Deep Learning-Based Crop Disease Detection Using EfficientNet-B3 for Smart Agriculture

Plant diseases substantially reduce global crop yields, and cotton production is particularly vulnerable to field-acquired variability in symptom appearance, background clutter, and illumination changes that limit the reliability and scalability of expert visual inspection. This study aimed to develop an accurate, computationally efficient, and explainable framework for real-time cotton leaf disease recognition that is suitable for deployment on resource-constrained edge devices. Using the SAR-CLD-2024 dataset (322 RGB images captured under natural agricultural conditions across seven categories, including healthy and diseased leaves), images were preprocessed via resizing and normalization and augmented online in the training set (random rotations, flips, brightness/contrast adjustments, and random cropping). An EfficientNet-B3 backbone initialized with ImageNet-pretrained weights was fine-tuned using categorical cross-entropy loss and Adam optimization, with early stopping, checkpointing, regularization, and a fixed-seed 70/15/15 train–validation–test partition to enhance reproducibility and reduce leakage. Performance was evaluated on an independent test set using accuracy, precision, recall, F1-score, MCC, balanced accuracy, Cohen’s kappa, confusion matrix, multi-class ROC/AUC, and precision–recall analysis, alongside computational benchmarking (parameters, FLOPs, memory, and inference latency) and comparative experiments against contemporary CNN, lightweight, and transformer-based models. The model showed stable convergence over 30 epochs with a small training–validation gap, predominantly correct predictions with limited confusion among visually similar classes, consistently high precision–recall behavior under moderate class imbalance, and stable performance across repeated runs with low variability and a tight confidence interval. Grad-CAM heatmaps localized necrotic lesions, discoloration, and infected tissues while largely ignoring background, and failure cases were associated with early-stage symptoms, occlusion, shadows, and inter-class similarity. Overall, the framework provides a reproducible, interpretable, and efficient solution for cotton leaf disease classification with practical implications for trustworthy, low-latency, on-device decision support in precision agriculture.