Detection of some weeds in wheat production of Tokat region by deep learning

The wheat is one of the important agricultural food sources produced in order to meet the rapidly increasing living population and accordingly the increasing nutritional needs. The presence of weeds is an important biotic factors that cause yield and quality losts in wheat production. It is necessary to determine the locations and types of weeds in order to combat weeds more effectively and to reduce the harmful effects and cost of the herbicides used. Such a sensitive determinations could be possible only with the opinion of an expert but it can be much faster and easier with the success of artificial intelligence algorithms in the recent period. In this study, charlock mustard (Sinapis arvensis L.), creeping thistle (Cirsium arvense (L.) Scop) and forking larkspur (Consolida regalis Gray) plants, which cause significant losses and poisoning in wheat production areas, were determined by deep learning method. 5 different phenelogical periods (cotyledon leaves period, 3-5 leaves period, pre-flowering period, flowering period and fruit and seed setting period) of each plant, which are important for control, were classified separately and 15 different classification processes were carried out. YOLOv5 deep learning architecture was used with a total of 145 792 labelled objects in the images of each class. All neural networks (Nano, Small, Medium, Large and ExLarge) that came with YOLOv5 were trained and the Precision, Recall, F-1 Score and AUC performances of the neural networks were evaluated. It has optained as the most successful neural network was YOLOv5s (small) with 98%, all neural networks showed their highest performance in KG2 (3-5 leaves period of creeping thistle). the lowest performance class of neural networks were with TH5 (fruit and seed setting period of forking larkspur) and YOLOv5s (small) with the highest rate of 45%, while the lowest value was YOLOv5n (nano) and th5 (seed setting period) with a value of 8%. This study is a first in terms of detecting weeds according to different fenelogical stages by computer vision.