Abstract Crop diseases significantly threaten global food security by directly affecting the crop yield and quality. The traditional diagnostic methods are labour intensive and human error prone. However, the existing deep learning solutions suffer with poor generalization due to the sharp loss landscapes. The proposed work addresses this limitation and optimizes the Convolutional Neural Network (CNN) using the Sharpness-Aware Minimization (SAM). This method minimizes both the training loss and loss landscape sharpness and enables the model to converge to a flatter-minima with improved generalization. The proposed work is evaluated on 60,000 corn leaf image samples for four classes with 15,000 balanced samples per class after augmentation. The optimized CNN model has achieved 99.66% test accuracy at 0.33% classification error rate and outperforms the conventional optimizers like Adam (98.44% accuracy) and the Stochastic Gradient Descent (SGD). The state-of-the-art analysis presents a 99% average precision rate along with 99.66% F1-score and 0.0013% mean squared error (MSE). The quantized model achieves an inference latency of 22.7 ms/image (≈44 FPS) on a Raspberry Pi 4 and reduces model overfitting and enhances feature discriminability. These results underscore the potential of SAM-based optimization in precision agriculture by driving a scalable automation of disease management. This work bridges the gap between theoretical advances in deep learning optimization and practical deployment in resource-constrained farming environments.
Sharma et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: