The detection of weeds with computer vision without the help of an expert is important for scientific studies and other purposes. The images used for the detection of weeds are recorded under controlled conditions and used in image processing-deep learning methods. In this study, the images of 3-4-leaf (true-leaf) periods of the wild mustard (Sinapis arvensis) plant, which is the critical process for chemical control, were recorded from its natural environment by a drone. The datasets were included 50-100-250-500 and 1 000 raw images and were augmented by image preprocessing methods. Totally 12 different augmentation methods used and datasets were examined for understand how to affects the numbers of images on training-validation performance. YOLOv5 was used as a deep learning method and results of the datasets were evaluated with the Confusion Matrix, Metrics-Precision, and Train-Object Loss. For results of Confusion Matrix where 1 000 images gave the highest results with TP (True Positive) 80% and FP (False Positive) 20%. The TP-FP ratios of 500, 250, 100 and 50 image numbers were respectively; 65%-35%, 43%-57%, 0%-100% and 0%-100%. With 100 and 50 images, the system did not show any TP success. The highest metrics- precision ratio was found 92.52% for 1 000 images set and for 500 and 250 image sets respectively; 88.34% and 79.87%. The 100 and 50 images datasets did not show any metrics-precision ratio. The minimum object loss ratio was 5% at 50th epochs in the 100 images dataset. This dataset was followed by other 50, 250, 500, and 1 000 images respectively; 5.4%, 6.14%, 6.16%, and 8.07%.