Dryland Farmland In Northern China Based On Improved YOLOv3 Research And Application Of Weed Detection

Shanxi region is the main grain-producing areas,northern dryland crop growth environment is relatively poor,farmland weed has become one of the key factors that affect the growth of crops in order to realize the organic dry production under the condition of green ecological prevention and control of weeds,to understand the common weeds in northern dry land of the categories of information such as proportion,weeds in the field.It is of great significance to detect,quantitatively evaluate the ecology and identify the dominant characteristics of weeds in dryland farmland.In the organic dry farming system of Shanxi Province,10 kinds of main weeds,which grow and develop simultaneously,such as corn,millet,sorghum and soybean,were taken as the research objects.In the early stage of crop growth,especially before crop growth and ridge sealing,the image data of weeds were collected,sorted out and labeled as a whole and part.Faster R-CNN,SSD(Single Shot Multi Box Detector),Efficientdet,YOLOv3 and the improved YOLOv3 algorithm were compared and evaluated for the detection effect of 10 weeds.And on the basis of improving YOLOv3 algorithm,we try to quantitatively identify many kinds of weeds.The main results and conclusions of this paper are as follows:(1)Constructing the weed dataset of dryland farming in northern China.Ten common weeds in northern dryland at different growth periods were collected.In order to ensure the detection quality,the weeds were labeled as a whole and labeled with local characteristics,and 6333 weed labeled samples were finally formed.(2)Determine the YOLOv3 model for improvement experiment.Faster R-CNN,SSD(Single Shot Multi Box Detector),Efficientdet and YOLOv3 were tested and compared on the dataset.As a whole,the detection results of the four models for these ten weeds showed relatively high accuracy(AP)of the four weeds(Amaranthus vulgaris,Amaranthus fovea,Quinoa and Thistle),which were all reflected in the YOLOv3 model.The average accuracy(MAP)of the four models were 61.26,70.11,66.49 and 76.83,respectively.The performance of the model on different crops was tested to further clarify the advantages of YOLOv3 in this experiment.(3)Weed detection based on the improved YOLOv3 algorithm.The results of the improved YOLOv3 for the detection of Scylla sinensis,Amaranthus indica,Thickle,Echinochloa palmifera,Scylla sinensis,Humulus scandens,Solanum sinensis,purslane and quinoa were 95%,92%,97%,87%,72%,80%,63%,68%,64%,89%,respectively,and the MAP result was 80.92%.Compared with the original YOLOv3 model,it is improved by 4.09%.And explore the local characteristics of weed detection.At the same time,the calculation of weed leaf coverage was extended.By improving YOLOv3 model to detect weeds in dryland,this method can be applied to organic dryland production system in Shanxi Province,which has a certain promoting effect.