Deep Learning-Based Classification of Apple Leaf Diseases under Field Conditions

บทคัดย่อ

Accurate identification of apple leaf diseases in field conditions is essential for sustaining crop yield and supporting precision agriculture. Variable illumination, cluttered backgrounds, and co-occurring symptoms complicate diagnosis in real orchards. This study applies a deep learning approach using a fine-tuned MobileNetV2 model to classify apple leaf diseases from a heterogeneous dataset derived from the Plant Pathology 2021 (FGVC8) benchmark. The original five labels were expanded by subdividing the "multiple disease" category into expert-defined compound subclasses, yielding 12 disease categories encompassing both single and compound infections. Data augmentation and transfer learning were employed to improve robustness, while interpretability was assessed through Grad-CAM and LIME visualizations. Results show that the model performs well on distinct single-disease categories such as rust, scab, and frogeye leaf spot, but struggles to detect overlapping or compound infections. These findings highlight both the potential and the challenges of lightweight CNN architectures for agricultural image classification. The study contributes evidence that explainable, compact deep learning models can support future efforts to build reliable tools for plant health monitoring in diverse field conditions.