Optimizing Crop Disease Detection: A Comparative Analysis of Transfer Learning Architectures for Precision Agriculture

Abstract

Crop diseases pose a substantial threat to global food security, yet existing detection methods often suffer from delays and labor-intensive processes. We present a comparative analysis of four transfer learning architectures—VGG16, ResNet50, MobileNetV2, and EfficientNetB0—for automated crop disease classification from leaf images. Our framework leverages a supervised deep learning approach trained on approximately 54,000 labeled images from six crop species (tomato, potato, maize, pepper, grapes, and apple) sourced from the PlantVillage database, covering seven major disease categories. All images were preprocessed to 224×224 pixels and augmented with rotation, flipping, zoom, and brightness adjustments. We employed a consistent architecture comprising a pretrained CNN feature extractor, global average pooling, a fully connected dense layer, and a softmax classifier. The Adam optimizer with a learning rate of 0.001 and categorical cross-entropy loss were used over 50 epochs with a batch size of 32. The dataset was split into 70% training, 15% validation, and 15% testing. Among all models, EfficientNetB0 achieved the highest accuracy of 97.3%, with precision, recall, and F1-score of 97.0%, 97.1%, and 97.0% respectively. ResNet50 attained 95.4% accuracy, MobileNetV2 reached 94.1%, and VGG16 yielded 93.2%. Confusion matrix analysis revealed minor misclassifications primarily among visually similar fungal diseases. Training curves demonstrated stable convergence with minimal overfitting. The superior performance of EfficientNetB0 highlights its suitability for real-time agricultural monitoring. This research contributes a systematic evaluation of transfer learning architectures for precision agriculture, demonstrating that pretrained models can effectively reduce computational complexity while maintaining high diagnostic accuracy.