Optimized deep learning framework for pomegranate disease detection using nature-inspired algorithms.

Background: Agriculture plays a pivotal role in global food security and socio-economic stability, yet crop productivity remains threatened by plant diseases that incur substantial economic losses. Pomegranate is an important fruit for both nutrition and business, but it is easily infected by pathogens that can lower yields by 20 to 40 percent. Traditional methods of finding these pathogens by hand are time-consuming, subjective, and not very effective, while existing deep learning models struggle with field noise, lighting variations, and computational inefficiency. To address these challenges, this study proposes an automated framework integrating a modified ResNet101 architecture with a Hybrid Genetic Algorithm-Particle Swarm Optimization (HGA-PSO) method. The approach employs dual-stream processing of original and noise-augmented images (Gaussian, salt-and-pepper, speckle) to enhance robustness.

Results: The framework achieved exceptional performance on a dataset of 5,000 images across five classes (four diseases, one healthy). Feature fusion from dual streams and HGA-PSO optimization reduced dimensionality by 50-70% while preserving discriminative power. Under rigorous 5-fold cross-validation, the Multi-Layer Perceptron (MLP) classifier attained 99.10% accuracy, a perfect ROC-AUC score (1.00), and high precision-recall metrics. Confusion matrices revealed near-zero misclassification, and real-world tests (single/batch images) confirmed strong generalization. Grad-CAM + + visualizations validated precise localization of disease regions. The model outperformed existing techniques (e.g., PSO-YOLOv8: 98.86%, Transformer models: 93.13%) in accuracy, precision, recall, and F1-score CONCLUSIONS: This research presents an optimized model for pomegranate disease detection by combining deep learning with nature inspired optimization. The dual-stream feature fusion and HGA-PSO significantly improves robustness again environment variability while reducing computation overhead. This framework offers a scalable solution for precision agriculture, enabling early disease intervention to mitigate economic losses. Future research could improve scalability and usefulness by looking into lightweight optimization methods, model interpretability, and how they can be used in limited-resource agricultural settings.