Maize disease classification using transfer learning and convolutional neural network with weighted loss.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Maize is a vital crop for food and animal feed, but cultivation is threatened by crop failures due to delayed disease management.
• Technological advancements, particularly in Machine Learning (ML), offer potential solutions for early disease detection and intervention in maize cultivation.

Purpose of the Study:

• To develop and evaluate Convolutional Neural Network (CNN) models for accurate classification of maize plant diseases.
• To compare the performance of different CNN architectures (ResNet-18, VGG16, EfficientNet) and optimization techniques for maize disease identification.

Main Methods:

• Utilized two distinct maize plant image datasets (4,144 images, 4 classes and 5,155 images, 7-8 classes).
• Addressed class imbalance in the second dataset using weighted cross-entropy loss.
• Experimented with ResNet-18, VGG16, and EfficientNet architectures, employing Stochastic Gradient Descent (SGD) optimization.

Main Results:

• VGG16 with a frozen layer achieved 97.146% accuracy on the first dataset.
• EfficientNet without a frozen layer, combined with weighted loss, attained 94.798% accuracy on the second, imbalanced dataset.
• Stochastic Gradient Descent (SGD) proved to be the optimal optimizer for both datasets.

Conclusions:

• CNN models demonstrate high efficacy in classifying maize plant diseases, aiding agricultural management.
• Model architecture selection and handling of data imbalance are critical for achieving optimal performance in maize disease identification.