A COMPARATIVE ANALYSIS OF CNN-BASED APPROACHES FOR PLANT LEAF DISEASE DETECTION

Abstract

Automated plant disease detection has emerged as a vital component of precision agriculture, offering the potential to improve crop yield while reducing reliance on manual inspection. This study presents a comparative evaluation of convolutional neural networks (CNNs) for vegetable disease classification. A dataset of over 20,000 images, encompassing 15 disease categories and healthy samples, was employed to assess both a custom CNN model and transfer learning architectures.

Four models were analyzed: a custom-designed CNN, VGG19, ResNet50, and Xception. The custom CNN achieved an accuracy of 87.50%, demonstrating the viability of lightweight models for resource-constrained applications. VGG19, leveraging transfer learning, improved performance to 89.52%, while ResNet50 achieved the highest accuracy of 94.86%, alongside strong precision, recall, and F1 scores, confirming its robustness and suitability for real-world deployment. In contrast, Xception underperformed, underscoring the importance of model selection and fine-tuning in plant disease recognition tasks.

The findings emphasize the effectiveness of transfer learning, particularly with deep architectures like ResNet50, in delivering accurate and reliable disease detection. At the same time, the custom CNN offers a practical alternative for environments with limited computational resources. This study provides a foundation for future exploration into hybrid and ensemble models, deployment on edge devices, and techniques to address class imbalance for enhanced model robustness and scalability.