Abstract: This paper proposed a method of turf weed recognition and accurate spraying based on deep learning. The research targets of this paper are bermudagrass lawn and its associated weeds,including dallisgrass,white clover,and purple nutsedge. By dividing the original image into several grid regions,the convolutional neural network model was used to identify the grid subimages,locate weed position,and determine the precise spraying area of herbicides. In order to explore the effects of different deep learning models on weed recognition,the VGGNet model,the GoogLeNet model,and the ShuffleNet model were selected to compare and analyze in terms of F₁ value,accuracy,and recognition speed. The F₁ value of all models under the verification set was higher than 0.97,indicating that the three models in this study have a good recognition effect on weeds. The GoogLeNet model was the most optimal weed recognition model,with the most balanced recognition rate and recognition speed. Its average accuracy and identification speed in the test set were 98.75% and 36.9 fps,respectively,which can be used in real-time weed recognition applications on turf. The results showed that the turf weed identification and herbicide precision spraying areas detection has highly feasible and excellent application effect,and can be used for turfgrass weed control based on precise herbicide spraying.

Key words: Turf weed, Weed recognition, Deep learning, Site-specific spraying, Image processing

摘要： 本文以狗牙根草坪及其伴生杂草毛花雀稗、白三叶以及莎草为研究对象，提出了一种基于深度学习的草坪杂草识别及除草剂喷施区域检测方法。通过将原图划分为若干格子区域，利用神经网络模型对格子图片进行杂草识别，实现杂草定位并进而确定除草剂喷施区域。为探究不同神经网络模型对杂草识别的效果，选取VGGNet模型、GoogLeNet模型和ShuffleNet模型，分别以F₁值、准确率和识别速度进行对比分析。验证集下所有模型的F₁值都高于0.97，表明本研究中的三个模型对于杂草都有较好的识别效果。其中，GoogLeNet模型为杂草识别最优模型，拥有最为均衡的识别率和识别速度。其在测试集的平均准确率和识别速度分别为98.75%和36.9 fps，能够用于草坪实时杂草识别应用。结果表明，本研究提出的草坪杂草识别与除草剂喷施区域检测方法具有高度的可行性和较优的应用效果，可用于基于除草剂精准喷施的草坪杂草防控。

关键词: 草坪杂草, 杂草识别, 深度学习, 喷施区域, 图像处理