A Comparative Study of Deep Learning Approaches for Apple Leaf Disease Classification

Apple production is declining due to diseases, particularly those affecting the leaves. These diseases spread rapidly, making accurate identification crucial for early control. This study compares the performance of four deep learning classification models for classifying apple leaf diseases. The datasets used are from three publicly available datasets that consist of different leaf conditions, diseased and healthy. The results of the experiment demonstrated that InceptionV3 achieved the best results across all datasets. The InceptionV3 had 98.89% accuracy on Apple Leaf Disease Symptoms dataset and 96.67% accuracy on the AppleLeaf9 dataset as well as 82.08% on the Kashmir dataset. Contrastively, InceptionV3 achieved the best accuracy ($82.08 \%$) with the Kashmir dataset compared to other models, ResNet50 at $80.00 \%$, EfficientNetB0 at $73.96 \%$, and MobileNetV2 at $66.04 \%$. Although the InceptionV3 is highly accurate, the computational power required for training is high, making it a challenge to consider especially if running on a less capable device. All datasets are publicly available, links are provided in the paper to support reproducibility. These findings provide a rigorous baseline and practical guidance for selecting architectures for orchard scale monitoring and motivate future work on lightweight models.