ResNet50-Based Robust Detection of Apple Leaf Diseases

Rapid and accurate detection of leaf diseases is vital for sustainable apple production, as early intervention can curb pathogen spread and reduce excessive pesticide application. In this study, we leverage the ResNet-50 convolutional neural network to categorize high-resolution apple leaf images into four prevalent health conditions: Healthy, Cedar Apple Rust, Black Rot, and Apple Scab. Our dataset, comprising over 2,000 images gathered under diverse orchard settings, was partitioned using stratified sampling to ensure proportional class representation. During training, we addressed residual class imbalances by incorporating classspecific weighting in the loss function. To mimic real-world variability, we applied on-the-fly data augmentation techniques adjusting brightness, introducing slight zoom levels, and performing random rotations. The model was trained for 50 epochs with an adaptive learning rate scheduler, achieving 99% validation accuracy and demonstrating robust performance across all categories. In addition to quantitative metrics, confusion matrix analysis validated balanced precision and recall scores. These findings underscore the adaptability of residual architectures for plant pathology tasks. Furthermore, our approach offers a scalable framework for integration into diagnostic tools, promising rapid decision support for growers and contributing to precision agriculture initiatives.