDIA Quadro RTX 5000 GPU was used to process video streams from four USB cameras simultaneously. YOLOv3 processed each video stream at an average framerate of 4.7 FPS using 9.2 GB of vRAM, while YOLOv3-Tiny processed each video stream at an average framerate of 20.5 FPS using 3.2 GB of vRAM. Initial results suggest that YOLOv3-Tiny can be deployed for use with a machine vision system to detect fescue and sheep sorrel in realtime for spot application of herbicide in wild blueberry fields. Using CNNs to selectively spray herbicide will appreciably reduce the volume of herbicides needed to manage wild blueberry fields, resulting in cost-savings for producers. Keywords—Deep learning; artificial intelligence; machine vision; precision agriculture; weed detection