Classifying Weed Development Stages Using Deep Learning Methods

The control of harmful weeds holds a significant place in the cultivation of agricultural products. A crucial criterion in this control process is identifying the development stages of the weeds. The technique to be used is determined based on the weed's growth stage. This study addresses the application of deep learning methods in classifying growth stages using images of various weed species to predict their development periods. Four different weed species, obtained from seeds collected in TurkeyAfyonkarahisar-Sinanpa & scedil;a Plain, were used in the study. The images were captured with a Nikon D7000 camera equipped with three different lenses, and the ROI extraction was performed using Lifex software. Using these ROI images, deep learning models such as DenseNet, EfficientNet, GoogleNet, Xception, and SqueezeNet were evaluated. Performance metrics including accuracy, F1 score, precision, and recall were employed. In the 4class dataset with ROI annotations, DenseNet and Xception achieved an accuracy of 86.57%, while EfficientNet demonstrated the highest performance with an accuracy of 89.55%. Following the initial tests, it was concluded that classes 3 and 4 exhibited extreme similarity caused most of the prediction errors. Merging the said classes significantly increased the accuracy and F1 scores across all models. In image classification tests, SqueezeNet and GoogleNet demonstrated the shortest processing times. However, while EfficientNet lagged slightly behind these models in terms of speed, it exhibited superior accuracy. In conclusion, although the use of ROI improved classification performance, class merging strategies resulted in a more significant performance enhancement.