Efficient Jute Disease Classification Using Hybrid Deep Learning Model

Abstract

Jute, being one of the most important economic crops in South Asia, is greatly impacted by leaf diseases such as Cercosporin Leaf Spot and Golden Mosaic, which traditional diagnosis tools are not well-equipped to address. While deep learning has revolutionized plant disease detection, existing frameworks for jute are hindered by small datasets, computational inefficiency, and unfiled applicability. In this paper, we propose an efficient deep learning architecture for autonomous jute disease classification, integrating novel preprocessing, best transfer learning, and real-time deployment. We introduce a robust dataset of 12,000 high-resolution images in three categories (healthy, Cercosporin Leaf Spot, Golden Mosaic), class-aware augmented with data sparsity and imbalance alleviation. Our method employs a hybrid preprocessing pipeline of wavelet-based denoising (Daubechies-4) and adaptive color normalization to disentangle leaf regions and make use of discriminative features. Using a ResNetRS50 model with transfer learning fine-tuning, we obtain state-ofthe-art results with 98.5% validation accuracy (4.3% improvement over existing literature) and 97.8% precision for challenging field images under varying illumination and occlusion. The model performs real-time inference at 42 FPS on NVIDIA T4 GPUs with a light footprint (1.2 GB VRAM) and is compatible with deployment on edge devices. Technical innovations include a dynamic augmentation method balancing minority classes through synthetic lesion generation, in-pipeline explain ability through Grad-CAM visualizations for farmer-friendly diagnosis, and a multi-stage training protocol combining progressive resizing and label smoothing to enhance generalization. Experimental validation on 1,850 samples from three geographies confirms 96.4% operating accuracy, surpassing human expert prediction by 23% in identifying early-stage disease. The applicability of the system is demonstrated by a simulated 18–22% reduction in crop loss through on-time intervention. We openly release the dataset and model to encourage further work, establishing a new benchmark for crop-specific AI technology in precision agriculture.