Modular CNN-Based Classification Framework for Multi-Fruit Agricultural Diseases Models

Main Article Content

Dr. Santosh Jagtap

Abstract

Early and accurate fruit disease detection is critical for sustainable horticulture, improved yield stability, and reduction of economic losses caused by fungal, bacterial, and physiological disorders in orchard crops (Strange & Scott, 2005). Deep-learning approaches, particularly convolutional neural networks (CNNs), have demonstrated strong performance in plant disease classification, often surpassing traditional machine-learning models through their ability to autonomously learn hierarchical visual features (Mohanty, Hughes, & Salathé, 2016; Brahimi et al., 2017). Despite these advances, most existing research focuses on single-crop systems, limiting their practicality in mixed-orchard contexts where growers cultivate multiple fruit species simultaneously (Ramesh & Vydeki, 2020). To address this gap, the present study introduces a hybrid cross-crop fruit disease detection architecture integrating EfficientNet-B0 for Mango disease classification and ResNet-18 for Pomegranate disease recognition. Both models are trained using real-world orchard imagery and optimized via transfer learning to generate robust disease embeddings. The hybrid two-branch framework achieves strong single-crop performance 60.81% accuracy for Mango and 95.63% accuracy for Pomegranate and demonstrates stable precision– recall values across all disease classes. Fourteen analytical figures, including confusion matrices, ROC curves, confidence distributions, and per-class accuracy visualizations, were examined to assess classification behavior and cross-crop generalization.

Article Details

Section
Articles