A fine tuned EfficientNet-B0 convolutional neural network for accurate and efficient classification of apple leaf...

Precise classification and detection of apple diseases are essential for efficient crop management and maximizing yield. This paper presents a fine-tuned EfficientNet-B0 convolutional neural network (CNN) for the automated classification of apple leaf diseases. The model builds upon a pre-trained EfficientNet-B0 base, enhanced through architectural modifications such as the integration of a global max pooling (GMP) layer, dropout, regularization, and full-model fine-tuning. To address class imbalance and improve generalization, the study adopts a holistic training strategy that integrates data augmentation, stratified data splitting, and class weighting, alongside transfer learning. The model is evaluated on the PlantVillage (PV) dataset and a curated Apple PV (APV) dataset and compared against EfficientNet-B0, EfficientNet-B3, Inception-v3, ResNet50, and VGG16 models. The fine-tuned model demonstrates outstanding test accuracies of 99.69% and 99.78% for classifying plant diseases using the APV and PV datasets, respectively. The fine-tuned model outperforms EfficientNet-B0, EfficientNet-B3, and VGG16 on both datasets and shows superior performance compared to Inception-v3 and ResNet-50 on the PV dataset. Both EfficientNet-B0 and the fine-tuned model demonstrate the lowest memory consumption and floating-point operations per second (FLOPs). Also, as compared to the EfficientNet-B0 model, the fine-tuned model achieves an 11% increase in accuracy on the APV dataset and a 49.5% accuracy improvement on the PV dataset, with approximately a 7-8% increase in both memory usage and FLOPs. The fine-tuned model thus emerges as an effective solution for plant leaf disease classification, delivering outstanding accuracy with optimized memory consumption and FLOPs, making it suitable for resource-constrained environments. This study demonstrates that fine-tuned CNN approaches, when combined with transfer learning, advanced data pre-processing, and architectural optimizations, can significantly enhance the accuracy of diseased leaf classification in crops with efficient implementation in limited-resource settings.