Apple leaf disease identification with high precision is one of the main challenges in precision agriculture. The datasets usually have class imbalance problems and environmental changes, which negatively impact deep learning approaches. In this paper, an ablation study is proposed to test three different scenarios: V1, a hybrid balanced dataset consisting of 10,028 images; V2, an imbalanced dataset as a baseline consisting of 14,582 original images; and V3, a 3× physical augmentation approach based on the 14,582 images. The classification performance of YOLOv11x was benchmarked against three state-of-the-art CNN architectures: ResNet-152, DenseNet-201, and EfficientNet-B1. The methodology incorporates controlled downsampling for dominant classes alongside scenario-based augmentation for minority classes, utilizing CLAHE-based texture enhancement, illumination simulation, and sensor noise generation. All the models were trained for up to 100 epochs under identical experimental conditions, with early stopping based on validation performance and an 80/20 train-validation split. The experimental results demonstrate that the impact of balancing strategies is model-dependent and does not universally improve performance. This highlights the importance of aligning data balancing strategies with architectural characteristics rather than applying uniform resampling approaches. YOLOv11x achieved its peak accuracy of 99.18% within the V3 configuration, marking a +0.62% improvement over the V2 baseline (99.01%). In contrast, EfficientNet-B1 reached its optimal performance in the V2 configuration (98.43%) without additional intervention. While all the models exhibited consistently high AUC values (≥99.94%), DenseNet-201 achieved the highest value (99.97%) across both V2 and V3 configurations. In fine-grained discrimination, the superior performance of YOLOv11x on challenging cases is verified, with only one incorrect classification (Rust to Scab), while ResNet-152 and DenseNet-201 incorrectly classified eight and seven samples, respectively. Degradation sensitivity analysis under controlled Gaussian noise and motion blur indicated that CNN baseline models maintained stable performance. High minority-class reliability, including a 96.20% F1-score for Grey Spot and 100% precision for Mosaic, further demonstrates effective fine-grained discrimination. Results indicate that data preservation with physically inspired augmentation (V3) is better than resampling-based balancing (V1), especially in terms of global accuracy and minority-class performance.
Fidan et al. (Mon,) studied this question.