Enhanced corn leaf disease detection using sharpness-aware minimization optimized CNNs.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Crop diseases pose a significant threat to global food security, impacting yield and quality.
• Traditional disease diagnostics are labor-intensive and prone to human error.
• Current deep learning models struggle with generalization due to sharp loss landscapes.

Purpose of the Study:

• To optimize Convolutional Neural Network (CNN) models for crop disease diagnosis using Sharpness-Aware Minimization (SAM).
• To address the generalization limitations of existing deep learning solutions in agriculture.
• To enhance the accuracy and efficiency of automated crop disease detection.

Main Methods:

• Implemented Sharpness-Aware Minimization (SAM) to optimize CNNs, minimizing both training loss and loss landscape sharpness.
• Trained and evaluated the optimized CNN on 60,000 augmented corn leaf images across four disease classes.
• Compared SAM's performance against conventional optimizers like Adam and Stochastic Gradient Descent (SGD).

Main Results:

• Achieved 99.66% test accuracy with a 0.33% classification error rate, surpassing Adam (98.44%) and SGD.
• Reported state-of-the-art metrics: 99% average precision, 99.66% F1-score, and 0.0013% Mean Squared Error (MSE).
• Quantized model achieved fast inference (22.7 ms/image on Raspberry Pi 4), reduced overfitting, and improved feature discriminability.

Conclusions:

• SAM-based CNN optimization significantly enhances generalization and accuracy in crop disease detection.
• The optimized model demonstrates practical viability for precision agriculture in resource-constrained environments.
• This work facilitates scalable automation of disease management, bridging deep learning advancements with agricultural needs.