Fig. 7
Area under the ROC curve (AUC) for each class across 10 independent training/testing splits using (left) DenseNet-121, (middle) ResNet-101, and (right) ResNet-50, all trained on augmented weed image datasets. All three models achieved excellent AUC values across most classes, with median scores consistently above 0.96. ResNet-101 displayed the most stable and uniformly high discrimination performance, while DenseNet-121 exhibited slightly more variation for the injured (IS) and standard mallow (SM) classes. These results align with the F1 score trends and reinforce the effectiveness of data augmentation in enhancing model generalization and separability.
