Leveraging Deep Learning to Identify Jute Leaf Diseases: A Pathway to Improved Agricultural Practices

This study presents a robust deep learning framework for automated detection of jute leaf diseases, focusing on Cercospora leaf spot, Golden Mosaic disease, and healthy leaves. Leveraging a Kaggle-sourced dataset of 920 images , we implement preprocessing steps including Gaussian normalization, resizing, and augmentation to enhance generalization. Four architectures are evaluated: a lightweight Custom CNN, ResNet50, VGG16, and a novel ensemble of MobileNetV2 and InceptionV3. Transfer learning, hyperparameter optimization , and regularization are applied to refine performance. The ensemble model achieves 99.89% accuracy, significantly outperforming standalone models (VGG16: 96.01%, Custom CNN: 70.13%, ResNet50: 54.31%), albeit with higher computational demands. Saliency maps validate model interpretability, highlighting disease-specific features like lesions and mosaic patterns. Results emphasize the trade-off between accuracy and resource efficiency, positioning VGG16 as a balanced choice for real-time deployment. This framework advances precision agriculture by enabling early disease diagnosis while addressing scalability and transparency through explainable AI.