Maize disease classification using transfer learning and convolutional neural network with weighted loss
Maize stands out as a versatile commodity, finding applications in food and animal feed industries. Notably, half of the total demand for maize is met through its utilization as animal feed. Despite its importance, maize cultivation often grapples with crop failures resulting from delayed disease ma...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2024-11-01
|
| Series: | Heliyon |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2405844024156009 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Maize stands out as a versatile commodity, finding applications in food and animal feed industries. Notably, half of the total demand for maize is met through its utilization as animal feed. Despite its importance, maize cultivation often grapples with crop failures resulting from delayed disease management or insufficient knowledge about these diseases, impeding timely intervention. The advent of technological advancements, particularly in Machine Learning, presents solutions to address these challenges. This research focuses on employing a Convolutional Neural Network (CNN) to classify maize plant diseases. Two datasets form the foundation of this study. The first dataset encompasses 4144 images distributed across 4 classes, while the second dataset comprises 5155 images distributed among 7 to 8 classes. The second dataset encounters issues related to imbalanced class distribution, where certain classes possess substantially more data than others. To mitigate this imbalance, the weighted cross-entropy loss method is employed. During experimentation, three distinct architectural models—ResNet-18, VGG16, and EfficientNet—are rigorously tested. Additionally, various optimizers are explored, with noteworthy results indicating that both datasets achieve peak accuracy through the use of the SGD (Stochastic Gradient Descent) optimization. For the first dataset, optimal results are obtained with the VGG16 architecture, leveraging a frozen layer in the classification stage and achieving an impressive accuracy of 97.146 %. Shifting the focus to the second dataset, the most favorable outcome is realized by employing the EfficientNet architecture without a frozen layer, coupled with the implementation of weighted loss to address the class imbalance, resulting in an accuracy of 94.798 %. |
|---|---|
| ISSN: | 2405-8440 |